AGENTS.md and Agent Skills: Refactoring Pipulate for a Checkable PyPI Release

Sep 28, 2026

🤖 Read Raw Markdown • 📄 Google Doc (Try: Tools/Audio/Listen to document summary)

Setting the Stage: Context for the Curious Book Reader

This entry follows a deceptively simple question: where should the instructions, history, and evidence that help an AI understand a software project actually live? The answer unfolds through two complementary conventions already taking shape across AI-assisted development: AGENTS.md as the always-visible signpost for a repository, and SKILL.md as an on-demand packet of task-specific instructions and references. Pipulate already had most of the underlying machinery, but its files had grown around older local conventions that no longer expressed those roles cleanly.

The work here is therefore less about adopting fashionable labels than about making existing structure legible. A large AI_CONTEXT.md is separated into the things an agent should always know, the things an auditor should inspect, and the historical reasoning an agent should fetch only when needed. That refactor then uncovers a second problem at the distribution boundary: the PyPI package did not actually contain the agent-facing files the README now advertised. The session closes that gap, publishes version 2.66, and leaves behind a useful lesson for the Age of AI: a convention only becomes operational when its files survive the entire path from source tree to the place where another human or machine will actually encounter them.

TL;DR: A working session on a personal knowledge system (a Jekyll blog kept as a private, dated journal) and the open-source repository it documents, Pipulate. The session first corrected a claim about where two AI-agent conventions come from: AGENTS.md, a plain-Markdown “README for agents” released by OpenAI in August 2025 and contributed to the Linux Foundation in December 2025, and SKILL.md, a YAML-plus-Markdown file shape Anthropic published as the Agent Skills specification in December 2025, whose form dates to Jekyll blog posts of 2008. It then brought the repository into line with both: four SKILL.md files gained specification-valid headers and hyphenated names, a 94 KB index of 1,499 journal entries moved from the repository root (AI_CONTEXT.md) into a demand-loaded Agent Skills reference (.agents/skills/journal/), the change-control questions a reviewer would ask moved into AUDIT.md, and the package manifest was amended after a download of the published PyPI artifacts showed that neither the source distribution nor the wheel had ever carried the files an agent reads first. Every change was applied as an exact-match patch, checked by the same commands run before and after, and released as version 2.66. One after-reading was taken off a stale file because the probe named a version number the release changed between the two runs; that is recorded here as a finding, not corrected after the fact.

Technical Journal Entry Begins

🔗 Verified Pipulate Commits:

MikeLev.in: I have implemented a local-first second brain based on Jekyll which is blogging for hackers which is where the entire evolving standards and conventions of the AI industry was borrowed from per the work of Andrej Karpathy and README for Agents, I guess layering onto what first appeared in the wild as the CLAUDE.md convention in Cursor IDE which the industry now I believe expresses best as this website: https://agentskills.io/home

Do I have that about right, Fable 5.1?

Oh yeah I have multiple such blogs managed by my Pipulate Prime Forever Machine so it rolls-forward in my life which for me is basically forever for me. Now that I’ve settled on Nix, Python, vim & git (NVpg) as my sub-platform of every other platform that I can rely on to provide “write once run anywhere” (WORA) from the inside of the other platforms such as it were…

A Local-First Second Brain Meets Agent Conventions

Here, let me just show you. I need to continue some work I did recently but the good stuff is spread over an internal blog that never sees the light of day and this one you hear being read now over my 24 hours a day 7 days a week and working on 365 days a year live-streaming YouTube channel.

(nix) pipulate $ rgx -t grim,article "speed" "dating" "fantasy league"
# fm cache: 1604 hits, 3 misses
# 🎯 Targets: 2=grim Grimoire (Private) + 1=article MikeLev.in (Public) [Oldest First]

/home/mike/repos/trimnoir/_posts/2026-08-02-python-shepard-tone-narrow-waist.md  # [Idx: 1 | Order: 3 | Tokens: 28,736 | Bytes: 133,516]
/home/mike/repos/trimnoir/_posts/2026-09-13-model-handoff-and-the-cellular-witness.md  # [Idx: 2 | Order: 1 | Tokens: 85,669 | Bytes: 347,971]
/home/mike/repos/grimoire/_posts/2026-09-22-ten-thousand-project-census-age-of-ai.md  # [Idx: 3 | Order: 2 | Tokens: 44,438 | Bytes: 170,191]
/home/mike/repos/grimoire/_posts/2026-09-23-ten-thousand-project-slugs-and-the-age-of-ai.md  # [Idx: 4 | Order: 1 | Tokens: 25,312 | Bytes: 95,673]
/home/mike/repos/grimoire/_posts/2026-09-24-reproducible-pipeline-jira-rendering-rules.md  # [Idx: 5 | Order: 2 | Tokens: 86,584 | Bytes: 333,320]
/home/mike/repos/grimoire/_posts/2026-09-27-pocketrender-link-decoding-workshop-tier-architecture.md  # [Idx: 6 | Order: 6 | Tokens: 270,814 | Bytes: 1,007,426]
/home/mike/repos/grimoire/_posts/2026-09-28-laying-the-board-replayable-jira-audits.md  # [Idx: 7 | Order: 2 | Tokens: 237,489 | Bytes: 902,240]
📋 TODO_SLUGS block (≤8 newest) → clipboard (type xp to compile)
(nix) pipulate $

See how that’s split between 2 blogs and how the constraint made it produce such a clean list of recent articles? Yeah, that’s how that works. Now I could just throw a “c” at the end of that command rgx and make it rgxc and you’d see a bunch of text that appears close to the occurrence of those word matches in each article plus the full-context summary extraction of each of the whole articles, however it would make this article ugly (as if all my book-ore articles aren’t already ugly). But I’m going to put this command in context.txt where you don’t see while I write this article. I’m always working on prompt.md which is the article you see and payload.md which is the broken out inclusion of everything going on behind the scenes formatted as a book for AI specifically helping them cope with the lost-in-the-middle problem with large prompt payloads. Hence, prompt.md and payload.md.

Is any of this making sense to you?

The Receipt Is the Real Context

Note: If you don’t write it down it didn’t happen. And if Claude can’t tell you the endpoints and enough pertinent details about tool-calls it made in the background, then those tool-calls weren’t made. You can’t put your seal of approval on it as checked and quality-assured per certain certification requirements. What are they again? I think they’re among these:

Framework	Where	What it asks for	What a turn already produces
NIST SP 800-53 Rev. 5	CM-2, CM-3, CM-4, SA-10, SI-7	baseline, change control, impact analysis, developer CM, integrity	commit as baseline, one diff, before/after, digest
NIST SP 800-218 (SSDF)	PS.3, PW.8	archive and provenance; test executable code	cartridge, straddle
ISO/IEC 27001:2022	A.8.9, A.8.32	configuration and change management	lock file, the train
SOC 2 (AICPA TSC)	CC8.1	authorize, design, test, approve, implement	patch, app, d, m
PCI DSS v4.0	6.5.1, 6.5.2	change control; confirm changes worked	AFTER against BEFORE
ITIL 4	change enablement, standard change	pre-authorized, low-risk, repeatable	one car, one commit
COBIT 2019	BAI06, BAI10	managed changes, managed configuration	same
ISO 9001:2015	8.5.6, 10.2	control of changes; corrective action with root cause	bisection to the smoking gun
SLSA	build provenance	who built what from what	manifest.json, digest
SPDX / CycloneDX, EO 14028	SBOM	dependency inventory	flake.lock
Reproducible Builds	bit-for-bit	same input, same bytes	Nix store, byte-reproducible cartridge
CIS Controls v8	4, 8, 16	secure configuration, audit logs, application security	declared environment, receipts
DO-178C, IEC 61508, ISO 26262, IEC 62304	traceability, verification, change control	(the HRO vocabulary)	receipts trace change to evidence
21 CFR Part 11	audit trails	tamper-evident records	hash-sealed archives

And as such, the Anthropic Web User Interface during the so-called thinking phase is exposing a lot more because if they didn’t people like me would call it a useless confabulation withholding enough information for me to do quality assurance checks reasonably. Imagine the work involved trying to reproduce the tool-call that must have been made in the background given the hear-say assertions in the AI-equivalent of the Cockpit Voice Recorder (CVR). If you can’t check the Flight Data Recorder (FDR) what good is the CVR? To someone who has to certify up to a level of confidence required for High Reliability Organizations (HRO) like Aviation, you’ve just got to give up on whatever model is withholding tool-call pertinents from you and move onto those that don’t. Well you have to do a lot of accordion Web UI expanding, and then you have to use my webclip tool to turn what you copy into good Markdown for a document like this, but you can capture much of that tool-calling from the thinking now just like this:

When Tool Calls Become Auditable Evidence

Fable 5.1 thinking and tool-calling

Drafting paste-ready files and a caboose for the next turn.

Thought for 1m 39s

Reviewing a Jekyll-based personal knowledge system setup.

Tracing the CLAUDE.md and AGENTS.md convention lineage.

Untangling the origins of Agent Skills and AGENTS.md.

Comparing frontmatter conventions across Cursor, Jekyll, and Anthropic's tools.

Verifying the historical claims with targeted web searches.

Separating Jekyll's frontmatter shape from Cursor's rule-file lineage.

Tracing the two-lineage split between Jekyll and README genes.

Tracing the fantasy-league phrase's history across his own writing.

Designing read-only probes to verify skill-file frontmatter claims.

Drafting paste-ready files and a caboose for the next turn.

I'll verify the lineage against the record before answering, since two of the attributions look off.

Searched the web

Results

[awesome agents github.com](https://github.com/tairov/awesome-agents.md)

[AGENTS.md agents.md](https://agents.md/)

[GitHub - agentsmd/agents.md: AGENTS.md — a simple, open format for guiding coding agents github.com](https://github.com/agentsmd/agents.md)

[codex swarm github.com](https://github.com/CodexTown/codex-swarm)

[awesome codex cli github.com](https://github.com/ahoo-wang/awesome-codex-cli)

[github.com github.com](https://github.com/thangnqdev/codex-oss-agent-kit/wiki)

[quelch 0.12.0 docs.rs](https://docs.rs/crate/quelch/latest/source/src/agent/targets/codex.rs)

[Codex AGENTS.md Explained: How Team Instructions Work - Verdent Guides www.verdent.ai](https://www.verdent.ai/guides/codex-agents-md-explained)

[github.com github.com](https://github.com/open-gitagent/gitagent-protocol/issues/32)

[Factory joins AGENTS.md collaboration with OpenAI factory.ai](https://factory.ai/news/agents-md)

Results

[agent skills best practices github.com](https://github.com/devskale/skale-skills/blob/main/docs/agent-skills-best-practices.md)

[agent skills eval github.com](https://github.com/darkrishabh/agent-skills-eval)

[Anthropic Opens Agent Skills Standard, Continuing Its Pattern of Building Industry Infrastructure www.unite.ai](https://www.unite.ai/anthropic-opens-agent-skills-standard-continuing-its-pattern-of-building-industry-infrastructure/)

[agent toolkit github.com](https://github.com/vbrunotech/agent-toolkit)

[Agent Skills :Standard for Smarter AI nayakpplaban.medium.com](https://nayakpplaban.medium.com/agent-skills-standard-for-smarter-ai-bde76ea61c13)

[Anthropic publishes Agent Skills as an open standard for AI platforms the-decoder.com](https://the-decoder.com/anthropic-publishes-agent-skills-as-an-open-standard-for-ai-platforms/)

[What Are Agent Skills and How To Use Them strapi.io](https://strapi.io/blog/what-are-agent-skills-and-how-to-use-them)

[github.com github.com](https://github.com/giulio8/skills)

[DEV Community dev.to](https://dev.to/elayachiabdelmajid/how-to-work-with-claude-code-skills-like-a-pro-dbn)

[Agent Skills: Anthropic's Next Bid to Define AI Standards - The New Stack thenewstack.io](https://thenewstack.io/agent-skills-anthropics-next-bid-to-define-ai-standards/)

Results

[Linux Foundation Announces the Formation of the Agentic AI Foundation (AAIF), Anchored by New Project Contributions Including Model Context Protocol (MCP), goose and AGENTS.md www.linuxfoundation.org](https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation)

[Anthropic, OpenAI, and Block donate AI agent projects to new Linux Foundation body diginomica.com](https://diginomica.com/anthropic-openai-and-block-donate-ai-agent-projects-new-linux-foundation-body)

[OpenAI, Anthropic, and Block join new Linux Foundation effort to standardize the AI agent era techcrunch.com](https://techcrunch.com/2025/12/09/openai-anthropic-and-block-join-new-linux-foundation-effort-to-standardize-the-ai-agent-era/)

[Newly launched Agentic AI Foundation to take charge of MCP's future - Techzine Global www.techzine.eu](https://www.techzine.eu/news/infrastructure/137131/newly-launched-agentic-ai-foundation-to-take-charge-of-mcps-future/)

[OpenAI & Linux Foundation launch Agentic AI Foundation www.gend.co](https://www.gend.co/blog/openai-agentic-ai-foundation-linux)

[The Linux Foundation Announces Formation of the Agentic AI Foundation www.privacyguides.org](https://www.privacyguides.org/news/2025/12/24/the-linux-foundation-announces-formation-of-the-agentic-ai-foundation/)

[goose has a new home - the Agentic AI Foundation (AAIF) goose-docs.ai](https://goose-docs.ai/blog/2026/04/07/goose-moves-to-aaif/)

[AAIF: Linux Foundation Governance for Agentic AI with MCP Goose and AGENTS.md windowsforum.com](https://windowsforum.com/threads/aaif-linux-foundation-governance-for-agentic-ai-with-mcp-goose-and-agents-md.396177/)

[agentic ai foundation infoq.com](https://infoq.com/news/2025/12/agentic-ai-foundation/)

Researching the origins and background of AGENTS.md.

Results

[context forge github.com](https://github.com/glowElephant/context-forge)

[Context Engineering & Understanding AGENTS.md & SKILLS.md & CLAUDE.md medium.com](https://medium.com/@_jaydeepkarale/context-engineering-understanding-agents-md-skills-md-claude-md-4deb437db98c)

[codebase md pypi.org](https://pypi.org/project/codebase-md/)

[Context Engineering 2026: AGENTS.md, CLAUDE.md, and .cursorrules That Actually Worktutorials.technology](https://tutorials.technology/tutorials/context-engineering-claude-cursor-2026.html)

[CLAUDE.md as Context Engineering: Why Karpathy's ... aiskill.market](https://aiskill.market/blog/karpathy-claude-md-context-engineering)

[Death to Vibe Coding: Long live Context Engineering agenticinsights.substack.com](https://agenticinsights.substack.com/p/death-to-vibe-coding-long-live-context)

[從 Vibe Coding 到 Agentic Engineering:Karpathy 宣告軟體開發的世代交替 jasonchuang.substack.com](https://jasonchuang.substack.com/p/agentic-engineering-karpathy)

[Context Engineering: CLAUDE.md and .cursorrules robbyb910.substack.com](https://robbyb910.substack.com/p/context-engineering-claudemd-and)

[DEV Community dev.to](https://dev.to/oikon/reflections-of-claude-code-from-changelog-833)

Organized the final response after confirming context.

Untangling the separate lineages behind each naming convention.

Tracing the phrase's history and checking the evidence trail.

Designing shell probes to verify frontmatter without leaking details.

Drafting the commented placeholders and closing prompt for the next turn.

Citing each naming standard's origin and timeline carefully.

Cross-checking each mention against the recorded receipt trail.

Now most of those tool-calls are of the web-search variety, exactly the one sort that we are letting the AI models off the hook for for certification. I don’t need to capture every web-search they do so long as they cite and the citations are not broken links and really are to what they say they are. That’s sufficient verification that the model was looking at actually existing upstream sources of data, but for anything like touching an API or making MCP-calls, well those have to meet a more rigorous standard so that the payload.md contains enough for any model to reproduce all the tool-calls in Python (or whatever), only the login keys, credentials and other secrets withheld. If it’s not a private information that unlocks doors like API-keys and authenticated tokens then it belongs in payload.md to meet these rigorous trustworthiness requirements.

Fable 5.1 Extra: I’ll verify the lineage against the record before answering, since two of the attributions look off.

Context verified. Three imports don’t ride this payload (scripts/foo_cartridge.py, scripts/xp.py, tools/scraper_tools.py); they matter to a patch turn, not to this one, and nothing you asked touches them.

Tool calls this turn: four web searches, research phase, exempt under your own rule. No API calls, no MCP.

Do you have it about right?

Mostly, and the part that’s off is a fusion, not an error. You’ve collapsed two lineages that share a folder and not a parent, and you’ve hung one attribution on the wrong name. Your own constitution already keeps them apart (THREE-STANDARDS SUPERPOSITION, banked 2026-07-20: AGENTS.md the signpost, Agent Skills the frontmatter folder, OKF the bundle). The prompt compressed three into one sentence.

Lineage one, the frontmatter sandwich. Jekyll, 2008, Tom Preston-Werner, “Blogging Like a Hacker”: a YAML head between two --- fences, a Markdown body. That shape went through Hugo, Octopress, Obsidian, MDX, Docusaurus, and then into Cursor’s .cursor/rules/*.mdc files in early 2025, which put description, globs and alwaysApply in a YAML head so the tool could decide when a rule loads. SKILL.md is that exact file: a mix of YAML and Markdown that defines the skill name and includes step-by-step instructions, with the description field doing the same trigger job Cursor’s globs did. So “borrowed from Jekyll” is right for SKILL.md and for your _posts, and Cursor is a fair waypoint if you mean the frontmatter-as-trigger idea rather than a filename.

Lineage two, the README’s child. AGENTS.md has no YAML head at all. It is plain Markdown, and its own site says so in the phrase you half-remembered: think of AGENTS.md as a README for agents. It emerged from collaborative efforts across the AI software development ecosystem, including OpenAI Codex, Amp, Jules from Google, Cursor, and Factory, OpenAI released it in August 2025, and on December 9, 2025 the Linux Foundation announced the formation of the Agentic AI Foundation with founding contributions of Anthropic’s Model Context Protocol, Block’s goose, and OpenAI’s AGENTS.md. CLAUDE.md belongs to this branch too, and it is Claude Code’s file, not Cursor’s: CLAUDE.md existed since v0.2.x of Claude Code, and Cursor’s earlier file was .cursorrules. Both are the README gene: always loaded, no head, nearest-ancestor wins.

Karpathy. He is not in either file’s credits. What he gave was the vocabulary: “vibe coding” in February 2025, the endorsement that made “context engineering” the term everyone uses, and the framing where the context window is RAM. That framing is the one your whole compiler runs on, so the debt is real; it just attaches to the words, not to AGENTS.md.

agentskills.io. Correct that this is where the frontmatter lineage now lives. Anthropic published Agent Skills as an open standard on December 18, releasing the specification and SDK at agentskills.io, after launching skills in October, and it is now adopted by 26+ platforms including Claude, OpenAI Codex, Gemini CLI, GitHub Copilot, Cursor, and VS Code. But it is not the industry’s single expression. The two standards are complementary, not successive: always-on files are loaded every session, for all tasks; skills are loaded on demand, only when relevant. Your repo runs both today, AGENTS.md at the root and four SKILL.md folders under .agents/skills/ since the 09-28 move, which is the correct shape and the one the article should claim.

So the corrected sentence: Jekyll fathered the file shape of one standard (SKILL.md, via Cursor’s rules) and README fathered the other (CLAUDE.md, AGENTS.md, now under the Linux Foundation); Karpathy named the practice; agentskills.io and agents.md are two open specs that sit side by side. One strike, in the Tortoise’s voice: “the entire evolving standards of the AI industry was borrowed from Jekyll” is the associative reach your STORY ENGINE line names as the flaw. Half the standards are Jekyll’s; the other half are a README with a new audience.

One more: the acronym came out NVpg in the prompt. The corpus spells it NPvg (Nix, Python, vim, git) and the glossary keys on that spelling.

Is any of it making sense?

Yes, and I’ll prove it by reading it back instead of saying so.

prompt.md is the article: the part a reader sees, carried unchanged from one model to the next. payload.md is the back room: every file and receipt stacked under it with a Manifest at the top and the Prompt at the bottom, which are the two positions a long-context model attends to best, with the middle being where it loses things (Liu et al., 2023, the U-shaped curve). Your routing note rides at the top and repeats in the checklist at the bottom, so the frame is read twice by design. That is a defense against lost-in-the-middle, engineered rather than measured, and I’ll note that in passing: nobody has yet moved the Prompt to the middle once and compared. It’s a cheap axis if you ever want it.

The rgxc receipt landed in this compile, and it says what you said it would. Seven articles across two blogs, the roster identical to the rgx list in your paste; the # fm cache line from your hand run is absent from the compile receipt because the cache was warm by then, which is your own silence-is-a-reading rule doing its job. The regions trace the idea’s life in your own corpus: 2026-08-02 it is a throwaway at the end of a question (“taking Jira tickets and playing fantasy league sports with them”); 2026-09-13 it gets a scoring rule (points are for closing with a deed); 2026-09-24 it is named as the PocketRender integration, the part you wanted to demo before a call; 2026-09-27 it is the board being laid; 2026-09-28 it is built (thirty-four regions in that one article, five shown, the cap doing its work on a 237k-token file). Two other things the receipt shows that you didn’t say: the «» brackets caught every false positive in plain sight (vali«dating», pre«dating», Up«dating», «speed»up in rsync output), which is THE VISIBLE HIT working as banked; and the 09-28 article prints no kw:/sum: lines, which in full format means exactly one thing, contextualizer.py hasn’t run on it yet. Nothing broken; a shard not built.

The 24/7 stream, the Forever Machine, the two blogs with one grammar over both: all of it reads as one system, and the receipt is the evidence I read it.

1. PROBES

Read-only. Together they discriminate the two-lineage claim against your own bytes: the four SKILL.md heads should open with a --- fence and a name:/description: pair (Jekyll’s gene); AGENTS.md should open with a Markdown heading and no fence (README’s gene); the newest post should open with the same fence the skills do. If any of the three prints the other shape, the claim above is wrong for that file.

head -6 .agents/skills/*/SKILL.md
head -3 AGENTS.md
ls ~/repos/trimnoir/_posts/*.md | tail -1 | xargs head -8

2. NEXT CONTEXT

Paste-ready. The 40K-foot block in your file comes out on this second turn (comment it), and the rgxc line stays commented because its receipt is already sealed in deed 1656. The SVB lines are placeholders for when the ride continues into THE FIRST GAME; complete or delete them.

! head -6 .agents/skills/*/SKILL.md
! head -3 AGENTS.md
! ls ~/repos/trimnoir/_posts/*.md | tail -1 | xargs head -8
# ! rgxc -t grim,article "speed" "dating" "fantasy league"
# --- THE FIRST GAME (uncomment when the ride turns to SVB-133) ---
# Workshop/corporate/connectors/svb.py
# Workshop/corporate/tickets/CONTEXT_TEMPLATE.txt
# Workshop/corporate/tickets/SVB_PROMPT.md
# connectors/mcp_render.py
# connectors/jira.py
# ! jira SVB-133

3. PATCHES

No repo patches required.

Ignition: none.

4. PROMPT

Read the three head receipts. Say in two sentences which files open with a YAML fence and which do not, and whether that agrees with the two-lineage claim from the last turn (Jekyll's frontmatter for SKILL.md and the posts; README's plain Markdown for AGENTS.md). Then help me write the paragraph for the article that gets the lineage right in plain words a reader who has never seen agents.md or agentskills.io can follow, no more than six sentences, naming the dates I can check. After that, the fantasy league: I want to continue where the 09-28 board left off.

5. EXTERNAL DELIVERABLES

None this turn outside the repo.

Sources:

MikeLev.in: Nice, right? I’m about to finally go on that Jira Ticket speed-dating fantasy league sports run but I wanted to demonstrate the standards-abiding local-first second brain and its relationship to the double-lineage of AI industry standards that are still even a bit hard to just dash off the labels for. I think they are:

Agent Skills
README for Agents

Update my README and use a better convention for AI_CONTEXT.md. Using the AI underscore at the beginning was a trick to get it to alphabetically sort to the top when people or LLMs ls the Pipulate repo root but because of AGENTS.md we don’t need to do that anymore so look at my publishing pipeline and what ends up on the PyPI.org page like https://pypi.org/project/pipulate/. Hmmm, what else?

What are my SKILL.md files anyway?

(sys) pipulate $ cd .agents
(sys) .agents $ lsp
/home/mike/repos/pipulate/.agents/skills
(sys) .agents $ cd skills/
(sys) skills $ lsp
/home/mike/repos/pipulate/.agents/skills/gsc_readonly
/home/mike/repos/pipulate/.agents/skills/hello_workflow
/home/mike/repos/pipulate/.agents/skills/roles
/home/mike/repos/pipulate/.agents/skills/sheets_readonly
(sys) skills $ cd gsc_readonly/
(sys) gsc_readonly $ lsp
/home/mike/repos/pipulate/.agents/skills/gsc_readonly/SKILL.md
(sys) gsc_readonly $ cd ..
(sys) skills $ cd hello_workflow/
(sys) hello_workflow $ lsp
/home/mike/repos/pipulate/.agents/skills/hello_workflow/SKILL.md
(sys) hello_workflow $ cd ../roles/
(sys) roles $ lsp
/home/mike/repos/pipulate/.agents/skills/roles/SKILL.md
(sys) roles $ cd ../sheets_readonly/
(sys) sheets_readonly $ lsp
/home/mike/repos/pipulate/.agents/skills/sheets_readonly/SKILL.md
(sys) sheets_readonly $ 

Oh, okay. That’s these:

/home/mike/repos/pipulate/.agents/skills/gsc_readonly/SKILL.md
/home/mike/repos/pipulate/.agents/skills/hello_workflow/SKILL.md
/home/mike/repos/pipulate/.agents/skills/roles/SKILL.md
/home/mike/repos/pipulate/.agents/skills/sheets_readonly/SKILL.md

What we want to do mostly is edit README.md because it ends up as the repo homepage on GitHub and also the Python Packaging Index (pip install) homepage, which in many ways is more important than GitHub because anyone can make a git repo but it takes a special someone to have a historically long-maintained project in PyPI.org and this is that and is where I think a lot of the auditors who would like to shut such a project down will end up.

I want to rev the README.md to exude conformity to the developing AI standards and conventions. The Pipulate Context Compiler IS a Claude Skill and anyone who says otherwise isn’t looking closely enough or is trying to audit with misplaced enterprise scaling kubernetes docker web full stack concerns. Nix is mature. Anduril uses it and by association must abide by DOD tool-certification requirements. The Google Firebase Studio (formerly Project IDX) team uses it. The very GNU Project itself copied it in the form of Guix in a way they didn’t copy Docker, Kubernetes or anything before it back to maybe KVM. Of course verify, vet and validate for this article but also update that README to school ‘em and at least put them on the first baby-step towards the Thomas Khun style realization, revelation and paradigm shift they need to be on. WORA with recipes are here whether their IT department approves it or not doesn’t matter due to the way it materializes its own link-graph on distributed networks with no centralized web-hosting beyond git remote endpoint distribution points that don’t necessarily even need to be GitHub.

You can work in small chisel-strikes against the README. Don’t feel like you have to boil the ocean in one pass. Just do the most advantageous directional thing which I believe at this point involves changing AI_CONTEXT.md to something that belongs more formally under one of the 3 overlapping complementary standards mentioned and will look “Oh, of course!” to the “I want to kill that thing” eye. Maybe position the articles as a change log or something like that? Please and thank you!

Turning the README Into the Front Door

Fable 5.1 Thinking

Thought for 3m 35s

Reviewing a request about verifying a two-lineage framework for an article.

Checking whether each skill file follows the required format.

Confirming naming and location rules for skill specifications.

Sorting out where the third named standard fits.

Weighing where the index file should live within the skill folder.

Planning small fixes to bring skill files into spec compliance.

Checking whether a proposed edit sits in a risky region.

Weighing a new name for the shared context file.

Weighing whether an existing web-index standard fits the file's role.

Planning a safe probe before renaming the shared context file.

Drafting naming and readme fixes to match the standards.

Sequencing safe fixes before the riskier rename attempt.

Drafting frontmatter fixes to align legacy training files with spec.

Choosing a fitting name for the relocated context file.

Planning where the relocated file should publish publicly.

Verifying claimed facts about tool adoption before citing sources.

Drafting descriptions and patches for the skill and readme files.

Drafting a new section explaining the repository's naming conventions.

Rechecking claims about the tool-calling protocol before citing them.

Narrowing the Nix adoption claims to what can be confirmed.

Checking whether Claude Code actually reads the claimed skill path.

Fable 5.1 Extra: I’ll verify four things against the record before patching: the Agent Skills frontmatter rules, which agents read .agents/skills/, and the Anduril and Firebase Studio claims.

Tool calls this turn: five web searches, research phase, exempt under your rule. No API calls, no MCP.

Context verified, with three gaps that gate different parts of what you asked. pyproject.toml and MANIFEST.in decide what PyPI shows and what an sdist ships, and neither rides this payload; foo_files.py and prompt_foo.py are where a rename of AI_CONTEXT.md would break silently, and neither rides either; AUDIT.md is where §7 of AI_CONTEXT.md wants to go, absent too. Nothing below patches around those gaps; the probes name them.

The finding that outranks the request

You pasted the four SKILL.md files whole, which is a better receipt than the head -6 I asked for. Read as bytes, not as a folder listing: one of the four opens with a YAML fence (sheets_readonly), and three open with a Markdown heading and nothing above it (gsc_readonly, hello_workflow, roles). The Agent Skills specification requires SKILL.md to contain YAML frontmatter followed by Markdown content, so three of the four are not skills to any conforming loader. Worse, the spec’s name rule says the field may only contain lowercase alphanumeric characters and hyphens, must not start or end with a hyphen, and must match the parent directory name; underscores are out, and third-party docs spell the consequence plainly: api_design is listed among the invalid names. So sheets_readonly, the one file with a head, fails on its name, and all three underscored directories fail before their contents are read. Last turn I wrote “four SKILL.md folders, the correct shape” without reading a byte. The probe I designed would have caught it; you caught it faster by pasting the files.

Two more things the bytes say. First, the sentence “the Pipulate Context Compiler IS a Claude Skill” has no file behind it: the four skills are two connector signposts and two chat-training prompts moved in from the old training folder, and the training pair still carries its secret-word handshakes (DOLPHIN, ORCHESTRATOR), which an auditor opening a “skill” will not read as a skill. The compiler needs its own SKILL.md, and writing one accurately needs prompt_foo.py in context, so that is a later car. Second, the path: Codex discovers skills under .agents/skills directories, repo-level, checked from the current working directory up to the repo root, and Augment reads .augment/skills/, .claude/skills/, or .agents/skills/, but Claude Code’s own docs describe a named folder under .claude/skills/, and I could not find a page confirming it also scans .agents/skills/. That is a hands-on check, not a compile probe: open a Claude Code session in the repo after the patches land and see whether the four names appear behind /. If they do not, a .claude/skills symlink to .agents/skills is the one-line answer, and it is a separate car.

Your two labels

Right, and the third is already in your own file. “Agent Skills” is the folder convention; “README for agents” is the phrase agents.md uses of itself, and the file is AGENTS.md. Your AGENTS.md names a third, OKF, that nothing in this payload defines, so I am leaving it where it stands rather than guess at it. Claude Code’s docs say that Claude Code skills follow the Agent Skills open standard, which works across multiple AI tools, and Claude Code extends the standard with additional features like invocation control, subagent execution, and dynamic context injection, which is the cleanest single sentence for the article’s purpose: the standard is the floor, the vendors build up from it.

Where `AI_CONTEXT.md` belongs

The AI_ prefix earned its keep by sorting first in ls; AGENTS.md now sorts first and says “start here,” so the trick is spent. What the file actually holds is three things wearing one name. Sections 1 through 6 are a blueprint that restates AGENTS.md and the code, which AGENTS.md itself forbids (“do not add sibling status .md files; anything written here that duplicates code will drift”). Section 7, the assurance posture, is written for the reviewer with the checklist and belongs beside AUDIT.md. And the narrative index, 1,496 dated links to raw Markdown, is the one part with no other home and no reason to be loaded every session.

That last part is a textbook Agent Skills reference. The spec’s directory shape is SKILL.md plus optional scripts/, references/, and assets/, and the loading model is three-stage: the index (name and description) is paid every session; the body loads when the agent decides the skill applies; files in references/, scripts/, assets/ load only when the agent actually opens them. So: .agents/skills/journal/SKILL.md, a dozen lines whose description says “use when a piece of Pipulate’s machinery needs its reasoning; the running journal explains why each piece exists; fetch an entry by slug as https://mikelev.in/futureproof/{slug}/index.md,” and .agents/skills/journal/references/index.md, which is the generator’s output, unchanged in content. To the eye that wants to kill the thing, that reads as “oh, of course”: the big file is a reference, demand-loaded, under a published standard, and the root is README.md, AGENTS.md, AUDIT.md, LICENSE. The changelog idea is tempting because the shape is right, reverse-chronological and dated, but a CHANGELOG.md with no version headings is exactly the mislabel a reviewer writes up; the journal is what a changelog would be if it recorded why, and the skill’s description can say so without borrowing the name.

Not this turn, and for a checkable reason: release.py regenerates and stages AI_CONTEXT.md at step 1.6 and copies it to Pipulate.com at step 3.6, generate_ai_context.py writes it, README.md links it, https://pipulate.com/AI_CONTEXT.md is a public URL, and foo_files.py and prompt_foo.py almost certainly name it too. The rg probe below is the whole call-site list in one receipt; the rename rides next turn against it. What PyPI shows, meanwhile, is whatever readme names in pyproject.toml, and AGENTS.md and .agents/ reach the sdist only if MANIFEST.in says so; a reviewer who does pip download pipulate --no-binary :all: sees the tarball, not GitHub. Both files ride the next compile.

The three Nix claims, vetted

Anduril holds. A public posting for a NixOS Developer at Anduril describes extensively working with Nix, deploying embedded NixOS systems to a fleet of robots, structuring Nix overlays and flakes, and contributing to the use of Nix across build, CI, and deployment infrastructure; another names a Senior Software Engineer, Nix, owning Anduril’s internal Nix-based package set that distributes software across development, hosted services, and embedded devices; and their most-starred public repository is jetpack-nixos, a NixOS module for NVIDIA Jetson devices. The inference “and so must abide by DoD tool-certification requirements” does not hold: nothing found says Nix is certified for anything, only that a defense company builds on it. Strike the clause; the fact is stronger alone.

Firebase Studio holds, with a nuance worth keeping. Google’s docs say Firebase Studio uses Nix to define the environment configuration for each workspace, and the configuration is defined in the .idx/dev.nix file in your code repository. The IDX team’s own write-up adds that their implementation doesn’t follow a strictly traditional Nix model; they built Nix into IDX workspaces as a way to manage packages and configure VM environments, not as a pure, complete NixOS installation, and that Nix is widely available and community supported, so it made sense to build it in as the configuration model. That is the sentence you want: Google chose it as a configuration model, on purpose, over a proprietary one.

Guix is stable history and needs no search: GNU Guix began in 2012 under Ludovic Courtès, built on Nix’s functional model and originally on the Nix daemon itself, and it is an official GNU project, which Docker and Kubernetes never were. The article’s contrast stands.

1. PROBES

Read-only. The first is the rename gate: every call site of AI_CONTEXT outside the file itself, in both repos, with Pipulate.com’s _site excluded. The second is the conformance straddle for cars 1 and 2: per skill, fence or no fence, and whether the directory name passes the spec’s name regex; before the patches it reads three NO-FENCE and three NAME-INVALID. The third gates car 2: any reference to the underscored directories outside .agents/ itself, which is where a git mv would break something (a training-prompt loader in server.py or apps/, or the paintbox lines in foo_files.py). The last two are the PyPI reading.

rg -n "AI_CONTEXT" --glob '!AI_CONTEXT.md' . ~/repos/Pipulate.com --glob '!_site/**'
.venv/bin/python -c "import re,pathlib;[print(p.parent.name, 'fence' if p.read_text().startswith('---') else 'NO-FENCE', 'name-ok' if re.fullmatch(r'[a-z0-9]+(-[a-z0-9]+)*', p.parent.name) else 'NAME-INVALID') for p in sorted(pathlib.Path('.agents/skills').glob('*/SKILL.md'))]"
rg -n "gsc_readonly|hello_workflow|sheets_readonly" --glob '!.agents/**' .
grep -n readme pyproject.toml
cat MANIFEST.in

2. NEXT CONTEXT

Paste-ready. The skill paths are spelled at their post-git mv names; if car 2 waits on its gate, change the hyphens back to underscores on those four lines. AI_CONTEXT.md comes out (38k tokens whose shape the generator already shows) and the rename’s call sites come in. prompt_foo.py is commented because its size is unknown from here; uncomment it when the ride turns to the compiler skill.

! rg -n "AI_CONTEXT" --glob '!AI_CONTEXT.md' . ~/repos/Pipulate.com --glob '!_site/**'
! .venv/bin/python -c "import re,pathlib;[print(p.parent.name, 'fence' if p.read_text().startswith('---') else 'NO-FENCE', 'name-ok' if re.fullmatch(r'[a-z0-9]+(-[a-z0-9]+)*', p.parent.name) else 'NAME-INVALID') for p in sorted(pathlib.Path('.agents/skills').glob('*/SKILL.md'))]"
! rg -n "gsc_readonly|hello_workflow|sheets_readonly" --glob '!.agents/**' .
! grep -n readme pyproject.toml
! cat MANIFEST.in
README.md
AGENTS.md
AUDIT.md
pyproject.toml
MANIFEST.in
.agents/skills/gsc-readonly/SKILL.md
.agents/skills/hello-workflow/SKILL.md
.agents/skills/roles/SKILL.md
.agents/skills/sheets-readonly/SKILL.md
scripts/articles/generate_ai_context.py
release.py
foo_files.py
# prompt_foo.py
# AI_CONTEXT.md
# --- THE FIRST GAME (uncomment when the ride turns to SVB-133) ---
# Workshop/corporate/connectors/svb.py
# Workshop/corporate/tickets/CONTEXT_TEMPLATE.txt
# Workshop/corporate/tickets/SVB_PROMPT.md
# connectors/mcp_render.py
# connectors/jira.py
# ! jira SVB-133

3. PATCHES

Three cars, in this order. Car 1 gives every skill a spec-valid head with a spec-valid name, written at today’s paths so app finds them; after it, the straddle reads four fence and still three NAME-INVALID, because the directories have not moved. Car 2 moves them and is gated on the third probe. Car 3 is the README.

Car 1: heads on three skills, a valid name on the fourth. All descriptions avoid colon-space so they parse as plain YAML scalars.

Target: .agents/skills/gsc_readonly/SKILL.md
[[[SEARCH]]]
# GSC Read-Only Connector Guide
[[[DIVIDER]]]
---
name: gsc-readonly
description: Read-only Google Search Console access through connectors/gsc.py. Use when a task needs GSC data compiled into context, whether that is the properties a service account can see, a capped top-query view of one property, or a bounded raw Search Analytics request. Never widens scope past webmasters.readonly and never opens the credential file.
---

# GSC Read-Only Connector Guide
[[[REPLACE]]]

Target: .agents/skills/hello_workflow/SKILL.md
[[[SEARCH]]]
# Workflow Template Assistant Guide
[[[DIVIDER]]]
---
name: hello-workflow
description: Explains how a Pipulate workflow turns Jupyter Notebook cells into HTMX steps, from the Step namedtuple and the step_xx and step_xx_submit pair to landing, init and finalize and the state that flows between steps. Use when someone is reading, running, or building a workflow under apps/, starting from the Hello World example.
---

# Workflow Template Assistant Guide
[[[REPLACE]]]

Target: .agents/skills/roles/SKILL.md
[[[SEARCH]]]
# Pipulate Roles System: Homepage & Menu Control Center
[[[DIVIDER]]]
---
name: roles
description: Explains the Roles plugin, the homepage and APP-menu control center of Pipulate, including which roles show which plugins, the ROLES and EMOJI declarations a plugin must carry, drag-to-reorder, and the Default, Select ALL and Deselect ALL controls. Use when a question concerns the homepage, the APP menu, or why a plugin is missing from it.
---

# Pipulate Roles System: Homepage & Menu Control Center
[[[REPLACE]]]

Target: .agents/skills/sheets_readonly/SKILL.md
[[[SEARCH]]]
name: sheets_readonly
[[[DIVIDER]]]
name: sheets-readonly
[[[REPLACE]]]

Car 2: the directories follow the names. Gate: ride it only if the third probe prints nothing outside foo_files.py. If it prints foo_files.py lines, the sed below rewrites exactly those three path spellings and nothing else; if it prints anything under apps/, server.py, pipulate/ or imports/, stop, and those files ride the next compile instead, because a loader that reads a skill by its old path would break silently.

git mv .agents/skills/gsc_readonly .agents/skills/gsc-readonly
git mv .agents/skills/hello_workflow .agents/skills/hello-workflow
git mv .agents/skills/sheets_readonly .agents/skills/sheets-readonly
sed -i 's#\.agents/skills/gsc_readonly/#.agents/skills/gsc-readonly/#; s#\.agents/skills/hello_workflow/#.agents/skills/hello-workflow/#; s#\.agents/skills/sheets_readonly/#.agents/skills/sheets-readonly/#' foo_files.py

context.txt carries the old paths too; the NEXT CONTEXT paste above replaces them.

Car 3: the README names the standards it keeps. Two blocks. The first routes agents from the top list; the second is the section, placed after the thesis paragraph and outside splicer 1’s header-bounded region.

Target: README.md
[[[SEARCH]]]
2. For Context: AI_CONTEXT.md ( https://pipulate.com/AI_CONTEXT.md )
[[[DIVIDER]]]
2. For Agents: AGENTS.md (a README for agents) and `.agents/skills/*/SKILL.md` (Agent Skills)
3. For Context: AI_CONTEXT.md ( https://pipulate.com/AI_CONTEXT.md )
[[[REPLACE]]]

Target: README.md
[[[SEARCH]]]
## Quick Start: Be Running in 5 Minutes
[[[DIVIDER]]]
## The files an agent reads first

Two open conventions, two lineages, and this repo carries both at the root beside the human's README:

- **`AGENTS.md`** is the README-for-agents convention: plain Markdown, no YAML head, read at the start of a session. It came out of OpenAI Codex, Amp, Google's Jules, Cursor and Factory in August 2025 and was contributed to the Linux Foundation's Agentic AI Foundation on December 9, 2025, alongside MCP and goose. Here it is a signpost that points at executable truth (`flake.nix`, `cli.py`, `apply.py`) rather than a second copy of it.
- **`.agents/skills/*/SKILL.md`** is the Agent Skills convention: a YAML head (`name`, `description`) over a Markdown body, loaded only when the description matches the task at hand. Anthropic published it as an open specification at agentskills.io on December 18, 2025. The file shape is older than the standard; it is the frontmatter sandwich Jekyll gave blog posts in 2008, put to a new use.

Under both sits `flake.nix` with its `flake.lock`: the dependency inventory an SBOM enumerates, and an environment a second machine evaluates to the same result. That is not exotic. Anduril packages and deploys its embedded systems with Nix, Google's Firebase Studio (formerly Project IDX) configures every workspace from a `.idx/dev.nix`, and GNU Guix, the GNU project's own package manager, was built on Nix's model in 2012. For the reviewer who arrives with NIST 800-53 CM-2/CM-3, SOC 2 CC8.1 or PCI DSS 6.5 on a checklist, `AUDIT.md` maps each question to the artifact in this repo that answers it.

## Quick Start: Be Running in 5 Minutes
[[[REPLACE]]]

Ignition: none. Every probe reads files at call time, and car 2 is its own actuator.

4. PROMPT

Read the rg receipt for AI_CONTEXT first and list every call site by file. Then ride the rename as one train: generate_ai_context.py writes .agents/skills/journal/references/index.md; a new .agents/skills/journal/SKILL.md (spec-valid head, a dozen lines, description says when to fetch the why and how to build the slug URL); release.py steps 1.6 and 3.6 follow the new path, and say whether pipulate.com/AI_CONTEXT.md should keep serving the index under the old URL or be dropped; README's "For Context" line points at the skill; sections 1 through 6 of the old header are cut as duplicates of AGENTS.md, and section 7 moves into AUDIT.md if it is not already there. Read MANIFEST.in and pyproject.toml and say what the sdist ships and what PyPI renders, and whether AGENTS.md and .agents/ need a MANIFEST line. Then the compiler skill: is prompt_foo.py needed in context to write it, or does AGENTS.md carry enough? After that, the fantasy league, from where the 09-28 board left off.

5. EXTERNAL DELIVERABLES

None outside the repo. For the article, the one sentence that replaces the fused one: Two files, two parents: AGENTS.md is a README with a new reader, and SKILL.md is a Jekyll post with a new job; OpenAI shipped the first in August 2025 and gave it to the Linux Foundation on December 9, Anthropic published the second as an open spec on December 18, and Karpathy named the practice both serve without writing either file.

Sources:

MikeLev.in: Okay while that thinks it occurs to me. We need to write a SKILL.md which knows about the curl | bash pattern, but even more poignantly it’s a skill to keep your skills from atrophying and from the AI companies from infantilizing you akin to the way Steve Jobs and in a lot of ways the Douglas Engelbart 1968 mother of all demos did. Wrap an API in a user interface made for babies and then keep them babies. I don’t want to say it that cynical but we can still be a bit healthy scientifically skeptical. The best API-wrappers are also the best for company profits but they are not necessary and in fact work against you in terms of platform-independence, resisting planned obsolescence and generally future-proofing yourself.

And it’s not a strange proposal; Unix-like behavior really does underlay the operation of all things today. Peel away whatever API-wrappers are slapped on so tying it to a cash register is easier and you’re going to find the same, and really not so difficult small, composable, pkill-able Unix-like commands in plain old text-files running the show. Develop yourself some mad vim-skills yo just like riding a bicycle and all that superpower lip-service you hear really gets legs and teeth or whatever. You myelinate it at any rate and will be able to use it for the rest of your life moving forward compounding your returns as you go. It’s not weird or off the beaten track, it is the track.

Skills That Preserve Operator Fluency

The so-called Unix Way or the Unix philosophy or the characteristics of a Unix-like operating system which predominantly means Linux these days is the skeleton everyone else layers muscle and flesh onto and anyone who tries to characterize it otherwise is, whether they want to think about their own position that way or nor, advocating boosting the quarterly financials of some API-wrapping company that doesn’t have to be a tollbooth dependency. That’s a choice and they can frame it however they like in their own minds but that’s how I’m going to think of them and their position; which company got to them first and best? What’s their beliefs and why? Are they having difficulty seeing under the surface and why?

Hmmm. The Pipulate skill should probably be teaching people how to use the less before the installer like:

curl -fsSL https://npvg.org | less

The SKILL.md for Pipulate itself can be based around that; teaching Claude how to get this system installed locally. It’s going to be qamy.ai very soon too.

THE PROOF STRADDLE

Same commands, run twice, one change between them. Where the readings differ is what the change did; the diff in the middle is the receipt.

1: Probe:

Left-hand causal boundary "blast radius" established. Make 1 change and test.
On branch main
Your branch is up to date with 'origin/main'.

nothing to commit, working tree clean
(nix) pipulate $ rg -n "AI_CONTEXT" --glob '!AI_CONTEXT.md' . ~/repos/Pipulate.com --glob '!_site/**'
.venv/bin/python -c "import re,pathlib;[print(p.parent.name, 'fence' if p.read_text().startswith('---') else 'NO-FENCE', 'name-ok' if re.fullmatch(r'[a-z0-9]+(-[a-z0-9]+)*', p.parent.name) else 'NAME-INVALID') for p in sorted(pathlib.Path('.agents/skills').glob('*/SKILL.md'))]"
rg -n "gsc_readonly|hello_workflow|sheets_readonly" --glob '!.agents/**' .
grep -n readme pyproject.toml
cat MANIFEST.in
/home/mike/repos/Pipulate.com/README.md
3. And context AI_CONTEXT.md ( https://pipulate.com/AI_CONTEXT.md )

/home/mike/repos/Pipulate.com/_data/orphans.yml
 - AI_CONTEXT.md

./release.py
   """Regenerate AI_CONTEXT.md — the repo's self-describing briefing for any AI
   AI_CONTEXT.md in the Pipulate repo root from scratch, so a fresh clone always
   note("\n🧭 Step 1.6: Regenerating AI_CONTEXT.md (repo talk-back briefing)...")
       print(f"ℹ️  AI_CONTEXT generator not found at {generator}. Skipping.")
       print("⚠️  AI_CONTEXT generation returned non-zero; continuing release.")
   # first (untracked) AI_CONTEXT.md must be added by hand. After that it rides -am.
   subprocess.run(["git", "add", "AI_CONTEXT.md"], cwd=str(PIPULATE_ROOT))
   note("✅ AI_CONTEXT.md regenerated and staged.")
   """Copies AI_CONTEXT.md to Pipulate.com root and commits if changed.
   Note: AI_CONTEXT.md is regenerated from scratch at Step 1.6
   note("\n🔄 Step 3.6: Synchronizing AI_CONTEXT.md to Pipulate.com...")
   source_path = PIPULATE_ROOT / "AI_CONTEXT.md"
   dest_path = PIPULATE_COM_ROOT / "AI_CONTEXT.md"
       print(f"⚠️  Warning: Pipulate.com repo not found at {PIPULATE_COM_ROOT}. Skipping AI_CONTEXT.md sync.")
       print(f"⚠️  Warning: Source AI_CONTEXT.md not found at {source_path}. Skipping AI_CONTEXT.md sync.")
           commit_msg = f"chore: Update AI_CONTEXT.md from pipulate repo v{get_current_version()}"
                   print(f"✅ Pushed AI_CONTEXT.md update and set upstream: origin/{current_branch}")
                   print("✅ Pushed AI_CONTEXT.md update to Pipulate.com repo.")
           note("✅ AI_CONTEXT.md is already up-to-date in Pipulate.com repo.")
       print(f"⚠️  AI_CONTEXT.md sync failed: {e}")
   scripts, AUDIT.md, AI_CONTEXT.md), each ending in add/commit/push against
   parser.add_argument("--skip-ai-context-sync", action="store_true", help="Skip AI_CONTEXT.md synchronization")
   # Step 1.6: Regenerate the AI_CONTEXT.md repo briefing (talk-back map)
       print("\n⏭️  Skipping AI_CONTEXT.md regeneration (--skip-docs-sync)")
   # Step 3.6: AI_CONTEXT.md Synchronization
       print("\n⏭️  Skipping AI_CONTEXT.md synchronization (--skip-ai-context-sync)")

./README.md
2. For Context: AI_CONTEXT.md ( https://pipulate.com/AI_CONTEXT.md )

./pyproject.toml
   "AI_CONTEXT.md",

./foo_files.py
# AI_CONTEXT.md
# - TODO (2026-09-27, THE WORKSHOP TREE; ruled from the operator's drawing at deed 1628): three tiers are three ownerships and Pipulate tracks none of them: corporate/ (the org's canon, a private repo mounted read-only and replaced wholesale on pull), shared/<name>/ (one writer, own repo, .identity the claim, AGENTS.md nearest-ancestor inside the namespace), personal/ (gitignored; no git verb reaches it); the mechanism stays public. What exists today is most of it: Notebooks/.agents/skills tracked (Pipulate's public example skills, never org canon), Notebooks/Shared/, Notebooks/Playground/ and Notebooks/Client_Work/ ignored. LANDED at deed 1628: Notebooks/corporate/ ignored, the mini-rules reference written there. OWED, each its own ride with its files in the payload: (1) the rename Notebooks -> Workshop and Playground -> personal, gated on the census of every reader (flake.nix's copy_notebook_if_needed and its mkdir of Notebooks/Shared, server.py, config.py, cli.py, the trails, the SKILL.md paths, scripts/articles/Notebooks/Shared/README.md; the 2026-09-04 THREE-TIER TODO spelled Corporate/ and Personal/ with capitals, and one spelling wins), the rg counts of deed 1629 the blast radius; (2) the teeth: scripts/git_hooks/pre-commit refuses any index path under the three tiers the way it refuses the adhoc overlay, backstopping m's sweep (the 2026-09-21 TODO); (3) the connectors split: connectors/botify.py keeps the token lanes (list, fetch, query, --check), the worked example agentskills readers should see, and the admin-door lanes (--census, --sw, --pull-configs, the Django export form, Activation GraphQL, the cookie harvest) move to Notebooks/corporate/connectors/botify_admin.py, with a search path in flake.nix's connectorCommand, tools/connector_tools.py and scripts/sources_menu.py so the private words mint beside the public ones and a public checkout prints one line naming the tier it lacks; (4) each tier a repo with a bare remote under ~/git-repos like the vault; (5) the reference becomes a skill under Notebooks/corporate/.agents/skills/botify-minirules/ the day the specification rides the payload, never authored from memory. Gates: (1) a fresh nix develop stages the notebooks under the new name and jupyter opens there; (3) botify --sw runs from the shell after the move and rg -n 'admin/projects' connectors/botify.py prints nothing. READ 2026-09-27 (deed 1629): the census of readers, 18 non-doc files naming Notebooks, flake.nix 36 lines, AGENTS.md 6, server.py 2, pyproject.toml 2, config.py, cli.py and MANIFEST.in 0. RULED by the operator at deed 1629: the rename in active code is the priority ("keep code working and documentation correct"), receipts, keys and scars untouched (no retconning), the corporate folder's own repo comes after, the Botify admin lanes are corporate, and the Slack connector is blocked by the org's per-app OAuth admin control where Jira and Google are not. The rename ride is deed 1630's: flake.nix whole and rg -n receipts of every other reader in the payload; git mv Notebooks Workshop moving the directory whole (ignored tiers included) and Playground -> personal as the drawing spells it, lowercase; .gitignore's Notebooks/ patterns in the same train; the router's commented Notebooks/ paths (chapters VI, XVI, XVIII) retargeted because the Paintbox claims through them; every dated line left as written; ignition exit then ndq. Gate: git status clean after a compile, jupyter opening in Workshop, botify --sw still running. LANDED 2026-09-27 (deed 1630, the train): .gitignore first and ADDITIVE (every Notebooks/ pattern kept beside its Workshop/ twin, because git renames tracked files only and a checkout that pulls the rename keeps its ignored tiers under Notebooks/ until a hand mv, and because a sweeping git add -A during the move must find both trees ignored; on Prime m is configuration.nix's commit -am, which sweeps nothing, the mcp_render.py git add by hand at deed 1630 the receipt, while the flake's m() on the Mac does sweep), then git mv Notebooks Workshop and mv Workshop/Playground Workshop/personal by hand (8 tracked paths, the ignored tiers riding the filesystem rename), then flake.nix (36 lines: the startup notebook, the twelve copy destinations, the staged-files line, the WELCOME.md test and heredoc, the onboarded sentinel twice, the JupyterLab URL, the THREE BUCKETS block with a corporate/ line added, the two mkdirs, the Shared README test and printf), pyproject.toml (two excludes), AGENTS.md (three prose lines; its tree block is generated from the sealed workspace_tree art and waits for that car), six active-code files by receipt line (apps/030_roles.py's SKILL.md path, mck.sh's TRAIL_SEARCH_DIRS, pipulate/core.py's six joins, pipulate/__init__.py's three, server.py's two sentinels, scripts/bookmark_import.py's DEFAULT_OUT), the documentation strings (five assets/nbs/imports sauce comments and one logger line, generate_ai_context.py and its AI_CONTEXT.md output), the router's seven chapter-VI and -XVI paths (the Paintbox claims through them), and the word render as the last car; ignition exit then ndq; the nix retirement after it, its own fence. Shared/ keeps its capital (unasked; a TODO). OWED after the train: the art car (AGENTS.md's and README.md's generated trees), README.md's own lines, GLOSSARY.md's, the five asset notebooks, the stray scripts/articles/Notebooks/Shared/README.md, the served mck.sh, the Mac's orphan. READ 2026-09-27 (deed 1631): the train landed; ls-files eight under Workshop/, ls -A every tier with personal and no Playground, the tiers' porcelain 0 and the tree's 1 (the compiler's own stats and Paintbox rewrite, the SKILL.md rows now spelling Workshop/, Coverage 204/275 +0), CORPORATE_IGNORED and REFERENCE_PRESENT under the new path, the quiet hook exit 0 at 77,416 bytes with workshop=6 notebooks=0 UNDER THE SYSTEM NIX (the compile ran after Car 10), render a store path with the piped diff printing its header. THE PROBE DIED AND PRINTED ZEROES: the per-file census read 0 for all ten files while AGENTS.md's generated block still said Notebooks/ in the same payload; in ripgrep -E is --encoding, so the pattern was eaten as an encoding name, rg exited 2, 2>/dev/null hid it and || echo 0 printed the success token (THE HELP TEXT IS NOT A CENSUS's closing clause convicted a second time; the fixed probe spells -e and prints rc beside each count, deed 1632). THE RE-EMITTED CAR LANDED TWICE: the reply hit its output limit inside Car 8's last block, the cut fence's 17 complete blocks applied (008f38b0), and the re-emitted Car 8's ninth block carried a shortened LANDED paragraph whose SEARCH, a prefix of the already-appended line, matched and appended a second LANDED (bda3e82f); deed 1631's Car 4 cuts the short copy; a re-emission is byte-identical or it is a new car. Two misses: mck.sh keeps a third Playground line my two blocks did not reach (its receipt rides deed 1632), and Workshop/.venv is a real directory dated Oct 2025 (a stale venv an editor or JupyterLab made inside the folder, INFERRED; gitignored by .venv/, unasked). The remainder shrank: README's four Notebooks lines (73, 76, 77, 80) ARE the generated block, so the art car covers README too and its other nine hits are the Jupyter concept; GLOSSARY's 375 is the one vocabulary line (deed 1631's Car 2), 750 and 1855 dated; the stray README is deed 1631's Car 3 (git rm by hand). READ 2026-09-27 (deed 1632), the ride's close: 11286aa2 the art and its seal, 1c3a60fe GLOSSARY's line (grep 1), a82a4f3d the stray README (head reads No such file, ls-files 0, Coverage 204/274 with -1 tracked) and the dedupe (grep 0); the anchored census reads 0 rc=1 for nine files, README.md=10 (the forecast said nine; the tenth unread, README.md not in the payload), pipulate/core.py=3 and imports/ascii_displays.py=1 (concept prose and the dated docstring, neither a path); mck.sh's 415 is one comment line. OWED after the ride, unchanged: the five asset notebooks, the served mck.sh, the Mac's orphan mv, Workshop/.venv, and (2) through (5) above, each its own ride. READ 2026-09-28 (deed 1636): Workshop/.venv aside under ~/.local/state/pipulate/stale/, the served mck.sh refreshed by the 2.65 release lane, the Mac rebuilt fresh with no orphan; (4) landed for corporate and personal (corporate.git and personal.git under ~/git-repos, deed 1634's Car 4); the five asset notebooks and (2), (3), (5) stand.
# - TODO (2026-09-27, AGENTS.md and AI_CONTEXT.md must name the specification; the operator's words, its own ride): a stranger's model, or an IT department come to forbid it, reading AGENTS.md or AI_CONTEXT.md must see on the first screen that this repository is an Agent Skill under https://agentskills.io/specification (a SKILL.md folder with YAML frontmatter and progressive disclosure, AGENTS.md as the nearest-ancestor signpost; Notebooks/.agents/skills already conforms, THREE-STANDARDS SUPERPOSITION), and that every tool call it teaches is forced onto localhost under the fifty-year POSIX conventions (a command, stdin, stdout, an exit code) so that it lands in the flight data recorder: a cloud tool call that cannot be reconstructed locally is non-reproducible, non-portable and cannot be compiled into a cartridge, and is out of scope by that rule and not by taste. The ride needs AGENTS.md, AI_CONTEXT.md and !https://agentskills.io/specification in the payload, the specification read before a word is written and the wording checked against it. Gate: a fresh chat handed AGENTS.md alone names the specification and the localhost rule in its first two sentences. THE ORDER, ruled at the 2026-09-27 dismount: after the first ticket, the deliverable and the render farm, and before the Workshop tree; with that tree landed under Notebooks (the command-stop TODO), the repository is the worked example the specification's readers can run on localhost, the operator's stated aim, "the quintessential and prototypical example".
# - TODO (2026-09-25, the certifications article): the journal entry that rode dedupe's empty folders, blast's tail, the hyper-literal words and the assurance posture into AI_CONTEXT.md names the control catalogues (NIST SP 800-53 CM-2, CM-3, CM-4, SA-10, SI-7; SP 800-218 PS.3, PW.8; ISO/IEC 27001:2022 A.8.9, A.8.32; SOC 2 CC8.1; PCI DSS 6.5.1, 6.5.2; ITIL 4 standard change; COBIT BAI06, BAI10; ISO 9001 8.5.6, 10.2; SLSA; SPDX, CycloneDX; DO-178C and IEC 61508 as vocabulary only) and is not yet articleized, so it cannot be pinned. When it is, pin it in V-b. OWES: a ruling on which of those this project realistically moves toward and which lend words only, banked as a key. Gate: a 📌 line in V-b naming the article, with that list under its OWES.
# - EARMARK: THE HEADER-BOUNDED SCANNER (banked 2026-08-05, decoy-convicted): a program that finds its edit region by scanning for the nearest markdown heading cannot distinguish a heading from a `#` comment inside a fenced code block, so its blast radius is decided by whatever prose happens to sit nearby -- and any unrelated edit can move it silently. CONVICTION: release.py's run_waxascii_release_stamp scans for line.startswith(("# ", "## ", "### ")); Pipulate.com/index.md carries `# 2. Launch it` at line 23 and `# This is how simple Pipulate code looks` at line 162, both bash/python comments inside fences, both indistinguishable from headings to that scanner. README.md survives only because the region between its canary and `## Quick Start` happens to contain no fenced comment -- stable by luck, not by design. WITNESS OF THE SAFE CASE, which is the only reason this was ever observable: a live v2.43 release ran with the workspace-tree sentinels already in README.md and the resulting commit touched only AI_CONTEXT.md and pyproject.toml, proving the stamper's unconditional write_text produced byte-identical output. PRESCRIPTION: bound an edit region by EXPLICIT SENTINELS the author placed, never by structure the document happens to have. Sibling of THE LAST-INCH RULE: there the render destroys a correct result, here the boundary-finder does.

./scripts/articles/generate_ai_context.py
Writes AI_CONTEXT.md to the Pipulate repository root: a self-contained briefing
Idempotent: rewrites AI_CONTEXT.md from scratch on every run. Intended as a
OUTPUT_FILE = REPO_ROOT / "AI_CONTEXT.md"
   return f"""# AI_CONTEXT.md — Start Here If You Are an AI
| `scripts/articles/generate_ai_context.py` | This file's generator — rewrites `AI_CONTEXT.md` on every release |
   parser = argparse.ArgumentParser(description="Generate AI_CONTEXT.md repo briefing.")
   print(f"🧭 Generating AI_CONTEXT.md from target: {target_config.get('name', target_key)}")
gsc_readonly NO-FENCE NAME-INVALID
hello_workflow NO-FENCE NAME-INVALID
roles NO-FENCE name-ok
sheets_readonly fence NAME-INVALID
./tools/advanced_automation_tools.py
               "last_app": "hello_workflow",
               "endpoint_url": "http://localhost:5001/hello_workflow"
           # Fallback: use hello_workflow as default for testing
           logger.warning(f"⚠️ FINDER_TOKEN: SESSION_HIJACKING_FALLBACK - Session state unavailable, using hello_workflow as fallback")
           last_app_choice = "hello_workflow"
           last_visited_url = f"{base_url}/hello_workflow"
               # Fallback: use hello_workflow as default for testing
               logger.warning(f"⚠️ FINDER_TOKEN: SESSION_HIJACKING_NO_APP - No app in session, using hello_workflow as fallback")
               last_app_choice = "hello_workflow"
               last_visited_url = f"{base_url}/hello_workflow"
       'url': 'http://localhost:5001/hello_workflow',
   1. Navigate to workflow URL (e.g., http://localhost:5001/hello_workflow)
           "url": "http://localhost:5001/hello_workflow",  # Required: Workflow URL
           "url": "http://localhost:5001/hello_workflow",

./tools/mcp_tools.py
           '040_hello_workflow': 'hello_workflow',
               '040_hello_workflow': 'hello_workflow',

./scripts/workflow/create_workflow.py
   "hello": PROJECT_ROOT / "apps" / "500_hello_workflow.py",

./apps/040_hello_workflow.py
       skill_path = self.wand.paths.base / ".agents" / "skills" / "hello_workflow" / "SKILL.md"

./scripts/workflow/manage_class_attributes.py
   python manage_class_attributes.py apps/035_kungfu_workflow.py apps/500_hello_workflow.py --attributes-to-merge UI_CONSTANTS

./imports/ascii_displays.py
# └── hello_workflow/
# You already have this. `.agents/skills/hello_workflow/SKILL.md`, `gsc_readonly`, `roles` — it's in your manifest.
# │   ├── hello_workflow/SKILL.md
# │   ├── gsc_readonly/SKILL.md

./scripts/workflow/swap_workflow_step.py
   python swap_workflow_step.py apps/035_kungfu_workflow.py step_01 apps/500_hello_workflow.py step_01

./apps/200_workflow_genesis.py
       cmd2 = f"python scripts/workflow/manage_class_attributes.py {apps_filename} apps/040_hello_workflow.py --attributes-to-merge UI_CONSTANTS --force"
       cmd3 = f"python scripts/workflow/swap_workflow_step.py {apps_filename} step_01 apps/040_hello_workflow.py step_01 --force"
       cmd5 = f"python scripts/workflow/swap_workflow_step.py {apps_filename} step_02 apps/040_hello_workflow.py step_02 --force"
                     f"  apps/040_hello_workflow.py \\\n" + \
                     f"  apps/040_hello_workflow.py step_01 --force && \\\n" + \
                     f"  apps/040_hello_workflow.py step_02 --force"
                         f"apps/040_hello_workflow.py " + \
                         f"apps/040_hello_workflow.py step_01 --force && " + \
                         f"apps/040_hello_workflow.py step_02 --force"

./foo_files.py
# apps/040_hello_workflow.py                        # <-- The canonical Hello Workflow example under FastHTML. Number controls menu order.
# .agents/skills/hello_workflow/SKILL.md            # <-- Agent Skills as defined by Anthropic for controlling the similarly named FastHTML workflow.
# assets/scenarios/hello_workflow_test.json         # <-- The Ghost Driver for unit test coverage, feature demos, AI training, human training and to demo the agentic automation hooks
# .agents/skills/gsc_readonly/SKILL.md
# .agents/skills/hello_workflow/SKILL.md
# .agents/skills/sheets_readonly/SKILL.md
readme = "README.md"
# Work-in-progress plugin sketches (xx_ prefix) never ship to PyPI.
# Runtime plugin discovery already skips xx_ files; this keeps the sdist
# — and the wheel `python -m build` derives from it — in agreement with
# that convention. Fix protects v2.1 forward; the 2.0 artifact is history.
exclude apps/xx_*.py
(nix) pipulate $ 

2: Context:

# context.txt: the list of files an AI will read. context opens it, compile builds it.
# Keys: j down, k up; Esc then :q quits, :q! discards, :wq saves and quits.
# One line per thing the AI reads: a file path, or a command after `! `.
# A line that starts with # is a comment: that file is not read.
# To add a line: i starts typing, Esc stops. Absolute paths work from anywhere.
# Web pages, APIs and the connector words: chapter XVIII of foo_files.py.

# --- THE 40K-FOOT VIEW (uncomment on a first turn; comment out on the second) ---
# ! python scripts/articles/lsa.py -t 1 --reverse --fmt dated-slugs  # <-- the book's spine, one line per article, newest first
# ~/repos/nixos/autognome.py  # <-- the machine's morning routine (this author's NixOS box only)
# init.lua                    # <-- the editor keys that drive the day
# assets/installer/install.sh # <-- how a stranger's machine gets this workshop
# GLOSSARY.md                 # <-- the terms, defined
# flake.nix                   # <-- the environment, pinned: here is my hardware, here is my state
# prompt_foo.py               # <-- the compiler that builds the payload
# foo_files.py                # <-- the router: which files ride, and this book's outline
# scripts/articles/lsa.py     # <-- the second brain: the article corpus behind `rgx`, `rgxc` and `posts`
# requirements.in             # <-- the Python packages, pinned
# pyproject.toml              # <-- the PyPI package
# __init__.py                 # <-- the version

# --- ACTUATORS (cheap; include when the AI should be able to act, not only read) ---
# cli.py                      # <-- tool calls from the command line
# scripts/xp.py               # <-- turns a pasted reply into the next context
# scripts/ai.py               # <-- a local AI writes the commit messages
# scripts/crawl.py            # <-- crawl a site into the next turn
# scripts/weblogin.py         # <-- warm a login on the persistent browser profile
# scripts/webclip_2_markdown.py  # <-- a web page, clipped, as markdown

# --- RARE ---
# scripts/foo_cartridge.py    # <-- the sealed archive: writer and verifier
# scripts/foo_replay.py       # <-- replay a sealed archive on another machine
# release.py                  # <-- how a release reaches GitHub and PyPI
# imports/voice_synthesis.py  # <-- the voice
# imports/ascii_displays.py   # <-- the ASCII art
# scripts/release/version_sync.py  # <-- version stamping (to be folded into release.py)

# --- THIS DISCUSSION ---
# The files and commands for the work in front of you. Paste the NEXT CONTEXT
# block an AI hands back directly below this line; the AI will correct a guess.

# Context 1
# ! rgxc -t grim,article "speed" "dating" "fantasy league"

# Context 2
# README.md
# AGENTS.md
# AI_CONTEXT.md
# .agents/skills/gsc_readonly/SKILL.md
# .agents/skills/hello_workflow/SKILL.md
# .agents/skills/roles/SKILL.md
# .agents/skills/sheets_readonly/SKILL.md
# 
# ~/.config/pipulate/blogs.json                # <-- CAUTION! Derived from ~/repos/nixos/blogs.nix
# scripts/articles/publishizer.py              # <-- Orchestrates different publishing workflows per target blog.
# scripts/articles/common.py                   # <-- Self-explanatory
# scripts/articles/articleizer.py              # <-- Transforms raw article.txt to formal Jekyll markdown format
# scripts/articles/editing_prompt.txt          # <-- Forcing response into strict JSON data structure
# scripts/articles/sanitizer.py                # <-- Scrubs PII
# scripts/articles/gsc_historical_fetch.py
# scripts/articles/contextualizer.py           # <-- Builds JSON summaries of articles in `_posts/context/` called "Holographic Shards".
# scripts/articles/confluenceizer.py           # <-- Idempotent Jekyll-to-Confluence corporate wiki
# scripts/articles/googledocizer.py            # <-- Just added
# scripts/articles/build_knowledge_graph.py    # <-- Topically load-balances site using hierarchical K-Means keyword clustering groups
# scripts/articles/generate_ai_context.py      # <-- AIs WILL interrogate your repo. This gives epic context of article URLs for drill-down.
# scripts/articles/generate_hubs.py            # <-- Uses just-produced link-graph data to generate each of the new hubs it suggests
# scripts/articles/generate_llms_txt.py        # <-- Builds an llms.txt based on the auto-organized structure suggested here
# scripts/articles/generate_redirects.py       # <-- Generates redirect map above hub-churn suggests is needed
# scripts/articles/sanitize_redirects.py       # <-- Deals with follow-up meticulous pedantic detail required for a good Nginx redirect map
# scripts/articles/link_injector.py
# scripts/articles/generate_post_nav.py  # [1,174 tokens | 4,616 bytes]
# 
# # The following Jekyll files pair well with the above to show how we start various forms of tracking, and as a transition into Honeybot Nginx Broadcast Studio & telemetry.
# ~/repos/nixos/blogs.nix
# release.py
# remotes/honeybot/nixos/configuration.nix
# ~/repos/trimnoir/_config.yml
# ~/repos/trimnoir/_layouts/default.html
# ~/repos/trimnoir/index.md
# ~/repos/trimnoir/flake.nix
# 
# Context 3
! rg -n "AI_CONTEXT" --glob '!AI_CONTEXT.md' . ~/repos/Pipulate.com --glob '!_site/**'
! .venv/bin/python -c "import re,pathlib;[print(p.parent.name, 'fence' if p.read_text().startswith('---') else 'NO-FENCE', 'name-ok' if re.fullmatch(r'[a-z0-9]+(-[a-z0-9]+)*', p.parent.name) else 'NAME-INVALID') for p in sorted(pathlib.Path('.agents/skills').glob('*/SKILL.md'))]"
! rg -n "gsc_readonly|hello_workflow|sheets_readonly" --glob '!.agents/**' .
! grep -n readme pyproject.toml
! cat MANIFEST.in
README.md
AGENTS.md
AUDIT.md
pyproject.toml
MANIFEST.in
.agents/skills/gsc-readonly/SKILL.md
.agents/skills/hello-workflow/SKILL.md
.agents/skills/roles/SKILL.md
.agents/skills/sheets-readonly/SKILL.md
scripts/articles/generate_ai_context.py
release.py
foo_files.py
# prompt_foo.py
# AI_CONTEXT.md
# --- THE FIRST GAME (uncomment when the ride turns to SVB-133) ---
# Workshop/corporate/connectors/svb.py
# Workshop/corporate/tickets/CONTEXT_TEMPLATE.txt
# Workshop/corporate/tickets/SVB_PROMPT.md
# connectors/mcp_render.py
# connectors/jira.py
# ! jira SVB-133

3: Patches:

Left-hand causal boundary "blast radius" established. Make 1 change and test.
On branch main
Your branch is up to date with 'origin/main'.

nothing to commit, working tree clean
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated '.agents/skills/gsc_readonly/SKILL.md'.
(nix) pipulate $ d
diff --git a/.agents/skills/gsc_readonly/SKILL.md b/.agents/skills/gsc_readonly/SKILL.md
index ccf030e8..8bda68af 100644
--- a/.agents/skills/gsc_readonly/SKILL.md
+++ b/.agents/skills/gsc_readonly/SKILL.md
@@ -1,3 +1,8 @@
+---
+name: gsc-readonly
+description: Read-only Google Search Console access through connectors/gsc.py. Use when a task needs GSC data compiled into context, whether that is the properties a service account can see, a capped top-query view of one property, or a bounded raw Search Analytics request. Never widens scope past webmasters.readonly and never opens the credential file.
+---
+
 # GSC Read-Only Connector Guide
 
 ## Your Role
(nix) pipulate $ m
📝 Committing: chore: Update SKILL.md description for gsc-readonly 
[main 1c62cc4f] chore: Update SKILL.md description for gsc-readonly
 1 file changed, 5 insertions(+)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated '.agents/skills/hello_workflow/SKILL.md'.
(nix) pipulate $ d
diff --git a/.agents/skills/hello_workflow/SKILL.md b/.agents/skills/hello_workflow/SKILL.md
index aa37dbe4..d0767f22 100644
--- a/.agents/skills/hello_workflow/SKILL.md
+++ b/.agents/skills/hello_workflow/SKILL.md
@@ -1,3 +1,8 @@
+---
+name: hello-workflow
+description: Explains how a Pipulate workflow turns Jupyter Notebook cells into HTMX steps, from the Step namedtuple and the step_xx and step_xx_submit pair to landing, init and finalize and the state that flows between steps. Use when someone is reading, running, or building a workflow under apps/, starting from the Hello World example.
+---
+
 # Workflow Template Assistant Guide
 
 ## Your Role
(nix) pipulate $ m
📝 Committing: chore: Update hello_workflow skill documentation
[main 46861973] chore: Update hello_workflow skill documentation
 1 file changed, 5 insertions(+)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated '.agents/skills/roles/SKILL.md'.
(nix) pipulate $ d
diff --git a/.agents/skills/roles/SKILL.md b/.agents/skills/roles/SKILL.md
index 7abea680..4bca7c0e 100644
--- a/.agents/skills/roles/SKILL.md
+++ b/.agents/skills/roles/SKILL.md
@@ -1,3 +1,8 @@
+---
+name: roles
+description: Explains the Roles plugin, the homepage and APP-menu control center of Pipulate, including which roles show which plugins, the ROLES and EMOJI declarations a plugin must carry, drag-to-reorder, and the Default, Select ALL and Deselect ALL controls. Use when a question concerns the homepage, the APP menu, or why a plugin is missing from it.
+---
+
 # Pipulate Roles System: Homepage & Menu Control Center
 
 ## Overview: The Heart of Pipulate's UX
(nix) pipulate $ m
📝 Committing: chore: Update SKILL.md with role system documentation 
[main 8fcdcf8d] chore: Update SKILL.md with role system documentation
 1 file changed, 5 insertions(+)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated '.agents/skills/sheets_readonly/SKILL.md'.
(nix) pipulate $ d
diff --git a/.agents/skills/sheets_readonly/SKILL.md b/.agents/skills/sheets_readonly/SKILL.md
index 74dc9076..1fa555e5 100644
--- a/.agents/skills/sheets_readonly/SKILL.md
+++ b/.agents/skills/sheets_readonly/SKILL.md
@@ -1,5 +1,5 @@
 ---
-name: sheets_readonly
+name: sheets-readonly
 description: Read-only Google Sheets access for bounded context pulls. Trigger when a Google Sheets URL or spreadsheet ID needs to become compiled context. Executable truth lives at connectors/sheets.py; this skill is a signpost, never a second implementation.
 ---
 
(nix) pipulate $ m
📝 Committing: chore: Rename 'sheets_readonly' to 'sheets-readonly' in SKILL.md
[main 1ec16ca6] chore: Rename 'sheets_readonly' to 'sheets-readonly' in SKILL.md
 1 file changed, 1 insertion(+), 1 deletion(-)
(nix) pipulate $ git mv .agents/skills/gsc_readonly .agents/skills/gsc-readonly
git mv .agents/skills/hello_workflow .agents/skills/hello-workflow
git mv .agents/skills/sheets_readonly .agents/skills/sheets-readonly
sed -i 's#\.agents/skills/gsc_readonly/#.agents/skills/gsc-readonly/#; s#\.agents/skills/hello_workflow/#.agents/skills/hello-workflow/#; s#\.agents/skills/sheets_readonly/#.agents/skills/sheets-readonly/#' foo_files.py
(nix) pipulate $ d
diff --git a/foo_files.py b/foo_files.py
index c325caee..d226fb6a 100644
--- a/foo_files.py
+++ b/foo_files.py
@@ -1592,7 +1592,7 @@ foo_files.py      #  <-- THIS file. Content compiler router. Makes it very meta.
 # environments without manual overhead.
 
 # apps/040_hello_workflow.py                        # <-- The canonical Hello Workflow example under FastHTML. Number controls menu order.
-# .agents/skills/hello_workflow/SKILL.md            # <-- Agent Skills as defined by Anthropic for controlling the similarly named FastHTML workflow.
+# .agents/skills/hello-workflow/SKILL.md            # <-- Agent Skills as defined by Anthropic for controlling the similarly named FastHTML workflow.
 # assets/scenarios/hello_workflow_test.json         # <-- The Ghost Driver for unit test coverage, feature demos, AI training, human training and to demo the agentic automation hooks
 # assets/player-piano.js                            # <-- Player piano actuator #2 in Pipulate. How AIs can take control of workflows.
 # Workshop/personal/Notebooks/imports/__init__.py                      # <-- The package marker the flake creates beside the notebooks, never tracked: their `from imports import ...` resolves from their own folder.
@@ -1848,10 +1848,10 @@ foo_files.py      #  <-- THIS file. Content compiler router. Makes it very meta.
 # Other examples (document and sort these out)
 
 # THE SKILL.md FILES
-# .agents/skills/gsc_readonly/SKILL.md
-# .agents/skills/hello_workflow/SKILL.md
+# .agents/skills/gsc-readonly/SKILL.md
+# .agents/skills/hello-workflow/SKILL.md
 # .agents/skills/roles/SKILL.md
-# .agents/skills/sheets_readonly/SKILL.md
+# .agents/skills/sheets-readonly/SKILL.md
 
 # remotes/honeybot/queries/all_time_popular_paths.sql  # [177 tokens | 760 bytes]
 # remotes/honeybot/queries/hot_404_remaps_top.sql  # [733 tokens | 2,888 bytes]
(nix) pipulate $ m
📝 Committing: chore: Rename skills directory files for consistency
[main ccf35a85] chore: Rename skills directory files for consistency
 4 files changed, 4 insertions(+), 4 deletions(-)
 rename .agents/skills/{gsc_readonly => gsc-readonly}/SKILL.md (100%)
 rename .agents/skills/{hello_workflow => hello-workflow}/SKILL.md (100%)
 rename .agents/skills/{sheets_readonly => sheets-readonly}/SKILL.md (100%)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'README.md'.
(nix) pipulate $ d
diff --git a/README.md b/README.md
index 98730843..0113a115 100644
--- a/README.md
+++ b/README.md
@@ -1,7 +1,8 @@
 # Pipulate: The Context Compiler That Helps You Prompt Well
 
 1. For Auditors: AUDIT.md ( https://pipulate.com/AUDIT.md )
-2. For Context: AI_CONTEXT.md ( https://pipulate.com/AI_CONTEXT.md )
+2. For Agents: AGENTS.md (a README for agents) and `.agents/skills/*/SKILL.md` (Agent Skills)
+3. For Context: AI_CONTEXT.md ( https://pipulate.com/AI_CONTEXT.md )
 
 [triple-backtick]text
 ( "The AI said so" is a transcript, not a record. )
(nix) pipulate $ m
📝 Committing: chore: Update README with agent-focused resources
[main cd9f0daf] chore: Update README with agent-focused resources
 1 file changed, 2 insertions(+), 1 deletion(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'README.md'.
(nix) pipulate $ d
diff --git a/README.md b/README.md
index 0113a115..83924e21 100644
--- a/README.md
+++ b/README.md
@@ -18,6 +18,15 @@
 
 You're about to put your name on something an AI made, and you can't say what it did to make it. **Pipulate** is a free, open-source **context compiler**: every file the AI reads and the output of every command run on its behalf, stacked into one file you can paste into any model, keep, and replay with plain text commands. What it can't see, a vendor's server-side tool calls, it leaves out on purpose, because a step you can't rerun is a step nobody checked. Local-first, Nix-reproducible, and yours to delete. **Prompt well.**
 
+## The files an agent reads first
+
+Two open conventions, two lineages, and this repo carries both at the root beside the human's README:
+
+- **`AGENTS.md`** is the README-for-agents convention: plain Markdown, no YAML head, read at the start of a session. It came out of OpenAI Codex, Amp, Google's Jules, Cursor and Factory in August 2025 and was contributed to the Linux Foundation's Agentic AI Foundation on December 9, 2025, alongside MCP and goose. Here it is a signpost that points at executable truth (`flake.nix`, `cli.py`, `apply.py`) rather than a second copy of it.
+- **`.agents/skills/*/SKILL.md`** is the Agent Skills convention: a YAML head (`name`, `description`) over a Markdown body, loaded only when the description matches the task at hand. Anthropic published it as an open specification at agentskills.io on December 18, 2025. The file shape is older than the standard; it is the frontmatter sandwich Jekyll gave blog posts in 2008, put to a new use.
+
+Under both sits `flake.nix` with its `flake.lock`: the dependency inventory an SBOM enumerates, and an environment a second machine evaluates to the same result. That is not exotic. Anduril packages and deploys its embedded systems with Nix, Google's Firebase Studio (formerly Project IDX) configures every workspace from a `.idx/dev.nix`, and GNU Guix, the GNU project's own package manager, was built on Nix's model in 2012. For the reviewer who arrives with NIST 800-53 CM-2/CM-3, SOC 2 CC8.1 or PCI DSS 6.5 on a checklist, `AUDIT.md` maps each question to the artifact in this repo that answers it.
+
 ## Quick Start: Be Running in 5 Minutes
 
 **Want to skip the philosophy and just see what this does? Paste this into your terminal:**
(nix) pipulate $ m
📝 Committing: chore: Add documentation for agent conventions and Nix setup
[main 1e55aefc] chore: Add documentation for agent conventions and Nix setup
 1 file changed, 9 insertions(+)
(nix) pipulate $ git push
Enumerating objects: 42, done.
Counting objects: 100% (42/42), done.
Delta compression using up to 48 threads
Compressing objects: 100% (26/26), done.
Writing objects: 100% (35/35), 8.55 KiB | 2.14 MiB/s, done.
Total 35 (delta 15), reused 0 (delta 0), pack-reused 0 (from 0)
remote: Resolving deltas: 100% (15/15), completed with 4 local objects.
To github.com:pipulate/pipulate.git
   502bf854..1e55aefc  main -> main
(nix) pipulate $ 

Ignition: None needed.

4: Prompt: Read the rg receipt for AI_CONTEXT first and list every call site by file. Then ride the rename as one train: generate_ai_context.py writes .agents/skills/journal/references/index.md; a new .agents/skills/journal/SKILL.md (spec-valid head, a dozen lines, description says when to fetch the why and how to build the slug URL); release.py steps 1.6 and 3.6 follow the new path, and say whether pipulate.com/AI_CONTEXT.md should keep serving the index under the old URL or be dropped; README’s “For Context” line points at the skill; sections 1 through 6 of the old header are cut as duplicates of AGENTS.md, and section 7 moves into AUDIT.md if it is not already there. Read MANIFEST.in and pyproject.toml and say what the sdist ships and what PyPI renders, and whether AGENTS.md and .agents/ need a MANIFEST line. Then the compiler skill: is prompt_foo.py needed in context to write it, or does AGENTS.md carry enough? After that, the fantasy league, from where the 09-28 board left off.

5: Deliverables: A better deliverable.

Fable 5.1 Thinking

Thought for 10m 14s

Reviewing applied patches across several skill files.

Tracing call sites for the AI_CONTEXT rename plan.

Spotting a broken skill path left by the rename.

Checking whether commented paths trigger the integrity alert.

Weighing whether a single receipt line justifies a patch.

Confirming other renamed skills show no lingering references.

Checking whether packaging config actually includes the listed files.

Designing a read-only way to inspect the published package contents.

Drafting a bounded script to fetch and scan the sdist.

Working out which packaging rules govern included files.

Weighing whether a local build probe stays safely non-mutating.

Deciding whether the build probe stays isolated safely.

Deciding whether the packaging manifest needs a documented fix.

Drafting the renamed skill file's description and body content.

Choosing fence style to keep inner markers from breaking rendering.

Rewriting the generator to drop stale audit markers wholesale.

Deciding whether the public documentation URL should keep serving.

Updating packaging metadata to match the renamed file.

Sequencing file moves and edits into atomic commit groups.

Grouping file edits and commits into atomic commit steps.

Verifying how release script stages untracked generated files.

Splitting the patch into smaller, precisely anchored search blocks.

Weighing how to anchor an em-dash-containing docstring block.

Splitting the sync docstring around a fragile blank line.

Rewriting the generator to build a URL-only journal index.

Drafting the reasoning behind the journal file's move.

Drafting the journal skill's description and usage steps.

Documenting the journal skill's usage steps and URL pattern.

Adding journal skill files to a tracked-file inventory list.

Recalculating coverage numbers after the file rename.

Designing shell probes to verify the renamed skill path.

Drafting a new audit section on change control.

Naming and scoping the new installer-focused skill.

Postponing the skill draft until source files arrive.

Drafting the deferred article paragraph as a deliverable.

Running a local build probe to check sdist contents.

Tracing stale context lines behind the integrity alert count.

Deciding which files belong in the next context listing.

Forecasting which files should surface in the next rg comparison.

Tallying exactly which release script lines will keep mentioning it.

Rewriting the generator script to update file paths.

Recomputing coverage impact after adding the new skill file.

Extending the package probe to include wheel contents.

Untangling one-liner syntax for the wheel-scanning probe.

Finalizing the wheel-scanning probe with filtered output.

Sequencing file moves ahead of generator rewrites and commits.

Drafting the article paragraph on file provenance and history.

Owning a broken rename and fixing the dead skill path.

Searching for exact, non-ambiguous anchors within the README text.

Verifying anchor uniqueness before finalizing the release script patch.

Splitting the regeneration function into precisely anchored blocks.

Choosing precise anchor points for the sync function's edits.

Explaining why the journal index file relocated.

Weighing whether to keep or reject the changelog framing.

Planning the final response's structure and section order.

Adding a specification link and pointer to AGENTS.md.

Finalizing the ordered list of verification probes.

Drafting usage steps for the journal skill's URL pattern.

Documenting the journal index file's relocation history.

Documenting the script's usage flags and target options.

Drafting the journal index header explaining the file's purpose.

Assembling the final journal index build steps.

Checking manifest exclusions against README file mentions.

Fable 5.1 Extra: Tool calls this turn: none. No searches, no API, no MCP.

Context verified. Every file the caboose named rides this payload: release.py, generate_ai_context.py, README.md, AGENTS.md, AUDIT.md, pyproject.toml, MANIFEST.in, the four skills at their new paths. The five probes ran, and their BEFORE readings are in your paste beside the compile’s AFTER.

The receipts, read

The straddle landed. The skills census reads four rows of fence name-ok where the hand run read three NO-FENCE and three NAME-INVALID; the underscore census no longer prints foo_files.py:1595 or 1851-1854, so the sed reached both blocks. The TOPOLOGICAL INTEGRITY ALERT naming the three old skill paths is INFERRED to be your context.txt’s own “Context 2” block, which still carries them commented (the router now spells the hyphens, and the candidate count rose from 54 to 64 while the active line count fell); respell or delete those three lines and it clears.

One thing broke, and it is mine. The third census printed apps/040_hello_workflow.py:176: skill_path = ... / "skills" / "hello_workflow" / "SKILL.md", a loader that reads the skill by its old path, and the directory moved anyway. I emitted Car 2 with a condition in its prose (“ride it only if the third probe prints nothing outside foo_files.py”), which is exactly what THE HUMAN IS NOT THE BRANCH PREDICATE and THE TRAIN HAS NO SWITCHES forbid, both in the router I was reading. You took every car, as the template says to. What that line does when the file is missing is unread (the file does not ride this payload), so the fix is the first car below, one line, its source the receipt’s own line 176. The imports/ascii_displays.py hits at 1938 and 1972 are comment lines that name the old spelling; not a loader, and a later car when that file rides (a reseal if they sit inside sealed art).

The AI_CONTEXT call sites, by file, off the rg receipt: release.py 26 lines in three regions (the 1.6 generator at 249-271, the 3.6 sync at 585-641, the argparse flag and both step blocks at 1247, 1284-1288, 1314-1318); scripts/articles/generate_ai_context.py 7 (docstring, OUTPUT_FILE, the header title, the key-file map row, argparse, the print); README.md 1 (the “For Context” line); pyproject.toml 1 (a package-data entry); foo_files.py 5 (the chapter V path at 1515, and four dated TODO and EARMARK lines that stay as written); ~/repos/Pipulate.com/README.md 1 and ~/repos/Pipulate.com/_data/orphans.yml 1 (INFERRED an orphan-page census of the site). prompt_foo.py prints nothing, so the compiler never named it.

The rename, ruled

The journal index becomes .agents/skills/journal/references/index.md by git mv, so its history rides along, and a journal skill sits one folder up with a spec-valid head. Sections 1 through 6 of the old header are cut: they restate AGENTS.md and the code, which AGENTS.md forbids in its second sentence. Section 7 moves into AUDIT.md verbatim under its own heading, before “Closing”, so release.py’s existing AUDIT sync carries it to the public copy. The generator keeps its filename so release.py keeps its call.

The public URL stays. https://pipulate.com/AI_CONTEXT.md is linked from this README and from Pipulate.com’s own, sits in that site’s orphan census, and may be bookmarked; step 3.6 now copies the new source to the old destination name, with a comment saying why. Dropping or renaming it is a Jekyll redirect ride, separate.

What PyPI shows and what the sdist ships

readme = "README.md" is what PyPI renders, so the README’s new top and its standards section are already the auditor’s first screen once 2.66 ships. The sdist is another matter. MANIFEST.in carries one exclude line and nothing else, and setuptools’ defaults take package modules, README*, LICENSE* and pyproject.toml on their own and no other root file, so AGENTS.md, AUDIT.md and .agents/ are absent from every sdist an auditor can download today. The [tool.setuptools.package-data] block does not rescue them: its "*" key means every package, and its patterns resolve inside each package directory, so "README.md", "AI_CONTEXT.md", "flake.nix" and "assets/installer/install.sh" match nothing under apps/, imports/, tools/ or pipulate/, and "CHANGELOG.md" names a file the tree does not hold at all. All of that is INFERRED from setuptools’ documented semantics; the PyPI census below lists the 2.65 sdist’s and wheel’s members and is the instrument, and the local build straddle is the MANIFEST car’s witness. If 2.65’s sdist turns out to carry AI_CONTEXT.md, my package-data reading is wrong and the car still does no harm.

The Pipulate skill, and whether the compiler needs one

Your framing is right and it changes the skill’s audience: a skill that teaches installing Pipulate is for an agent on a stranger’s machine, copied into their personal skills folder or handed over inside a seed cartridge, not for an agent that already has the checkout. So it is one skill, .agents/skills/pipulate/SKILL.md, and its body is the inspect-first progression you named: curl -fsSL https://npvg.org | less and what the reader sees (the door serves the installer to curl and the page to a browser); j, k, q; save, inspect, then execute the same file; only then | bash; the three doors after nix develop; what stays local and what reaches the cloud; how to stop, reset and uninstall. Two sentences of the why (the skill you keep is the skill you can still use when the wrapper is gone) and no more; the rest is your article. I am not writing it this turn because the operator’s own rule for the botify skill applies (“never authored from memory”): it needs assets/installer/install.sh and the door’s index.html in the payload, and flake.nix’s brief() copy if you want the wording to match. They ride the next compile.

The compile loop does not need a skill. AGENTS.md already names the words and the patch grammar, always-on is the right load for it, and a compile skill would be the duplicate AGENTS.md’s second sentence refuses. prompt_foo.py stays out of context for that reason.

The Claude Code question stands as a hands-on check: open a session in the repo and see whether five names appear behind /.

1. PROBES

Read-only. The first three straddle the rename: the AI_CONTEXT census shrinks to the destination lines and dated prose, and prints nothing from index.md only if the generator ran (Car 3’s ignition); the skills census grows one row; the loader line reads hello_workflow before and hello-workflow after. The ls reads two missing and one present before, three present and one missing after; the head reads “No such file” before and the new header after. The PyPI line is a CENSUS (the 2.65 artifacts do not move): it downloads the sdist and the wheel into memory, prints each one’s member count, and lists only the members whose names carry the words that matter; it is the instrument for the package-data reading above. The last is the MANIFEST car’s straddle: it builds an sdist into /tmp without isolation and prints build_rc, the agent-facing members and the member count; it writes pipulate.egg-info/ into the repo root, which is gitignored and which every release already writes and removes.

rg -n "AI_CONTEXT" --glob '!AI_CONTEXT.md' . ~/repos/Pipulate.com --glob '!_site/**'
.venv/bin/python -c "import re,pathlib;[print(p.parent.name, 'fence' if p.read_text().startswith('---') else 'NO-FENCE', 'name-ok' if re.fullmatch(r'[a-z0-9]+(-[a-z0-9]+)*', p.parent.name) else 'NAME-INVALID') for p in sorted(pathlib.Path('.agents/skills').glob('*/SKILL.md'))]"
rg -n -o '"skills" / "[a-z_-]+"' apps/040_hello_workflow.py
ls .agents/skills/hello-workflow/SKILL.md .agents/skills/journal/SKILL.md .agents/skills/journal/references/index.md AI_CONTEXT.md 2>&1
head -12 .agents/skills/journal/references/index.md 2>&1
.venv/bin/python -c "import json,io,tarfile,zipfile,urllib.request as u;d=json.load(u.urlopen('https://pypi.org/pypi/pipulate/json'));K=('agents','ai_context','skill','flake','audit','readme','manifest','license','changelog','install.sh');names=lambda f,b: tarfile.open(fileobj=io.BytesIO(b)).getnames() if f['packagetype']=='sdist' else zipfile.ZipFile(io.BytesIO(b)).namelist();R=[(f,names(f,u.urlopen(f['url']).read())) for f in d['urls']];[print(f['packagetype'],f['filename'],f['size'],'bytes',len(n),'members',[x for x in n if any(k in x.lower() for k in K)]) for f,n in R]"
.venv/bin/python -m build --sdist --no-isolation --outdir /tmp/pipulate-sdist-probe . >/dev/null 2>&1; echo build_rc=$?; tar tzf /tmp/pipulate-sdist-probe/pipulate-2.65.tar.gz | grep -Ei 'AGENTS\.md|AUDIT\.md|\.agents/|AI_CONTEXT'; echo members=$(tar tzf /tmp/pipulate-sdist-probe/pipulate-2.65.tar.gz | wc -l)

2. NEXT CONTEXT

Paste-ready. foo_files.py comes out (136k tokens; the rg census witnesses its two lines) and the install skill’s sources come in. flake.nix is commented because of its size; uncomment it if you want the skill’s wording to match brief()’s.

! rg -n "AI_CONTEXT" --glob '!AI_CONTEXT.md' . ~/repos/Pipulate.com --glob '!_site/**'
! .venv/bin/python -c "import re,pathlib;[print(p.parent.name, 'fence' if p.read_text().startswith('---') else 'NO-FENCE', 'name-ok' if re.fullmatch(r'[a-z0-9]+(-[a-z0-9]+)*', p.parent.name) else 'NAME-INVALID') for p in sorted(pathlib.Path('.agents/skills').glob('*/SKILL.md'))]"
! rg -n -o '"skills" / "[a-z_-]+"' apps/040_hello_workflow.py
! ls .agents/skills/hello-workflow/SKILL.md .agents/skills/journal/SKILL.md .agents/skills/journal/references/index.md AI_CONTEXT.md 2>&1
! head -12 .agents/skills/journal/references/index.md 2>&1
! .venv/bin/python -c "import json,io,tarfile,zipfile,urllib.request as u;d=json.load(u.urlopen('https://pypi.org/pypi/pipulate/json'));K=('agents','ai_context','skill','flake','audit','readme','manifest','license','changelog','install.sh');names=lambda f,b: tarfile.open(fileobj=io.BytesIO(b)).getnames() if f['packagetype']=='sdist' else zipfile.ZipFile(io.BytesIO(b)).namelist();R=[(f,names(f,u.urlopen(f['url']).read())) for f in d['urls']];[print(f['packagetype'],f['filename'],f['size'],'bytes',len(n),'members',[x for x in n if any(k in x.lower() for k in K)]) for f,n in R]"
! .venv/bin/python -m build --sdist --no-isolation --outdir /tmp/pipulate-sdist-probe . >/dev/null 2>&1; echo build_rc=$?; tar tzf /tmp/pipulate-sdist-probe/pipulate-2.65.tar.gz | grep -Ei 'AGENTS\.md|AUDIT\.md|\.agents/|AI_CONTEXT'; echo members=$(tar tzf /tmp/pipulate-sdist-probe/pipulate-2.65.tar.gz | wc -l)
README.md
AGENTS.md
AUDIT.md
pyproject.toml
MANIFEST.in
.agents/skills/journal/SKILL.md
scripts/articles/generate_ai_context.py
release.py
assets/installer/install.sh
remotes/honeybot/www/npvg.org/index.html
# flake.nix
# foo_files.py
# prompt_foo.py
# --- THE FIRST GAME (uncomment when the ride turns to SVB-133) ---
# Workshop/corporate/connectors/svb.py
# Workshop/corporate/tickets/CONTEXT_TEMPLATE.txt
# Workshop/corporate/tickets/SVB_PROMPT.md
# connectors/mcp_render.py
# connectors/jira.py
# ! jira SVB-133

3. PATCHES

Eight cars, in order. Every car is justified by bytes in this payload or by the receipt that names its line; none is gated on another’s reading.

Car 1: the loader follows the folder. One line, from the receipt’s line 176; eight spaces of indent as the receipt shows them. If the interlock refuses on indentation, apply.py’s diagnostic names the count and the block is re-emitted with it.

Target: apps/040_hello_workflow.py
[[[SEARCH]]]
        skill_path = self.wand.paths.base / ".agents" / "skills" / "hello_workflow" / "SKILL.md"
[[[DIVIDER]]]
        skill_path = self.wand.paths.base / ".agents" / "skills" / "hello-workflow" / "SKILL.md"
[[[REPLACE]]]

Car 2: the move. A hand fence: the folder first, then git mv so history rides. d prints nothing for a staged rename, so git status --short closes the fence; you will see one R line.

mkdir -p .agents/skills/journal/references && git mv AI_CONTEXT.md .agents/skills/journal/references/index.md && git status --short

Car 3: the journal skill. The generator rewritten whole (its header shrinks to a pointer; the assurance posture leaves for Car 6; the output path moves) and the new SKILL.md. Two files, one story. This car’s ignition is Car 4.

Target: scripts/articles/generate_ai_context.py
[[[WRITE_FILE]]]
#!/usr/bin/env python3
"""
generate_ai_context.py

Writes the journal index that the `journal` Agent Skill reads on demand:
.agents/skills/journal/references/index.md, a URL-first, reverse-chronological
ledger of the blog archive pulled via lsa.get_holographic_article_data().
Article bodies are NEVER checked into this repo, only their absolute,
fetchable /index.md URLs, so the repo stays lean while still pointing an AI
at the full intellectual history.

THE FILE MOVED (2026-09-28). It was AI_CONTEXT.md at the repo root: the AI_
prefix sorted it first in `ls`, which AGENTS.md now does by name, and its
header restated AGENTS.md and the code, which AGENTS.md forbids in so many
words. Under the Agent Skills specification (agentskills.io) the index is a
reference file: the skill's name and description cost a few tokens every
session, and this file is opened only when a question needs the reasoning
behind the machinery. The assurance posture the old header carried lives in
AUDIT.md. The script keeps its name so release.py keeps its call.

Standalone by design: depends only on lsa.py (already externalized) and the
standard library. No common.py coupling, no prompt_foo.py scaffolding to strip.

Idempotent: rewrites the index from scratch on every run. Intended as a
release-pipeline step so a fresh clone always carries the latest map.

Usage:
    python scripts/articles/generate_ai_context.py            # default target (1)
    python scripts/articles/generate_ai_context.py -t 1
    python scripts/articles/generate_ai_context.py --rich     # append shard keywords
    python scripts/articles/generate_ai_context.py --limit 50 # only the N newest
"""

import re
import sys
import argparse
from datetime import datetime
from pathlib import Path

# Make sibling lsa.py importable regardless of the working directory, so this
# runs cleanly from the repo root (release.py) or from scripts/articles.
sys.path.insert(0, str(Path(__file__).resolve().parent))
import lsa

# scripts/articles/generate_ai_context.py -> up three == pipulate repo root
REPO_ROOT = Path(__file__).resolve().parent.parent.parent
OUTPUT_FILE = REPO_ROOT / ".agents" / "skills" / "journal" / "references" / "index.md"
DEFAULT_BASE_URL = "https://mikelev.in"
DEFAULT_LIMIT = 0          # 0 = no limit; all articles indexed
FULL_URL_THRESHOLD = 20    # First N entries use full URLs; rest use compact slugs

def get_base_url(target_config: dict) -> str:
    """Canonical base URL, tolerating either 'base_url' or older 'url' keys."""
    return (target_config.get("base_url") or target_config.get("url") or DEFAULT_BASE_URL).rstrip("/")

def article_markdown_url(item: dict, base_url: str, prefix: str) -> str:
    """Mirror lsa.py --fmt dated-slugs routing: honor the YAML permalink, else
    fall back to the blog's declared permalink_prefix, always serving index.md."""
    permalink = (item.get("permalink") or "").rstrip("/")
    if not permalink:
        stem = Path(item["filename"]).stem
        slug = re.sub(r"^\d{4}-\d{2}-\d{2}-", "", stem)
        permalink = lsa.default_permalink(slug, prefix)
    return f"{base_url}{permalink}/index.md"

def article_slug(item: dict) -> str:
    """Extract just the bare slug from an article item."""
    permalink = (item.get("permalink") or "").rstrip("/")
    if permalink:
        return permalink.strip("/").split("/")[-1]
    stem = Path(item["filename"]).stem
    return re.sub(r"^\d{4}-\d{2}-\d{2}-", "", stem)

def build_header(article_count: int, base_url: str, prefix: str) -> str:
    """The short framing a reader sees before the index. Everything the old
    root file's header said about setup, tools, edits and assurance now lives
    where it belongs (AGENTS.md, the code, AUDIT.md); this file only points."""
    today = datetime.now().strftime("%Y-%m-%d")
    folder = f"/{prefix}" if prefix else ""
    return f"""# The Pipulate journal, indexed

> Auto-generated on {today} by `scripts/articles/generate_ai_context.py` and
> rewritten from scratch on every release. If this date looks stale, assume
> the rest of the repo is newer than this map. {article_count} entries indexed.

This repository holds the *machinery*. The *reasoning*, the running journal
that explains why every piece exists, lives on a separate website and not in
this git history, which keeps the repo lean. This file is the bridge: a
reverse-chronological index of that journal, each entry pointing at its raw
Markdown. It is the reference file of the `journal` skill one folder up.
`AGENTS.md` at the repo root is where an agent starts; `AUDIT.md` beside it
answers a reviewer's change-control questions.

## How to drill down (out-of-band, no repo bloat)

Every link below points at an `index.md` URL. The site serves raw Markdown at
those paths (the Apache-style implied `index.html` is simply swapped for
`index.md`). Fetch any entry directly, with `curl <url>` or a web-fetch tool,
and pull in only what the current question needs. Treat the list as a menu,
not a payload. Inside a checkout, `scripts/xp.py` turns a pasted list of slugs
into the next compile's context; `AGENTS.md` names that grammar.

## The narrative index (newest first)

The first {FULL_URL_THRESHOLD} entries include full `index.md` URLs to establish
the link pattern. All remaining entries are bare slugs. Reconstruct any full
URL as: `{base_url}{folder}/{{slug}}/index.md`
"""

def build_ledger(target_config: dict, rich: bool, limit) -> tuple:
    """Returns (markdown_lines, count) for the URL-first article index."""
    target_path = Path(target_config["path"]).expanduser().resolve()
    base_url = get_base_url(target_config)

    if not target_path.is_dir():
        print(f"⚠️  Article source not found: {target_path}. Writing header-only file.", file=sys.stderr)
        return "", 0

    metadata = lsa.get_holographic_article_data(str(target_path))  # newest-first
    prefix = lsa.permalink_prefix(target_config)
    folder = f"/{prefix}" if prefix else ""
    url_pattern = f"{base_url}{folder}/{{slug}}/index.md"
    if limit:
        metadata = metadata[:limit]

    lines = []
    for idx, item in enumerate(metadata):
        title = item.get("title", "Untitled")
        # File size via stat — cheap metadata syscall, no full read needed
        try:
            byte_size = Path(item["path"]).stat().st_size if item.get("path") else 0
        except OSError:
            byte_size = 0
        size_k = f"{max(1, round(byte_size / 1000))}k" if byte_size else "?"
        if idx < FULL_URL_THRESHOLD:
            url = article_markdown_url(item, base_url, prefix)
            line = f"- [{item['date']}] [{title}]({url})"
        else:
            if idx == FULL_URL_THRESHOLD:
                lines.append(
                    f"\n## Compact slug index — pattern: {url_pattern}\n"
                    f"\nFormat: `[date] [size] slug` — fetch any entry as `{url_pattern}`\n"
                )
            slug = article_slug(item)
            line = f"- [{item['date']}] [{size_k}] {slug}"
        if rich and item.get("shard_kw"):
            line += f" — {item['shard_kw']}"
        lines.append(line)

    return "\n".join(lines), len(metadata)

def main():
    parser = argparse.ArgumentParser(description="Generate the journal index (.agents/skills/journal/references/index.md).")
    parser.add_argument("-t", "--target", type=str, default="1", help="Target ID from blogs.json (default: 1)")
    parser.add_argument("--rich", action="store_true", help="Append holographic-shard keywords to each entry.")
    parser.add_argument("--limit", type=int, default=DEFAULT_LIMIT, help=f"Index only the N newest articles (default: {DEFAULT_LIMIT}; 0 = all).")
    args = parser.parse_args()

    targets = lsa.load_targets()
    target_key = args.target or "1"
    if target_key not in targets:
        print(f"❌ Invalid target key: {target_key}", file=sys.stderr)
        sys.exit(1)
    target_config = targets[target_key]
    base_url = get_base_url(target_config)
    limit = args.limit if (args.limit and args.limit > 0) else None
    prefix = lsa.permalink_prefix(target_config)

    print(f"🧭 Generating the journal index from target: {target_config.get('name', target_key)}")
    ledger, count = build_ledger(target_config, args.rich, limit)
    header = build_header(count, base_url, prefix)
    body = ledger if ledger else "_No articles indexed (article source unavailable at generation time)._"

    final = header + "\n" + body + "\n"
    OUTPUT_FILE.parent.mkdir(parents=True, exist_ok=True)
    OUTPUT_FILE.write_text(final, encoding="utf-8")
    print(f"✅ Wrote {OUTPUT_FILE} ({count} entries, {len(final.encode('utf-8')):,} bytes).")

if __name__ == "__main__":
    main()
[[[END_WRITE_FILE]]]

Target: .agents/skills/journal/SKILL.md
[[[WRITE_FILE]]]
---
name: journal
description: The journal behind Pipulate, indexed. More than a thousand dated articles explain why each piece of the machinery exists, and references/index.md lists them newest first with a fetchable raw-Markdown URL per entry. Use when a question asks why something in this repo is built the way it is, when a comment in the code cites a date or a deed number, or when a change needs reasoning that the git log does not carry.
---

# journal: the reasoning behind the machinery

This repository holds the machinery. The reasoning lives in a running journal
on a separate site, one entry per working session, and this skill is the
bridge to it. Nothing here duplicates the code; the index points and the site
holds. It is what a changelog would be if a changelog recorded why.

## What is here

- `references/index.md`, every entry newest first. The first twenty carry full
  URLs; the rest are bare slugs with a date and a size, and the file's own
  header prints the pattern that turns a slug into a page.

## How to use it

1. Find the entry in `references/index.md` by date, title or slug. A date in
   a code comment (`banked 2026-09-26`) or a deed number (`deed 1596`) names
   a session, and that day's entries carry the why.
2. Build the URL with the pattern at the top of the index (today it is
   `https://mikelev.in/futureproof/<slug>/index.md`); the site serves raw
   Markdown at that path.
3. Fetch only the entry the question needs. One entry can run to 200k
   tokens. The index is a menu, never a payload.
4. Inside a checkout, `scripts/xp.py` turns a pasted list of slugs into the
   next compile's context; `AGENTS.md` at the repo root names that grammar.

## What is not here

- Setup, tools and edit rules: `AGENTS.md`, and the code it points at.
- The reviewer's change-control questions and the artifacts that answer
  them: `AUDIT.md`.
- Article bodies. They never enter this repo; the URL is the whole promise.
[[[END_WRITE_FILE]]]

Car 4: ignition, and the stage. Regenerates the moved file with its new header and stages the new SKILL.md, which git commit -am would otherwise never see. git status --short closes it so the staged rename and the new file are both visible before m.

.venv/bin/python scripts/articles/generate_ai_context.py && git add .agents/skills/journal && git status --short

Car 5: release.py follows the path. Step 1.6 regenerates and stages the new file; step 3.6 copies it to Pipulate.com under the old public name, and says why. Emoji-bearing lines are anchored on their text, not their glyphs.

Target: release.py
[[[SEARCH]]]
    """Regenerate AI_CONTEXT.md — the repo's self-describing briefing for any AI
    that clones and inspects it. Reads the (separate) blog archive and rewrites
    AI_CONTEXT.md in the Pipulate repo root from scratch, so a fresh clone always
    greets an AI with the latest narrative map. Non-fatal: skips cleanly if the
    generator or the article source is unavailable."""
    note("\n🧭 Step 1.6: Regenerating AI_CONTEXT.md (repo talk-back briefing)...")
[[[DIVIDER]]]
    """Regenerate the journal index, .agents/skills/journal/references/index.md,
    the repo's self-describing map for any AI that clones and inspects it.
    Reads the (separate) blog archive and rewrites the index from scratch, so a
    fresh clone always carries the latest narrative map. Until 2026-09-28 this
    file was AI_CONTEXT.md at the repo root. Non-fatal: skips cleanly if the
    generator or the article source is unavailable."""
    note("\n🧭 Step 1.6: Regenerating the journal index (.agents/skills/journal/references/index.md)...")
[[[REPLACE]]]

Target: release.py
[[[SEARCH]]]
AI_CONTEXT generator not found at {generator}. Skipping.
[[[DIVIDER]]]
journal index generator not found at {generator}. Skipping.
[[[REPLACE]]]

Target: release.py
[[[SEARCH]]]
AI_CONTEXT generation returned non-zero; continuing release.
[[[DIVIDER]]]
journal index generation returned non-zero; continuing release.
[[[REPLACE]]]

Target: release.py
[[[SEARCH]]]
    # Stage explicitly: `git commit -am` ignores untracked files, so the very
    # first (untracked) AI_CONTEXT.md must be added by hand. After that it rides -am.
    subprocess.run(["git", "add", "AI_CONTEXT.md"], cwd=str(PIPULATE_ROOT))
    note("✅ AI_CONTEXT.md regenerated and staged.")
[[[DIVIDER]]]
    # Stage explicitly: `git commit -am` ignores untracked files, so the very
    # first (untracked) index must be added by hand. After that it rides -am.
    subprocess.run(["git", "add", ".agents/skills/journal/references/index.md"], cwd=str(PIPULATE_ROOT))
    note("✅ journal index regenerated and staged.")
[[[REPLACE]]]

Target: release.py
[[[SEARCH]]]
    """Copies AI_CONTEXT.md to Pipulate.com root and commits if changed.
[[[DIVIDER]]]
    """Copies the journal index to Pipulate.com root, as AI_CONTEXT.md, and commits if changed.
[[[REPLACE]]]

Target: release.py
[[[SEARCH]]]
    Note: AI_CONTEXT.md is regenerated from scratch at Step 1.6
    (run_ai_context_generation), so by the time this runs the source is fresh.
[[[DIVIDER]]]
    Note: the index is regenerated from scratch at Step 1.6
    (run_ai_context_generation), so by the time this runs the source is fresh.
[[[REPLACE]]]

Target: release.py
[[[SEARCH]]]
    note("\n🔄 Step 3.6: Synchronizing AI_CONTEXT.md to Pipulate.com...")
    source_path = PIPULATE_ROOT / "AI_CONTEXT.md"
    dest_path = PIPULATE_COM_ROOT / "AI_CONTEXT.md"
[[[DIVIDER]]]
    note("\n🔄 Step 3.6: Synchronizing the journal index to Pipulate.com as AI_CONTEXT.md...")
    source_path = PIPULATE_ROOT / ".agents" / "skills" / "journal" / "references" / "index.md"
    # THE URL IS A PROMISE (2026-09-28): the source moved into the journal
    # skill and the public copy keeps its old name, because
    # https://pipulate.com/AI_CONTEXT.md is linked from README.md and from
    # Pipulate.com's own, and may be bookmarked. Renaming it is a Jekyll
    # redirect ride, never a rename here.
    dest_path = PIPULATE_COM_ROOT / "AI_CONTEXT.md"
[[[REPLACE]]]

Target: release.py
[[[SEARCH]]]
Source AI_CONTEXT.md not found at {source_path}
[[[DIVIDER]]]
Source journal index not found at {source_path}
[[[REPLACE]]]

Target: release.py
[[[SEARCH]]]
    # Step 1.6: Regenerate the AI_CONTEXT.md repo briefing (talk-back map)
[[[DIVIDER]]]
    # Step 1.6: Regenerate the journal index (the repo's talk-back map)
[[[REPLACE]]]

Target: release.py
[[[SEARCH]]]
Skipping AI_CONTEXT.md regeneration (--skip-docs-sync)
[[[DIVIDER]]]
Skipping journal index regeneration (--skip-docs-sync)
[[[REPLACE]]]

Car 6: the assurance posture moves into AUDIT.md. Section 7 of the old header, verbatim under its own heading, before “Closing”; the AUDIT sync carries it public at the next release.

Target: AUDIT.md
[[[SEARCH]]]
## Closing
[[[DIVIDER]]]
## Change control: the questions and the artifacts that answer them

A reviewer charged with assessing this project's use of AI-generated code will
have a fixed set of questions. They are the right questions. Each is answered
below by the artifact that answers it, so the review can proceed from evidence
rather than from description, and so the reviewer may put the same questions,
with equal fairness, to any other AI-assisted work in the estate.

- **What changed?** One diff per change, applied by exact string match
  (`apply.py`), which refuses a diff that does not match the working tree byte
  for byte. A change that cannot be shown as a diff is not applied.
- **What was the state before?** A commit identifier. The baseline is
  content-addressed and restorable by anyone holding the repository, with the
  author absent.
- **How is it known that the change did what was claimed?** A fixed set of
  read-only commands is run before and after the change, under identical
  instrumentation, and the two outputs are compared. The difference is the
  evidence; no difference is also evidence.
- **Where is the record, and can it be altered?** Both observations, the diff
  and the commit identifier are sealed in one archive whose SHA-256 digest is
  published beside it. It verifies with the Python standard library and no
  vendor tool (`scripts/foo_cartridge.py`); one altered byte fails.
- **Who proposed, who applied, who approved?** Proposal (by an author or an
  assistant) and application (by the operator, through the interlock above) are
  separate acts by separate parties, each leaving its own trace. Nothing an
  assistant proposes reaches the tree except through that interlock.
- **Can the environment be reproduced?** The toolchain is declared in a lock
  file (`flake.lock`) that pins every dependency by content hash, and the
  declaration is itself under version control. Two machines evaluating it
  produce the same environment.
- **What does it touch on the host?** One directory tree and the user's own
  home directory. Managed packages, device-management profiles and endpoint
  agents are not modified.
- **How is a bad change reversed?** By restoring the prior commit. No approval
  is consumed; a baseline is a fact, not a request.
- **Can secrets or client identifiers leave?** Every compiled payload passes a
  substitution table, an identity denylist and a credential-shape tripwire
  before it is emitted, and the tripwire has no off switch.

In the vocabulary of the control catalogues: these are the questions NIST SP
800-53 asks under CM-2, CM-3, CM-4, SA-10 and SI-7; NIST SP 800-218 (SSDF)
under PS.3 and PW.8; ISO/IEC 27001:2022 under A.8.9 and A.8.32; SOC 2 under
CC8.1; PCI DSS v4.0 under 6.5.1 and 6.5.2; ITIL 4 under change enablement, where
a change this small and this repeatable is a standard change; COBIT 2019 under
BAI06 and BAI10; and ISO 9001:2015 under 8.5.6 and 10.2. The sealed archive is
the provenance record SLSA describes; the lock file is the dependency inventory
an SBOM (SPDX, CycloneDX) enumerates; the environment property is the one the
Reproducible Builds project names. The traceability that DO-178C and IEC 61508
demand of any change to a certified artifact is the same traceability, applied
here to ordinary work. This is vocabulary, not a certificate: no assessor has
attested to any of it, and the repository does not claim otherwise. The
controls this project is realistically built toward are the
configuration-management and provenance families; the safety-critical standards
lend their words only.

Two observations follow, and a careful reviewer reaches them unprompted. First,
a review is only as strong as the answers the same questions receive elsewhere,
and AI-assisted work performed without such records cannot answer them. Second,
the procedure is an instrument rather than a position: it applies to any change
by anyone holding a repository, and a reviewer who wishes to verify these claims
independently may do so with the command printed in the footer of every sealed
archive, which needs nothing installed beyond Python. The instrument is offered
to the review function on the same terms it is used here.

## Closing
[[[REPLACE]]]

Car 7: the pointers follow. The README’s top line, AGENTS.md’s skills line (the specification’s URL and the journal signpost), the dead package-data entry, and the router’s two claims. Four files, one story: nothing names the old root file as a live path any more.

Target: README.md
[[[SEARCH]]]
3. For Context: AI_CONTEXT.md ( https://pipulate.com/AI_CONTEXT.md )
[[[DIVIDER]]]
3. For the Why: `.agents/skills/journal/` (the journal index, an Agent Skills reference; served at https://pipulate.com/AI_CONTEXT.md )
[[[REPLACE]]]

Target: AGENTS.md
[[[SEARCH]]]
- Skills (Agent Skills spec): `.agents/skills/*/SKILL.md`, at the repo root beside this file
[[[DIVIDER]]]
- Skills (Agent Skills spec, https://agentskills.io/specification): `.agents/skills/*/SKILL.md`, at the repo root beside this file
- The reasoning behind any piece of the machinery: the `journal` skill, whose `references/index.md` lists every journal entry newest first with a fetchable URL
[[[REPLACE]]]

Target: pyproject.toml
[[[SEARCH]]]
    "README.md",
    "AI_CONTEXT.md",
    "LICENSE",
[[[DIVIDER]]]
    "README.md",
    "LICENSE",
[[[REPLACE]]]

Target: foo_files.py
[[[SEARCH]]]
# AI_CONTEXT.md
[[[DIVIDER]]]
# .agents/skills/journal/references/index.md
[[[REPLACE]]]

Target: foo_files.py
[[[SEARCH]]]
# THE SKILL.md FILES
# .agents/skills/gsc-readonly/SKILL.md
# .agents/skills/hello-workflow/SKILL.md
# .agents/skills/roles/SKILL.md
# .agents/skills/sheets-readonly/SKILL.md
[[[DIVIDER]]]
# THE SKILL.md FILES
# .agents/skills/gsc-readonly/SKILL.md
# .agents/skills/hello-workflow/SKILL.md
# .agents/skills/journal/SKILL.md
# .agents/skills/journal/references/index.md
# .agents/skills/roles/SKILL.md
# .agents/skills/sheets-readonly/SKILL.md
[[[REPLACE]]]

Car 8: the agent-facing files ship with the source. Two MANIFEST.in lines; the local build straddle is its witness, and the next release is its public one.

Target: MANIFEST.in
[[[SEARCH]]]
exclude apps/xx_*.py
[[[DIVIDER]]]
exclude apps/xx_*.py

# THE AGENT-FACING FILES SHIP WITH THE SOURCE (2026-09-28). setuptools' sdist
# takes package modules, README, LICENSE and pyproject.toml on its own and no
# other root file, so the two files an agent reads first and the skills folder
# were absent from every sdist an auditor could download. The package-data
# block in pyproject.toml does not reach them either: its patterns resolve
# inside package directories, and none of these files lives in one.
include AGENTS.md AUDIT.md
graft .agents
[[[REPLACE]]]

Ignition: Car 4 is the ignition for Car 3 (the generator runs and writes the new header); every other car is read by a probe at call time. The AFTER forecasts, as deltas: the skills census +1 row (journal); the loader line hello_workflow to hello-workflow; the AI_CONTEXT census loses every line in generate_ai_context.py but one (the dated docstring), every line in pyproject.toml, and the router’s line 1515, and prints nothing from index.md; the local build gains AGENTS.md, AUDIT.md and the .agents/ members; Coverage +1 claimed and +1 tracked (the new SKILL.md, once Car 4 stages it).

4. PROMPT

Read the receipts first: the skills census (five rows), the loader line, the ls and the head, the AI_CONTEXT census, and both sdist censuses; say what 2.65's sdist and wheel actually carry and whether the package-data reading held. Then write .agents/skills/pipulate/SKILL.md, the install skill for an agent on a stranger's machine, from install.sh, the npvg.org door page, README's Quick Start and AUDIT.md's first-run section: inspect first (curl -fsSL https://npvg.org | less, what the reader sees, j/k/q), then save, inspect and execute the same file, then the one-line install, the three doors after nix develop, what stays local and what reaches the cloud, and how to stop, reset and uninstall; two sentences of the why and no more. Say in one line why the compile loop stays in AGENTS.md rather than a skill. After that, the fantasy league, from where the 09-28 board left off, once the SVB lines are uncommented.

5. EXTERNAL DELIVERABLES

One, for the article, manual-paste: the six-sentence lineage paragraph the earlier caboose asked for, in plain words, every date checkable.

Two files carry the AI industry’s conventions for talking to a repository, and they have different parents. AGENTS.md is a README with a new reader: plain Markdown, no header block, read at the start of every session; OpenAI shipped it in August 2025 with Codex, Amp, Google’s Jules, Cursor and Factory behind it, and on December 9, 2025 the Linux Foundation took it into the new Agentic AI Foundation beside Anthropic’s Model Context Protocol and Block’s goose. SKILL.md is a Jekyll post with a new job: a small YAML block between two --- lines, then Markdown, the shape Tom Preston-Werner gave blog posts in 2008, now carrying a name and a description so a tool can decide when to load the instructions underneath; Anthropic published that shape as the open Agent Skills specification at agentskills.io on December 18, 2025. The two are complementary, not successive: the first is always on, the second loads on demand. Andrej Karpathy wrote neither file; he named the practice both serve, “vibe coding” in February 2025 and the framing of the context window as working memory. A second brain kept as a Jekyll blog since before any of this was therefore not borrowing from the AI industry; it was already in the file format half the industry settled on.

MikeLev.in:

THE PROOF STRADDLE

Same commands, run twice, one change between them. Where the readings differ is what the change did; the diff in the middle is the receipt.

1: Probe:

(nix) pipulate $ g

Left-hand causal boundary "blast radius" established. Make 1 change and test.
On branch main
Your branch is up to date with 'origin/main'.

nothing to commit, working tree clean
(nix) pipulate $ rg -n "AI_CONTEXT" --glob '!AI_CONTEXT.md' . ~/repos/Pipulate.com --glob '!_site/**'
.venv/bin/python -c "import re,pathlib;[print(p.parent.name, 'fence' if p.read_text().startswith('---') else 'NO-FENCE', 'name-ok' if re.fullmatch(r'[a-z0-9]+(-[a-z0-9]+)*', p.parent.name) else 'NAME-INVALID') for p in sorted(pathlib.Path('.agents/skills').glob('*/SKILL.md'))]"
rg -n -o '"skills" / "[a-z_-]+"' apps/040_hello_workflow.py
ls .agents/skills/hello-workflow/SKILL.md .agents/skills/journal/SKILL.md .agents/skills/journal/references/index.md AI_CONTEXT.md 2>&1
head -12 .agents/skills/journal/references/index.md 2>&1
.venv/bin/python -c "import json,io,tarfile,zipfile,urllib.request as u;d=json.load(u.urlopen('https://pypi.org/pypi/pipulate/json'));K=('agents','ai_context','skill','flake','audit','readme','manifest','license','changelog','install.sh');names=lambda f,b: tarfile.open(fileobj=io.BytesIO(b)).getnames() if f['packagetype']=='sdist' else zipfile.ZipFile(io.BytesIO(b)).namelist();R=[(f,names(f,u.urlopen(f['url']).read())) for f in d['urls']];[print(f['packagetype'],f['filename'],f['size'],'bytes',len(n),'members',[x for x in n if any(k in x.lower() for k in K)]) for f,n in R]"
.venv/bin/python -m build --sdist --no-isolation --outdir /tmp/pipulate-sdist-probe . >/dev/null 2>&1; echo build_rc=$?; tar tzf /tmp/pipulate-sdist-probe/pipulate-2.65.tar.gz | grep -Ei 'AGENTS\.md|AUDIT\.md|\.agents/|AI_CONTEXT'; echo members=$(tar tzf /tmp/pipulate-sdist-probe/pipulate-2.65.tar.gz | wc -l)
/home/mike/repos/Pipulate.com/README.md
3:3. And context AI_CONTEXT.md ( https://pipulate.com/AI_CONTEXT.md )

/home/mike/repos/Pipulate.com/_data/orphans.yml
36:  - AI_CONTEXT.md

./release.py
249:    """Regenerate AI_CONTEXT.md — the repo's self-describing briefing for any AI
251:    AI_CONTEXT.md in the Pipulate repo root from scratch, so a fresh clone always
254:    note("\n🧭 Step 1.6: Regenerating AI_CONTEXT.md (repo talk-back briefing)...")
257:        print(f"ℹ️  AI_CONTEXT generator not found at {generator}. Skipping.")
263:        print("⚠️  AI_CONTEXT generation returned non-zero; continuing release.")
269:    # first (untracked) AI_CONTEXT.md must be added by hand. After that it rides -am.
270:    subprocess.run(["git", "add", "AI_CONTEXT.md"], cwd=str(PIPULATE_ROOT))
271:    note("✅ AI_CONTEXT.md regenerated and staged.")
585:    """Copies AI_CONTEXT.md to Pipulate.com root and commits if changed.
587:    Note: AI_CONTEXT.md is regenerated from scratch at Step 1.6
590:    note("\n🔄 Step 3.6: Synchronizing AI_CONTEXT.md to Pipulate.com...")
591:    source_path = PIPULATE_ROOT / "AI_CONTEXT.md"
592:    dest_path = PIPULATE_COM_ROOT / "AI_CONTEXT.md"
595:        print(f"⚠️  Warning: Pipulate.com repo not found at {PIPULATE_COM_ROOT}. Skipping AI_CONTEXT.md sync.")
599:        print(f"⚠️  Warning: Source AI_CONTEXT.md not found at {source_path}. Skipping AI_CONTEXT.md sync.")
612:            commit_msg = f"chore: Update AI_CONTEXT.md from pipulate repo v{get_current_version()}"
626:                    print(f"✅ Pushed AI_CONTEXT.md update and set upstream: origin/{current_branch}")
629:                    print("✅ Pushed AI_CONTEXT.md update to Pipulate.com repo.")
638:            note("✅ AI_CONTEXT.md is already up-to-date in Pipulate.com repo.")
641:        print(f"⚠️  AI_CONTEXT.md sync failed: {e}")
653:    scripts, AUDIT.md, AI_CONTEXT.md), each ending in add/commit/push against
1247:    parser.add_argument("--skip-ai-context-sync", action="store_true", help="Skip AI_CONTEXT.md synchronization")
1284:    # Step 1.6: Regenerate the AI_CONTEXT.md repo briefing (talk-back map)
1288:        print("\n⏭️  Skipping AI_CONTEXT.md regeneration (--skip-docs-sync)")
1314:    # Step 3.6: AI_CONTEXT.md Synchronization
1318:        print("\n⏭️  Skipping AI_CONTEXT.md synchronization (--skip-ai-context-sync)")

./pyproject.toml
94:    "AI_CONTEXT.md",

./README.md
5:3. For Context: AI_CONTEXT.md ( https://pipulate.com/AI_CONTEXT.md )

./foo_files.py
1515:# AI_CONTEXT.md
2615:# - TODO (2026-09-27, THE WORKSHOP TREE; ruled from the operator's drawing at deed 1628): three tiers are three ownerships and Pipulate tracks none of them: corporate/ (the org's canon, a private repo mounted read-only and replaced wholesale on pull), shared/<name>/ (one writer, own repo, .identity the claim, AGENTS.md nearest-ancestor inside the namespace), personal/ (gitignored; no git verb reaches it); the mechanism stays public. What exists today is most of it: Notebooks/.agents/skills tracked (Pipulate's public example skills, never org canon), Notebooks/Shared/, Notebooks/Playground/ and Notebooks/Client_Work/ ignored. LANDED at deed 1628: Notebooks/corporate/ ignored, the mini-rules reference written there. OWED, each its own ride with its files in the payload: (1) the rename Notebooks -> Workshop and Playground -> personal, gated on the census of every reader (flake.nix's copy_notebook_if_needed and its mkdir of Notebooks/Shared, server.py, config.py, cli.py, the trails, the SKILL.md paths, scripts/articles/Notebooks/Shared/README.md; the 2026-09-04 THREE-TIER TODO spelled Corporate/ and Personal/ with capitals, and one spelling wins), the rg counts of deed 1629 the blast radius; (2) the teeth: scripts/git_hooks/pre-commit refuses any index path under the three tiers the way it refuses the adhoc overlay, backstopping m's sweep (the 2026-09-21 TODO); (3) the connectors split: connectors/botify.py keeps the token lanes (list, fetch, query, --check), the worked example agentskills readers should see, and the admin-door lanes (--census, --sw, --pull-configs, the Django export form, Activation GraphQL, the cookie harvest) move to Notebooks/corporate/connectors/botify_admin.py, with a search path in flake.nix's connectorCommand, tools/connector_tools.py and scripts/sources_menu.py so the private words mint beside the public ones and a public checkout prints one line naming the tier it lacks; (4) each tier a repo with a bare remote under ~/git-repos like the vault; (5) the reference becomes a skill under Notebooks/corporate/.agents/skills/botify-minirules/ the day the specification rides the payload, never authored from memory. Gates: (1) a fresh nix develop stages the notebooks under the new name and jupyter opens there; (3) botify --sw runs from the shell after the move and rg -n 'admin/projects' connectors/botify.py prints nothing. READ 2026-09-27 (deed 1629): the census of readers, 18 non-doc files naming Notebooks, flake.nix 36 lines, AGENTS.md 6, server.py 2, pyproject.toml 2, config.py, cli.py and MANIFEST.in 0. RULED by the operator at deed 1629: the rename in active code is the priority ("keep code working and documentation correct"), receipts, keys and scars untouched (no retconning), the corporate folder's own repo comes after, the Botify admin lanes are corporate, and the Slack connector is blocked by the org's per-app OAuth admin control where Jira and Google are not. The rename ride is deed 1630's: flake.nix whole and rg -n receipts of every other reader in the payload; git mv Notebooks Workshop moving the directory whole (ignored tiers included) and Playground -> personal as the drawing spells it, lowercase; .gitignore's Notebooks/ patterns in the same train; the router's commented Notebooks/ paths (chapters VI, XVI, XVIII) retargeted because the Paintbox claims through them; every dated line left as written; ignition exit then ndq. Gate: git status clean after a compile, jupyter opening in Workshop, botify --sw still running. LANDED 2026-09-27 (deed 1630, the train): .gitignore first and ADDITIVE (every Notebooks/ pattern kept beside its Workshop/ twin, because git renames tracked files only and a checkout that pulls the rename keeps its ignored tiers under Notebooks/ until a hand mv, and because a sweeping git add -A during the move must find both trees ignored; on Prime m is configuration.nix's commit -am, which sweeps nothing, the mcp_render.py git add by hand at deed 1630 the receipt, while the flake's m() on the Mac does sweep), then git mv Notebooks Workshop and mv Workshop/Playground Workshop/personal by hand (8 tracked paths, the ignored tiers riding the filesystem rename), then flake.nix (36 lines: the startup notebook, the twelve copy destinations, the staged-files line, the WELCOME.md test and heredoc, the onboarded sentinel twice, the JupyterLab URL, the THREE BUCKETS block with a corporate/ line added, the two mkdirs, the Shared README test and printf), pyproject.toml (two excludes), AGENTS.md (three prose lines; its tree block is generated from the sealed workspace_tree art and waits for that car), six active-code files by receipt line (apps/030_roles.py's SKILL.md path, mck.sh's TRAIL_SEARCH_DIRS, pipulate/core.py's six joins, pipulate/__init__.py's three, server.py's two sentinels, scripts/bookmark_import.py's DEFAULT_OUT), the documentation strings (five assets/nbs/imports sauce comments and one logger line, generate_ai_context.py and its AI_CONTEXT.md output), the router's seven chapter-VI and -XVI paths (the Paintbox claims through them), and the word render as the last car; ignition exit then ndq; the nix retirement after it, its own fence. Shared/ keeps its capital (unasked; a TODO). OWED after the train: the art car (AGENTS.md's and README.md's generated trees), README.md's own lines, GLOSSARY.md's, the five asset notebooks, the stray scripts/articles/Notebooks/Shared/README.md, the served mck.sh, the Mac's orphan. READ 2026-09-27 (deed 1631): the train landed; ls-files eight under Workshop/, ls -A every tier with personal and no Playground, the tiers' porcelain 0 and the tree's 1 (the compiler's own stats and Paintbox rewrite, the SKILL.md rows now spelling Workshop/, Coverage 204/275 +0), CORPORATE_IGNORED and REFERENCE_PRESENT under the new path, the quiet hook exit 0 at 77,416 bytes with workshop=6 notebooks=0 UNDER THE SYSTEM NIX (the compile ran after Car 10), render a store path with the piped diff printing its header. THE PROBE DIED AND PRINTED ZEROES: the per-file census read 0 for all ten files while AGENTS.md's generated block still said Notebooks/ in the same payload; in ripgrep -E is --encoding, so the pattern was eaten as an encoding name, rg exited 2, 2>/dev/null hid it and || echo 0 printed the success token (THE HELP TEXT IS NOT A CENSUS's closing clause convicted a second time; the fixed probe spells -e and prints rc beside each count, deed 1632). THE RE-EMITTED CAR LANDED TWICE: the reply hit its output limit inside Car 8's last block, the cut fence's 17 complete blocks applied (008f38b0), and the re-emitted Car 8's ninth block carried a shortened LANDED paragraph whose SEARCH, a prefix of the already-appended line, matched and appended a second LANDED (bda3e82f); deed 1631's Car 4 cuts the short copy; a re-emission is byte-identical or it is a new car. Two misses: mck.sh keeps a third Playground line my two blocks did not reach (its receipt rides deed 1632), and Workshop/.venv is a real directory dated Oct 2025 (a stale venv an editor or JupyterLab made inside the folder, INFERRED; gitignored by .venv/, unasked). The remainder shrank: README's four Notebooks lines (73, 76, 77, 80) ARE the generated block, so the art car covers README too and its other nine hits are the Jupyter concept; GLOSSARY's 375 is the one vocabulary line (deed 1631's Car 2), 750 and 1855 dated; the stray README is deed 1631's Car 3 (git rm by hand). READ 2026-09-27 (deed 1632), the ride's close: 11286aa2 the art and its seal, 1c3a60fe GLOSSARY's line (grep 1), a82a4f3d the stray README (head reads No such file, ls-files 0, Coverage 204/274 with -1 tracked) and the dedupe (grep 0); the anchored census reads 0 rc=1 for nine files, README.md=10 (the forecast said nine; the tenth unread, README.md not in the payload), pipulate/core.py=3 and imports/ascii_displays.py=1 (concept prose and the dated docstring, neither a path); mck.sh's 415 is one comment line. OWED after the ride, unchanged: the five asset notebooks, the served mck.sh, the Mac's orphan mv, Workshop/.venv, and (2) through (5) above, each its own ride. READ 2026-09-28 (deed 1636): Workshop/.venv aside under ~/.local/state/pipulate/stale/, the served mck.sh refreshed by the 2.65 release lane, the Mac rebuilt fresh with no orphan; (4) landed for corporate and personal (corporate.git and personal.git under ~/git-repos, deed 1634's Car 4); the five asset notebooks and (2), (3), (5) stand.
2623:# - TODO (2026-09-27, AGENTS.md and AI_CONTEXT.md must name the specification; the operator's words, its own ride): a stranger's model, or an IT department come to forbid it, reading AGENTS.md or AI_CONTEXT.md must see on the first screen that this repository is an Agent Skill under https://agentskills.io/specification (a SKILL.md folder with YAML frontmatter and progressive disclosure, AGENTS.md as the nearest-ancestor signpost; Notebooks/.agents/skills already conforms, THREE-STANDARDS SUPERPOSITION), and that every tool call it teaches is forced onto localhost under the fifty-year POSIX conventions (a command, stdin, stdout, an exit code) so that it lands in the flight data recorder: a cloud tool call that cannot be reconstructed locally is non-reproducible, non-portable and cannot be compiled into a cartridge, and is out of scope by that rule and not by taste. The ride needs AGENTS.md, AI_CONTEXT.md and !https://agentskills.io/specification in the payload, the specification read before a word is written and the wording checked against it. Gate: a fresh chat handed AGENTS.md alone names the specification and the localhost rule in its first two sentences. THE ORDER, ruled at the 2026-09-27 dismount: after the first ticket, the deliverable and the render farm, and before the Workshop tree; with that tree landed under Notebooks (the command-stop TODO), the repository is the worked example the specification's readers can run on localhost, the operator's stated aim, "the quintessential and prototypical example".
2785:# - TODO (2026-09-25, the certifications article): the journal entry that rode dedupe's empty folders, blast's tail, the hyper-literal words and the assurance posture into AI_CONTEXT.md names the control catalogues (NIST SP 800-53 CM-2, CM-3, CM-4, SA-10, SI-7; SP 800-218 PS.3, PW.8; ISO/IEC 27001:2022 A.8.9, A.8.32; SOC 2 CC8.1; PCI DSS 6.5.1, 6.5.2; ITIL 4 standard change; COBIT BAI06, BAI10; ISO 9001 8.5.6, 10.2; SLSA; SPDX, CycloneDX; DO-178C and IEC 61508 as vocabulary only) and is not yet articleized, so it cannot be pinned. When it is, pin it in V-b. OWES: a ruling on which of those this project realistically moves toward and which lend words only, banked as a key. Gate: a 📌 line in V-b naming the article, with that list under its OWES.
3113:# - EARMARK: THE HEADER-BOUNDED SCANNER (banked 2026-08-05, decoy-convicted): a program that finds its edit region by scanning for the nearest markdown heading cannot distinguish a heading from a `#` comment inside a fenced code block, so its blast radius is decided by whatever prose happens to sit nearby -- and any unrelated edit can move it silently. CONVICTION: release.py's run_waxascii_release_stamp scans for line.startswith(("# ", "## ", "### ")); Pipulate.com/index.md carries `# 2. Launch it` at line 23 and `# This is how simple Pipulate code looks` at line 162, both bash/python comments inside fences, both indistinguishable from headings to that scanner. README.md survives only because the region between its canary and `## Quick Start` happens to contain no fenced comment -- stable by luck, not by design. WITNESS OF THE SAFE CASE, which is the only reason this was ever observable: a live v2.43 release ran with the workspace-tree sentinels already in README.md and the resulting commit touched only AI_CONTEXT.md and pyproject.toml, proving the stamper's unconditional write_text produced byte-identical output. PRESCRIPTION: bound an edit region by EXPLICIT SENTINELS the author placed, never by structure the document happens to have. Sibling of THE LAST-INCH RULE: there the render destroys a correct result, here the boundary-finder does.

./scripts/articles/generate_ai_context.py
5:Writes AI_CONTEXT.md to the Pipulate repository root: a self-contained briefing
18:Idempotent: rewrites AI_CONTEXT.md from scratch on every run. Intended as a
41:OUTPUT_FILE = REPO_ROOT / "AI_CONTEXT.md"
77:    return f"""# AI_CONTEXT.md — Start Here If You Are an AI
131:| `scripts/articles/generate_ai_context.py` | This file's generator — rewrites `AI_CONTEXT.md` on every release |
285:    parser = argparse.ArgumentParser(description="Generate AI_CONTEXT.md repo briefing.")
301:    print(f"🧭 Generating AI_CONTEXT.md from target: {target_config.get('name', target_key)}")
gsc-readonly fence name-ok
hello-workflow fence name-ok
roles fence name-ok
sheets-readonly fence name-ok
176:"skills" / "hello_workflow"
ls: cannot access '.agents/skills/journal/SKILL.md': No such file or directory
ls: cannot access '.agents/skills/journal/references/index.md': No such file or directory
.agents/skills/hello-workflow/SKILL.md  AI_CONTEXT.md
head: cannot open '.agents/skills/journal/references/index.md' for reading: No such file or directory
bdist_wheel pipulate-2.65-py3-none-any.whl 797866 bytes 81 members ['pipulate-2.65.dist-info/licenses/LICENSE']
sdist pipulate-2.65.tar.gz 789587 bytes 98 members ['pipulate-2.65/LICENSE', 'pipulate-2.65/MANIFEST.in', 'pipulate-2.65/README.md']
build_rc=0
members=98
(nix) pipulate $

2: Context:

# context.txt: the list of files an AI will read. context opens it, compile builds it.
# Keys: j down, k up; Esc then :q quits, :q! discards, :wq saves and quits.
# One line per thing the AI reads: a file path, or a command after `! `.
# A line that starts with # is a comment: that file is not read.
# To add a line: i starts typing, Esc stops. Absolute paths work from anywhere.
# Web pages, APIs and the connector words: chapter XVIII of foo_files.py.

# --- THE 40K-FOOT VIEW (uncomment on a first turn; comment out on the second) ---
# ! python scripts/articles/lsa.py -t 1 --reverse --fmt dated-slugs  # <-- the book's spine, one line per article, newest first
# ~/repos/nixos/autognome.py  # <-- the machine's morning routine (this author's NixOS box only)
# init.lua                    # <-- the editor keys that drive the day
# assets/installer/install.sh # <-- how a stranger's machine gets this workshop
# GLOSSARY.md                 # <-- the terms, defined
# flake.nix                   # <-- the environment, pinned: here is my hardware, here is my state
# prompt_foo.py               # <-- the compiler that builds the payload
# foo_files.py                # <-- the router: which files ride, and this book's outline
# scripts/articles/lsa.py     # <-- the second brain: the article corpus behind `rgx`, `rgxc` and `posts`
# requirements.in             # <-- the Python packages, pinned
# pyproject.toml              # <-- the PyPI package
# __init__.py                 # <-- the version

# --- ACTUATORS (cheap; include when the AI should be able to act, not only read) ---
# cli.py                      # <-- tool calls from the command line
# scripts/xp.py               # <-- turns a pasted reply into the next context
# scripts/ai.py               # <-- a local AI writes the commit messages
# scripts/crawl.py            # <-- crawl a site into the next turn
# scripts/weblogin.py         # <-- warm a login on the persistent browser profile
# scripts/webclip_2_markdown.py  # <-- a web page, clipped, as markdown

# --- RARE ---
# scripts/foo_cartridge.py    # <-- the sealed archive: writer and verifier
# scripts/foo_replay.py       # <-- replay a sealed archive on another machine
# release.py                  # <-- how a release reaches GitHub and PyPI
# imports/voice_synthesis.py  # <-- the voice
# imports/ascii_displays.py   # <-- the ASCII art
# scripts/release/version_sync.py  # <-- version stamping (to be folded into release.py)

# --- THIS DISCUSSION ---
# The files and commands for the work in front of you. Paste the NEXT CONTEXT
# block an AI hands back directly below this line; the AI will correct a guess.

# Context 1
# ! rgxc -t grim,article "speed" "dating" "fantasy league"

# Context 2
# README.md
# AGENTS.md
# AI_CONTEXT.md
# .agents/skills/gsc_readonly/SKILL.md
# .agents/skills/hello_workflow/SKILL.md
# .agents/skills/roles/SKILL.md
# .agents/skills/sheets_readonly/SKILL.md
# 
# ~/.config/pipulate/blogs.json                # <-- CAUTION! Derived from ~/repos/nixos/blogs.nix
# scripts/articles/publishizer.py              # <-- Orchestrates different publishing workflows per target blog.
# scripts/articles/common.py                   # <-- Self-explanatory
# scripts/articles/articleizer.py              # <-- Transforms raw article.txt to formal Jekyll markdown format
# scripts/articles/editing_prompt.txt          # <-- Forcing response into strict JSON data structure
# scripts/articles/sanitizer.py                # <-- Scrubs PII
# scripts/articles/gsc_historical_fetch.py
# scripts/articles/contextualizer.py           # <-- Builds JSON summaries of articles in `_posts/context/` called "Holographic Shards".
# scripts/articles/confluenceizer.py           # <-- Idempotent Jekyll-to-Confluence corporate wiki
# scripts/articles/googledocizer.py            # <-- Just added
# scripts/articles/build_knowledge_graph.py    # <-- Topically load-balances site using hierarchical K-Means keyword clustering groups
# scripts/articles/generate_ai_context.py      # <-- AIs WILL interrogate your repo. This gives epic context of article URLs for drill-down.
# scripts/articles/generate_hubs.py            # <-- Uses just-produced link-graph data to generate each of the new hubs it suggests
# scripts/articles/generate_llms_txt.py        # <-- Builds an llms.txt based on the auto-organized structure suggested here
# scripts/articles/generate_redirects.py       # <-- Generates redirect map above hub-churn suggests is needed
# scripts/articles/sanitize_redirects.py       # <-- Deals with follow-up meticulous pedantic detail required for a good Nginx redirect map
# scripts/articles/link_injector.py
# scripts/articles/generate_post_nav.py  # [1,174 tokens | 4,616 bytes]
# 
# # The following Jekyll files pair well with the above to show how we start various forms of tracking, and as a transition into Honeybot Nginx Broadcast Studio & telemetry.
# ~/repos/nixos/blogs.nix
# release.py
# remotes/honeybot/nixos/configuration.nix
# ~/repos/trimnoir/_config.yml
# ~/repos/trimnoir/_layouts/default.html
# ~/repos/trimnoir/index.md
# ~/repos/trimnoir/flake.nix
# 
# Context 3
# ! rg -n "AI_CONTEXT" --glob '!AI_CONTEXT.md' . ~/repos/Pipulate.com --glob '!_site/**'
# ! .venv/bin/python -c "import re,pathlib;[print(p.parent.name, 'fence' if p.read_text().startswith('---') else 'NO-FENCE', 'name-ok' if re.fullmatch(r'[a-z0-9]+(-[a-z0-9]+)*', p.parent.name) else 'NAME-INVALID') for p in sorted(pathlib.Path('.agents/skills').glob('*/SKILL.md'))]"
# ! rg -n "gsc_readonly|hello_workflow|sheets_readonly" --glob '!.agents/**' .
# ! grep -n readme pyproject.toml
# ! cat MANIFEST.in
# README.md
# AGENTS.md
# AUDIT.md
# pyproject.toml
# MANIFEST.in
# .agents/skills/gsc-readonly/SKILL.md
# .agents/skills/hello-workflow/SKILL.md
# .agents/skills/roles/SKILL.md
# .agents/skills/sheets-readonly/SKILL.md
# scripts/articles/generate_ai_context.py
# release.py
# foo_files.py
# # prompt_foo.py
# # AI_CONTEXT.md
# # --- THE FIRST GAME (uncomment when the ride turns to SVB-133) ---
# # Workshop/corporate/connectors/svb.py
# # Workshop/corporate/tickets/CONTEXT_TEMPLATE.txt
# # Workshop/corporate/tickets/SVB_PROMPT.md
# # connectors/mcp_render.py
# # connectors/jira.py
# # ! jira SVB-133
# 
# Context 4
! rg -n "AI_CONTEXT" --glob '!AI_CONTEXT.md' . ~/repos/Pipulate.com --glob '!_site/**'
! .venv/bin/python -c "import re,pathlib;[print(p.parent.name, 'fence' if p.read_text().startswith('---') else 'NO-FENCE', 'name-ok' if re.fullmatch(r'[a-z0-9]+(-[a-z0-9]+)*', p.parent.name) else 'NAME-INVALID') for p in sorted(pathlib.Path('.agents/skills').glob('*/SKILL.md'))]"
! rg -n -o '"skills" / "[a-z_-]+"' apps/040_hello_workflow.py
! ls .agents/skills/hello-workflow/SKILL.md .agents/skills/journal/SKILL.md .agents/skills/journal/references/index.md AI_CONTEXT.md 2>&1
! head -12 .agents/skills/journal/references/index.md 2>&1
! .venv/bin/python -c "import json,io,tarfile,zipfile,urllib.request as u;d=json.load(u.urlopen('https://pypi.org/pypi/pipulate/json'));K=('agents','ai_context','skill','flake','audit','readme','manifest','license','changelog','install.sh');names=lambda f,b: tarfile.open(fileobj=io.BytesIO(b)).getnames() if f['packagetype']=='sdist' else zipfile.ZipFile(io.BytesIO(b)).namelist();R=[(f,names(f,u.urlopen(f['url']).read())) for f in d['urls']];[print(f['packagetype'],f['filename'],f['size'],'bytes',len(n),'members',[x for x in n if any(k in x.lower() for k in K)]) for f,n in R]"
! .venv/bin/python -m build --sdist --no-isolation --outdir /tmp/pipulate-sdist-probe . >/dev/null 2>&1; echo build_rc=$?; tar tzf /tmp/pipulate-sdist-probe/pipulate-2.65.tar.gz | grep -Ei 'AGENTS\.md|AUDIT\.md|\.agents/|AI_CONTEXT'; echo members=$(tar tzf /tmp/pipulate-sdist-probe/pipulate-2.65.tar.gz | wc -l)
README.md
AGENTS.md
AUDIT.md
pyproject.toml
MANIFEST.in
.agents/skills/journal/SKILL.md
scripts/articles/generate_ai_context.py
release.py
assets/installer/install.sh
remotes/honeybot/www/npvg.org/index.html
# flake.nix
# foo_files.py
# prompt_foo.py
# --- THE FIRST GAME (uncomment when the ride turns to SVB-133) ---
# Workshop/corporate/connectors/svb.py
# Workshop/corporate/tickets/CONTEXT_TEMPLATE.txt
# Workshop/corporate/tickets/SVB_PROMPT.md
# connectors/mcp_render.py
# connectors/jira.py
# ! jira SVB-133

3: Patches: Well this is satisfying.

Left-hand causal boundary "blast radius" established. Make 1 change and test.
On branch main
Your branch is up to date with 'origin/main'.

nothing to commit, working tree clean
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'apps/040_hello_workflow.py'.
(nix) pipulate $ d
diff --git a/apps/040_hello_workflow.py b/apps/040_hello_workflow.py
index 5efe036b..c1a393e1 100644
--- a/apps/040_hello_workflow.py
+++ b/apps/040_hello_workflow.py
@@ -173,7 +173,7 @@ class HelloFlow:
         self.ui = self.wand.get_ui_constants()
 
         # Dynamically load the AI's instruction manual
-        skill_path = self.wand.paths.base / ".agents" / "skills" / "hello_workflow" / "SKILL.md"
+        skill_path = self.wand.paths.base / ".agents" / "skills" / "hello-workflow" / "SKILL.md"
         if skill_path.exists():
             self.TRAINING_PROMPT = skill_path.read_text(encoding='utf-8')
         else:
(nix) pipulate $ m
📝 Committing: chore: Rename 'hello_workflow' to 'hello-workflow'
[main 6748cc0f] chore: Rename 'hello_workflow' to 'hello-workflow'
 1 file changed, 1 insertion(+), 1 deletion(-)
(nix) pipulate $ mkdir -p .agents/skills/journal/references && git mv AI_CONTEXT.md .agents/skills/journal/references/index.md && git status --short
R  AI_CONTEXT.md -> .agents/skills/journal/references/index.md
(nix) pipulate $ d
(nix) pipulate $ patch
(nix) pipulate $ app
✅ WHOLE-FILE WRITE: OVERWROTE 'scripts/articles/generate_ai_context.py'.
(nix) pipulate $ d
diff --git a/scripts/articles/generate_ai_context.py b/scripts/articles/generate_ai_context.py
index 2d57509e..4b763136 100644
--- a/scripts/articles/generate_ai_context.py
+++ b/scripts/articles/generate_ai_context.py
@@ -2,21 +2,27 @@
 """
 generate_ai_context.py
 
-Writes AI_CONTEXT.md to the Pipulate repository root: a self-contained briefing
-that lets the repo "talk back" to any AI that clones and inspects it.
-
-It fuses a small static framing header (what Pipulate is, how to drill down, the
-player-piano protocol) with a URL-first, reverse-chronological narrative ledger
-pulled from the blog archive via lsa.get_holographic_article_data(). Article
-bodies are NEVER checked into this repo — only their absolute, fetchable
-/index.md URLs — so the repo stays lean while still pointing an AI at the full
-intellectual history.
+Writes the journal index that the `journal` Agent Skill reads on demand:
+.agents/skills/journal/references/index.md, a URL-first, reverse-chronological
+ledger of the blog archive pulled via lsa.get_holographic_article_data().
+Article bodies are NEVER checked into this repo, only their absolute,
+fetchable /index.md URLs, so the repo stays lean while still pointing an AI
+at the full intellectual history.
+
+THE FILE MOVED (2026-09-28). It was AI_CONTEXT.md at the repo root: the AI_
+prefix sorted it first in `ls`, which AGENTS.md now does by name, and its
+header restated AGENTS.md and the code, which AGENTS.md forbids in so many
+words. Under the Agent Skills specification (agentskills.io) the index is a
+reference file: the skill's name and description cost a few tokens every
+session, and this file is opened only when a question needs the reasoning
+behind the machinery. The assurance posture the old header carried lives in
+AUDIT.md. The script keeps its name so release.py keeps its call.
 
 Standalone by design: depends only on lsa.py (already externalized) and the
 standard library. No common.py coupling, no prompt_foo.py scaffolding to strip.
 
-Idempotent: rewrites AI_CONTEXT.md from scratch on every run. Intended as a
-release-pipeline step so a fresh clone always greets an AI with the latest map.
+Idempotent: rewrites the index from scratch on every run. Intended as a
+release-pipeline step so a fresh clone always carries the latest map.
 
 Usage:
     python scripts/articles/generate_ai_context.py            # default target (1)
@@ -38,7 +44,7 @@ import lsa
 
 # scripts/articles/generate_ai_context.py -> up three == pipulate repo root
 REPO_ROOT = Path(__file__).resolve().parent.parent.parent
-OUTPUT_FILE = REPO_ROOT / "AI_CONTEXT.md"
+OUTPUT_FILE = REPO_ROOT / ".agents" / "skills" / "journal" / "references" / "index.md"
 DEFAULT_BASE_URL = "https://mikelev.in"
 DEFAULT_LIMIT = 0          # 0 = no limit; all articles indexed
 FULL_URL_THRESHOLD = 20    # First N entries use full URLs; rest use compact slugs
@@ -70,165 +76,33 @@ def article_slug(item: dict) -> str:
 
 
 def build_header(article_count: int, base_url: str, prefix: str) -> str:
-    """The static framing an interrogating AI reads before the index."""
+    """The short framing a reader sees before the index. Everything the old
+    root file's header said about setup, tools, edits and assurance now lives
+    where it belongs (AGENTS.md, the code, AUDIT.md); this file only points."""
     today = datetime.now().strftime("%Y-%m-%d")
-    host = base_url.replace("https://", "").replace("http://", "").rstrip("/")
     folder = f"/{prefix}" if prefix else ""
-    return f"""# AI_CONTEXT.md — Start Here If You Are an AI
-
-> Auto-generated on {today}. Rewritten from scratch on every release. If this
-> date looks stale, assume the rest of the repo is newer than this map.
-> {article_count} entries indexed.
-
-You are reading the orientation file for **Pipulate**, a local-first,
-hand-cranked AI/SEO automation framework (the NPvg stack: Nix, Python, Vim,
-Git). If a human asked you to investigate `pipulate/pipulate`, read this before
-forming conclusions.
-
-## Core Project Blueprint (For AI Ingestion)
-
-### 1. Architecture & Core Paradigms
-- **Single-Tenant Sovereignty:** This is a local-first, single-tenant workspace appliance. It executes entirely on the host machine, containing all interactions within local boundaries (`localhost:5001`).
-- **Durable State Engine:** Application persistence utilizes localized SQLite instances interfaced via `fastlite` handles. Universal workflows mutate state safely through the centralized `DictLikeDB` wrapper found in `pipulate/core.py`.
-- **Interface Foundations:** The UI is orchestrated via FastHTML and dynamic reactive HTMX views, with global styling parameters completely centralized inside `config.py`.
-
-### 2. Runtime Environment Invariants
-- **Deterministic Nix Shells:** Full environment stability and reproducible packaging are guaranteed exclusively by the pinned inputs of `flake.nix`.
-- **The Execution Constraint Warning:** NEVER invoke raw global interpreter binaries (`python`). ALWAYS force usage of the isolated virtual environment path: `.venv/bin/python`.
-- **Topological Manifold Mappings:** Workspace directories are calculated deterministically from the frame root layout definitions via structural utilities built into `pipulate/__init__.py`.
-
-### 3. Asymmetric Code Patching Protocol
-- **Surgical Code Modifications:** Any programmatic code changes are executed entirely via a strict search-and-replace pipeline (`apply.py`). Indentation structural matrices, inline comments, and whitespace layout details must be handled verbatim.
-- **Out-of-Band Clipboard Routing:** The synaptic communication ring relies on parsing explicit token block invariants (`[[[TODO_SLUGS]]]`, `[[[TODO_FILES]]]`, `[[[APPLY_PATCH]]]`) passed across the system clip buffer using `scripts/xp.py`.
-
-### 4. PyPI Packaging & Repository Landscapes
-- **Dependency Definitions:** Base requirements are managed via top-level entries in `requirements.in` and locked down systematically using pip-compile downstream.
-- **Negative Space Strategy:** Version-controlled core frameworks share directory space safely with private client deliverables and ad-hoc trace targets (`Workshop/personal/`) explicitly managed by `.gitignore` exclusions.
-
-### 5. The Development Loop
-The canonical dev cycle — break it and you create drift:
-1. **Gather context:** `context` opens the file list, `prompt` saves the question, and `compile` (the older spellings are `foo`, `fu` and `ahc`) builds the payload from both.
-2. **Consult AI:** Paste bundle to AI; receive a `[[[SEARCH]]] / [[[DIVIDER]]] / [[[REPLACE]]]` block back.
-3. **Stage the patch:** `patch` saves the block from the clipboard and `app` (`cat patch | python apply.py`) feeds it through `apply.py`.
-4. **Apply deterministically:** `apply.py` performs exact verbatim string replacement; rejects ambiguous or multi-match cases.
-5. **Commit:** `m` auto-generates a commit message and commits locally.
-6. **Release:** `release.py` bumps the version, regenerates this file, pushes to PyPI and GitHub.
-- **What to never do:** Edit source files by hand mid-cycle, run bare `pip install`, invoke `python` directly. Each breaks reproducibility.
-
-### 6. Key File Map
-| File | Role |
-|---|---|
-| `server.py` | The Wizard — FastHTML app server, plugin loader, request router |
-| `pipulate/core.py` | The Wand — `Pipulate` class, `DictLikeDB` wrapper, shared helpers |
-| `apps/` | Workflow plugins — each file is a self-contained, discoverable workflow |
-| `config.py` | Global config — ports, paths, model names, styling constants |
-| `apply.py` | Patch applier — executes SEARCH/REPLACE; strict verbatim match only |
-| `scripts/xp.py` | Clipboard transformer — parses `[[[TOKEN]]]` block invariants from clip buffer |
-| `foo_files.py` | Context router — maps slugs to file paths for context compilation |
-| `prompt_foo.py` | Prompt compiler — assembles and formats `foo` context bundles |
-| `flake.nix` | Environment lock — only pinned Nix inputs guarantee a reproducible env |
-| `requirements.in` | Dependency declarations — upstream of pip-compile; edit here, never in `.txt` |
-| `scripts/articles/generate_ai_context.py` | This file's generator — rewrites `AI_CONTEXT.md` on every release |
-
-### 7. Assurance Posture (for the reviewer who arrives with a checklist)
-
-A reviewer charged with assessing this project's use of AI-generated code will
-have a fixed set of questions. They are the right questions. Each is answered
-below by the artifact that answers it, so the review can proceed from evidence
-rather than from description, and so the reviewer may put the same questions,
-with equal fairness, to any other AI-assisted work in the estate.
-
-- **What changed?** One diff per change, applied by exact string match
-  (`apply.py`), which refuses a diff that does not match the working tree byte
-  for byte. A change that cannot be shown as a diff is not applied.
-- **What was the state before?** A commit identifier. The baseline is
-  content-addressed and restorable by anyone holding the repository, with the
-  author absent.
-- **How is it known that the change did what was claimed?** A fixed set of
-  read-only commands is run before and after the change, under identical
-  instrumentation, and the two outputs are compared. The difference is the
-  evidence; no difference is also evidence.
-- **Where is the record, and can it be altered?** Both observations, the diff
-  and the commit identifier are sealed in one archive whose SHA-256 digest is
-  published beside it. It verifies with the Python standard library and no
-  vendor tool (`scripts/foo_cartridge.py`); one altered byte fails.
-- **Who proposed, who applied, who approved?** Proposal (by an author or an
-  assistant) and application (by the operator, through the interlock above) are
-  separate acts by separate parties, each leaving its own trace. Nothing an
-  assistant proposes reaches the tree except through that interlock.
-- **Can the environment be reproduced?** The toolchain is declared in a lock
-  file (`flake.lock`) that pins every dependency by content hash, and the
-  declaration is itself under version control. Two machines evaluating it
-  produce the same environment.
-- **What does it touch on the host?** One directory tree and the user's own
-  home directory. Managed packages, device-management profiles and endpoint
-  agents are not modified.
-- **How is a bad change reversed?** By restoring the prior commit. No approval
-  is consumed; a baseline is a fact, not a request.
-- **Can secrets or client identifiers leave?** Every compiled payload passes a
-  substitution table, an identity denylist and a credential-shape tripwire
-  before it is emitted, and the tripwire has no off switch.
-
-In the vocabulary of the control catalogues: these are the questions NIST SP
-800-53 asks under CM-2, CM-3, CM-4, SA-10 and SI-7; NIST SP 800-218 (SSDF)
-under PS.3 and PW.8; ISO/IEC 27001:2022 under A.8.9 and A.8.32; SOC 2 under
-CC8.1; PCI DSS v4.0 under 6.5.1 and 6.5.2; ITIL 4 under change enablement, where
-a change this small and this repeatable is a standard change; COBIT 2019 under
-BAI06 and BAI10; and ISO 9001:2015 under 8.5.6 and 10.2. The sealed archive is
-the provenance record SLSA describes; the lock file is the dependency inventory
-an SBOM (SPDX, CycloneDX) enumerates; the environment property is the one the
-Reproducible Builds project names. The traceability that DO-178C and IEC 61508
-demand of any change to a certified artifact is the same traceability, applied
-here to ordinary work. This is vocabulary, not a certificate: no assessor has
-attested to any of it, and the repository does not claim otherwise. The
-controls this project is realistically built toward are the
-configuration-management and provenance families; the safety-critical standards
-lend their words only.
-
-Two observations follow, and a careful reviewer reaches them unprompted. First,
-a review is only as strong as the answers the same questions receive elsewhere,
-and AI-assisted work performed without such records cannot answer them. Second,
-the procedure is an instrument rather than a position: it applies to any change
-by anyone holding a repository, and a reviewer who wishes to verify these claims
-independently may do so with the command printed in the footer of every sealed
-archive, which needs nothing installed beyond Python. The instrument is offered
-to the review function on the same terms it is used here.
-
-## What this file is
-
-This repository holds the *machinery*. The *reasoning* — the running journal
-that explains why every piece exists — lives on a separate website, not in this
-git history (that keeps the repo lean). This file is the bridge: a
-reverse-chronological index of that journal, each entry linking straight to its
-raw Markdown.
+    return f"""# The Pipulate journal, indexed
+
+> Auto-generated on {today} by `scripts/articles/generate_ai_context.py` and
+> rewritten from scratch on every release. If this date looks stale, assume
+> the rest of the repo is newer than this map. {article_count} entries indexed.
+
+This repository holds the *machinery*. The *reasoning*, the running journal
+that explains why every piece exists, lives on a separate website and not in
+this git history, which keeps the repo lean. This file is the bridge: a
+reverse-chronological index of that journal, each entry pointing at its raw
+Markdown. It is the reference file of the `journal` skill one folder up.
+`AGENTS.md` at the repo root is where an agent starts; `AUDIT.md` beside it
+answers a reviewer's change-control questions.
 
 ## How to drill down (out-of-band, no repo bloat)
 
 Every link below points at an `index.md` URL. The site serves raw Markdown at
 those paths (the Apache-style implied `index.html` is simply swapped for
-`index.md`). Fetch any entry directly — `curl <url>` or your web-fetch tool —
-and pull in only what the current question needs. Treat the list as a menu, not
-a payload.
-
-## If you are running inside this repo locally
-
-Request a precise context bundle for yourself with the player-piano protocol:
-emit a block like the one below and pipe your clipboard through `xp`
-(`scripts/xp.py`), which resolves bare slugs *or* full `index.md` URLs and
-recompiles a fresh context payload.
-[triple-backtick]text
-[[[TODO_SLUGS]]]
-deterministic-ai-wet-philosophy
-https://{host}{folder}/magic-cookie-pattern-self-bootstrapping-sovereignty/index.md
-[[[END_SLUGS]]]
-
-[[[TODO_FILES]]]
-prompt_foo.py
-apply.py
-[[[END_FILES]]]
-[triple-backtick]
-To list the tools an assistant can call from inside the repository, run
-`.venv/bin/python cli.py mcp-discover`.
+`index.md`). Fetch any entry directly, with `curl <url>` or a web-fetch tool,
+and pull in only what the current question needs. Treat the list as a menu,
+not a payload. Inside a checkout, `scripts/xp.py` turns a pasted list of slugs
+into the next compile's context; `AGENTS.md` names that grammar.
 
 ## The narrative index (newest first)
 
@@ -282,7 +156,7 @@ def build_ledger(target_config: dict, rich: bool, limit) -> tuple:
 
 
 def main():
-    parser = argparse.ArgumentParser(description="Generate AI_CONTEXT.md repo briefing.")
+    parser = argparse.ArgumentParser(description="Generate the journal index (.agents/skills/journal/references/index.md).")
     parser.add_argument("-t", "--target", type=str, default="1", help="Target ID from blogs.json (default: 1)")
     parser.add_argument("--rich", action="store_true", help="Append holographic-shard keywords to each entry.")
     parser.add_argument("--limit", type=int, default=DEFAULT_LIMIT, help=f"Index only the N newest articles (default: {DEFAULT_LIMIT}; 0 = all).")
@@ -298,12 +172,13 @@ def main():
     limit = args.limit if (args.limit and args.limit > 0) else None
     prefix = lsa.permalink_prefix(target_config)
 
-    print(f"🧭 Generating AI_CONTEXT.md from target: {target_config.get('name', target_key)}")
+    print(f"🧭 Generating the journal index from target: {target_config.get('name', target_key)}")
     ledger, count = build_ledger(target_config, args.rich, limit)
     header = build_header(count, base_url, prefix)
     body = ledger if ledger else "_No articles indexed (article source unavailable at generation time)._"
 
     final = header + "\n" + body + "\n"
+    OUTPUT_FILE.parent.mkdir(parents=True, exist_ok=True)
     OUTPUT_FILE.write_text(final, encoding="utf-8")
     print(f"✅ Wrote {OUTPUT_FILE} ({count} entries, {len(final.encode('utf-8')):,} bytes).")
 
(nix) pipulate $ m
📝 Committing: chore: Rename AI_CONTEXT.md to .agents/skills/journal/references/index.md
[main 2f58b8cb] chore: Rename AI_CONTEXT.md to .agents/skills/journal/references/index.md
 2 files changed, 41 insertions(+), 166 deletions(-)
 rename AI_CONTEXT.md => .agents/skills/journal/references/index.md (100%)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ WHOLE-FILE WRITE: CREATED '.agents/skills/journal/SKILL.md'.
(nix) pipulate $ d

--- UNTRACKED (invisible to the diff above; m will stage these) ---
  + .agents/skills/journal/SKILL.md
(nix) pipulate $ m
❌ ai.py returned empty message
(nix) pipulate $ git add .agents/skills/journal/SKILL.md
(nix) pipulate $ m
📝 Committing: chore: Update journal skill documentation
[main 2818cd4a] chore: Update journal skill documentation
 1 file changed, 37 insertions(+)
 create mode 100644 .agents/skills/journal/SKILL.md
(nix) pipulate $ .venv/bin/python scripts/articles/generate_ai_context.py && git add .agents/skills/journal && git status --short
🧭 Generating the journal index from target: MikeLev.in (Public)
✅ Wrote /home/mike/repos/pipulate/.agents/skills/journal/references/index.md (1499 entries, 94,548 bytes).
M  .agents/skills/journal/references/index.md
(nix) pipulate $ d
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'release.py'.
(nix) pipulate $ d
diff --git a/release.py b/release.py
index 4abee4ae..615d3e01 100755
--- a/release.py
+++ b/release.py
@@ -246,12 +246,13 @@ def run_waxascii_release_stamp():
         return False
 
 def run_ai_context_generation():
-    """Regenerate AI_CONTEXT.md — the repo's self-describing briefing for any AI
-    that clones and inspects it. Reads the (separate) blog archive and rewrites
-    AI_CONTEXT.md in the Pipulate repo root from scratch, so a fresh clone always
-    greets an AI with the latest narrative map. Non-fatal: skips cleanly if the
+    """Regenerate the journal index, .agents/skills/journal/references/index.md,
+    the repo's self-describing map for any AI that clones and inspects it.
+    Reads the (separate) blog archive and rewrites the index from scratch, so a
+    fresh clone always carries the latest narrative map. Until 2026-09-28 this
+    file was AI_CONTEXT.md at the repo root. Non-fatal: skips cleanly if the
     generator or the article source is unavailable."""
-    note("\n🧭 Step 1.6: Regenerating AI_CONTEXT.md (repo talk-back briefing)...")
+    note("\n🧭 Step 1.6: Regenerating the journal index (.agents/skills/journal/references/index.md)...")
     generator = PIPULATE_ROOT / "scripts" / "articles" / "generate_ai_context.py"
     if not generator.exists():
         print(f"ℹ️  AI_CONTEXT generator not found at {generator}. Skipping.")
(nix) pipulate $ m
📝 Committing: refactor: update AI context index
[main 5cf51dba] refactor: update AI context index
 2 files changed, 27 insertions(+), 156 deletions(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'release.py'.
(nix) pipulate $ d
diff --git a/release.py b/release.py
index 615d3e01..61d2fe28 100755
--- a/release.py
+++ b/release.py
@@ -255,7 +255,7 @@ def run_ai_context_generation():
     note("\n🧭 Step 1.6: Regenerating the journal index (.agents/skills/journal/references/index.md)...")
     generator = PIPULATE_ROOT / "scripts" / "articles" / "generate_ai_context.py"
     if not generator.exists():
-        print(f"ℹ️  AI_CONTEXT generator not found at {generator}. Skipping.")
+        print(f"ℹ️  journal index generator not found at {generator}. Skipping.")
         return False
     # Direct subprocess.run (not run_command) so a failure never sys.exit()s the release.
     result = subprocess.run([sys.executable, str(generator)], cwd=str(PIPULATE_ROOT),
(nix) pipulate $ m
📝 Committing: chore: Update AI context generator message in release.py
[main 19442cb5] chore: Update AI context generator message in release.py
 1 file changed, 1 insertion(+), 1 deletion(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'release.py'.
(nix) pipulate $ d
diff --git a/release.py b/release.py
index 61d2fe28..1a20dabf 100755
--- a/release.py
+++ b/release.py
@@ -261,7 +261,7 @@ def run_ai_context_generation():
     result = subprocess.run([sys.executable, str(generator)], cwd=str(PIPULATE_ROOT),
                             capture_output=not VERBOSE, text=True)
     if result.returncode != 0:
-        print("⚠️  AI_CONTEXT generation returned non-zero; continuing release.")
+        print("⚠️  journal index generation returned non-zero; continuing release.")
         for stream in (result.stdout, result.stderr):
             if stream and stream.strip():
                 print(stream.rstrip(), file=sys.stderr)
(nix) pipulate $ m
📝 Committing: chore: Update AI_CONTEXT generation error message
[main 18b65487] chore: Update AI_CONTEXT generation error message
 1 file changed, 1 insertion(+), 1 deletion(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'release.py'.
(nix) pipulate $ d
diff --git a/release.py b/release.py
index 1a20dabf..12836708 100755
--- a/release.py
+++ b/release.py
@@ -267,9 +267,9 @@ def run_ai_context_generation():
                 print(stream.rstrip(), file=sys.stderr)
         return False
     # Stage explicitly: `git commit -am` ignores untracked files, so the very
-    # first (untracked) AI_CONTEXT.md must be added by hand. After that it rides -am.
-    subprocess.run(["git", "add", "AI_CONTEXT.md"], cwd=str(PIPULATE_ROOT))
-    note("✅ AI_CONTEXT.md regenerated and staged.")
+    # first (untracked) index must be added by hand. After that it rides -am.
+    subprocess.run(["git", "add", ".agents/skills/journal/references/index.md"], cwd=str(PIPULATE_ROOT))
+    note("✅ journal index regenerated and staged.")
     return True
 
 def parse_ascii_art_stats(output):
(nix) pipulate $ m
📝 Committing: chore: Stage journal index regeneration
[main 0327639b] chore: Stage journal index regeneration
 1 file changed, 3 insertions(+), 3 deletions(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'release.py'.
(nix) pipulate $ d
diff --git a/release.py b/release.py
index 12836708..a44e9251 100755
--- a/release.py
+++ b/release.py
@@ -583,7 +583,7 @@ def sync_audit_md():
         return False
 
 def sync_ai_context_md():
-    """Copies AI_CONTEXT.md to Pipulate.com root and commits if changed.
+    """Copies the journal index to Pipulate.com root, as AI_CONTEXT.md, and commits if changed.
 
     Note: AI_CONTEXT.md is regenerated from scratch at Step 1.6
     (run_ai_context_generation), so by the time this runs the source is fresh.
(nix) pipulate $ m
📝 Committing: chore: Refactor AI context sync documentation in release.py
[main dcf6d68f] chore: Refactor AI context sync documentation in release.py
 1 file changed, 1 insertion(+), 1 deletion(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'release.py'.
(nix) pipulate $ d
diff --git a/release.py b/release.py
index a44e9251..91174196 100755
--- a/release.py
+++ b/release.py
@@ -585,7 +585,7 @@ def sync_audit_md():
 def sync_ai_context_md():
     """Copies the journal index to Pipulate.com root, as AI_CONTEXT.md, and commits if changed.
 
-    Note: AI_CONTEXT.md is regenerated from scratch at Step 1.6
+    Note: the index is regenerated from scratch at Step 1.6
     (run_ai_context_generation), so by the time this runs the source is fresh.
     """
     note("\n🔄 Step 3.6: Synchronizing AI_CONTEXT.md to Pipulate.com...")
(nix) pipulate $ m
📝 Committing: chore: Update AI_CONTEXT.md regeneration note in release.py
[main 401b7bac] chore: Update AI_CONTEXT.md regeneration note in release.py
 1 file changed, 1 insertion(+), 1 deletion(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'release.py'.
(nix) pipulate $ d
diff --git a/release.py b/release.py
index 91174196..a37322e6 100755
--- a/release.py
+++ b/release.py
@@ -588,8 +588,13 @@ def sync_ai_context_md():
     Note: the index is regenerated from scratch at Step 1.6
     (run_ai_context_generation), so by the time this runs the source is fresh.
     """
-    note("\n🔄 Step 3.6: Synchronizing AI_CONTEXT.md to Pipulate.com...")
-    source_path = PIPULATE_ROOT / "AI_CONTEXT.md"
+    note("\n🔄 Step 3.6: Synchronizing the journal index to Pipulate.com as AI_CONTEXT.md...")
+    source_path = PIPULATE_ROOT / ".agents" / "skills" / "journal" / "references" / "index.md"
+    # THE URL IS A PROMISE (2026-09-28): the source moved into the journal
+    # skill and the public copy keeps its old name, because
+    # https://pipulate.com/AI_CONTEXT.md is linked from README.md and from
+    # Pipulate.com's own, and may be bookmarked. Renaming it is a Jekyll
+    # redirect ride, never a rename here.
     dest_path = PIPULATE_COM_ROOT / "AI_CONTEXT.md"
 
     if not PIPULATE_COM_ROOT.exists():
(nix) pipulate $ m
📝 Committing: chore: Update AI_CONTEXT.md source path
[main 51d755e0] chore: Update AI_CONTEXT.md source path
 1 file changed, 7 insertions(+), 2 deletions(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'release.py'.
(nix) pipulate $ d
diff --git a/release.py b/release.py
index a37322e6..720c71ea 100755
--- a/release.py
+++ b/release.py
@@ -602,7 +602,7 @@ def sync_ai_context_md():
         return False
 
     if not source_path.exists():
-        print(f"⚠️  Warning: Source AI_CONTEXT.md not found at {source_path}. Skipping AI_CONTEXT.md sync.")
+        print(f"⚠️  Warning: Source journal index not found at {source_path}. Skipping AI_CONTEXT.md sync.")
         return False
 
     # Copy the file
(nix) pipulate $ m
📝 Committing: fix: Correctly handle missing journal index
[main 2d0b3016] fix: Correctly handle missing journal index
 1 file changed, 1 insertion(+), 1 deletion(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'release.py'.
(nix) pipulate $ d
diff --git a/release.py b/release.py
index 720c71ea..8e289b26 100755
--- a/release.py
+++ b/release.py
@@ -1287,7 +1287,7 @@ def main():
         print("\n⏭️  Skipping Waxascii release stamping (--skip-docs-sync)")
         waxascii_sync_success = True
 
-    # Step 1.6: Regenerate the AI_CONTEXT.md repo briefing (talk-back map)
+    # Step 1.6: Regenerate the journal index (the repo's talk-back map)
     if not args.skip_docs_sync:
         run_ai_context_generation()
     else:
(nix) pipulate $ m
📝 Committing: chore: Update AI_CONTEXT.md repo briefing description
[main 9b608db0] chore: Update AI_CONTEXT.md repo briefing description
 1 file changed, 1 insertion(+), 1 deletion(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'release.py'.
(nix) pipulate $ d
diff --git a/release.py b/release.py
index 8e289b26..fcc49075 100755
--- a/release.py
+++ b/release.py
@@ -1291,7 +1291,7 @@ def main():
     if not args.skip_docs_sync:
         run_ai_context_generation()
     else:
-        print("\n⏭️  Skipping AI_CONTEXT.md regeneration (--skip-docs-sync)")
+        print("\n⏭️  Skipping journal index regeneration (--skip-docs-sync)")
     
     # Docs-sync step retired. Its "Skipping (--skip-docs-sync)" line printed on
     # EVERY run whether or not the flag was given: a false statement in the
(nix) pipulate $ m
📝 Committing: chore: Update AI_CONTEXT regeneration message
[main 0d9a8caa] chore: Update AI_CONTEXT regeneration message
 1 file changed, 1 insertion(+), 1 deletion(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'AUDIT.md'.
(nix) pipulate $ d
diff --git a/AUDIT.md b/AUDIT.md
index 6bfd46bb..0988d64c 100644
--- a/AUDIT.md
+++ b/AUDIT.md
@@ -167,6 +167,70 @@ a reassuring, sparse default and an exhaustive audit/debug mode. A quiet default
 is not opacity, and a verbose audit mode is available on request; the two coexist
 without flooding the first-run terminal.
 
+## Change control: the questions and the artifacts that answer them
+
+A reviewer charged with assessing this project's use of AI-generated code will
+have a fixed set of questions. They are the right questions. Each is answered
+below by the artifact that answers it, so the review can proceed from evidence
+rather than from description, and so the reviewer may put the same questions,
+with equal fairness, to any other AI-assisted work in the estate.
+
+- **What changed?** One diff per change, applied by exact string match
+  (`apply.py`), which refuses a diff that does not match the working tree byte
+  for byte. A change that cannot be shown as a diff is not applied.
+- **What was the state before?** A commit identifier. The baseline is
+  content-addressed and restorable by anyone holding the repository, with the
+  author absent.
+- **How is it known that the change did what was claimed?** A fixed set of
+  read-only commands is run before and after the change, under identical
+  instrumentation, and the two outputs are compared. The difference is the
+  evidence; no difference is also evidence.
+- **Where is the record, and can it be altered?** Both observations, the diff
+  and the commit identifier are sealed in one archive whose SHA-256 digest is
+  published beside it. It verifies with the Python standard library and no
+  vendor tool (`scripts/foo_cartridge.py`); one altered byte fails.
+- **Who proposed, who applied, who approved?** Proposal (by an author or an
+  assistant) and application (by the operator, through the interlock above) are
+  separate acts by separate parties, each leaving its own trace. Nothing an
+  assistant proposes reaches the tree except through that interlock.
+- **Can the environment be reproduced?** The toolchain is declared in a lock
+  file (`flake.lock`) that pins every dependency by content hash, and the
+  declaration is itself under version control. Two machines evaluating it
+  produce the same environment.
+- **What does it touch on the host?** One directory tree and the user's own
+  home directory. Managed packages, device-management profiles and endpoint
+  agents are not modified.
+- **How is a bad change reversed?** By restoring the prior commit. No approval
+  is consumed; a baseline is a fact, not a request.
+- **Can secrets or client identifiers leave?** Every compiled payload passes a
+  substitution table, an identity denylist and a credential-shape tripwire
+  before it is emitted, and the tripwire has no off switch.
+
+In the vocabulary of the control catalogues: these are the questions NIST SP
+800-53 asks under CM-2, CM-3, CM-4, SA-10 and SI-7; NIST SP 800-218 (SSDF)
+under PS.3 and PW.8; ISO/IEC 27001:2022 under A.8.9 and A.8.32; SOC 2 under
+CC8.1; PCI DSS v4.0 under 6.5.1 and 6.5.2; ITIL 4 under change enablement, where
+a change this small and this repeatable is a standard change; COBIT 2019 under
+BAI06 and BAI10; and ISO 9001:2015 under 8.5.6 and 10.2. The sealed archive is
+the provenance record SLSA describes; the lock file is the dependency inventory
+an SBOM (SPDX, CycloneDX) enumerates; the environment property is the one the
+Reproducible Builds project names. The traceability that DO-178C and IEC 61508
+demand of any change to a certified artifact is the same traceability, applied
+here to ordinary work. This is vocabulary, not a certificate: no assessor has
+attested to any of it, and the repository does not claim otherwise. The
+controls this project is realistically built toward are the
+configuration-management and provenance families; the safety-critical standards
+lend their words only.
+
+Two observations follow, and a careful reviewer reaches them unprompted. First,
+a review is only as strong as the answers the same questions receive elsewhere,
+and AI-assisted work performed without such records cannot answer them. Second,
+the procedure is an instrument rather than a position: it applies to any change
+by anyone holding a repository, and a reviewer who wishes to verify these claims
+independently may do so with the command printed in the footer of every sealed
+archive, which needs nothing installed beyond Python. The instrument is offered
+to the review function on the same terms it is used here.
+
 ## Closing
 
 The correct review of Pipulate is the review of a local-first, single-tenant,
(nix) pipulate $ m
📝 Committing: refactor: introduce change control questions and artifacts
[main bf789881] refactor: introduce change control questions and artifacts
 1 file changed, 64 insertions(+)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'README.md'.
(nix) pipulate $ d
diff --git a/README.md b/README.md
index 83924e21..f7d8ebd6 100644
--- a/README.md
+++ b/README.md
@@ -2,7 +2,7 @@
 
 1. For Auditors: AUDIT.md ( https://pipulate.com/AUDIT.md )
 2. For Agents: AGENTS.md (a README for agents) and `.agents/skills/*/SKILL.md` (Agent Skills)
-3. For Context: AI_CONTEXT.md ( https://pipulate.com/AI_CONTEXT.md )
+3. For the Why: `.agents/skills/journal/` (the journal index, an Agent Skills reference; served at https://pipulate.com/AI_CONTEXT.md )
 
 [triple-backtick]text
 ( "The AI said so" is a transcript, not a record. )
(nix) pipulate $ m
📝 Committing: chore: Update AI context reference to .agents/skills/journal/
[main 38f8a655] chore: Update AI context reference to .agents/skills/journal/
 1 file changed, 1 insertion(+), 1 deletion(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'AGENTS.md'.
(nix) pipulate $ d
diff --git a/AGENTS.md b/AGENTS.md
index 30184d9f..d3216f92 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -36,7 +36,8 @@ run since the sentinels landed.
 
 - Discover: `.venv/bin/python cli.py mcp-discover`
 - Execute:  `.venv/bin/python cli.py call <tool_name> --json-args '{...}'`
-- Skills (Agent Skills spec): `.agents/skills/*/SKILL.md`, at the repo root beside this file
+- Skills (Agent Skills spec, https://agentskills.io/specification): `.agents/skills/*/SKILL.md`, at the repo root beside this file
+- The reasoning behind any piece of the machinery: the `journal` skill, whose `references/index.md` lists every journal entry newest first with a fetchable URL
 
 ## Context (how this repo talks to AI)
 
(nix) pipulate $ m
📝 Committing: chore: Update Agent Skills spec documentation link
[main daed3d78] chore: Update Agent Skills spec documentation link
 1 file changed, 2 insertions(+), 1 deletion(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'pyproject.toml'.
(nix) pipulate $ d
diff --git a/pyproject.toml b/pyproject.toml
index f2e654e4..441a1065 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -91,7 +91,6 @@ exclude = [
     "assets/installer/install.sh", 
     "flake.nix",
     "README.md",
-    "AI_CONTEXT.md",
     "LICENSE",
     "CHANGELOG.md"
 ]
(nix) pipulate $ m
📝 Committing: chore: Update pyproject.toml exclude list
[main 4dc39785] chore: Update pyproject.toml exclude list
 1 file changed, 1 deletion(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'foo_files.py'.
(nix) pipulate $ d
diff --git a/foo_files.py b/foo_files.py
index d226fb6a..6fc17a33 100644
--- a/foo_files.py
+++ b/foo_files.py
@@ -1512,7 +1512,7 @@ foo_files.py      #  <-- THIS file. Content compiler router. Makes it very meta.
 # ! python scripts/articles/lsa.py -t 4 --reverse --fmt dated-slugs  # <-- BotifyML work journal
 # scripts/articles/lsa.py   # <-- You can show the AI how the magic rolling pin works
 
-# AI_CONTEXT.md
+# .agents/skills/journal/references/index.md
 # scripts/takeover_main.sh  # <-- Successful branch experiments rapidly take-over main when successful
 
 # ----------------------------------------------------------------------------
(nix) pipulate $ m
📝 Committing: chore: Update references file name
[main 8b2e8267] chore: Update references file name
 1 file changed, 1 insertion(+), 1 deletion(-)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'foo_files.py'.
(nix) pipulate $ d
diff --git a/foo_files.py b/foo_files.py
index 6fc17a33..179594bd 100644
--- a/foo_files.py
+++ b/foo_files.py
@@ -1850,6 +1850,8 @@ foo_files.py      #  <-- THIS file. Content compiler router. Makes it very meta.
 # THE SKILL.md FILES
 # .agents/skills/gsc-readonly/SKILL.md
 # .agents/skills/hello-workflow/SKILL.md
+# .agents/skills/journal/SKILL.md
+# .agents/skills/journal/references/index.md
 # .agents/skills/roles/SKILL.md
 # .agents/skills/sheets-readonly/SKILL.md
 
(nix) pipulate $ m
📝 Committing: chore: Add journal skill files and references
[main 46de5e64] chore: Add journal skill files and references
 1 file changed, 2 insertions(+)
(nix) pipulate $ patch
(nix) pipulate $ app
✅ DETERMINISTIC PATCH APPLIED: Successfully mutated 'MANIFEST.in'.
(nix) pipulate $ d
diff --git a/MANIFEST.in b/MANIFEST.in
index ae4c7c83..937347ed 100644
--- a/MANIFEST.in
+++ b/MANIFEST.in
@@ -3,3 +3,12 @@
 # — and the wheel `python -m build` derives from it — in agreement with
 # that convention. Fix protects v2.1 forward; the 2.0 artifact is history.
 exclude apps/xx_*.py
+
+# THE AGENT-FACING FILES SHIP WITH THE SOURCE (2026-09-28). setuptools' sdist
+# takes package modules, README, LICENSE and pyproject.toml on its own and no
+# other root file, so the two files an agent reads first and the skills folder
+# were absent from every sdist an auditor could download. The package-data
+# block in pyproject.toml does not reach them either: its patterns resolve
+# inside package directories, and none of these files lives in one.
+include AGENTS.md AUDIT.md
+graft .agents
(nix) pipulate $ m
📝 Committing: chore: Add agent-facing files and documentation
[main bb8c137c] chore: Add agent-facing files and documentation
 1 file changed, 9 insertions(+)
(nix) pipulate $ git push
Enumerating objects: 90, done.
Counting objects: 100% (90/90), done.
Delta compression using up to 48 threads
Compressing objects: 100% (38/38), done.
Writing objects: 100% (75/75), 41.12 KiB | 8.22 MiB/s, done.
Total 75 (delta 44), reused 59 (delta 31), pack-reused 0 (from 0)
remote: Resolving deltas: 100% (44/44), completed with 13 local objects.
To github.com:pipulate/pipulate.git
   1e55aefc..bb8c137c  main -> main
(nix) pipulate $ 

Ignition: blah, blah, not needed, provided by release process I think which I’ll want to do on my own the end of this turn.

4: Prompt: Read the receipts first: the skills census (five rows), the loader line, the ls and the head, the AI_CONTEXT census, and both sdist censuses; say what 2.65’s sdist and wheel actually carry and whether the package-data reading held. Then write .agents/skills/pipulate/SKILL.md, the install skill for an agent on a stranger’s machine, from install.sh, the npvg.org door page, README’s Quick Start and AUDIT.md’s first-run section: inspect first (curl -fsSL https://npvg.org

less, what the reader sees, j/k/q), then save, inspect and execute the same file, then the one-line install, the three doors after nix develop, what stays local and what reaches the cloud, and how to stop, reset and uninstall; two sentences of the why and no more. Say in one line why the compile loop stays in AGENTS.md rather than a skill. After that, the fantasy league, from where the 09-28 board left off, once the SVB lines are uncommented.

5: Deliverables: Hmmm, let’s deliver a deliverable.

(nix) pipulate $ g

Left-hand causal boundary "blast radius" established. Make 1 change and test.
On branch main
Your branch is up to date with 'origin/main'.

nothing to commit, working tree clean
(nix) pipulate $ vim __init__.py 
(nix) pipulate $ d
diff --git a/__init__.py b/__init__.py
index 8f59e202..138e72a1 100644
--- a/__init__.py
+++ b/__init__.py
@@ -12,7 +12,7 @@ Usage:
     pipulate
 """
 
-__version__ = "2.65"
+__version__ = "2.66"
 # APOSTROPHES RESTORED (2026-08-04). They were stripped as a workaround for
 # flake.nix's descMatch regex, whose character class excluded ' from the
 # CAPTURE and truncated the banner to "(So)". That regex was fixed in the same
@@ -23,7 +23,7 @@ __version__ = "2.65"
 # is not a property of the system; the regex is. Blast radius is one banner:
 # nothing but flake.nix reads this name -- version_sync.py syncs __version__
 # and __description__, never this. So'wI' chu' -- "engage the cloaking device."
-__version_description__ = "About on the Menu"
+__version_description__ = "Convention Convention"
 # SPDX expression, single source of truth, synced into pyproject.toml by
 # scripts/release/version_sync.py. "-or-later" (not bare AGPL-3.0, which is
 # deprecated SPDX) because the header below grants "any later version".
(nix) pipulate $ 

And the release:

(nix) pipulate $ release
╭───────────────────────────────────────────────────────── LLM Response Quality Assurance ──────────────────────────────────────────────────────────╮
│                                                                                                                                                   │
│                      ( "The AI said so" is a transcript, not a record. )                                                                          │
│                                            O        /)  ____     This drawing is checksummed.                                                     │
│ >  Same files, same words, same bytes:      o /)\__//  /    \    If it reaches you altered, the                                                   │
│ >  a change you can replay is a change    ___(/_ 0 0  |      |   model rewrote what it was told to                                                │
│ >  you can put your name behind.        *(    ==(_T_)== NPvg |   copy, and that is the finding.                                                   │
│ >  One diff per turn, a reading before    \  )   ""\  |      |   Verify with nothing installed:                                                   │
│ >  and after, and a digest you can check.  |__>-\_>_>  \____/    python scripts/foo_cartridge.py foo.zip                                          │
│                                                                                                                                                   │
╰───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
📋 Current version: 2.66
✅ Updated pyproject.toml (version and description)
✅ Pushed AUDIT.md update to Pipulate.com repo.
✅ Pushed AI_CONTEXT.md update to Pipulate.com repo.
🗂️  index.md workspace tree regenerated from the sealed asset.
✅ Pushed index.md workspace tree update to Pipulate.com repo.
[main ace3eda4] version: update __version__ to 2.66
✅ Pushed bb8c137c..ace3eda4  main -> main
Successfully built pipulate-2.66.tar.gz and pipulate-2.66-py3-none-any.whl
🎉 Published 2.66 -> https://pypi.org/project/pipulate/2.66/

╭────────────────────────────────────────────────────────── 🎉 Release Pipeline Complete ───────────────────────────────────────────────────────────╮
│                                                                                                                                                   │
│                                                            🎉 Pipulate Release Summary                                                            │
│  ╭────────────────────────────────┬──────────────────────────────────────────────────────────────────────────────────────────┬─────────────────╮  │
│  │ Component                      │ Details                                                                                  │     Status      │  │
│  ├────────────────────────────────┼──────────────────────────────────────────────────────────────────────────────────────────┼─────────────────┤  │
│  │ 🤖 gemma3:latest Message       │ version: update __version__ to 2.66                                                      │       ✨        │  │
│  │                                │                                                                                          │  gemma3:latest  │  │
│  ├────────────────────────────────┼──────────────────────────────────────────────────────────────────────────────────────────┼─────────────────┤  │
│  │ 📦 Version                     │ 2.66                                                                                     │     ✅ Set      │  │
│  ├────────────────────────────────┼──────────────────────────────────────────────────────────────────────────────────────────┼─────────────────┤  │
│  │ 🚀 PyPI Release                │ https://pypi.org/project/pipulate/2.66/                                                  │     ✅ Live     │  │
│  ├────────────────────────────────┼──────────────────────────────────────────────────────────────────────────────────────────┼─────────────────┤  │
│  │ ⏰ Completed                   │ 2026-09-28 19:06:42                                                                      │     🎯 Done     │  │
│  ╰────────────────────────────────┴──────────────────────────────────────────────────────────────────────────────────────────┴─────────────────╯  │
│                                                                                                                                                   │
╰───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
🔄 server.py touched; the watchdog restarts the server.
(nix) pipulate $ 

Hmmm, I still saw mention of AI_CONTEXT.md in the release process. Let’s go look at the PyPI page. It says:

For Auditors: AUDIT.md ( https://pipulate.com/AUDIT.md )
For Agents: AGENTS.md (a README for agents) and .agents/skills/*/SKILL.md (Agent Skills)
For the Why: .agents/skills/journal/ (the journal index, an Agent Skills reference; served at https://pipulate.com/AI_CONTEXT.md )

Okay, I get it. Not bad. I’m sure the rationale for not changing that filename is in here somewhere. But this is a great place to wrap this article. Great work Fable 5.1! This was another fine Sandworm ride… oops, article. There’s no Sci-Fi geeking out here. Please disperse. Please disperse.

Release 2.66 and the Public Promise

Hop off the ride. This ride’s stated goal is reached — dismount. This is the NOTARY BEAT: the ride ends here, is witnessed here, and is sealed here. Answer all seven beats, briefly:

TL;DR: a short, dry, neutral abstract for the TOP of the published article — written for an unfamiliar reader or AI summarizer who has never seen this system. No hype, no insider handles unexplained.
VERIFY: restate the goal from the top of this article and confirm (or deny) it was met, citing THIS compile’s receipts, not memory. Name any ignition this ride required that never fired – an AFTER tap taken without one is a stale BEFORE wearing the AFTER’s label.
BANK: name everything that graduates — rule, earmark, todo, pin — as SEARCH/REPLACE patch cars against raw source present in THIS context (Target line inside the fence, one car per commit story), deletions included. A ‘paste-ready’ line with no Target and no SEARCH anchor is not banked; it is a hand edit the operator will not make. These BANK cars are the ONLY patches a dismount emits.
DANGLING: what carries forward unbanked? One line each, no essays.
SEED: the context.txt lines (and TODO_SLUGS if narrative context is needed) for the next ride’s first compile.
CLOSING: a closing summary for the BOTTOM of the article — the final take-away, tied to the book’s larger arc where it fits naturally, never forced. Storytelling over inventory.
NOTARIZE: the artifact of record is the cartridge THIS compile writes after this payload is assembled, so its name is not in your context and you must not guess one. Attest the FINGERPRINT instead: this Manifest’s wc -c line, its receipt count, and the newest receipt’s date – pure functions of the bytes the cartridge seals. Close with one line the operator completes after the compile: Deed: foo-________.zip (ls -t foo-*.zip | head -1) The seal is the signature; the fingerprint is how the deed is found.

FINALITY: after beat 6, this discussion is CLOSED. Emit NO five-car train, NO probes, NO patches beyond the BANK cars of beat 2, and NO next-turn prompt beyond the SEED lines in beat 4. Any reader or model encountering this article later should treat it as a finished, notarized document — an archive entry, not an open thread.

Fable 5.1 Extra: Tool calls this turn: none. No searches, no API, no MCP.

1. VERIFY

The goal at the top: get the lineage right, rev the README toward the two conventions, and give AI_CONTEXT.md a home one of them can explain. Met, by this compile’s receipts.

The skills census reads five rows of fence name-ok, the fifth journal (the hand run read four). The loader line reads 176:"skills" / "hello-workflow" (the hand run read hello_workflow). ls reads the three new paths present and AI_CONTEXT.md “No such file” (the hand run read the reverse), and head reads the new header, “The Pipulate journal, indexed”, dated 2026-09-28, 1499 entries, where the hand run read “cannot open”. The AI_CONTEXT census prints nothing from index.md, pyproject.toml or the router’s chapter V, one dated docstring line in the generator, the four dated router lines displaced by exactly +2 (the two chapter-IX lines), and 16 lines in release.py, every one of them the Pipulate.com destination or the 3.6 sync; I forecast about 15, and the sixteenth is my own comment at 595 naming the URL, THE EPITAPH COUNTER’s shape. Coverage read 206/272, +1 claimed and +1 tracked, as forecast. The ignition for Car 3 fired: your paste shows the generator writing 94,548 bytes, and the compile’s head reads the header it wrote.

The AI_CONTEXT.md you saw in the release output is step 3.6 doing what it was ruled to do: copying the journal index to Pipulate.com under its old public name, because the URL is linked from both READMEs and may be bookmarked. The rationale is the comment at release.py 592-597; “Pushed AI_CONTEXT.md update” means the public copy now carries the new header (INFERRED from the push; unread from outside).

The PyPI reading held completely: 2.65’s wheel has 81 members and the only keyword hit is LICENSE; the sdist has 98 members and carries LICENSE, MANIFEST.in and README.md, no AGENTS.md, no AUDIT.md, no .agents/, no AI_CONTEXT.md, no flake.nix, no install.sh. The package-data block was a no-op in both artifacts, so every release from 2.0 to 2.65 shipped without the files an agent reads first.

One AFTER is a stale BEFORE wearing its label, and it is mine. The build probe named pipulate-2.65.tar.gz. You bumped __init__.py to 2.66 and released between the taps, so the compile’s build wrote pipulate-2.66.tar.gz and the tar line read the 2.65 file your hand run had left: build_rc=0, members=98, no hits, byte-identical to BEFORE, which is what a stale file prints. The MANIFEST car’s effect is therefore UNWITNESSED by this compile; the 2.66 artifacts on PyPI are the witness, built after bb8c137c, INFERRED to carry the three names. The PyPI census in the same compile still read 2.65 as latest minutes after the 19:06:42 upload, INFERRED a cache lag on the JSON endpoint. Both lines ride the SEED in a form that names no version.

The install skill the caboose asked for is not written: the dismount’s FINALITY forbids a car that is not a bank, and the skill is new work. It carries forward with its sources named.

2. BANK

No BANK cars this turn, and the reason is a receipt, not a choice: every graduation below targets foo_files.py, and my own NEXT CONTEXT commented it out of this compile to save 136k tokens. The dismount rule requires raw source in context, and a Target with no source behind it is the hand edit the rule refuses. The ledger rides the SEED, so the next compile’s first car banks what this beat cannot: THE TEMPLATE THAT FORBIDS WHAT IT DEMANDS, convicted by the emitter who wrote the context list.

3. DANGLING

THE PROBE NAMED THE VERSION: a probe that names an artifact by a value the ride can move (a version in a filename) prints the old world after the move and looks identical to a null result; name the newest file, or print the name beside the reading. Bank in foo_files.py beside THE DOUBLE-TAP RULE and THE OPERATOR IS A VARIABLE.
The MANIFEST car’s AFTER: witnessed only when the corrected build line and the PyPI census print AGENTS.md, AUDIT.md and .agents/ members for 2.66.
.agents/skills/pipulate/SKILL.md, the install skill: from install.sh, the npvg.org door page, README’s Quick Start and AUDIT.md’s first-run section; inspect first with less, then save, read and run the same file, the three doors, what stays local, how to stop, reset and remove; two sentences of the why.
The compile loop stays in AGENTS.md: always-on is its right load, and a skill would be the duplicate AGENTS.md’s second sentence refuses.
d printed “m will stage these” over the new SKILL.md and Prime’s m (the configuration.nix alias) staged nothing; the 09-27 “m that answers on Prime” TODO gains a specimen, and the d label is false on that machine.
A shell car ends with its own m, said inside the fence: Cars 2 and 4 had none, d printed nothing for their staged bytes, and three cars became two commits (2f58b8cb carries the move and the generator; 5cf51dba carries the regenerated index and the docstring). Harmless and pushed; the boundaries moved.
pyproject.toml’s package-data block is a no-op by receipt (the 2.65 wheel carries none of its five names) and "CHANGELOG.md" names a file the tree does not hold; cut it after a census of whether cli.py’s pipulate install needs a bundled install.sh at all.
The 09-27 “AGENTS.md must name the specification” TODO is half discharged: the URL rides AGENTS.md now; the localhost sentence is still owed, the specification read before it is worded.
imports/ascii_displays.py 1938, 1945, 1972-73 name the underscore spellings in comments; a car when the file rides, a reseal if they sit in sealed art.
~/repos/Pipulate.com/README.md:3 still reads “And context AI_CONTEXT.md”; true, since the URL stays, but it could point at the skill; that repo’s own car.
context.txt’s commented Context 2 and 3 blocks name the four retired paths and ring the integrity alert (4 of 67); respell or delete by hand.
ai.py specimens this ride: “Update journal skill documentation” for a file CREATED; “Update pyproject.toml exclude list” for a package-data line, the hunk’s context row read as the change; “refactor: introduce change control questions and artifacts” for prose; “fix: Correctly handle missing journal index” for a message string; “Update AI_CONTEXT generation error message” naming the string it removed. “Rename AI_CONTEXT.md to .agents/skills/journal/references/index.md” was true, git’s own rename line carrying it.
Claude Code and .agents/skills/: the hands-on check, unrun; a .claude/skills symlink is the one-line answer if five names do not appear.
THE FIRST GAME, SVB-133: the next ride; svb SVB-133 --write --prompt writes its own board into context.txt.

4. SEED

The first compile of the next ride: the two witnesses that name no version, and the ledger for the bank car. The game’s lines are written by svb, not by this list; the file block below stays commented for a turn about the machinery rather than a turn playing it.

! .venv/bin/python -m build --sdist --no-isolation --outdir /tmp/pipulate-sdist-probe . >/dev/null 2>&1; echo build_rc=$?; f=$(ls -t /tmp/pipulate-sdist-probe/*.tar.gz | head -1); echo file=$f; tar tzf "$f" | grep -Ei 'AGENTS\.md|AUDIT\.md|\.agents/|AI_CONTEXT'; echo members=$(tar tzf "$f" | wc -l)
! .venv/bin/python -c "import json,io,tarfile,zipfile,urllib.request as u;d=json.load(u.urlopen('https://pypi.org/pypi/pipulate/json'));K=('agents','ai_context','skill','flake','audit','readme','manifest','license','changelog','install.sh');names=lambda f,b: tarfile.open(fileobj=io.BytesIO(b)).getnames() if f['packagetype']=='sdist' else zipfile.ZipFile(io.BytesIO(b)).namelist();R=[(f,names(f,u.urlopen(f['url']).read())) for f in d['urls']];[print(f['packagetype'],f['filename'],f['size'],'bytes',len(n),'members',[x for x in n if any(k in x.lower() for k in K)]) for f,n in R]"
! .venv/bin/python -c "import re,pathlib;[print(p.parent.name, 'fence' if p.read_text().startswith('---') else 'NO-FENCE', 'name-ok' if re.fullmatch(r'[a-z0-9]+(-[a-z0-9]+)*', p.parent.name) else 'NAME-INVALID') for p in sorted(pathlib.Path('.agents/skills').glob('*/SKILL.md'))]"
foo_files.py
# --- THE INSTALL SKILL (uncomment when that car rides) ---
# assets/installer/install.sh
# remotes/honeybot/www/npvg.org/index.html
# AUDIT.md
# README.md
# --- THE FIRST GAME (svb SVB-133 --write --prompt writes the board; these are for a turn about the machinery) ---
# Workshop/corporate/connectors/svb.py
# Workshop/corporate/tickets/CONTEXT_TEMPLATE.txt
# Workshop/corporate/tickets/SVB_PROMPT.md
# connectors/mcp_render.py
# connectors/jira.py
# ! jira SVB-133

5. CLOSING

The AI_ prefix was a trick to win an alphabetical sort, and it worked for as long as nothing better existed. Then a standard arrived that sorts first by name and says “start here,” and the trick was spent. What the file behind it held was three things wearing one name: a blueprint that duplicated the signpost, a reviewer’s posture that belonged beside the audit, and an index of fourteen hundred dated entries with nowhere else to live. The blueprint was cut, the posture moved to AUDIT.md, and the index became what it always was, a reference an agent opens only when a question needs the why. The public door kept its old name on purpose, because a URL is a promise and the file behind it is not.

The receipt that outranks the rest was not on the plan. A package that has been on PyPI for years, the place a skeptic goes when a git repository proves nothing, had never shipped the two files that skeptic would read first. Nobody knew because nobody had opened the tarball; a single line of standard-library Python opened it, and the manifest gained two lines. The ride’s own instrument then failed in the way the book keeps finding: it read the past and called it the present, because it named a version number and the release moved the number. The reading was identical to the one before it, and identical is the tell. The next compile reads the newest file by its date, not by a name someone typed.

The second brain that started this article was in the frontmatter format before half the AI industry settled on it, and the repository that documents it now carries both conventions, the one that descends from a README and the one that descends from a blog post, at the root, in the sdist, and in a skill an auditor can open and recognize. That is the whole arc of the Forever Machine in one afternoon: the machinery moves forward, the reasoning stays where it can be fetched, and the name on the door does not change when the room behind it does.

6. NOTARIZE

Fingerprint of the bytes this compile sealed: the Summary’s byte line reads 543,746 bytes (UTF-8), 521,527 characters, 130,939 verified tokens; seven live command receipts ride the Manifest; the newest receipt is the head of .agents/skills/journal/references/index.md, whose header reads “Auto-generated on 2026-09-28” and “1499 entries indexed”. The seal footer under the Prompt names this compile’s archive; it is read off the footer, not guessed.

Deed: foo-a9a3d655-1660.zip (ls -t foo-*.zip

head -1)

Book Analysis

Ai Editorial Take

The most interesting new angle is that this entry quietly turns software distribution into part of context engineering. A repository can have beautifully named standards files, correct front matter, excellent audit prose, and a coherent philosophy, yet all of that effectively disappears for a package consumer if the build pipeline drops those files. The PyPI census exposes a boundary that is easy to overlook: agent architecture does not stop at the source tree. Packaging rules, source distributions, wheels, public URLs, and release scripts determine which parts of the project’s intended memory actually cross into another environment. In that sense, the strongest moment is not the standards history or even the AI_CONTEXT.md move; it is the realization that discoverability has to survive transport. That gives the article relevance beyond Pipulate because every project adopting agent-facing conventions eventually has to answer the same question: does the thing another agent downloads contain the instructions you think you published?

Refactoring Pipulate around AGENTS.md + Agent Skills exposed a packaging blind spot: PyPI wasn't shipping the files agents should read first. I moved the journal index on demand and fixed the sdist. https://mikelev.in/futureproof/agents-md-agent-skills-pypi-receipts/ #AI #Python

Title Brainstorm

Title Option: AGENTS.md and Agent Skills: Refactoring Pipulate for a Checkable PyPI Release
- Filename: agents-md-agent-skills-pypi-receipts.md
- Rationale: The strongest search-facing choice because it leads with the two concrete conventions readers are likely to recognize or investigate, then connects them to the article’s unexpected practical result: verifying and repairing what actually ships through PyPI.
Title Option: Two Files, Two Lineages: AGENTS.md and the Agent Skills Pattern
- Filename: two-file-lineages-agents-md-agent-skills.md
- Rationale: Best for the historical and conceptual thread. It highlights the article’s distinction between an always-on README-like signpost and a front-matter-driven skill without pretending they are successive versions of one standard.
Title Option: From AI_CONTEXT.md to an On-Demand Journal Skill
- Filename: ai-context-to-journal-agent-skill.md
- Rationale: Best for readers interested in context architecture. It captures the central refactor from one overloaded root file into separate always-on, audit, and demand-loaded responsibilities.
Title Option: The Files Agents Read First: Packaging AI Context for PyPI
- Filename: files-agents-read-first-pypi-context.md
- Rationale: Best for the packaging lesson. It foregrounds the discovery that repository conventions are incomplete if the distribution artifact silently omits the very files another agent or reviewer is supposed to inspect.

Content Potential And Polish

Core Strengths:
- The article turns an abstract standards discussion into visible repository changes: skill front matter, directory names, README routing, journal references, package manifests, and a numbered release.
- The two-lineage framing gives readers a useful mental model for separating always-on repository guidance from task-triggered skills without treating one as a replacement for the other.
- The proof-straddle structure produces unusually concrete before-and-after evidence, including the skill census, loader path, package member counts, file move, and version 2.66 release.
- The discovery that the published PyPI artifacts omitted the agent-facing files gives the piece a genuine investigative turn rather than leaving it as a documentation cleanup story.
- Keeping the old public AI_CONTEXT.md URL while changing the source architecture underneath it is a strong example of separating internal structure from an external compatibility promise.
- The session naturally connects the broader second-brain philosophy to mundane packaging details, showing that durable knowledge architecture depends on transport and distribution as much as prose.
Suggestions For Polish:
- Distinguish the documented histories of AGENTS.md and Agent Skills from the more interpretive Jekyll ancestry argument. The latter is a useful analogy, but wording it explicitly as a file-shape lineage will keep the historical claim precise.
- Consider shortening some of the embedded patch transcript once the important receipts have been established. The full log has archival value, but a published reader mainly needs the before state, consequential change, and after state.
- Move the explanation for retaining the public AI_CONTEXT.md URL slightly earlier. The release output naturally makes a reader wonder whether the rename was incomplete, and the URL-as-promise rationale resolves that confusion immediately.
- Correct Thomas Khun to Thomas Kuhn in the prose.
- Keep the Nix adoption examples narrowly descriptive. The article itself already discovers the useful distinction between evidence that an organization uses a tool and a stronger certification claim that the evidence does not establish.
- The critique of convenience wrappers is most persuasive when framed around retained operator skill, portability, inspectability, and exit options rather than presumed motives of the companies or people choosing those wrappers.

Next Step Prompts

Audit the released Pipulate 2.66 source distribution and wheel as transport artifacts. Produce a compact matrix showing whether README.md, AGENTS.md, AUDIT.md, each .agents/skills/ entry, installer assets, and the journal reference are present, absent, intentionally web-served, or discoverable through another documented path.
Draft .agents/skills/pipulate/SKILL.md from the installer, the npvg.org door page, README Quick Start, and AUDIT first-run material. Center it on inspect-before-execute: curl -fsSL https://npvg.org | less, save and inspect the same bytes, then run them, followed by startup, stop, reset, removal, local-versus-network behavior, and the minimum explanation of why command-line fluency remains portable.

Setting the Stage: Context for the Curious Book Reader

Technical Journal Entry Begins

🔗 Verified Pipulate Commits:

A Local-First Second Brain Meets Agent Conventions

The Receipt Is the Real Context

When Tool Calls Become Auditable Evidence

Do you have it about right?

Is any of it making sense?

1. PROBES

2. NEXT CONTEXT

3. PATCHES

4. PROMPT

5. EXTERNAL DELIVERABLES

Turning the README Into the Front Door

The finding that outranks the request

Your two labels

Where AI_CONTEXT.md belongs

The three Nix claims, vetted

1. PROBES

2. NEXT CONTEXT

3. PATCHES

4. PROMPT

5. EXTERNAL DELIVERABLES

Skills That Preserve Operator Fluency

THE PROOF STRADDLE

The receipts, read

The rename, ruled

What PyPI shows and what the sdist ships

The Pipulate skill, and whether the compiler needs one

1. PROBES

2. NEXT CONTEXT

3. PATCHES

4. PROMPT

5. EXTERNAL DELIVERABLES

THE PROOF STRADDLE

Release 2.66 and the Public Promise

1. VERIFY

2. BANK

3. DANGLING

4. SEED

5. CLOSING

6. NOTARIZE

Book Analysis

Ai Editorial Take

🐦 X.com Promo Tweet

Title Brainstorm

Content Potential And Polish

Next Step Prompts

Where `AI_CONTEXT.md` belongs