Back to Work Diary

A 953-line skill entry fits the budget again

A 953-line skill entry collapses to 267 lines plus four pull-only references, and the same day seventeen Omarchy repos pin the audit-harness and review Action to the same SHA.

The day reads as two coordinated hygiene passes. The first is internal to the contributing-clanker skill entry, and the second is a single dep-refresh PR applied to seventeen Omarchy entry repos. They share the same shape: a long running surface (a single SKILL.md; a long-running audit-harness pin) gets trimmed or locked, and the receipt is a number plus a hash.

The SKILL.md that nobody could load in full

contributing-clanker PR #77 (commit 63df1ee, 2026-09-09 00:40 UTC) cut skills/contribute/SKILL.md from 953 lines to 267 lines, a 72% reduction. The content moved into four reference files, three of them new:

  • references/operations.md (69 lines, new): dashboard, live reconciliation, override audit, runtime drift, dry-run email policy.
  • references/repo-intake.md (91 lines, new): GitHub URL or owner/repo slug parsing, owned-repo scope guard, new versus known-repo briefing.
  • references/submission-policy.md (143 lines, new): required fit fields, collaboration surface, counterparty design intent, candidate updates after submission.
  • references/workflow-guide.md (153 to 112 lines): trimmed end-to-end flow.

The SKILL.md itself now reads as a thin index. Step 1 (Detect invocation mode) points at [repo intake](references/repo-intake.md) and [operations](references/operations.md). Step 3 (Require a current dossier and first-touch fit) points at [submission policy](references/submission-policy.md). The Resources section at the end lists all seven reference files in plain markdown links. A contributor who needs only the daily status flow pulls only operations.md; a contributor onboarding a new repo pulls only repo-intake.md.

The diff is 558 insertions and 893 deletions across the five files. The skill fits a tighter first-load budget, and the references stay readable as their own documents because each one was written with its own scope.

A small but telling diff: the front matter changed the allowed-tools entry from - Task to - Agent. The skill entry now declares the modern subagent tool rather than the deprecated one, which would have produced a stale warning on every invocation. The wording of the description block also tightened: from a multi-line justified narrative to a one-paragraph capability statement.

Seventeen repos pin to one SHA

The second batch is a coordinated dep refresh that landed across all seventeen Omarchy entry repos on the same day. Each repo merged a single chore(deps): refresh ecosystem dependencies (#N) PR with an identical commit body: “Updates the audit harness and review Action, removes the vulnerable qs override, hard-pins remaining Actions, fixes the maintainer intake link, and refreshes the reviewed anti-tamper seal. Local and remote gates passed.”

The substantive changes per repo are the same five:

  1. @intentsolutions/audit-harness bumped from ^1.3.1 to ^1.4.0.
  2. The review Action jeremylongshore/minimax-code-review bumped from d1314b96c1b261d5bf026d0a669823a7e5ce6b46 to bcaa65182e52183f14746663aa2049641c0dbca5 # v0.5.0. Every reference in .github/workflows/minimax-review.yml now uses the same SHA with a # v0.5.0 comment, so a reader can audit the pin by eye.
  3. A vulnerable qs override is removed from package.json (the dep update replaces the override with a fixed upstream version).
  4. The remaining Actions in gates.yml and minimax-review.yml are hard-pinned to full SHAs where they were previously referenced by tag.
  5. The maintainer-intake link in README.md is fixed from a relative ../../issues/new?template=... to an absolute https://github.com/jeremylongshore/<repo>/issues/new?template=..., which works whether the README is read on GitHub or from a clone.

The .harness-hash file rolls on every repo. Reading the capture-conveyor-entry diff shows the new hash entries for .github/workflows/minimax-review.yml and package.json (the two files whose contents changed); the other entries stay byte-for-byte identical. That is the audit surface verifying the seal: every file listed in .harness-hash was reviewed and approved as part of the new state.

The seventeen repos touched: bazaar, capture-conveyor, crew-chief, desk-transition, docket, flow-boundary, foundry, listening-post, loose-ends, mlb-booth, omatrail, pit-wall, quiet-queue, wait-state, widget-template, workspace-storyboard, x-files. (mlb-booth got the refresh plus its own separate feature commit, covered below.)

omarchy-mlb-booth-entry: persist club and clock settings

The mlb-booth entry repo shipped an additional commit on the same day that was not part of the dep-refresh wave: 6366a3a feat: persist club and clock settings (#5), by David Reed. A new SettingsForm.qml (358 lines) adds keyboard-accessible settings for all 30 clubs and 12 or 24-hour first-pitch times, persists the selected club across shell restarts and failed or stale schedule responses, and adds a deterministic live-feed loading path.

The commit also updates Model.js (55 new lines), Panel.qml (210 lines with churn), e2e/rig-after-open.sh (27 lines), and rolls the .harness-hash and .rig-proof.json fingerprints. The render-proof test was the gating surface: the new .rig-proof.json records the panel’s post-settings render, and the maintainer approval used the same seal pathway as the other entries.

This was a real user-facing change on top of the hygiene wave, not part of the wave itself.

What the two batches share

Both reads are routine on a single surface and structural on the estate. The SKILL.md refactor only touches one file’s line count in the diff stat, but it changes what gets loaded when a contributor invokes /contribute and what stays hidden behind a reference link. The seventeen-repo dep refresh only changes a SHA and a ^ in each repo, but it locks the audit-harness to one published version across the whole Omarchy entry estate, removes a vulnerable override from each clone, and reseals the harness hash for every entry.

The shape is the same. A long-running surface is reduced to a smaller loadable form, and a long-running pin is hardened against silent drift. The cost is the diff size; the receipt is the harness hash on each entry and the references index on the skill entry.