DECISIONS.md as short as a rebuild needs; SKILL.md proven as written

[suleimansh]

The three DECISIONS.md (agent-data, skill-branches, skill-tickets) and both SKILL.md, each read against the code by a fresh agent seeing only its package, round after round, until a read found nothing the code does that the file does not say and nothing the file says that the code does not do. The rule for a line: it stays only if a rebuild of the package from the file alone would differ without it; bullets that say one thing are merged; no decision is cut. One commit per round (62 in all), so the history shows what each read caught.

The files grew, not shrank. The first pass shortened every bullet; every later read then surfaced a decision the code makes that the file had never stated, each a line a rebuild would otherwise get wrong. Where a read caught my own previous edit overshooting, the next commit says so.

• agent-data/DECISIONS.md: 84 → 76 lines, four reads. Restored three meanings the shortening lost; added where reads look (the repository a worktree was made from), the lock's scope (in memory, per repository and branch), the output buffer, a failed write dropping the op's new files, the exclude rule written once, the library's own commit messages; corrected "every write pushes" and "a pull runs on its own".
• skill-branches/DECISIONS.md: 93 → 150 lines, twenty reads. Among what the reads added: a branch gone only locally comes back from origin; the session-name charset and the leading-dash rule; forced removal; remove/prune push by default; links follow any branch, are made by the package's own runs, are judged by their target's name; what the reclaim pushes and when it does not; the holds-nothing rule as reachability, not merge state; what the skill has the agent do; every command's result and refusal shape, the JSON keys, the removal order.
• skill-tickets/DECISIONS.md: 109 → 164 lines, twenty-two reads. Among what the reads added: the two callers and the executable; what the skill tells the agent; the full row and its keys; the sync's three steps and where the exclude rules are written; the command's detached write versus the program's checkout cycle; the refusal order per command; the queue file the command creates itself; placement details; the batch claim's message; a queue edit counting only once pushed.
• skill-branches/SKILL.md: 48 → 50 lines, five reads, and skill-tickets/SKILL.md: 142 → 150 lines, five reads. Proven first: every command in both run as written in a throwaway repository, from a plain agent-* clone and from a .branches/agent-<id> checkout made the daemon's way, and one real agent run per skill (branches: name, commit, status; tickets: list, claim, plan, queue, release), with no old names or product words left. Then sentence by sentence. Added: a continued agent keeps the branch and the name it was given; status and name are the agent's only commands; the finish gate and the user's untracked files; put replaces the whole file, empty stdin included; close leaves the queue entry; release before stopping unless closed; a rename or branch switch changes the holder; the refusal shapes as printed. Both SKILL.SPEC.md follow.

Code findings that are not decisions are in #1757: the priority string, the branch budget scan, the npm pack tarballs, Source: lines, release's output, the queue's first unranked entry, the reclaim pushing a user's branch, remove by name versus by id, and more.

🤖 curated · Fable 5.1, effort high — the agent proposed every change and ran the reads and the rig; each round's picks are mine.