commit

Commit — Session-Isolated, Doc-Sync-Gated (gateway-inline)

Commits the changes this Claude session produced, leaving prior uncommitted work alone. The gateway runs the flow inline — it holds the session's diff, file list, and change intent, so the mechanics carry no closure a fresh container would hold. Only the doc-sync scan dispatches, and only when the diff touches the doc surface: it is the one step that earns a cold, context-clean pass — cold-eyes catch staleness the author is blind to.

Execution (gateway-inline)

1. Session file list — from this conversation, the files this session edited/wrote. This is the session-isolation ground truth; a file you don't remember touching stays off it. Stage by explicit path only.

2. Gather state — git status, git diff, git diff --staged (a staged path outside the §1 list is another session's — early warning only; §5's readback is the gate), git log --oneline -10 (recent style).

3. Doc-sync grep-gate — mechanical, no judgment:

   • Deferred mode — check first (drain-worktree isolated commit): doc-sync belongs to the merge boundary — skip all of §3, go to §5.
   • Frozen-provenance exemption — card lifecycle, history dimension: every changed path inside docs/work/ (card-thread appends, card deletions, work-README high-water bump), plus every changed docs/outward/OUT-###.md (thread appends, entry deletions) and the outward README's high-water bump → skip the gate, go to the link check below then §5. Both are append-only threads, self-contained; cross-card ID mentions are frozen provenance, not behavior narration — so hits and refs leave every docs/work/ card thread and every docs/outward/OUT-###.md thread out of the scan scope too, whatever term they name or anchor they cite; each folder's standing README.md narrates its thread contract and stays in both lanes. A doc whose leading frontmatter declares dimension: history is frozen provenance the same way — a change to it yields no term (terms skips it), hits and refs leave it out of the scan scope, and the link check below still covers it. Mixed diff → run the gate on the other paths only.
   • Premise lane — product-anchor paths: the diff touches the product anchor (docs/overview.md § Problem / § User, or a dedicated product doc where the repo splits one out) → route the anchor portion through §3b — its closure set is the whole doc surface, so it already covers everything this gate would reach for the anchor; the rest of the diff continues through this gate.
   • Terms → grep hits — one script call each: <skill-base>/assets/doc-links.sh terms <changed paths…> prints the gate's terms; scaffolding paths (bench, tests, fixtures, goldens, session ledgers), asset extensions, generic basenames, terms under four characters, and the machine state harness-bootstrap seeds under .claude/ (the runway receipt, templates/) yield none. Nothing printed → the grep lane is empty, skip to the citer lookup. Otherwise <skill-base>/assets/doc-links.sh hits <terms…> | grep -vxF "$(printf '%s\n' <changed files>)" prints the doc-surface files mentioning a term in code shape — a whole word inside backticks, or a path segment — so bare prose stays out. Harness-seeded hub stems (overview, techstack, decisions, the runway receipt, the worktree template) count only as a bare token inside backticks: a path mention of a hub doc is a citation the anchored citer lane already reaches, not a narration. Hit files join the scan scope; the agent judges them cold (§4). hits walks the markdown surface — a manifest description field the diff changes is read here directly.
   • Link-hit (reverse citers) — mechanical, beside the grep: for each changed docs/** / root README.md / plugin-README (plugins/*/README.md) path, take its post-image hunk ranges from git diff -U0 -- <path> (@@ -a,b +c,d @@ → +c,d) and run <skill-base>/assets/doc-links.sh anchors <path> <ranges…>. (top) among the slugs means whole-file grain → <skill-base>/assets/doc-links.sh refs <path>; otherwise refs <path>#<slug>… in one call of the same script, which prints the union. Exclude the changed files themselves. Every file returned is a declared citer of the changed truth — collect them as the citer read-set.
   • Link-target extraction (forward links) — mechanical, beside the reverse lookup: from the diff's added lines, extract every markdown link target (grep -oE '\]\([^)]+\.md[^)]*\)' over + lines), resolve each relative to its linking file. Each target is a doc the new prose claims agreement with — add the target files to the scan scope. A target that is itself a changed file still joins: its unchanged sections are exactly where a same-commit contradiction hides.
   • Scope ceiling — count before dispatching: the scan scope is citers ∪ grep hits ∪ link-targets. Above 8 docs, surface it to the user first — the count, the list grouped by lane (citer / hit / link-target), and the options: split the commit by concern, narrow to the anchors actually changed, or proceed over the full set. A hub doc fans out visibly or not at all. At or under 8 → dispatch as below.
   • Any grep hit OR any citer OR any link-target found → dispatch the doc-sync scan (§4), the scan scope — citer read-set + grep-hit files + link-target files — riding the prompt.
   • Neither → the diff narrates nothing; go to §5.
   • Link integrity (every non-deferred commit, exemption included): run <skill-base>/assets/doc-links.sh check from the repo root — <skill-base> is the Base directory for this skill: path surfaced at invocation; zero model tokens. Broken links (path or anchor) surface to the user with the commit: fix or explicitly acknowledge before landing; never silently skip. One class is exempt by construction — the consumed-card-link skip: a link from a card thread (docs/work/{BUG,DEBT,GAP}-###.md) to an absent card ID is consumed provenance — the sibling resolved — so check skips it uncounted, while every other target from a card, and any dangling card link from a non-card doc, stays strict.