wiki-watch

name: wiki-watch description: "Run the wiki-watcher (W1) — one idempotent cycle that detects doc-worthy changes in a watched repo (+ its active PLAN.md/design/ROADMAP.md), judges significance, and dispatches the documenter to author the update PR-default. Triggers when the operator says 'run the wiki watcher', 'watch the wiki', 'check for doc-worthy changes', or drives /wiki-watch on a loop (/loop or cron). One invocation = ONE cycle (NOT a daemon — DC-W3): cooldown-gated + cursor-backed so repeated runs never drop or double-dispatch. Opt-in: device toggle in .agentm-config.json + a per-repo .harness/wiki-watch.json marker. PR-default (a human merges); direct-commit opt-in per trusted repo. Graceful-skip when disabled / unconfigured / unregistered / gh unavailable. The engine is cross-host; the scheduling wiring is Claude-first (see Antigravity-Limitations)." kind: skill supported_hosts: [claude-code, antigravity] version: 0.1.0 install_scope: project

wiki-watch — run the wiki-watcher (W1)

The continuous-mode surface of wiki-maintenance: instead of running the documenter one-shot at a phase boundary, run it on a watch loop so a wiki stays in sync with a repo's doc-worthy changes. Ships as W1 — an idempotent single-cycle engine the operator drives (DC-W3), not a daemon: one invocation runs one poll → detect → significance → dispatch → audit cycle and exits. Cooldown-gated + cursor-backed, so re-running it on the host's loop (/loop or cron) never drops a change or double-dispatches.

Position vs. the phase-boundary documenter. developer-workflows' phase commands dispatch the documenter once, at /work//release boundaries. wiki-watch dispatches the same documenter (reused as-is — never a second writer) repeatedly, off a change feed, between boundaries. Same write-executor; different trigger.

Opt-in (two gates) + safety

The watcher dispatches autonomously, so it stays off until two opt-ins line up (see Wiki-Watch-Config):

1. Device — wiki_watch.enabled: true (or wiki_watch_enabled: true) in <install-prefix>/.agentm-config.json.
2. Per-repo — a <repo>/.harness/wiki-watch.json marker (its presence is the opt-in) carrying watch_sources + dispatch_mode (pr default | direct).

PR is the default autonomous boundary (DC-W1): the PR is the async preview that reconciles the documenter's own preview-before-write gate with autonomous mode — a human merges. direct commits straight to the default branch and is an explicit per-repo opt-in for a trusted repo only. The significance gate, the audit log, and PR-as-preview bound the blast radius. Public-repo PII guardrails gate every push (the pre-push hook is the hard enforcer; the engine also runs a PII pre-check before any push). The audit log is local, never committed.

Running one cycle

The deterministic engine is scripts/wiki_watch_cycle.py. Run it, then act on its report:

python3 scripts/wiki_watch_cycle.py run --repo <repo-root> [--slug <slug>] [--no-cooldown]

It prints a JSON CycleReport. The skill then:

1. Read the report. If skipped: true (disabled / unconfigured / cooldown / unregistered / no commits), stop — clean no-op. Otherwise it carries token, wiki_target, dispatch_mode, the plan (pr/direct/skip + branch), and candidates — each with a deterministic classification (doc-source/code/minor) and recommendation.
2. Judge significance (the agent step). The recommendation is the deterministic floor — start conservative:
   • dispatch (doc-source): doc-worthy by default — include it.
   • skip (minor — tests/CI/config): drop it.
   • judge (code): you decide — read the changed file's diff and ask "does this change what a reader of the wiki needs to know?" (new/changed public API, behavior, architecture → yes; refactor/rename/internal → no). Promote to dispatch only when genuinely doc-worthy. Bias toward fewer, higher-quality updates. If nothing survives the judgment, advance the cursor (finalize_cycle) and stop — the changes were seen, just not doc-worthy.
3. Plan the landing. If plan.action == "skip" (PR-default but gh unavailable), report it and stop — do not fall back to direct-commit. For pr: prepare_branch (creates plan.branch). For direct: stay on the current branch.
4. Dispatch the documenter (reuse). Spawn the documenter sub-agent with the context from build_documenter_context(...) (its familiar /work-style task+diff contract): the changed sources + the wiki target. It authors/updates only the affected pages, in the operator's voice, Diátaxis single-mode, keeping check-wiki.py --strict green. Never author wiki content yourself — the documenter is the write-executor.
5. Land + audit. Run finalize_pr (commit → PII guard → push → gh pr create) or finalize_direct (commit → PII guard → push). On success, call mark_and_audit_dispatch(...) per candidate (records the dispatch + PR/commit link, prevents re-dispatch on the next cycle); on failure, record_dispatch_failure(...) (retry/backoff). When every candidate for this token is handled, finalize_cycle(...) advances the cursor.

Re-invoking is always safe: the cursor + dispatched-set make a re-run a no-op for already-handled changes (idempotent), and the cooldown gate no-ops a too-soon re-run.

Scheduling (honest: Claude-first — DC-W4)

wiki-watch is a single-cycle engine; you drive the loop. There is no cross-host plugin scheduler — so the engine ships cross-host but the scheduling wiring is Claude-first:

• Claude Code — drive it with /loop (interactive) or a cron line invoking the /wiki-watch command headless (e.g. claude -p "/wiki-watch"). The cooldown gate (--cooldown, default 15 min) keeps a tight loop from over-firing.
• Antigravity — there is no installable trigger path for a shipped plugin (native every()/on_file_change() only; no SessionStart hook). Run /wiki-watch manually; true scheduling is deferred to the agentm V7 scheduled-sidecar. This gap is tracked in the growing Antigravity-Limitations register — revisit when the host ships a trigger surface.

The hands-off idle-chain hook (auto-fire on SessionStart) and always-on (W3) are deferred follow-ons; both would just call this same single-cycle engine.

Graceful-skip (clean no-ops, never hard-fail)

• Device toggle off / no per-repo marker → skip with a reason.
• Repo unregistered in repo_registry or no wiki_path (and no <root>/wiki) → skip.
• Within the cooldown window → skip.
• Not a git repo / no commits → skip.
• gh unavailable in PR-default mode → skip (never silently direct-commits).
• agentm kernel unreachable (no repo_registry.py / harness_memory.py) → state falls back to repo-local <repo>/.harness/wiki-watch/; an unresolved wiki target skips the repo.

Companion surfaces

• documenter sub-agent — the write-executor wiki-watch dispatches (reused as-is; wiki/**-scoped).
• /wiki-watch command — the thin slash entry for /loop/cron (Claude Code).
• Run the wiki-watcher — the operator how-to.
• recent-wiki-changes — the read-only "what changed in my wikis" view (this skill is the write loop).