Nightly Docs And Changelog

nightly-docs-and-changelog

A configuration of loop-controller. That skill supplies the loop machinery — primitive selection, the full guardrail stack, state externalization. This skill supplies the three things specific to "keep an existing repo's docs+changelog from rotting on a nightly cadence": the per-night recipe (it diffs since the last run and authors the doc + changelog updates for that surface itself, following the documentation conventions docs-agent defines — docs-agent is the orchestrator's build-phase role and is not user-invocable on a schedule), the mechanical proof (docs match the changed surface AND every user-relevant change is recorded, ending in a reviewable PR), and the HITL boundary (it opens a PR and stops — it never merges docs). Read loop-controller for the guardrails; they're inherited, not repeated here.

Why disable-model-invocation: this loop edits docs, writes the changelog, commits, and opens a PR on its own, on a schedule. It is user-scheduled — you set it up once (or type /nightly-docs-and-changelog), not have Claude silently start rewriting docs because some code changed during the day.

The 5-part contract

The proof: docs current AND changelog complete, default-FAIL

"Current" is not "the docs looked fine last week." It is two conditions observed together against the diff since the last run: the docs covering the changed surface match the implementation, and every user-relevant change since the marker has a CHANGELOG.md entry (a quiet night gets an explicit no-change record so a skipped night is distinguishable from a missed one). Assume stale until a fresh sweep proves both — that's the default-FAIL stance, and it's why every night re-diffs from the marker rather than trusting yesterday.

Name the artifact explicitly: the docs/changelog diff plus the PR the sweep opens. The PR is the proof a human reviews; an empty diff with a no-change record is a valid (and common) outcome.

Step 1 — Diff since the last run

The whole sweep is scoped by the last-run marker. Read it, diff marker..HEAD, and that diff — nothing else — defines tonight's work surface. No marker yet (first run) means bootstrap from the last release tag or a sensible recent baseline, not the whole history. The marker mechanism (git tag vs .claude/ file), how to advance it atomically only after the PR opens, and how to recover a missed night (the marker, not the clock, defines the window — a laptop-off gap is just a wider diff next run) are in references/scheduling.md.

Step 2 — Classify user-relevant vs internal, then write the updates

Two distinct outputs come off the same diff:

• Docs that must track the code — for the changed surface, author the doc updates yourself (you hold Write/Edit), following the documentation conventions docs-agent establishes for this repo. docs-agent is the orchestrator's build-phase doc role — disable-model-invocation, dispatched only inside a contract-first build with a lead, contracts, and an ownership map — so a user-scheduled nightly sweep on an already-shipped repo reuses its conventions, not its dispatch. This loop adds the scheduling, diff-scoping, and proof, and matches docs-agent's house style rather than inventing a second one.
• The changelog — only user-relevant changes get an entry. A new flag, a behavior change, a fixed user-facing bug, a breaking change: yes. An internal refactor, a test-only change, a lint fix, a dependency bump with no user-visible effect: no. The exact user-relevant-vs-internal rubric (with examples) is in references/scheduling.md.

Step 3 — Open ONE PR, then stop (HITL)

Commit the docs+changelog changes and open one PR via git-pr titled for the sweep date and summarizing what moved. Then stop. This loop never auto-merges — docs and changelog wording is a human judgment call, and a PR opened at 02:00 is meant to be reviewed at 09:00, not merged in the dark. One PR per night keeps the review unit small; don't batch multiple nights into one PR.

How this differs from its neighbors

• vs docs-agent — docs-agent is a build-phase role: it writes the docs for a build in progress, dispatched by the orchestrator alongside the other role agents. This skill is the recurring nightly sweep that keeps an already-shipped repo's docs and changelog from rotting over time. It owns the cadence, the diff-since-last-run window, the changelog discipline, the default-FAIL proof, and the PR/HITL boundary — and authors the doc updates itself in docs-agent's house style (docs-agent is orchestrator-only and not user-invocable, so this loop reuses its conventions, not its dispatch). Because the nightly sweep runs outside a parallel orchestrated build — there is never a concurrent docs-agent — editing the docs/ + CHANGELOG.md that docs-agent owns at build time creates no live ownership conflict; the two never run at once, and this loop follows the same conventions. Remove this skill and docs-agent still can't keep your repo current on its own; nothing schedules it or scopes it to "what changed since last night."
• vs git-pr — git-pr is the PR mechanics; this loop calls it as its final step, it doesn't reinvent PR creation.
• vs a bare /loop — see Step 4: /loop is the wrong primitive for a nightly cadence, which is the whole reason this skill schedules differently.

Step 4 — Schedule it to survive laptop-off (the routine, NOT /loop)

Per loop-controller Step 1, the primitive choice is load-bearing here and it is not /loop. /loop is session-scoped, expires after ~3 days, and does not catch up on missed fires — a laptop closed overnight kills it, which is exactly the window this loop must run in. So the primitive is a cloud routine / Desktop scheduled task (via /schedule) that fires on a real cron independent of any open session and survives the machine being off. The full /loop-vs- scheduled-routine reasoning lives in loop-controller's references/primitives.md — don't re-document it. The exact /schedule setup, the cron line, and what the routine prompt invokes are in references/scheduling.md.

Guardrails specific to this loop

Inherits the full stack from loop-controller → references/safety.md. The caps this loop sets:

• Pass cap per night — one sweep = one PR; cap the within-night iterations (default ~10) so a pathological diff can't burn the whole budget before morning. Hitting it is a stop-and-open-a-partial-PR-and-flag, not a license to hand-wave the docs.
• No-progress detection — if the same changed surface keeps drifting across 3 consecutive nights (docs re-edited but never converge, or the same file shows up undocumented night after night), stop and surface it. Three nights on one surface means it needs a human, not a fourth automated pass.
• Budget cap — a nightly routine runs unattended for weeks; enforce a token/cost ceiling that terminates the sweep (read it from .claude/profile.yaml when present), not just warns.
• HITL is the exit, not a checkpoint mid-loop. Opening the PR is where this loop hands off to a human — it never merges, never force-pushes, and never rewrites a published CHANGELOG entry. Drafting docs + changelog and opening a PR is reversible and AFK-safe; merging is the human's call.
• Never cheat the proof. Don't advance the marker without opening a PR (that silently swallows a night's changes), don't drop a user-relevant change to keep the changelog short, and don't mark docs "current" without checking them against the diff. A clean night that came from skipping the diff is a finding, not a win.

Choosing the driver primitive

The scheduled routine (/schedule) is the outer engine — it fires the sweep on cron and survives laptop-off. Within a night, the sweep pushes to a finish line (docs current + changelog complete + PR opened), provable from the diff and the gh PR result, so the inner pass can run under auto mode unattended inside the reversible boundary (draft + PR only). The loop as a whole has no /goal finish line across nights — it watches the repo evolve and sweeps each night until you unschedule it.

Reference files

• references/scheduling.md — the per-night algorithm and the doc-writing conventions; the last-run marker mechanism (git tag vs .claude/docs-sweep-last, advancing it atomically, recovering a missed night); the user-relevant-vs- internal changelog rubric with examples; the no-change-night record format; and the /schedule routine setup (cron line, routine prompt, what survives laptop-off).