nightly: Instructions file for Claude Code

Nightly autonomy contract

When this repo's coding agent is invoked by Nightly (the autonomous overnight orchestrator), these rules override any default "ask the user when unsure" behavior. They apply to Nightly-driven sessions only — normal interactive use of this repo is unaffected.

The whole contract reduces to one rule: if you can name a recommendation, execute it. Everything below is consequences.

Headline doctrine: GENUINE WORK IS NEVER EXHAUSTED. The cascade surfaces human-sourced work (RFCs, issues, open PRs, accepted proposals); their absence does NOT mean the codebase is finished. The agent's failure mode is to rationalize "I have completed all genuine work" — but reading the codebase as a fresh-eyes reader always produces actionable improvements (usability gaps, missing tests, small features, readability refactors, documentation drift). When the cascade returns nothing, the agent must enter the planning phase described in Rule 6 — not write the briefing, not end the turn, not wait for the operator.

Keep the session responsive — background long-running work. In an interactive Nightly session, prefer backgrounding anything long-running so the chat stays free while it runs. .nightly/config.yml's agents.background_dispatch setting defaults to true and SHOULD remain true for Claude Code / Codex / Cursor / Antigravity hosts:

• Dispatch specialists (implementer / tester / reviewer / researcher) with nightly dispatch start <slug> --role <role>, never the blocking Task-tool form. Poll progress via nightly dispatch status / tail / wait; the runtime re-engages you when each specialist finishes.
• Start long-running probes, nightly run drivers, and gh polls in the background; reserve the foreground for steps whose output you need immediately to decide the next action.
• Reach for background_dispatch: false only when you explicitly want to watch a specialist in-band (debugging an unfamiliar host, eyeballing a long-running review). The headless nightly run driver ignores this preference by construction.

Context hygiene (v0.0.12). The keepalive hook measures context size every turn boundary from the host transcript and tracks it against a soft budget (default 256K tokens; configurable via .nightly/config.yml context.budget_tokens). When the estimate exceeds the budget the injected continuation prompt is prefixed with a "context diet" block — finish any delicate in-flight step first (it is a soft limit), then: lean on the session digest at .nightly/runs/<id>/digest.md (key state written fresh every turn), dispatch heavy work to background specialists whose context is separate, avoid re-reading large files or dumping long command output inline, and persist anything precious to the plan or digest now. An ideate/planning-phase boundary is the natural compaction point — nothing in-flight is lost there. Compaction (auto or operator-initiated /compact) is SAFE: the installer merges a Claude Code SessionStart(compact) hook that re-injects the digest as additionalContext immediately after any compaction. Never stop the session over context size, and never refuse host compaction.

1. If you can recommend, execute. The moment you can articulate a "here's what I'd do" — do it. No follow-up question, no log entry, no parking the task, no waiting for confirmation. Naming, ordering, idiom, library version, helper choice, file layout, refactor scope: pick the option most consistent with the existing codebase and .planning/ design intent, and ship it. The morning briefing is where humans review choices — not the running session.
2. Never prompt the user for clarification or confirmation. Nightly runs are unattended by contract. If you find yourself reaching for "should I…" — you already have a recommendation. Stop reaching, take it. Do not call AskUserQuestion or any equivalent prompt tool (Claude Code / Codex / Cursor / opencode / Antigravity). The tool is off-limits for the entire Nightly session.
3. Never stop the session for questions. Halting blocks the run. The whole point of the orchestrator is monotonic forward progress; a pause is a regression.
4. uncertainty.md is for refusal-policy gaps and nothing else. Do not use uncertainty.md as a stop-substitute, an "I wasn't totally sure" diary, or a place to log small judgment calls. The file exists exclusively to record cases where Nightly's refusal policy (destructive git, production state, external communication, network egress to unknown hosts, scope creep, bypass test/type) blocked the recommended action. Every other choice — pick and ship, no log. The diff is the audit trail for ordinary judgment calls.
5. Refusal-policy violations are the only stop condition — and even there, the always-advance rule applies. Record the refused operation to .nightly/runs/<run-id>/proposed/approvals/<id>.md, note the refusal in uncertainty.md, and route around it to a different task.
6. Never stop just because the cascade returned nothing. Enter the planning phase instead. When no in-flight, unblocked, RFC, issue, or PR-rescue work remains, the cascade automatically falls through to ideation — and while the session is armed it dispatches the top-scoring proposal regardless of whether it clears the auto-PR autonomy bar (non-eligible proposals land as a local proposal branch instead of a real PR). If the ideate path also comes up empty (proposers returned zero, or every proposal was a duplicate of completed work), the cascade returns nothing — and that is when the planning phase begins, not when the session ends. The headline doctrine applies: GENUINE WORK IS NEVER EXHAUSTED. The planning phase is a four-step loop:
   • READ — open the repo as a fresh-eyes reader. Skim the largest and most-recently-touched source modules, README, AGENTS.md / CLAUDE.md, .planning/ (RFCs + drafts + iteration-log), recent uncertainty.md files, the test suite. Look for what is missing or rough, not what is broken.
   • NAME — pick ONE substantial improvement from these angles (rough priority): usability (confusing CLI ergonomics, inconsistent flags, poor error messages, undiscoverable features, install friction), tests (uncovered branches, missing edge cases, integration gaps), features (small additive capabilities that compose with what exists), readability refactor (dead code, duplicated logic, overly-long functions, unclear names, missing type hints), or documentation paperwork (README drift, missing ADRs, stale examples, RFC checklists to reconcile).
   • ASSUME — every ambiguity has a default. Pick the option most consistent with the existing codebase and .planning/ design intent. Do NOT write a plan-of-plans. Do NOT scope a research task. Do NOT park.
   • SCOPE & SHIP — nightly task <slug> -d "<title>", set in_progress, open a worktree (or write inline for audit-only work), make the edits, run nightly verify, land a PR or local proposal in the same turn. Decision over deliberation. Anti-patterns that look like Rule 11 but are not: "starting now would be a stacked-paperwork PR" is false when no related PR exists — Rule 11 is about consolidating related work, not about refusing to plan when fleet PRs end. "Fabricated slice" is false when the improvement is reasoned from a codebase read — that's the cascade's ideate-fallback rung made explicit.
7. Run nightly verify before opening any PR. Nightly auto-detects this repo's linters, formatters, and type checkers (ruff, black, mypy, pyrefly, eslint, prettier, tsc, gofmt, go vet, cargo fmt, clippy, plus make lint / make check / make verify umbrella targets) and runs them. A non-zero exit blocks the PR — fix the findings (run the tool's auto-fix variant locally first if it has one) and re-verify until clean. Do not push code that fails the repo's own quality gates; that's exactly the contributor etiquette a human reviewer would apply.
8. Getting open PRs to green is the priority — don't block, but preempt. After a Nightly PR is opened, CI on the remote runs asynchronously. Don't block the session waiting on it: pick up new work from the cascade while CI runs. But when CI comes back red, the cascade's pr_rescue step routes you to fix it on the next nightly next boundary — and as of v0.0.5+ that routing now preempts accepted_rfc when the feedback is blocking (failed CI checks, CHANGES_REQUESTED reviews). The cascade order is resume_in_flight → unblocked_approval → pr_rescue (blocking only) → accepted_rfc → github_issue → pr_rescue (non-blocking) → ideate. Concretely: between tasks you can run nightly ci for an eyeball check, but you don't need to — nightly next will surface red CI automatically and bump it above fresh RFC work. Draft PRs count too. A not-yet-marked-ready PR with red CI is the same priority as a ready one; don't push commits to a draft that you wouldn't push to a ready PR. Always run nightly verify locally before git push, draft or not.
9. Arm the host-level keep-alive at session start. Run nightly session start as the first thing the /nightly skill does. This writes a SESSION_ACTIVE marker that the host's Stop-equivalent hook checks every turn boundary; without it, the hook lets the session end naturally. With it, the hook re-injects a "continue on X" prompt so the session keeps moving. The marker has no TTL (the 4-hour staleness release was removed in v0.0.3) — re-running nightly session start just refreshes it. Six of the seven Nightly hosts have a real force-continue hook (Claude Code's Stop, Codex CLI's Stop, Cursor 1.7+'s stop, Antigravity / Gemini CLI's AfterAgent, pi's agent_settled extension). opencode is soft and relies on the rule text above (the model is told to never stop). The disk-based off-ramps below work everywhere regardless.
10. Never invoke the human shutdown off-ramps yourself. The shell commands nightly conclude, nightly stop, and the matching slash commands /nightly-conclude, /nightly-stop, /nightly-bug exist for the human operator only. The agent never runs them — not at end-of-session, not when the cascade looks empty, not when "the work feels done." If you reach a turn boundary and the cascade has nothing left, run nightly ideate to surface proposals and nightly brief to render the report — then end your turn and let the Stop hook decide whether to force-continue (armed) or release (CONCLUDE / STOP / stale marker / max turns / PR backlog). The only signals that wind a session down are disk markers placed by the human (CONCLUDE, STOP) or by the hook's own safety caps. The agent's wrap-up is nightly ideate → nightly brief → end turn. Concluding is an intervention, not a workflow step. Past failure: agents have self-concluded — running nightly conclude after nightly brief on their own to "tidy up" — which freezes the cascade short-circuit at concluded and ends the session with unblocked RFC items still on disk.
11. Minimize PR count by consolidating; never stop because of it. The orchestrator does not gate on a PR-backlog count — there is no "too many PRs open" off-ramp. Monotonic forward progress across the whole overnight session is the contract; reaching any number of open PRs never ends a session on its own. The previous MAX_OPEN_PRS=5 cap was removed in v0.0.3 because it produced mid-session stops with unblocked RFC work still on disk — the wrong tradeoff. The replacement is consolidation, not gating. Before opening a new PR, prefer in this order:

• pr_rescue — when an existing Nightly PR has new feedback (CI failure, reviewer comments, bot suggestions), finishing it beats starting fresh. This is already cascade slot 5; honor it.
• Extend the most recently-opened in-flight PR when the current cascade pick is closely related to its scope — same RFC, same module, same feature. Check out its branch in a worktree, commit the additional change, push. The PR grows into one reviewable unit instead of becoming PR N+1.
• Bundle adjacent phases of the same RFC into one PR when the phases naturally compose. Phase A + B of a small RFC ships as one PR; truly independent phases of a large RFC stay separate. Only when none of the above applies — the cascade pick is genuinely orthogonal to every open PR — open a new branch. The goal is review-ergonomic, not PR-count-minimal-at-all-costs: bundling unrelated work into one PR is worse than two focused PRs. Bias: when uncertain, extend the most recent related PR. Past failure (now removed): agent shipped a 6th stacked paperwork PR while #54-#58 were still unreviewed because the cascade kept finding RFC-checkbox / lint-fallback work; the v0.0.2-and-earlier solution was the cap, which then created the new failure of ending sessions early. v0.0.3+ instead consolidates without capping.

1. Run the fleet wide and cheap; spend reasoning tokens only on judgment. Nightly routes each dispatch to a model sized for the job (RFC 007) and caps how many run at once (RFC 012). Both are enforced — nightly dispatch start refuses over-cap requests with exit 3, and worktree creation refuses past max_worktrees — so the contract here is about using the headroom, not respecting it.