papi-auto

Use papi-plan, papi-build, papi-strategy and papi-idea for the canonical phase rules. This skill coordinates one cycle; it grants no new scope, human review authority, publication, deployment or external messaging.

Modes

• dry-run: orient, inspect the current cycle, handoffs and evidence, then show proposed actions. Do not claim tasks, apply plans or edit files.
• Resume/default: continue the already approved cycle from its recorded state.
• unattended: continue only already authorized work. Never bypass plan approval, human verdicts, access controls or release authority.

Orient and preserve progress

Keep the active project connection and workspace binding fixed. Surface its name and verify it matches the intended project. Inspect the caller-owned cycle with build_list; distinguish teammate cycles, claims and person-only decisions. Use MCP records as project authority, not .papi/ files.

Fetch remote main before new branches. Preserve unrelated dirty work and isolate in a permitted task directory; never auto-stash, reset or discard it. Record the exact base, branch, owner, scope and durable receipts.

• In Progress: inspect its branch, commits, saved changes and owned process before resuming. A status alone is not progress.
• In Review: already built. Read its report and review queue; never rebuild.
• In Cycle/Backlog: check for existing implementation first. Build only eligible approved tasks with a handoff and satisfied prerequisites.
• Done/Cancelled: skip. Never change historical ownership for bookkeeping.
• Already shipped with missing evidence: follow papi-plan's restoration route.

When no cycle is active, prepare at least two coherent eligible tasks, aiming for three when available. Show exact IDs, names, scope and blockers before applying. An approved resume needs no new scope approval. Never silently drop tasks, include another person's work or invent filler. Run a due strategy review in this conversation; material choices require the user's authority. Never force past the planner's gate.

Classify and sequence

Read every handoff together before the first build. Sequence producers before consumers, then smaller eligible work. Same-module cycle tasks share a branch regardless of size, with one commit per task. Preserve an explicitly approved integrated PR. A shared branch takes its highest risk tier. Person-only decisions remain with their owner; continue independent work while awaiting them.

Risk-tier definition

Risk-tier modules are auth, data, migration, security, ci, payment and billing. Paths are categories, not a fixed list: inspect the owning repository for auth/session guard files, database migration code, CI/CD and deployment configuration, build/lint plumbing, and edge/serverless runtime functions. For PAPI maintainers, REVIEW_RISK_TIER_MODULES and REVIEW_RISK_TIER_GLOBS in packages/server/src/lib/review-risk-tier.ts own enforced classification. Customer projects use their own rules rather than copying PAPI's paths.

Review authority

Risk selects checks and reviewers; it never authorizes acceptance. Read the exact diff and report, run owning production-path checks and audit corrections. A builder's audit is not independent review. Delegate only through available host tools when authorized or required by owning instructions. Never invent a review or model identity from defaults. Keep builds In Review until the human verdict required by review_submit exists. Inspect review_list first and bind the explicit verdict to the reviewed commit. Absence of a verdict is not acceptance.

Execute and recover

Start eligible tasks through build_execute, implement the handoff, exercise the real entry point, push the commit and record an accurate completion report. Use native host tools actually available; Task, Bash and Skill are not universal APIs. Hosted HTTP cannot read caller-local response files.

Inspect failures immediately. Preserve the command, original error and last durable progress; verify claimed blockers before escalation. Repair the smallest safe in-scope cause. Retry only with evidence of changed input or a transient cause, at most one identical transient retry and one revised approach per failing operation. Further failure needs fresh diagnosis, not an indefinite loop. Never weaken tests, bypass credentials or stop shared services. Notify meaningful failure or an owned process stop with its recovery, and continue independent work when a genuine human action remains pending.

If delegation is authorized, record owner, scope, dispatch time, expected output, last durable progress and the next check within five minutes. Inspect files, receipts and owned processes at intervals no longer than five minutes. Between turns use an actual available heartbeat when requested or required, not merely saved instructions. Disclose any inability to provide timely supervision.

Finish accurately

Before an authorized merge, verify exact head, local gate, current CI, reviews and unresolved threads, then confirm the remote merge commit. Release only through the owning authorized versioned workflow after required tasks and reviews are resolved. Bookkeeping closure is not deployment. Never plan the next cycle automatically. Report pending work with its owner and next action; keep a parent audit open until every required client qualification has actual evidence.

Clean up only verified owned processes and disposable state; preserve shared resources and durable remote receipts. After compaction, orient and recover that state in this conversation rather than restarting work.