atlas-settle - SKILL.md Agent Skill

name: atlas-settle description: Provides close-out protocols and judgment procedures for S4 Close and S5 Learn under the ATLAS Loop, releasing advisory locks and logging final feature performance metrics.

Atlas Settle Skill

This skill provides final close-out and settlement protocols for tasks and features executing under the ATLAS Loop (specifically stages S4 Close and S5 Learn).

Deterministic tasks (like S1 Verify and S3 Validate) are automated via utility scripts in this skill, while interpretive close-out gates (PR sign-offs, retro reflection, and memory updates) follow the judgment procedures detailed here.

1. Stage S4: Close (Mandatory Cleanup)

The purpose of S4 Close is to ensure that the local workspace is returned to a clean state, releasing all advisory locks to prevent path collisions with later work in the same session. (These locks are single-machine advisory only — see docs/flow/artifact-ownership.md.)

1.1. Lock & Claim Release Judgment

As a developer or agent finishing a feature branch, you must verify that no residual resource blocks remain:

Diff Hygiene: Before releasing claims or leases, compare changed files with the active claim scope. Warn about out-of-scope changes, but do not block Settle in this lightweight gate.
Work Claims: Verify that the advisory claim lock created during Stage A: Align is removed from .ai/claims/.
File Leases: Verify that all advisory leases acquired during Stage L: Lay on specific files are removed from .ai/locks/.
State Cleanliness: Inspect the flow-state.json file. The locks array must be empty [] and status should be transitioned to a completed/settling state.

1.2. Verification Action

Execute the lock release script:

node .claude/skills/atlas-settle/scripts/release-locks.js

Confirm that the command exits 0 and outputs a passing gate-result JSON envelope.

2. Stage S5: Learn (Retrospective & Observability)

The purpose of S5 Learn is to capture task metadata, log duration and retry attempt metrics, and extract long-term learnings for the codebase before concluding.

2.1. Retrospective Reflection (Interpretive)

Before archiving a phase or plan, you MUST compile a retrospective report outlining:

Key Decisions: Document any design patterns, library integrations, or architectural choices made during the phase.
Issues Encountered: Detail any blockages, command failures, or test regressions, and how they were resolved.
Lessons Learned: List technical takeaways or recommendations for future phases.

These are written to the phase's {phase}-{plan}-SUMMARY.md file.

Durable memory updates ("Vệt ốc sên" compaction): do not distill the session into .ai/memory/* inline in this (expensive) session. Run saf compact-memory to prep the compaction pack and handoff scaffold, then delegate the distillation to a subagent on the prescribed fast model (see the ## Snail Trail — Memory Compaction at Settle section of your instruction file) running the saf-handoff skill. Gate the result with saf handoff.

2.2. Metric Logging Verification (Deterministic)

Log the task's final attempt count and latency metrics to the central period observability ledger by executing:

node .claude/skills/atlas-settle/scripts/signal-log.js

Verify that the signal log script appends a valid JSONL entry to .ai/signals/current-period.jsonl conforming to the entities.schema.json schema contract.

2.3. Handoff & State Synchronization

Verify that the flow-state.json is updated and all local state changes are synchronized with the git history.