Session End

Session End Skill

Platform Note: State files (STATE.md, wave-scope.json) live in the platform's native directory: .claude/ (Claude Code), .codex/ (Codex CLI), .cursor/ (Cursor IDE), or .pi/ (Pi). All references to .claude/ below should use the platform's state directory. Shared metrics live in .orchestrator/metrics/. See skills/_shared/platform-tools.md.

Project-instruction file: CLAUDE.md and AGENTS.md (Codex CLI) are transparent aliases — see skills/_shared/instruction-file-resolution.md. All references to CLAUDE.md in this skill resolve via that precedence rule.

Phase 0: Bootstrap Gate

Read skills/_shared/bootstrap-gate.md and execute the gate check. If the gate is CLOSED, invoke skills/bootstrap/SKILL.md and wait for completion before proceeding. If the gate is OPEN, continue to Phase 1.

Phase 0.5: Parallel-Aware Preamble

Skip silently when persistence: false in Session Config.

Before Phase 1, run the parallel-aware preamble per skills/_shared/parallel-aware-preamble.md. The preamble detects other active sessions in the worktree-family via findPeers(repoRoot, { mySessionId }), classifies the caller's mode via classifyMode(callerMode) against the exclusivity-matrix, and fires the appropriate AUQ on conflict.

Outcome handling:

• PASS_THROUGH → continue to Phase 1
• EXCLUSIVE_BLOCKED → exit Phase 0 cleanly per the AUQ outcome
• PROMOTION_OFFER → user picks Worktree-Promotion (see parallel-aware-auq.md outcome-handling — calls enterWorktree()), in-place + Deviation, or Abbrechen

For session-end specifically: the preamble is DETECTION-ONLY. The lock-release path in later phases keeps its current behavior — releasing the OWN session's lock requires no matrix consultation.

Implementation reference: skills/_shared/parallel-aware-preamble.md § Implementation. AUQ reference: skills/_shared/parallel-aware-auq.md.

Phase 1: Plan Verification

Read back the session plan that was agreed at the start. For EACH planned item:

1.1 Done Items

• Verify with evidence: read the changed files, check git diff, run relevant test
• Confirm acceptance criteria are met
• Mark as completed

1.2 Partially Done Items

• Document what was completed and what remains
• Create a VCS issue for the remaining work with:
  • Title: [Carryover] <original task description>
  • Labels: priority:<original>, status:ready
  • Description: what's done, what's left, context for next session
• Link to original issue if applicable

1.3 Not Started Items

• Document WHY (blocked? de-scoped? out of time?)
• If still relevant: ensure original issue remains status:ready
• If no longer relevant: close with comment explaining why

1.3a Optional /goal Backlog-Drain (opt-in — #636)

Advisory-only continuation anchor at the session-end backlog seam. Never auto-invokes /goal, never blocks the close. /goal is a user slash-command; the operator decides whether to drain now or carry over.

Gate conditions — ALL must be true for this nudge to surface:

1. goal-integration.enabled: true in Session Config (default: false).
2. session-end-backlog is listed in goal-integration.seams.

When any gate condition is false, skip this step silently — no surfaced suggestion, no STATE.md write, no AUQ.

What it does — when the gate fires AND ≥1 still-relevant Not-Started (§1.3) or Partially-Done (§1.2) item exists AND the operator would rather drain the backlog now than carry it to a future session, surface ONE suggested /goal command as an advisory bullet. Example:

/goal Drain the remaining backlog items <list>; done when each item's acceptance check passes as shown by 'npm test' output in this turn AND 'npm run typecheck' prints 0 errors in this turn, or stop after 20 turns.

Advisory-only contract: this step never auto-invokes /goal, never blocks the close, raises no AskUserQuestion, and writes nothing to STATE.md. It is informational prose only — the operator copies the command if they want it. The deterministic Phase 2 Quality Gate of session-end remains the completion authority: /goal keeps the loop alive across turns, but npm test / npm run typecheck / npm run lint and their exit codes decide whether the drained work is correct.

The /goal evaluator reads the transcript only and runs NO tools — it anchors CONTINUATION, never JUDGMENT. The suggested condition therefore references freshly-run gate output "in this turn's output" and embeds a bound ("or stop after N turns"). Cross-reference .claude/rules/loop-and-monitor.md § LM-008 for the full /goal continuation-vs-judgment contract rather than restating it here.

One goal per session: only ONE /goal can be active at a time. This backlog seam and the inter-wave fix-loop seam (wave-loop.md § /goal Continuation Anchor) cannot both hold an active goal simultaneously — the operator picks one.

1.4 Emergent Work

• Tasks that were NOT in the plan but were done (fixes, discoveries)
• Document and attribute to relevant issues
• If new issues were identified: create them on the VCS platform

1.5 Discovery Scan (if enabled)

Read skills/session-end/discovery-scan.md for embedded discovery dispatch and findings triage.

1.6 Safety Review

Skip if persistence is false in Session Config (STATE.md won't exist).

Review safety metrics from the session. This is informational — it does NOT block the session close.

1. Read <state-dir>/STATE.md to extract:

   • Circuit breaker activations: agents that hit maxTurns (PARTIAL), agents that spiraled (SPIRAL), agents that failed (FAILED)
   • Worktree status: which agents used worktree isolation, any fallbacks or merge conflicts

2. Read enforcement hook logs from stderr (if captured): count of scope violations blocked/warned, command violations blocked/warned

3. Summarize:

   Safety review:
   - Agents: [X] complete, [Y] partial (hit turn limit), [Z] spiral/failed
   - Enforcement: [N] scope violations, [M] command blocks
   - Isolation: [K] agents in worktrees, [J] fallbacks
   

4. If any agents were SPIRAL or FAILED, ensure carryover issues exist (cross-reference with Phase 1.2)

5. Carryover validation fallback (#261): Walk each Wave History entry in STATE.md. For every agent whose status is SPIRAL or FAILED, check whether the line ends with a → issue #NNN suffix (or → existing #NNN). If the suffix is absent, the auto-create call in wave-executor did not run (e.g. a consumer-project #251 V0.x.y-close incident where the session crashed before dispatch completed, or the CLI was offline at detection time). Retroactively file the carryover via createSpiralCarryoverIssue:

   import { createSpiralCarryoverIssue } from '${PLUGIN_ROOT}/scripts/lib/spiral-carryover.mjs';
   
   // For each SPIRAL/FAILED agent missing the "→ issue #NNN" suffix:
   const result = await createSpiralCarryoverIssue({
     taskDescription: '<agent task from Wave History>',
     kind: 'SPIRAL', // or 'FAILED'
     context: '<Deviations / error context from STATE.md>',
     priority: 'high',
     vcs: '<from Session Config>'
   });
   // result.created → note new issue id in Final Report under "New Issues Created"
   // result.skipped === 'duplicate' → an earlier session already filed one; record the existing id
   // result.skipped === 'error' → log in Final Report as "⚠ carryover filing failed for <task>: <error>" and continue (do NOT block close)
   

   The module is idempotent via its task-hash dedup marker, so re-running the fallback across sessions will not create duplicates.

1.6.6 Record "What Not To Retry" entries (#623)

Skip if persistence is false (STATE.md won't exist).

For every SPIRAL or FAILED agent surfaced in the walk above, ALSO append a cross-session "What Not To Retry" entry to STATE.md. This is the durable, human-readable continuity slot that the NEXT session-start surfaces as a forced-read block (session-start Phase 6.5.1) so a future session does not re-attempt the same failed approach. Unlike a carryover issue (which captures unfinished work), this captures the approach that should not be repeated.

import { appendWhatNotToRetryOnDisk } from '${PLUGIN_ROOT}/scripts/lib/state-md.mjs';

// `parsed` = parseStateMd(STATE.md); session id from the `session:` frontmatter field.
const sessionId = parsed.frontmatter.session ?? 'unknown-session';
const today = new Date().toISOString().slice(0, 10); // YYYY-MM-DD

// For each SPIRAL/FAILED agent from the Wave History walk:
await appendWhatNotToRetryOnDisk(repoRoot, {
  approach: '<agent task description from Wave History>',
  why_failed: '<SPIRAL|FAILED> — <one-line context from Deviations / error>',
  session_id: sessionId,
  date: today,
});

The helper is lock-guarded (PSA-005) and prunes the section FIFO to the 10 most-recent entries on each append. Optional coordinator entry: if the session abandoned an approach for reasons NOT captured by a SPIRAL/FAILED agent (e.g. a design that proved unworkable mid-session), the coordinator MAY add a free-text entry through the SAME appendWhatNotToRetryOnDisk helper with a descriptive approach + why_failed. Recording is informational and does NOT block the close.

1.7 Metrics Collection

Read skills/session-end/metrics-collection.md for JSONL schema and conditional field rules.

1.8 Session Review

Dispatch the session-reviewer agent to verify implementation quality before the quality gate:

On Codex CLI, dispatch via the session-reviewer agent role defined in .codex-plugin/agents/session-reviewer.toml.

1. Invoke subagent_type: "session-orchestrator:session-reviewer" with:
   • Scope: all files changed this session (from git diff --name-only against the base branch)
   • Context: the session plan (issues, acceptance criteria) and all wave results from STATE.md
2. Wait for the reviewer's Verdict:

   • PROCEED — continue to Phase 2

   • FIX REQUIRED — disposition each listed item by severity:

     Finding class	Disposition
     HIGH+ / blocking review finding	Fix inline if quick (<2 min); else create an issue (priority:high, status:ready) and note it in the Final Report
     MED / LOW review finding	Fold in-session if quick; else record under "Unresolved Review Findings" in the Final Report — DO NOT create an issue (#617)
     Planned-carryover (item was in the plan, not finished)	ALWAYS create a [Carryover] issue per Phase 1.2 — unchanged
     SPIRAL / FAILED agent carryover	ALWAYS file via createSpiralCarryoverIssue per Phase 1.6 — unchanged

1.9 Mission-Status Classification (when mission-status present in STATE.md)

Skip if persistence is false in Session Config, or if mission-status: is absent from STATE.md frontmatter. When absent, fall back to binary checkbox detection in 1.1–1.4 unchanged — full backward compat.

When STATE.md frontmatter contains a mission-status: array (set by session-plan + wave-executor per #340), use the enum values to classify items into the 1.1–1.4 buckets. Read the array via parseMissionStatus(frontmatter) from scripts/lib/state-md.mjs.

Classification mapping:

• status: completed → 1.1 Done Items (item finished; verify with evidence per 1.1)
• status: testing or status: in-dev → 1.2 Partially Done (carryover; document what remains)
• status: validated or status: brainstormed → 1.3 Not Started (carryover; check if still relevant)
• Items NOT present in the mission-status: array → fall back to binary checkbox detection per 1.1–1.4 unchanged

Backward compat: When mission-status: is absent from STATE.md (pre-#340 STATE.md files, or sessions where session-plan did not emit the block), behave exactly as before — enum classification is skipped entirely and 1.1–1.4 binary checkbox logic runs as the sole classification mechanism.

1.10 Mission Status Breakdown (when mission-status present)

Skip if mission-status: is absent from STATE.md frontmatter (backward compat — no breakdown emitted).

After classifying items in Phase 1.9, produce a Mission Status breakdown subsection as part of the closed/carryover summary output. Count the number of tasks at each enum value across ALL waves:

### Mission Status Breakdown
- completed:    <N> tasks
- testing:      <N> tasks
- in-dev:       <N> tasks
- validated:    <N> tasks
- brainstormed: <N> tasks
- Total:        <N> tasks across <W> waves

Rules:

• Count each task-id entry from the mission-status: frontmatter array by its current status value.
• completed maps to Phase 1.1 (Done). testing + in-dev map to Phase 1.2 (Partial). validated + brainstormed map to Phase 1.3 (Not Started).
• Include this block in the Phase 6 Final Report under ### Carried Over or as a standalone subsection immediately after the Completed/Carried Over/New Issues lists.
• When all tasks are completed, the breakdown still appears (confirms clean session state).

Phase 2: Quality Gate

Verification Reference: See verification-checklist.md in this skill directory for the full quality gate checklist.

Run ALL checks listed in the verification checklist. If any check fails: fix if quick (<2 min), otherwise create a priority:high issue. Do NOT commit broken code.

Phase 2.0a: Echo-Stub Detection (GH #42)

gate-full.mjs emits a top-level stubbed: {} map in its JSON result, keyed by check name (typecheck, test, lint); value is { kind: 'echo'|'noop' }. When any check was short-circuited as a stub, runCheck() already returned status: 'pass' — so the overall gate verdict is green, but the result is meaningless.

Detection: immediately after parsing the gate-full JSON result, evaluate:

const stubbedEntries = Object.entries(result.stubbed ?? {});

If stubbedEntries.length > 0, surface a HIGH WARN block in the close summary:

⚠ QUALITY GATE STUBBED — <N> command(s) are echo/noop stubs, not real checks:
  - <check-name>: <kind> stub  (configured: "<command string>")
Re-configure with a real test command in CLAUDE.md Session Config before /close,
OR document this exception in /close --reason.

Behavior by enforcement mode:

• enforcement: strict — block /close. Treat as a Phase 2 failure. Present the WARN block and exit without committing.
• enforcement: warn (default) — continue, but write quality-gate-stubbed: true to STATE.md Deviations so the metrics writer captures it.
• enforcement: off — silent. Emit a single-line stderr log only (echo-stub detected: <check-name>).

Recipe: for container-based test runners (e.g. EspoCRM PHPUnit) where an echo-stub was the historical workaround, see docs/recipes/quality-gate-container-pattern.md.

Source issue: GH #42 (root cause: a consumer-project #251 V0.15.7-close incident — silent false-positive close-verdicts from echo-stub test commands).

2.1 Vault Validation (if configured)

Read skills/session-end/vault-operations.md for validator bash contract and reporting matrix.

2.2 CLAUDE.md (or AGENTS.md) Drift Check (if configured)

Read skills/session-end/drift-operations.md for checker bash contract and reporting matrix. Complements 2.1: vault-sync validates frontmatter inside the vault tree; drift-check validates narrative claims (paths, counts, issue refs, session-file refs) in top-level repo docs.

2.3 Vault Staleness Check (if configured)

Skip this subsection if vault-staleness.enabled is not true (default: false).

Step 1 — Resolve mode

Read vault-staleness.mode from $CONFIG (default: warn). Valid values: off | warn | strict.

If mode === 'off', skip Phase 2.3 entirely.

Step 2 — Invoke staleness probes

Both probes already ship in skills/discovery/probes/. Invoke each via Node import (no shell-out):

import { runProbe as runStaleness } from '$REPO_ROOT/skills/discovery/probes/vault-staleness.mjs';
import { runProbe as runNarrative }  from '$REPO_ROOT/skills/discovery/probes/vault-narrative-staleness.mjs';

const projectStaleness = await runStaleness(projectRoot, config);
const narrativeStaleness = await runNarrative(projectRoot, config);

Each probe returns { findings: Array, metrics: Object, duration_ms: Number } and auto-appends a JSONL summary record to its respective metrics file.

Step 3 — Aggregate and route by mode

totalFindings = projectStaleness.findings.length + narrativeStaleness.findings.length

• mode === 'warn' (default): report findings to closing report Docs Health line. Never block close.
• mode === 'strict':
  • If totalFindings === 0: continue, log Vault staleness: clean (mode=strict).
  • If totalFindings > 0: BLOCK the close. Present the findings list and offer override:
    • On Claude Code: AskUserQuestion with options:
      1. "Fix and retry Phase 2.3" (Recommended) — exit close, let user investigate
      2. "Override and close" — proceed, log a Deviation entry in STATE.md ## Deviations: - [<ISO timestamp>] Phase 2.3: Vault staleness strict-mode findings overridden by user. Findings: <count> (projects: <N>, narratives: <M>).
      3. "Abort close" — exit close without writing
    • On Codex CLI / Cursor IDE: same options as numbered Markdown list.

Step 4 — Surface to closing report

Pass the aggregated counts and mode forward to Phase 6 Final Report (Docs Health line — see Phase 6 below).

Phase 2.5: Custom Phases (#637)

Opt-in. Skip this phase entirely if custom-phases in $CONFIG is absent or [] (the default).

Repos declare deterministic close/housekeeping phases as a contract (not the freeform special: convention): each phase runs a command with exit-code gating and Final-Report reporting. The block is parsed by scripts/lib/config/custom-phases.mjs; each record is { name, when, command, mode, review } (already validated — unsafe records were dropped at parse time).

Step 1 — Read + filter by when

Read custom-phases from $CONFIG and the session-type from STATE.md frontmatter (feature | deep | housekeeping | none):

• If session-type === 'housekeeping': keep phases with when ∈ {housekeeping, both}.
• Otherwise (feature/deep/any other): keep phases with when ∈ {session-end, both}.

If no phases remain after filtering, skip to Phase 3.

Step 2 — Run each phase in declaration order

For each kept phase:

• mode === 'off' ⇒ skip silently (do not run the command).
• Otherwise run command via Bash. Capture the exit code and the last ~10 lines of stdout (these become the report summary — do NOT inline the full output).
• If review is set, read that file after the command as the review step and note its path in the report.

Step 3 — Route by mode

• mode === 'warn' (default): record the result (name, exit code, summary) for the Phase 6 Final Report "Custom Phases" line. Never block the close — even on a non-zero exit.
• mode === 'hard':
  • exit code 0 ⇒ continue; record <name>: pass (mode=hard).
  • exit code ≠ 0 ⇒ BLOCK the close using the same routing pattern as Phase 2.3 strict-mode. Present the phase name + captured summary and offer override:
    • On Claude Code: AskUserQuestion with options:
      1. "Fix and retry Phase 2.5" (Recommended) — exit close, let the user investigate.
      2. "Override and close" — proceed, log a Deviation entry in STATE.md ## Deviations: - [<ISO timestamp>] Phase 2.5: custom-phase '<name>' (mode=hard) exited <code>, overridden by user.
      3. "Abort close" — exit close without writing.
    • On Codex CLI / Cursor IDE: same options as a numbered Markdown list.

A hard-fail (whether overridden or not) ALWAYS appends its result line to STATE.md ## Deviations; warn-mode results do not.

Step 4 — Surface to closing report

Pass each phase result (name, mode, exitCode, summary, review?) forward to the Phase 6 Final Report "Custom Phases" line (see Phase 6 below).

Phase 3: Documentation Updates

Final heartbeat (#590-3) — at Phase 3 entry, refresh the session-lock heartbeat BEFORE the multi-minute close-out chain (vault-mirror, dialectic, durable-commit, metrics). A long-idle deep session may not have had PostToolBatch activity for >4h; without a refresh the 4h-TTL lock would lapse mid-close and appear stale to a concurrent session. Place this call BEFORE Phase 3.8 Session Lock Release (which deletes the lock — refreshing a deleted lock is a no-op). Best-effort: a failure must NOT block the close.

// Final heartbeat (#590-3) — refresh before the multi-minute close-out (vault-mirror, dialectic, durable-commit)
// so a long-idle deep session's 4h-TTL lock does not lapse mid-close.
// BEFORE Phase 3.8 lock-release (which deletes the lock).
import { updateHeartbeat } from 'scripts/lib/session-lock.mjs';
updateHeartbeat({ sessionId, repoRoot: process.cwd() });

Skip silently if persistence: false in Session Config (no session.lock exists in that mode).

3.0 Defensive Cleanup

Delete <state-dir>/wave-scope.json if it still exists:

rm -f <state-dir>/wave-scope.json

This should have been cleaned up by wave-executor after the final wave, but crashed sessions or interrupted executions may leave it behind. A stale scope manifest from a previous session could incorrectly restrict the next session's enforcement hooks.

3.1 SSOT Files

• Update STATUS.md / STATE.md if they exist (metrics, dates, status)
• Update CLAUDE.md (or AGENTS.md on Codex CLI) if patterns or conventions changed during this session
• Check <state-dir>/rules/ — if a new pattern was established, suggest a new rule file

3.2 Docs Verification (docs-orchestrator integration)

Skip this subsection if docs-orchestrator.enabled config is not true (default: false). Also skip entirely if docs-orchestrator.mode is off.

Reads docs-tasks from STATE.md frontmatter (written by wave-executor Pre-Wave 1b), computes CHANGED_FILES via git diff --name-only "$SESSION_START_REF..HEAD", and runs a per-task verification loop (outcome: ok/partial/gap). In warn mode logs results non-blocking; in strict mode blocks on any gap and presents an AskUserQuestion override prompt. Emits a ### Documentation Coverage (docs-orchestrator) block for inclusion in the Phase 6 Final Report.

See phase-3-2-docs-verification.md for full details.

3.2a Session Handover (for significant sessions)

If this session made substantial changes, create or update:

• <state-dir>/session-handover/ doc with: tasks completed, resume point, metrics changed, issues opened/closed
• Or update <state-dir>/STATE.md with session digest

3.3 Claude Rules Freshness

Review <state-dir>/rules/ files that are relevant to this session's work:

• Are the rules still accurate after this session's changes?
• Should any rule be updated with new patterns?
• Should a new path-scoped rule be created?
• Suggest changes but DO NOT modify without user confirmation

3.4 Update STATE.md

Ownership Reference: See skills/_shared/state-ownership.md. session-end is authorized to set status: completed plus the optional updated timestamp (#184), and — as of Phase A of Epic #271 — the 5 Recommendation fields written by Phase 3.7a. No other fields.

Runtime Ordering Note (Epic #271 Phase A): Phase 3.4's status: completed write executes LAST in Phase 3, AFTER Phase 3.7 (sessions.jsonl) and Phase 3.7a (Compute and Write Recommendations). The ordinal position here (3.4) is kept for historical compatibility; the canonical runtime order is 3.1 → 3.2 → 3.3 → 3.4a → 3.5 → 3.5a → 3.6 → 3.6.5 → 3.6.7 → 3.6.8 → 3.7 → 3.7a → 3.7b → 3.7c → 3.4. Rationale: Phase 3.7a reads in-memory session metrics and writes the 5 Recommendation fields via updateFrontmatterFields; that write must complete BEFORE the STATE.md frontmatter is finalized with status: completed so the Recommendation fields are visible to the next session-start while STATE.md is still status: active. Crash-resilience: if /close aborts between 3.7a and 3.4, STATE.md carries status: active + Recommendations; session-start Phase 1.5 offers resume (and the banner renders). If the reverse ordering were used (status: completed first), a crash would leave status: completed without Recommendations — the Reader would silently no-op the banner, losing the handoff.

Gate: Only run if persistence is enabled in Session Config and <state-dir>/STATE.md exists.

1. Set frontmatter status: completed
2. Record final wave count and completion time in the frontmatter
3. Touch updated: <ISO 8601 UTC> in the frontmatter (issue #184). Use scripts/lib/state-md.mjs → touchUpdatedField for safety:

   node --input-type=module -e "
   import {readFileSync, writeFileSync} from 'node:fs';
   import {touchUpdatedField} from '${PLUGIN_ROOT}/scripts/lib/state-md.mjs';
   const p = '<state-dir>/STATE.md';
   writeFileSync(p, touchUpdatedField(readFileSync(p, 'utf8'), new Date().toISOString()));
   "
   

   Silent no-op if the file has no frontmatter.
4. Keep the file as a record — do NOT delete it (next session-start reads it)

If STATE.md doesn't exist, skip this subsection.

3.4a Coordinator Snapshot Cleanup (#196)

Pre-dispatch snapshots (refs/so-snapshots/<sessionId>/wave-*) are created by wave-executor before each wave dispatch so that session-start can offer recovery if a session is interrupted mid-wave. On a clean close those snapshots are no longer needed and should be deleted. In addition, orphaned refs from older sessions that were never cleaned up (e.g. after a hard crash) are garbage-collected using an age-based policy (14 days).

Gate: Only run if persistence is true in Session Config. Skip entirely when persistence is off (snapshots are never written in that mode).

node --input-type=module -e "
import { listSnapshots, deleteSnapshot, gcSnapshots } from '${PLUGIN_ROOT}/scripts/lib/coordinator-snapshot.mjs';

// Step A: delete this session's snapshots (clean close → we don't need them)
const mine = await listSnapshots({ sessionId: '${SESSION_ID}' });
for (const s of mine) {
  const r = await deleteSnapshot({ refName: s.ref });
  if (!r.ok) console.error('snapshot cleanup:', r.error);
}

// Step B: GC orphans older than 14 days (non-fatal)
const gc = await gcSnapshots({ olderThanDays: 14 });
console.log(\`snapshot cleanup: deleted \${mine.length} from this session + \${gc.deletedCount} expired orphans (scanned \${gc.scanned}).\`);
"

Failures in either step are logged to stderr but do not block session close — a missed cleanup is self-healing via the 14-day GC on the next session.

This cleanup is the counterpart to the session-start Phase 1.5 recovery prompt: once a session closes cleanly, future sessions must not be offered recovery for its snapshots.

3.5 Session Memory

Gate: Only run if persistence is enabled in Session Config AND platform is Claude Code (session memory at ~/.claude/projects/ is Claude Code-only). Learnings (Phase 3.5a) and metrics (Phase 3.7) still write to .orchestrator/metrics/ on all platforms.

1. Create ~/.claude/projects/<project>/memory/session-<YYYY-MM-DD>.md with:
   • Frontmatter: name, description (1-line summary), type: project
   • ## Outcomes — per-issue status (completed / partial / not started) with evidence
   • ## Learnings — patterns discovered, architectural insights, gotchas
   • ## Next Session — priority recommendations, suggested session type, blockers
2. Update ~/.claude/projects/<project>/memory/MEMORY.md:
   • Under a ## Sessions heading (create if missing), add: - [Session <date>](session-<date>.md) — <one-line summary>

3.5a Learning Extraction + 3.6 Memory Cleanup & Learnings Write

Read skills/session-end/learning-patterns.md for extraction heuristics, confidence updates, passive decay, and JSONL write procedure.

3.6.3 Memory Proposals Collection (#501, F2.1)

Gate: Skip this phase entirely when ANY of:

• persistence is false in Session Config
• memory.proposals.enabled is false (default: true)
• .orchestrator/metrics/proposals.jsonl does not exist OR contains zero entries

After learnings are written (Phase 3.6) and BEFORE auto-dream dispatch (Phase 3.6.5), collect agent-proposed memory entries written during this session and present them to the operator via AskUserQuestion multiSelect. Approved entries flow to learnings.jsonl with _provenance: agent-proposed@<wave-id>. Rejected entries are archived to .orchestrator/proposals.rejected.log.

The proposals queue is populated mid-session by wave-executor agents calling node scripts/memory-propose.mjs --type ... --subject ... --insight ... --evidence ... --confidence .... The CLI enforces:

• Quota per wave (default 5, configurable via memory.proposals.quota-per-wave)
• Confidence floor (default 0.5, configurable via memory.proposals.confidence-floor)
• Wrong-context guard (CLI exits non-zero when STATE.md status is not active)

Coordinator-direct procedure

1. Read Session Config: memory.proposals.enabled (default true), memory.proposals.quota-per-wave (default 5), memory.proposals.confidence-floor (default 0.5), auto-dream.min-confidence (default 0.5 — issue #566; SECOND gate above the write-time memory.proposals.confidence-floor).

2. Invoke collectProposals from scripts/lib/memory-proposals/collector.mjs, passing the collect-emit confidence floor from Session Config:

   import { collectProposals } from '${PLUGIN_ROOT}/scripts/lib/memory-proposals/collector.mjs';
   const { queue, stats, perWaveSummaries } = await collectProposals({
     repoRoot: process.cwd(),
     // Issue #566: collect-emit confidence floor. Records with
     // `record.confidence < minConfidence` are dropped from `queue` (but
     // counted in stats). When the key is absent, defaults to 0.5 via the
     // `_parseAutoDream` parser.
     minConfidence: config['auto-dream']?.['min-confidence'],
   });
   

3. If queue.length === 0: log memory-proposals: queue empty (stats: ${JSON.stringify(stats)}) and continue.

4. AUQ pagination logic: partition the queue into FIFO batches of 4 inline:

   • Empty queue → silent skip (no AUQ rendered).
   • 1-4 items → single multiSelect call with all items as options.
   • 5+ items → sequential multiSelect calls in batches of 4 (FIFO order; final batch may have < 4 items).

   // Inlined from former scripts/lib/memory-proposals/auq-partition.mjs (PRD F2.2 #502 closed; see #558 M2).
   const BATCH_SIZE = 4;
   const batches = [];
   if (Array.isArray(queue) && queue.length > 0) {
     for (let i = 0; i < queue.length; i += BATCH_SIZE) {
       batches.push(queue.slice(i, i + BATCH_SIZE));
     }
   }
   

   Then iterate batches and emit one AskUserQuestion per batch with header: "Memory — Confirm Proposals (Batch N of M)". Option label format: [<type-12>] | <subject-40> | conf=X.XX. Option description: evidence: <first 60 chars of insight>. multiSelect: true.

5. After all batches answered, partition the queue into approved (any option selected across all batches) and rejected (all unselected).

6. Invoke writeApproved and archiveRejected from scripts/lib/memory-proposals/sink.mjs:

   import { writeApproved, archiveRejected, clearProposalsJsonl } from '${PLUGIN_ROOT}/scripts/lib/memory-proposals/sink.mjs';
   const writeResult = await writeApproved({ approved, repoRoot, sessionId });
   const archiveResult = await archiveRejected({ rejected, repoRoot, reason: 'user-declined' });
   await clearProposalsJsonl({ repoRoot });
   

7. Log outcome for Phase 6 Final Report: memory.proposals: <queued> queued → <approved> approved, <rejected> rejected (dropped: <dropped> quota, <below_floor> below-floor).

Failure modes

• If collectProposals fails (fs error): log warning ⚠ memory-proposals: collect failed (${err}) — skipping, do not block session close.
• If writeApproved reports errors per-record: log each, but continue (per-record fault isolation per sink contract).
• If clearProposalsJsonl fails: log warning; do not block. The file may be re-collected at the next session-end, idempotent.

Cross-references

• PRD: docs/prd/2026-05-21-learning-memory-modernization.md § F2.1
• Modules: scripts/lib/memory-proposals/{schema,store,collector,sink}.mjs
• CLI: scripts/memory-propose.mjs (agents call this)
• Hook: hooks/pre-bash-memory-propose-audit.mjs (audit trail)
• Coordinator AUQ spec: agents/memory-proposal-collector.md (reference doc)
• Sibling phases: 3.6.5 Auto-Dream (#502), 3.6.6 Skill-Applied Judge (#645 L3), 3.6.7 Auto-Dialectic (#506)
• Issue: #501

3.6.5 Auto-Dream Dispatch (#502, F2.2)

Skip this phase if memory-cleanup-threshold: 0 (kill-switch per PRD F2.2). Also skip on non-Claude-Code platforms (memory dir at ~/.claude/projects/ is Claude Code-only, mirrors Phase 3.5 gate).

After learnings are written (Phase 3.6), determine whether to emit a manual-cadence nudge to run /memory-cleanup --dry-run in the next session. The decision uses MEMORY.md line count and a sessions-since-last-cleanup signal. There is no memory-cleanup agent in the registry, so the historical auto-dream subagent dispatch never fired (see #614) — the nudge replaces it. A manually-run /memory-cleanup --dry-run writes a unified-diff proposal to .orchestrator/pending-dream.md for the session after that to apply via /memory-cleanup --apply-pending.

1. Read memory-cleanup-threshold (default 5) and memory-cleanup-soft-limit (default 180) from $CONFIG.

2. Invoke shouldDispatchAutoDream from scripts/lib/auto-dream.mjs:

   import { shouldDispatchAutoDream } from '${PLUGIN_ROOT}/scripts/lib/auto-dream.mjs';
   import { resolveMemoryDir } from '${PLUGIN_ROOT}/scripts/lib/memory-paths.mjs';
   const memoryDir = resolveMemoryDir();
   const decision = await shouldDispatchAutoDream({
     repoRoot: process.cwd(),
     memoryDir,
     threshold: config['memory-cleanup-threshold'] ?? 5,
     softLimit: config['memory-cleanup-soft-limit'] ?? 180,
   });
   

3. If decision.trigger === false: log auto-dream: not triggered (${decision.reason}) and continue. Emit no nudge.

4. If decision.trigger === true: do not dispatch a subagent — there is no memory-cleanup agent in agents/, so the historical Agent({…}) dispatch pointed at the agent name memory-cleanup (a subagent type that was never built) and never fired (see #614). Instead, emit a manual-cadence nudge and continue:

   auto-dream: cadence reached (${decision.reason}) — run /memory-cleanup --dry-run manually in the next session, then apply the proposal with /memory-cleanup --apply-pending.

   The shouldDispatchAutoDream decision helper and scripts/lib/auto-dream.mjs lib stay in use: they compute the signal that drives this nudge and back the manual /memory-cleanup path (writePendingDream / readPendingDream / applyPendingDream).

5. Record the outcome (skipped / nudge-emitted) so Phase 6 Final Report can surface a line: auto-dream: manual /memory-cleanup --dry-run recommended (cadence reached) — apply with /memory-cleanup --apply-pending next session.

The pending-dream sidecar at .orchestrator/pending-dream.md is intentionally outside the vault tree — vault-mirror (Phase 3.7) must exclude it from its scope so the proposal survives the session close without being mirrored into 50-sessions/.

Cross-reference: PRD F2.2 acceptance criteria; scripts/lib/auto-dream.mjs API (shouldDispatchAutoDream, readDreamSignals, writePendingDream, readPendingDream, applyPendingDream).

3.6.6 Skill-Applied Judge (#645, L3)

Default OFF. Skip this phase — with NO module import and NO sidecar created — unless BOTH gates pass (evaluated in this order):

1. config['skill-evolution'].judge !== true → skip (the judge: key in the top-level skill-evolution: block; default false).
2. persistence === false in Session Config → skip.

When skipped, log skill-judge: disabled (skill-evolution.judge=false) (or persistence=false) and return. This is the disabled-path guarantee: with the judge off, only L1 (skill-invocations.jsonl, written by the PreToolUse hook) and L2 (scripts/lib/skill-health/join.mjs) records exist — no judgment, no error, zero L3 code executes. Do NOT import scripts/lib/skill-judge.mjs on the disabled path.

After learnings are written (Phase 3.6) and the auto-dream decision is made (Phase 3.6.5), and when the judge is enabled, run a bounded, read-only LLM-judge over this session's selected skills to emit ADVISORY per-skill applied/completed judgments to .orchestrator/metrics/skill-judgments.jsonl.

The #614 distinction (the whole point of L3's Design A): unlike the 3.6.5 / 3.6.7 nudge-only paths — which cannot dispatch a live subagent because the target read-only agents (memory-cleanup, dialectic-deriver) cannot write their own sidecars — L3 performs a LIVE read-only dispatch. This is #614-safe because the read-only skill-applied-judge agent RETURNS JSON and the COORDINATOR writes the sidecar, not the agent. A read-only agent that returns judgments is allowed; a read-only agent that must write a file is the #614 trap.

Advisory-only: the judge output is written with advisory: true (schema-rejected otherwise) and NEVER feeds an auto-action gate — not a sunset decision, not a C2 repair (scripts/lib/skill-evolution/*), not a promotion. Per #645 R9(b) the C2 repair gate stays deterministic; L3 is a signal for humans/dashboards only.

1. Read config['skill-evolution'].judge (default false), config['skill-evolution']['judge-budget-tokens'] (default 8000), and persistence. Apply the two skip gates above.

2. Determine the judged set — only THIS session's selected skills. Read .orchestrator/metrics/skill-invocations.jsonl and collect the distinct skill values whose session_id matches the current session id. If the judged set is empty, runSkillJudge returns status: 'empty-input' (no dispatch) — log and continue.

3. Invoke runSkillJudge from scripts/lib/skill-judge.mjs, wiring the real dispatch via the DI seam:

   import { runSkillJudge } from '${PLUGIN_ROOT}/scripts/lib/skill-judge.mjs';
   import { appendSkillJudgment } from '${PLUGIN_ROOT}/scripts/lib/skill-judgments-schema.mjs';
   import path from 'node:path';
   
   const budgetTokens = config['skill-evolution']['judge-budget-tokens'] ?? 8000;
   const result = await runSkillJudge({
     // Claude Code path: wire the real read-only haiku subagent as dispatchAgent.
     dispatchAgent: ({ model, prompt, maxTokens }) =>
       Agent({ subagent_type: 'skill-applied-judge', model: 'haiku', prompt, max_tokens: maxTokens }),
     repoRoot: process.cwd(),
     sessionId,
     transcriptTail,                 // recent session transcript excerpt (UNTRUSTED — fenced by the lib)
     selectedSkills,                 // distinct skills from step 2
     model: 'haiku',
     budget: { input: budgetTokens, output: 4000 },
   });
   

   • Claude Code path: dispatchAgent wraps the real Agent({ subagent_type: 'skill-applied-judge', model: 'haiku', … }). The agent is sandbox-tier: read-only and RETURNS one fenced ```json block — it never writes files.
   • **Cod

…(truncated)