Single fire-and-forget command that runs the full session learning pipeline: insights → reflect → knowledge → improve → action-candidates → report. All machines run local Phases 1–9 against logs/orchestrator/ and commit derived state. dev-primary additionally runs Phase 10a (cross-machine compilation) and Phase 10 (report). Safe for cron scheduling. Use when session ends, nightly cron fires, or you want to harvest learnings from recent sessions. Replaces running 4 skills manually.
Single fire-and-forget command that runs the full session learning pipeline: insights → reflect → knowledge → improve → action-candidates → report. All machines run local Phases 1–9 against logs/orchestrator/ and commit derived state. dev-primary additionally runs Phase 10a (cross-machine compilation) and Phase 10 (report). Safe for cron scheduling. Use when session ends, nightly cron fires, or you want to harvest learnings from recent sessions. Replaces running 4 skills manually.
dev-primary git pull in Phase 10a picks up all machines' committed derived state.
Pipeline Summary
Run phases sequentially. Non-mandatory phases log failure and continue. Fatal failures
in Phases 1 or 4 set _PIPELINE_EXIT=1. Phase 10 always runs via trap EXIT.
Phase
Name
Mandatory
Short description
1
Insights
✓
Extract skill usage, tool patterns, session-quality signals from all log sources
1b
Drift Detection
dev-primary
Detect python_runtime/file_placement/git_workflow violations in yesterday's log
2
Reflect
—
Invoke /reflect against reflect-history/ and trends/
3
Knowledge
—
Invoke /knowledge; update patterns/
3b
Memory Compaction
—
Compact MEMORY.md + topic files
3c
Memory Curation
—
Promote stable patterns, expire stale entries
4
Improve
✓
Invoke /improve; update skills + rules from candidates/
5
Correction Trends
—
Analyze corrections/ for recurring failure patterns
6
WRK Feedback + Ecosystem
—
WRK quality review + skill usage frequency + ecosystem health
7
Action Candidates
—
Convert candidates/ entries to WRK items
8
Report Review
—
Review learning report for coherence
9
Skill Coverage Audit
weekly
Audit skill coverage via identify-script-candidates.sh + skill-coverage-audit.sh; include tier distribution from skill-tier-report.py (A/B/C/D per config/skills/quality-tiers.yaml)
Sessions are pure multi-agent execution engines — all brain directed at the task.
Analysis, maintenance, and learning are deferred to the nightly run.
In-session
Nightly pipeline
WRK gate check + active-wrk set
All insight/reflect/knowledge/improve runs
Multi-agent implementation
Correction trend analysis
Fast signal capture (hooks write raw signals)
Candidate → WRK auto-creation
/session-start context load
Memory and skill file updates
Cross-review (Codex gate)
Ecosystem health checks
Commit + push
Session archive rsync
Must NOT run standalone during sessions:/insights, /reflect, /knowledge,
/improve, consume-signals.sh heavy analysis, ecosystem-health-check.sh,
session-end-evaluate.sh scoring.
Stop hooks: one hook only, raw write, < 1 second. See WRK-304.
scripts/planning/ — ensemble planning outputs harvested by Planning Quality Loop
Exit-Handoff Boundary
When the user asks to "document and prepare to exit," do not run the heavyweight comprehensive-learning pipeline in-session. Instead, create or update the task-specific durable handoff/report, verify commit/push/clean-state evidence, and leave deeper insights/reflect/knowledge/improve processing to the nightly pipeline. The exit response should be concise: handoff path, pushed commit(s), known dirty-state exceptions, external-action status, and remaining next steps.
For the concrete closeout checklist, use references/exit-handoff-closeout.md. Key requirements: write the handoff under docs/session-handoffs/ when no task-specific location exists, include final clean/sync proof for every touched tier-1 repo, commit and push the handoff unless blocked, inspect any hook-generated dirt before claiming clean state, and explicitly state that no external send/action was performed unless the user approved it.
For travel-planning sessions captured primarily as GitHub issues/comments rather than repo files, use references/github-issue-backed-travel-exit-closeout.md: verify issue/comment URLs live, write the control-repo handoff when the issue repo has no local checkout, stage only the handoff, and report synced-but-dirty control repo state precisely.
Iron Law
No learning pipeline phase (/insights, /reflect, /knowledge, /improve) shall run standalone during an active work session — learning is deferred to the nightly pipeline, always.
Explicit Skill-Library Update Requests
If the user explicitly asks to "review the conversation and update the skill library," do not hide behind the nightly deferral rule. Perform a targeted skill update using skill_manage against the currently loaded class-level skill or the closest existing umbrella. This is a bounded library-maintenance action, not the heavyweight comprehensive-learning pipeline. If a referenced support file is missing, create it immediately under references/ and keep SKILL.md pointing to it.
Use references/conversation-review-skill-update-mode.md as the operating checklist for this mode. Key rules: be active by default, patch loaded/governing class-level skills first, prefer support files under existing umbrellas over narrow one-session skills, and treat user corrections about style/format/workflow as first-class skill-library signals.
After targeted skill-library edits, treat the skill ledger as part of the closeout artifact set. A skill_manage patch/write can create or later append tracked ledger entries such as logs/orchestrator/hermes/skill-patches.jsonl; inspect that dirt after the primary skill commit, commit it separately if it is intentional metadata, then fetch/verify final HEAD == origin/<branch>. Do not claim a clean exit immediately after the first commit if hooks or skill tooling generated follow-up ledger dirt.
Rationalization Defense
Excuse
Reality
"I'll just run a quick /reflect to capture this insight"
/reflect consumes significant context and token budget. The nightly pipeline captures the same signals from hooks and logs — for free.
"The session is almost over, might as well run /improve now"
"Almost over" is when context is most valuable. Defer to nightly; hooks already captured the raw signals.
"This learning will be lost if I don't process it now"
Stop hooks write raw signals in < 1 second. The nightly pipeline processes them. Nothing is lost by deferring.
"The nightly cron might not run tonight"
Fix the cron job, do not work around it by running learning mid-session. Two problems are worse than one.
Red Flags
These phrases signal you are about to violate the Iron Law:
