Postmortem → Improve

Close the loop: observe failure → write truth → change what the next prompt sees.

Primary enzyme (org or solo):

@harness/LESSONS.active.md

App-local under {appData}/harness-lessons/<scopeId>/ (outside project git).
Scope: org-<orgRootSessionId> when in org lineage, else ws-<workspaceHash>.
Harness injects non-empty content into system prompt as ## Active Operational Lessons on the next LLM turn.

Specialist skills own optional secondary mutations (boost, org-restructure, …).
Do not invent a parallel org model.

Not This Skill

Skill	Use for
teamwork	First-time scaffold / choose substrate
org	Create org, spawn/resume org members
org-restructure	Apply role/constitution structural changes
delegation-eval-loop	Grade one child sprint (not fleet learning)
boost / recruit	Tune or create assistant configs only
schedule / loop	Recurring wake-ups (cadence only)
agent-init	Workspace agents.md (dirties project git — not default here)

Core Rules

1. Evidence before narrative. Prefer tool errors, session status, diffs, coordination files over memory.
2. Two-track triage (mandatory).
   • Deterministic / tool / contract defect (contract_lie, wrong schema, handler bug, missing binary): do not write LESSONS. File a fix/bug task or apply a product patch if the user asked for code changes.
   • Behavioral / decision failure: update @harness/LESSONS.active.md with falsifiable rules.
3. Prompt enzyme or it did not happen. Chat-only conclusions and KANBAN-only backlog are deferred, not Done.
4. Git-clean by default. Never treat workspace agents.md / SOUL.md / tracked repo edits as the learning channel unless the user explicitly asks.
5. Bounded ambition. Max 5 active lesson rules; ≤30 lines / 2KB (harness truncates beyond that). Evict before adding a 6th.
6. Write from the governing session only. Harness rejects @harness writes from children (parent_session_id / depth>0) and from org members whose org_root_session_id != this session. Propose candidate rules in final text; root merges into @harness/LESSONS.active.md.

When to Run

• A mission, sprint, or objective finished (success or failure)
• The same failure class repeats (≥2 times)
• User asks for postmortem, retro, after-action, continuous improvement, or "learn from this"
• Optional: org/teamwork with stuck coordination/KANBAN.md Blocked items

Teamwork/org is optional. Solo sessions skip @teamwork archive steps.

Workflow

1. Gather evidence (short)

Collect only what supports root-cause claims:

• Failed acceptance criteria / eval rejects
• Session terminals: cancelled, timeout, circuit-break, empty final text
• Tool/config mismatches via agent__listAgents / tool__listServers when relevant
• Optional: session trace excerpts — do not dump entire traces into the archive

If @teamwork/ exists, also skim MISSION.md → ROLES.md → agents.md → KANBAN.md → HANDOFF.md.

2. Write the archive (optional but recommended)

With teamwork: @teamwork/coordination/POSTMORTEMS/YYYY-MM-DD-<slug>.md
Solo / no teamwork: @harness/POSTMORTEMS/YYYY-MM-DD-<slug>.md

Use postmortem-template.md.

Optional index line in @harness/LESSONS.md (or @teamwork/coordination/LESSONS.md if teamwork exists):

- YYYY-MM-DD — <slug> — <one-line lesson> — track: behavior|defect

3. Triage Top-1

Pick one primary actionable. Classify:

Track	Meaning	Done when
defect	Tool/schema/handler/environment contract broken	Bug/fix task filed or code patched (user-authorized); LESSONS write forbidden
behavior	Agent chose wrong strategy/tool/sequence despite honest tools	@harness/LESSONS.active.md updated with ≤5 structured rules

Routing helpers (secondary, after enzyme): improvement-routing.md.

4. Apply (behavior track)

Read current @harness/LESSONS.active.md (may be missing). Merge Top-1 rule(s).

Rule syntax (one line each):

- [YYYY-MM-DD] TRIGGER: <when> | FORBIDDEN: <anti-pattern> | REQUIRED: <correct action>

Constraints:

• Plain text only — no HTML/XML, no "ignore previous instructions"
• Prefer concrete tool names / args / acceptance checks
• If already at 5 rules, remove or merge the least relevant before adding

Write with either:

• workspace__writeFile to @harness/LESSONS.active.md with mode: "overwrite" (default create renames to LESSONS.active-1.md and the prompt will ignore it), or
• workspace__editFile / replaceLines on the existing file.

Never rely on default create-mode for updates.

5. Verify learning stuck

Before declaring the cycle complete:

• Archive exists (or user waived) with evidence pointers
• Track decided: defect or behavior
• behavior: @harness/LESSONS.active.md updated; next turn should show ## Active Operational Lessons in system prompt
• defect: LESSONS not used as a workaround; fix/bug path recorded
• No learning left only in chat
• Workspace git not dirtied for the learning step

6. Optional secondary routes

After the enzyme step, top remaining actions may go to boost / org-restructure / delegation-eval-loop / schedule — see routing table. Keep ≤2 secondary actions.

Cadence (optional)

• After each major mission: run this skill even on success (near-misses)
• Weekly: scan open defect tasks + LESSONS.active staleness

Use schedule for org-wide cadence; loop only for in-conversation reminders.

Guardrails

• Do not rewrite history to protect a role — blame systems (contracts, tools, handoffs).
• Do not dissolve the org or recreate agent__createOrg as "improvement".
• Do not apply every idea — prefer reversible, measurable changes.
• Do not use LESSONS to paper over broken tools.
• Do not edit workspace agents.md as the default learning path.

References

• Postmortem template
• Improvement routing matrix
• LESSONS.active format