coalesce-labs/ticket-compound

ticket-compound — engineering compound loop

Capture what a ticket taught us into the shared store (thoughts/ + ADRs), so future agents on any machine make better decisions. This is research-curate evolved from inventory to action. Read reference.md (this dir) for the learnings-store schema before writing anything.

Two authority levels (hard rule):

• Autonomous — write/append/update/delete in thoughts/shared/learnings/ and thoughts/shared/CONCEPTS.md, and prune stale notes in thoughts/shared/{research,plans}/.
• Propose only (APPROVE-gated) — any change to docs/adrs.md. Never edit an ADR directly; queue it for the morning ritual to approve.

Invocation

/catalyst-dev:ticket-compound <TICKET> [mode:headless]

• <TICKET> — Linear key (e.g. CTL-619). If omitted, detect from the branch / CATALYST_TICKET.
• mode:headless — non-interactive: apply all unambiguous autonomous actions silently, mark ambiguous learnings status: stale, never block on a prompt, end with the sentinel line. This is what the morning ritual (and, later, the daemon) use. Default is interactive.

Step 1 — Gather raw signal (the orchestrator reads; do NOT delegate writes)

For <TICKET>, collect:

1. Friction — every ## Friction / friction: block the phase agents left in their artifacts: thoughts/shared/{research,plans}/*<TICKET>*.md and the worker signal files ~/catalyst/workers/<TICKET>/*.json.
   • thoughts/shared/friction/<TICKET>.md — the dedicated per-phase friction log (primary friction source).
2. The diff — git log --oneline origin/main..HEAD and git diff --stat origin/main..HEAD (or the merged SHA from phase-monitor-merge.json).
3. Ticket — read via direct SQL against the replica (title, description, final state, estimate); see the linearis skill's "Reading Linear" section.
4. Event trail (optional) — the ticket's lines in ~/catalyst/events/YYYY-MM.jsonl.

Capture learnings from failed/abandoned tickets too — the dead-ends are high-signal ("what didn't work" is a first-class section).

Step 2 — Three TEXT-ONLY sub-agents (parallel; they return text, never write files)

Spawn via Task, all at once. Each returns text to you; you do the single write in Step 3 (avoids partial-write races):

• Context Analyzer — from the diff + ticket + friction, pick the track (bug vs knowledge), problem_type, category (subdir), component, severity, and a filename slug. Returns the frontmatter skeleton (validate against reference.md).
• Solution Extractor — write the entry body per the track's template (bug: Problem/Symptoms/What Didn't Work/Solution/Why This Works/Prevention; knowledge: Context/Guidance/Why This Matters/When to Apply/Examples). Ground every claim in the diff/friction — no invention.
• Related-Learnings Finder — rg -li "<keywords>" thoughts/shared/learnings/**/*.md, read the frontmatter of hits, and score overlap (problem, root cause, component, files, prevention). Returns HIGH (4–5) / MODERATE (2–3) / LOW (0–1) plus the matched paths.

Step 3 — Assemble + write ONE learnings entry (overlap routing)

• HIGH overlap → update the existing entry in place (merge new detail, set last_updated:), rather than creating a near-duplicate. This is the anti-bloat rule.
• MODERATE/LOW → create thoughts/shared/learnings/<category>/<slug>.md.

Then validate frontmatter (fail loud on the YAML traps in reference.md):

bash "${CLAUDE_PLUGIN_ROOT:-plugins/dev}/scripts/compound/validate-learnings.sh" "<written-path>"

Step 4 — Curate the store (five outcomes, autonomous in thoughts/)

For the related entries the Finder surfaced, classify and act — autonomously (this is the thoughts/ layer): | Outcome | When | Action | |---|---|---| | Keep | accurate, refs valid | none | | Update | core correct, refs/paths drifted | targeted in-place edit | | Consolidate | 2+ heavily overlap, both correct | merge into the canonical, delete the subsumed | | Replace | core guidance now misleading | write successor, delete old | | Delete | implementation gone AND domain gone AND no inbound links | remove (git history preserves) |

Same five-outcome pass applies to stale thoughts/shared/{research,plans}/ notes this ticket contradicted (the "off-track context that got us redirected"). In mode:headless, only act on unambiguous cases; mark the rest status: stale + stale_reason.

Step 5 — Vocabulary → CONCEPTS.md (autonomous)

Scan the diff + friction for Catalyst domain terms not yet in thoughts/shared/CONCEPTS.md (e.g. "reclaim", "revive-budget", "orphan", "signal ownership"). Append concise definitions. Create the file if absent.

Step 6 — ADR changes → PROPOSE, never apply

If a learning rises to a standing rule (something every agent must always do/avoid), do NOT edit docs/adrs.md. Append a proposal to the approval queue:

thoughts/shared/compound/pending/<TICKET>.md

Each proposal: the target ADR (new / amend ADR-NNN / supersede ADR-NNN), the exact proposed text, and a one-line rationale + evidence (ticket + learning path). The morning ritual surfaces these; a human approves via briefing-followup's action-compound handler, which is the only thing that writes docs/adrs.md.

Step 7 — Discoverability check (pointer only)

Confirm CLAUDE.md teaches agents that the learnings store exists, its shape, and when to grep it. If not, propose (interactive) / apply (headless) a minimal pointer — never the learnings themselves:

thoughts/shared/learnings/ — past problem→solution entries (grep by component/tags/problem_type).
Search before implementing or debugging in a known area. Curated by /catalyst-dev:ticket-compound.

Step 8 — Report

• Interactive: summarize entry written/updated, curation actions, CONCEPTS additions, and any queued ADR proposals (with the approval command).
• Headless: structured block ending with the grep-able sentinel:

✓ ticket-compound complete (headless)
ticket: CTL-619
entry: thoughts/shared/learnings/orchestrator-issues/daemon-false-dead-first-commit.md (created)
curated: 1 updated, 0 deleted ; concepts: +2 ; adr-proposals: 1 (pending approval)
ticket-compound complete

Out of scope (this slice)

• The daemon firing this automatically after monitor-deploy (manual / morning-ritual triggered for now).
• The estimation loop (separate slice — compound-estimate owns estimation numbers).
• ticket-retro (the cross-ticket view — separate slice).