weave

/weave

Version: v1.3

At convergence, fan out deep cross-session mining, then prove each finding turned into a real change — or kill it.

/weave is an orchestrator-only convergence skill. Workers do not self-invoke it. It mines across recent Claude and Codex sessions, verifies the findings, and routes every finding through an action ledger. The ledger's strict conversion-to-change metric is the anti-waste spine: fan-out without disposition and verified change is not convergence.

Non-negotiable law

• A weave run uses the formal skill through skillcreator; no ad-hoc "lite weave."
• The harness, high-importance routed conclusions, conversion metrics, registry transitions, and retro are committed artifacts. Scratch digests and findings may be regenerated; conclusions may not live only in docs.local/.
• Gather runs off-Claude; Claude performs judgment. /agent-routing references/model-and-effort.md § Decide From the Mission owns model/effort choice; quota-probe before launch, preserve worker lineage. For Cursor, use T3 delegate_task with target: {providerInstanceId: "cursor", model: "default"} (Auto). One pointer brief per child; each reports to the collab. No AutoCursor, headless CLI fan-out or background Cursor loops.
• New/manual dispatches give each miner K sessions with a STATIC-first prompt (centerpieces keep a dedicated miner; cache shape), centerpieces first, in bounded waves. The retained legacy cursor runner is not a route for new work (see /cursor-workflows). Expand one flat frontier per depth; never nest parallel() calls.
• Rounds create rounds; stop after 2 non-empty dry rounds, the second a confirmation sweep (or explicit confirmation exhaustion); synthesis last; caps are backstops (§2a). A zero-agent work batch is a coverage defect.
• Disk artifacts outrank agent receipts. Missing miner outputs are re-mined.
• Every finding needs cited evidence and a disposition. "Never landed" claims require live gh/git verification.
• After synthesis, run the mandatory Red/Blue/Red fact-check until dry, apply corrections everywhere, discharge every skill edit, update the rule registry, and let the final boss—not the workflow—decide whether the loop may finish.
• Re-score the prior run against what actually landed. Re-poll terminal state at write time; stamps come from date; every write gets same-turn read-back.
• The successor reads the entire weave, corrections, discharge table, and every BROKEN-OPEN registry row, then ACKs them item by item.
• Keep this frontmatter description at 1024 characters or fewer; an over-long description silently skip-loads the skill (weave's own 1272-char description tripped this). [g1:86 → 2026-05-31.md:69]

Trigger and decision table

Situation	Action
"weave"	Arm the skill, run scripts/convergence-gate.sh with the fleet's mode, and wait. Never auto-fire. In a T3 fleet that means the §1.1 capture, then --fleet t3 --t3-state <file> --t3-projects "<ids>". Inside T3 with cmux on PATH, auto checks both fleets.
"weave now"	Fire immediately, skipping the gate. Operator override — use only when the operator knows the fleet is quiet.
Gate path	Require 0 open fleet PRs, idle and settled workers, no in-flight Codex, RAM headroom, and explicit demo approval. Prefer orc's SENTINEL over raw PR count when present.
T3 fleet (leads/workers are T3 threads or delegate_task children)	Capture the weave-t3-fleet/v1 snapshot (references/run-lifecycle.md §1.1): t3_thread_list for every fleet project, task_status per task, t3_thread_read per child the listing misses, t3_pending_request_list for your own thread. Then run the gate with --fleet t3 --t3-state <file> --t3-projects "<ids>". It passes only when convergence is proven: any working, waiting or unknown state, missing field, duplicate row, unswept project or missing child blocks. A result_available task whose worker thread is unsettled blocks too: settle it first.
cmux fleet	Run the gate with --fleet cmux; it reads cmux list-status pane idleness.
Gate cannot detect a fleet	BLOCKED. Pass --fleet and its input; never treat "nothing to check" as converged.
Small corpus (about 10 sessions or fewer)	Flat fan-out may be appropriate; still verify on disk and loop until dry.
Large or centerpiece-heavy corpus	Use staged mining, with centerpiece digests first and later rounds seeded by earlier findings.
Finding is a failure	Add evidence-cited cause_layer and route the fix to code, instructions, or tool description.
Skill edit/new candidate	End APPLIED, DISPATCHED with acknowledged owner, or TRACKED with evidence; no fourth state.
Live sessions grow during the run	Run the delta wave and re-aggregate before close.
Roadmap-only model-change tracking	Do not block the run or claim it is implemented.

Phase gates

1. Enter through the fleet-canon-#7 claim/guard contract; /collab-monitor owns the concrete watcher.
2. Converge, discover, prepare, batch, mine, and verify files on disk.
3. Aggregate the ledger; report both strict and refined conversion metrics.
4. Re-score the prior run, red-team every load-bearing claim, and propagate corrections.
5. Discharge every edit, emit verified intent briefs and grill carryover, then run the delta wave.
6. Update the committed retro and rule registry, perform full relay, and exit only through the final boss.

Read the relevant reference

• Running the gate (T3 snapshot schema: §1.1; cmux fleet: §1.2), discovery, mining, dispatch, delta wave, or miner contract? Read references/run-lifecycle.md.
• Aggregating, re-scoring, red-teaming, correcting, discharging edits, or computing correction rates? Read references/verification-and-discharge.md.
• Emitting track intents, self-QA gates, grill carryover, name claims, or monitors? Read references/outputs-and-coordination.md.
• Writing retros, model-change roadmap notes, registry transitions, or checking the artifact inventory? Read references/continuity-and-registry.md.
• Choosing flat versus staged topology, batch width, or centerpiece ordering? Read references/topology.md.

Cross-skill routing

• /skill-creator owns the session-mining engine and skill evaluation.
• /agent-routing § Review routing owns review topology; never start a reviewer yourself.
• /large-plan receives mined track intents; domain leads author the plans.
• /never-fabricate owns evidence and verification discipline.
• /pr-loop owns PR conversion and review handling.
• /orc owns convergence detection and miner dispatch.
• Fleet canon #7 owns lead/orchestrator monitor law; /collab-monitor owns the concrete watcher.