yunzeforbetter/castflow/origin-evolve-skill

Origin Evolve

Turn pending memory snapshots in .castflow-runtime/traces/trace.md into approved skill updates. A project skill is written under castflow-skills/<name>/, including a later update of that skill. Never write a project skill under .castflow-runtime/skills/ (cold start replaces that tree). Never write a factory-owned skill. Never write adapter mirrors. After writes: python .castflow/manager.py sync.

Trigger: origin evolve (or the same intent). Never run unprompted.

Yield

• Install / UI -> castflow.bat / manager.py launch (not a skill)
• Create a new module skill -> the pasted /goal prompt (MODULE_MARK_SYSTEM_PROMPT.md)
• Requirement to a runnable long task -> goal-loop-creator
• Feature work in a module -> that programmer-*-skill

Nav

Trace (schema:4)

Hook-generated ledger. Fields: timestamp, type, validated, quality, gate_hint, memory_snapshots, plus MEMORY subblocks with skill, anchors, quality. Read-only. Ignore legacy score / modules / pipeline_run_id / pending-pipeline.

Eligible (MEMORY grain)

validated does not grant eligibility. It only sorts already-eligible items and counts the TRACE toward notify.

eligible(m) =
  (m.type == feedback AND m.quality == ok)
  OR homologous_count(m) >= 2

homologous_count(m) is the size of m's connected component from python .castflow/manager.py homology (includes self). Singleton size is 1.

Do not invent Jaccard in this skill. Do not spawn one Python process per pair.

Flow

Step 0 Flush if needed -> lock -> Step 1 Triage -> Step 2 Distill -> Step 3 Propose -> Step 4 Approve -> Step 5 Write

Step 0. If .castflow-runtime/traces/.trace_memory_snapshots exists, run python .castflow/manager.py flush before taking .trace_lock. Then create .trace_lock. If the snapshot store still exists, abort (do not call flush while holding the lock).

Step 1. Keep pending only. Do not expire on validated:_. Leftover pending-pipeline -> invalid. Abort if any schema:N with N > 4.

Cluster MEMORY subblocks with one manager.py homology call. Priority among eligible only: P0 TRACE validated:false; P1 feedback+ok; P2 clustered waiting. Skip ineligible.

Step 2. Read each eligible <!-- MEMORY -->. Verify named APIs still exist. Dedup against the target SKILL_MEMORY.

Step 3. Each proposal: Append / Merge / Retire, target skill+file, full text, evidence timestamps + slugs, risk. Proposal sentences use .castflow-runtime/config.json language (missing or en is English, zh is Chinese). A non-programmer entry keeps the Anchors: and Related: labels. A programmer-*-skill entry is a rule One generation pass would keep: an implicit rule, a convention, or a file locator the examples do not already state. Do not write a call shape, a line number, or a signature-gap body into that skill. Update a locator only when its script file is added, removed, renamed, or the feature moves to another file, not because the implementation inside an existing script changed. If SKILL_MEMORY.md is absent, the first real entry creates it. Do not create it empty, and do not invent a rule so the file exists. Capacity 2000 words (SKILL_MEMORY / cross-cutting). Retire a non-programmer entry, or a programmer rule, when grep shows its anchors are gone. Retire a programmer locator when its script file is gone. Attribution: (1) whitelisted skill: field (2) anchors hit exactly one project skill (3) 1 and 2 disagree -> user pick (4) anchors hit >=2 project skills -> .castflow-runtime/rules/cross-cutting.md (5) none -> do not write, leave waiting. Never write .claude/rules/ business rules. Never write GLOBAL, AGENTS.md, CLAUDE.md, hooks, or this skill. A project-wide convention is a suggestion to edit the AGENTS.md project section, not a write.

Step 4. One proposal at a time. Rejection writes EVOLVE_REJECTION.

Step 5. Atomic write of a project skill only, under castflow-skills/<name>/ (create the directory if this is the first file). Do not write .castflow-runtime/skills/<name>/, and do not write a factory-owned name there or in castflow-skills/. A same-named castflow-skills/ directory does not replace a factory skill. Replace analyzed entries with <!-- PROCESSED ts:... entries:N proposals:M -->. Drop .trace_lock in finally. Then python .castflow/manager.py sync.

No score calibration. If snapshots look noisy, tell the user it is a hook/config issue.