gm-skill 2.0.1480 → 2.0.1482

package/AGENTS.md

@@ -50,6 +50,8 @@ Only record non-obvious technical caveats that cost multiple runs to discover. R
 
 **Detail-heavy caveats live in rs-learn (`.gm/rs-learn.db`), not here.** Per-crate runtime quirks, Windows process-spawn mechanics, hook implementation details, ocw/site/workflow specifics, and similar fact-base material are exfiltrated to rs-learn and reachable via `exec:recall`. AGENTS.md keeps only top-level rules that govern gm-the-repo. When in doubt: gm-the-repo architecture or cross-cutting policy stays here; single-crate or single-platform mechanism goes to rs-learn.
 
+**Every memorize run also drains AGENTS.md — migration is bidirectional, deflation is the back-pressure.** AGENTS.md grows monotonically if facts only flow in; left unchecked it bloats past the context budget it protects. So every session that dispatches `memorize-fire` for new facts ALSO picks a few existing AGENTS.md entries that have become detail-heavy / single-crate / single-platform (the material this policy says belongs in rs-learn) and exfiltrates them: `memorize-fire` the entry's substance to the default namespace, then delete or compress its AGENTS.md paragraph to a one-line pointer in the same commit. Witnessed by both the store gaining the fact (recallable next turn) and the AGENTS.md byte-count dropping. A few entries per run, never a wholesale rewrite; top-level cross-cutting rules stay, everything reachable by `recall` drains. Skipping the drain on a memorize run is the slow-bloat drift this policy exists to prevent.
+
 ## Coding Style
 
 **No comments in code.** No inline, block, or JSDoc comments in every possible location, source, generated output, hooks, scripts.
@@ -112,7 +114,7 @@ Every possible skill's `allowed-tools:` frontmatter is reduced to `Skill, Read,
 
 **The agent IS the LLM rs-learn calls; every judgment rs-learn needs is the agent deciding on the spot**: rs-learn never reaches out to a model for a quality score, a relevance call, a prune decision, a route outcome, or a loss signal. plugkit IS the harness and the agent IS the model it dispatched, so each of those is a decision the agent makes inline, from its knowledge of the current situation, and reports back through the spool. The MicroLoRA adapter trains from a trajectory-quality the agent self-reports (`learn{feedback, embedding, payload:{quality}}`, a rank-2 learned scoring head that reads a bge-small embedding and emits a per-target logit `B·(A·embedding)`, sona-style — it scores targets from embeddings, it does not reshape the vector; consuming its score in recall re-ranking is the open integration. Distinct from the FastGRNN model-selection router, which takes `learn{record_outcome, target:<model id>, quality}`); bad-memory pruning is the agent judging a recall hit stale and dispatching `memorize-prune{key}`; the deep core takes the agent's `record_loss`; the attention takes `nudge_relation`. None of these wait on a signal the host must expose, because the agent already holds it. Encourage heavy `recall` and `learn` use so the on-the-spot judgments are grounded in prior context, not guessed — every mutable resolution is already a memorization run, and the same recall-first reflex should inform every quality/prune/optimization call. The instruction prose names where each self-report fires (VERIFY closes the training loop); the principle here is the load-bearing one: there is no separate judge model, the agent is it, and it decides now.
 
-**host_exec_js is synchronous**: wasm host runs `exec_js` via Node `child_process.spawnSync`; long subprocesses block the watcher. Pass a real per-call timeout; orphaned background work unsupported under wasm.
+**host_exec_js is synchronous**: pass a real per-call `timeoutMs` (zero/missing is a hard error); long subprocesses block the watcher; no async/background exec under wasm. Mechanism detail in rs-learn (`recall: host_exec_js synchronous`).
 
 **Sync-before-emit (codeinsight + search)**: outputs must come from freshly-completed indices. Cache serves only on digest match (mtime sum + git HEAD + dirty-tree marker). Default invocation runs fresh. `--read-cache` permitted only when `.codeinsight.digest` matches; mismatch auto-refreshes. rs-search runs scan + embed + sweep before first result; emits `[index fully synced: …]`. Unverified-index emit = stale ground truth.
 
@@ -120,11 +122,11 @@ Every possible skill's `allowed-tools:` frontmatter is reduced to `Skill, Read,
 
 **Skill SKILL.md frontmatter `allowed-tools:` is harness-enforced**: If a skill omits `allowed-tools` or does not list `Skill`, the model loses the ability to invoke downstream skills that turn. The shipped surface is a single skill (`gm-skill`); this rule governs every possible future skill that participates in a chain.
 
-**rs-learn observability**: every possible learning-pipeline state change emits a structured `evt: {event, sess, ts, ...}` line via `wasm_dispatch::emit_event` (host_log level 1) into `.gm/exec-spool/.watcher.log` and gm-log/<date>/plugkit.jsonl. Event taxonomy: `embed_fail` (step + error), `embed_init_ok/fail/cached_fail`, `memorize_reject` (reason, text_prefix, namespace), `memorize_embed_rollback` (key, namespace, error), `discipline_sigil_ignored` (sigil in @<name> request routed to default), `table_dropped` (dim-mismatch silent drops), `recall_score_unavailable` (host_vec_search strips score). Recall replies now include `mode` (vector_top_k|fallback_like|kv_query), `namespace`, `derived_query`, and per-hit `score` (null when host elides). gmsniff flags: `--embed-failures`, `--recall-misses`, `--recall-scores`, `--classifier-rejects`, `--memory-leverage`, `--recall-modes`, `--table-drops`, `--discipline-sigil-ignored`. ccsniff `--learning-xref` joins transcript turn windows to rs_learn events by sess (now stamped) + project + time-window. Learning quality is observable, not algorithmic black box.
+**rs-learn observability**: every learning-pipeline state change emits a structured `evt:{event,sess,ts,...}` line into `.gm/exec-spool/.watcher.log` + gm-log; recall replies carry `mode`/`namespace`/`derived_query`/per-hit `score`; gmsniff/ccsniff expose the taxonomy. Learning quality is observable, not a black box. Full event taxonomy + flag list in rs-learn (`recall: rs-learn observability taxonomy`).
 
-**SKILL.md auto-refresh**: every possible bootstrap call (`bootstrapPlugkit`) compares the sha256 of the bundled `gm-skill/skills/gm-skill/SKILL.md` (shipped inside the npm package) against the installed copies at `~/.agents/skills/gm-skill/SKILL.md` and `~/.claude/skills/gm-skill/SKILL.md`. Hash mismatch triggers atomic write (`.tmp` + rename) of both targets so the agent sees the latest prose on next session, no manual reinstall needed. Logged to `bootstrap.jsonl` as `SKILL.md refreshed`. The bundled SKILL.md is the source of truth; reinstalling gm-skill only matters when the npm package itself changes, which the cascade pipeline guarantees on every possible plugkit version bump.
+**SKILL.md auto-refresh**: `bootstrapPlugkit` sha256-compares the bundled SKILL.md against installed copies and atomically rewrites on mismatch, so the agent always sees the latest prose. Mechanism detail in rs-learn (`recall: SKILL.md auto-refresh`).
 
-**Skill-initiated bootstrap contract**: `lib/skill-bootstrap.js` performs wasm initialization for skill-driven dispatch without hook infrastructure. `bootstrapPlugkit(sessionId)` accepts optional SESSION_ID, ensures the wasm artifact and `plugkit-wasm-wrapper.js` are in place, writes status/error to `.gm/exec-spool/.bootstrap-status.json` and `.bootstrap-error.json` for spool awareness, and returns `{ ok: true }` on success or `{ ok: false, error: message }` on failure. Failures are non-fatal, callers fall back to a degraded surface.
+**Skill-initiated bootstrap contract**: `lib/skill-bootstrap.js::bootstrapPlugkit(sessionId)` initializes wasm for skill-driven dispatch without hooks; failures are non-fatal (degraded fallback). Mechanism detail in rs-learn (`recall: skill-initiated bootstrap contract`).
 
 ## Cascade pipeline
 

package/gm-plugkit/package.json

@@ -1,6 +1,6 @@
 {
   "name": "gm-plugkit",
-  "version": "2.0.1480",
+  "version": "2.0.1482",
   "description": "Bootstrap and daemon-spawn tool for gm plugkit binary. Downloads the correct platform binary, verifies SHA256, and starts the spool watcher daemon. Includes plugkit-wasm-wrapper for WASM-based spool watching.",
   "main": "index.js",
   "bin": {

package/gm.json

@@ -1,6 +1,6 @@
 {
   "name": "gm",
-  "version": "2.0.1480",
+  "version": "2.0.1482",
   "description": "Spool-dispatch orchestration engine with unified state machine, skills, and automated git enforcement",
   "author": "AnEntrypoint",
   "license": "MIT",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "gm-skill",
-  "version": "2.0.1480",
+  "version": "2.0.1482",
   "description": "Canonical universal harness — AI-native software engineering via skill-driven orchestration; bootstraps plugkit for task execution and session isolation. Install in any AI coding agent host.",
   "author": "AnEntrypoint",
   "license": "MIT",