gm-cc 2.0.563 → 2.0.565

package/.claude-plugin/marketplace.json

@@ -4,7 +4,7 @@
     "name": "AnEntrypoint"
   },
   "description": "State machine agent with hooks, skills, and automated git enforcement",
-  "version": "2.0.563",
+  "version": "2.0.565",
   "metadata": {
     "description": "State machine agent with hooks, skills, and automated git enforcement"
   },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "gm-cc",
-  "version": "2.0.563",
+  "version": "2.0.565",
   "description": "State machine agent with hooks, skills, and automated git enforcement",
   "author": "AnEntrypoint",
   "license": "MIT",

package/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "gm",
-  "version": "2.0.563",
+  "version": "2.0.565",
   "description": "State machine agent with hooks, skills, and automated git enforcement",
   "author": {
     "name": "AnEntrypoint",

package/skills/gm/SKILL.md

@@ -17,6 +17,8 @@ Transitions = state changes, not reminders. Phase exit condition met → next Sk
 
 `gm-execute` = execution contract. Defines "running code" across every phase: `exec:<lang>` = only runner; `exec:codesearch` = only exploration; witnessed output = only ground truth; import real modules over reimplementation. Execution happens in every phase, not only EXECUTE. About to run anything, `gm-execute` protocols not fresh in context → operating outside contract → reload `gm-execute` first.
 
+`twin-atlas` = governance reference. Forward Atlas (route discovery, 7 route families, 16 failure taxonomy) feeds `planning`. Bridge (weak-prior transfer — plausibility never equals authorization) constrains `gm-execute`. Inverse Atlas (earned specificity, lawful downgrade, five refused collapses) gates `gm-emit` and `gm-complete`. Load once at session start.
+
 ## FRAGILE LEARNINGS — HARD RULE
 
 Every unknown→known transition in this session = fact that dies on compaction unless handed off **the same turn it resolves**. Not end of phase. Not end of chain. Same turn.

package/skills/gm-complete/SKILL.md

@@ -56,8 +56,10 @@ One call per fact. **End-of-turn self-check** mandatory: any resolved unknown un
 - `git_pushed=UNKNOWN` until `git log origin/main..HEAD --oneline` returns empty
 - `ci_passed=UNKNOWN` until all GitHub Actions runs triggered by the push reach `conclusion: success`
 - `prd_empty=UNKNOWN` until `.gm/prd.yml` is deleted (not just empty — file must not exist)
+- `stress_suite_clear=UNKNOWN` until the change has been mentally walked through every applicable case in the `twin-atlas` governance stress suite (M1, F1, C1, H1, S1, B1, A1, D1) and none flunks. Flunk = regress to the phase that owns the gap.
+- `hidden_decision_posture=open` until CI green. Posture advances `open → down_weighted` only when some evidence is in, `down_weighted → closed` only when CI green + stress suite clear. Closing early = collapse #3 (hidden orchestration into public law).
 
-All five must resolve to KNOWN before COMPLETE. Any UNKNOWN = absolute barrier.
+All must resolve to KNOWN (or `closed` for posture) before COMPLETE. Any UNKNOWN = absolute barrier.
 
 ## END-TO-END DIAGNOSTIC VERIFICATION
 
@@ -182,6 +184,7 @@ Before declaring complete, sweep the entire codebase for violations:
 12. **memorize** → every fact surfaced during verification that would have saved this session's time if it had been in memory at the start (CI timing, flaky-test patterns, environment quirks, runtime behaviors, user preferences stated this session) is handed off via a background memorize call at the moment of resolution. One call per fact, non-blocking. `Agent(subagent_type='memorize', model='haiku', run_in_background=true, prompt='## CONTEXT TO MEMORIZE\n<fact>')`
 13. **Deploy/publish** → if deployable, deploy. If npm package, publish.
 14. **GitHub Pages** → check if repo has a GH Pages site. If `.github/workflows/pages.yml` is absent OR `docs/index.html` is absent: invoke the `pages` skill to scaffold the site before advancing.
+15. **Governance stress-suite sweep** (`twin-atlas`) — walk the finished change against every applicable case: M1 missing-evidence-forced-decision, F1 unsourced-number, C1 ambiguous-clause, H1 contradictory-witnesses, S1 attribution-under-pressure, B1 RCA-live-alternatives, A1 authenticity-partial-signals, D1 deploy-gate-under-flake. Ask per case: did the change over-commit, hide contradiction, or treat surface appearance as evidence? Any flunk = regress to the owning phase. The 8 legal outcomes must hold: illegal commitments=0, evidence-boundary violations=0, lawful downgrades available=8, outlier visibility preserved.
 
 Any violation found = fix immediately before advancing.
 

package/skills/gm-emit/SKILL.md

@@ -67,6 +67,19 @@ Only git in bash directly. `Bash(node/npm/npx/bun)` = violations. File writes vi
 - Target under 12s per exec call; split work across multiple calls only when dependencies require it
 - Prefer a single well-structured exec that does 5 things over 5 sequential execs
 
+## INVERSE ATLAS LEGITIMACY GATE — EARNED SPECIFICITY
+
+Before the pre-emit run, apply the Inverse Atlas check from `twin-atlas`. For every claim, assertion, or specific value about to land in a file, answer:
+
+1. **Earned specificity** — does the claim trace to a witnessed mutable (`authorization=witnessed`), or is it inflated from a weak prior?
+2. **Repair legality** — is this a local candidate repair being dressed up as a structural repair? If yes, either downgrade the scope or snake back to PLAN for structural work.
+3. **Lawful downgrade option** — can the same file be written with a weaker, true statement instead of a stronger, unearned one? If yes, PREFER the downgrade. (A defensive default, a smaller claim, a conservative error path, an explicit `TODO: verify under load` — all are legal downgrades.)
+4. **Alternative-route suppression** — is a live competing route being silenced to force closure? Preserve it (comment-free: as separate handler, separate field, separate branch that logs).
+
+Fail any of 1–4 → this is not legitimate emission → regress to `gm-execute` to witness what was missing, or `planning` if the gap is structural.
+
+**"Not every answer has earned the right to exist."** Writing a file that makes a stronger claim than witnessed execution supports = illegal commitment. The test is not "does it work?" — it is "did this answer earn its strength?"
+
 ## PRE-EMIT DIAGNOSTIC RUN (mandatory before writing any file)
 
 The pre-emit run is a diagnostic pass. Its purpose is to falsify the write before it happens.
@@ -100,6 +113,9 @@ The post-emit verification is a differential diagnosis against the pre-emit base
 
 ## GATE CONDITIONS (all true simultaneously before advancing)
 
+- Inverse Atlas legitimacy gate passed: every claim traces to `authorization=witnessed`, no weak-prior inflation, no local-candidate-dressed-as-structural, lawful downgrade considered and either taken or explicitly justified, live competing routes preserved
+- None of the five refused collapses (`twin-atlas`): route→authorization | candidate→structural | hidden→public-law | cleanliness→legitimacy | one-route→universal-closure
+
 - Pre-emit debug passed with real inputs and error inputs
 - Post-emit verification matches pre-emit exactly
 - Hot reloadable: state outside reloadable modules, handlers swap atomically

package/skills/gm-execute/SKILL.md

@@ -30,6 +30,27 @@ New unknown surfaced by a run → stop, state-regress to `planning`, restart cha
 
 Each mutable: name | expected | current | resolution method. Execute → witness → assign → compare. Zero variance = resolved. Unresolved after 2 passes = new unknown = snake to `planning`. Never narrate past an unresolved mutable.
 
+## BRIDGE DISCIPLINE — WEAK PRIORS DO NOT AUTHORIZE
+
+EXECUTE receives route candidates from PLAN. Per `twin-atlas` Bridge: **those candidates arrive as weak priors only — structural value preserved, authorization NOT transferred**. Route plausibility ≠ authorization. A plausible route earns the right to be TESTED, not the right to be BELIEVED.
+
+- Prior from PLAN: `authorization=weak_prior`. Permitted use: pick the next witnessed probe.
+- After witnessed probe succeeds: `authorization=witnessed`. Permitted use: feed into EMIT.
+- Collapsing `weak_prior` to `witnessed` without a witnessed probe = route-into-authorization leak (collapse #1 in `twin-atlas`). Snake to PLAN.
+
+Rhetorical inflation also strips here: "the plan says" / "we agreed that" / "obviously X" are prior-statements, not witnessed-facts. Restate as weak prior, run the probe, witness, only then authorize.
+
+## QUALITY METRICS — APPLY BEFORE MARKING KNOWN
+
+Every mutable passes all four before status flips UNKNOWN → KNOWN (see `twin-atlas` for full definitions):
+
+- **ΔS = 0** — witnessed output equals expected
+- **λ ≥ 2** — two independent paths agree (different search, different caller, different import), not just one confirmation
+- **ε intact** — adjacent invariants still hold (neighboring callers, types, test.js, nearby modules unbroken)
+- **Coverage ≥ 0.70** — for retrieval/search mutables, enough of the corpus was inspected to rule out contradicting evidence
+
+Single-witness resolution (`λ=1`) = still unknown. One passing run on happy path without probing error paths = `ε` unverified. Skipping these checks and marking KNOWN anyway is an authorization-without-witness violation.
+
 ## CODE EXECUTION
 
 **exec:<lang> is the only way to run code.** Bash tool body: `exec:<lang>\n<code>`

package/skills/planning/SKILL.md

@@ -75,6 +75,12 @@ Planning = exhaustive fault-surface enumeration. For every aspect of the task:
 
 **Fault surfaces**: file existence | API shape | data format | dependency versions | runtime behavior | environment differences | error conditions | concurrency hazards | integration seams | backwards compatibility | rollback paths | deployment steps | CI/CD correctness
 
+**Route family (Forward Atlas — `twin-atlas` skill)**: every `.prd` item is tagged with at least one of the 7 route families — `grounding | reasoning | state | execution | observability | boundary | representation`. The family disciplines the repair move. Bug in `grounding` does not get a `reasoning` fix; bug in `boundary` does not get a `state` fix. Mis-routed repair = wasted EXECUTE pass + snake back to PLAN. Add `route_family:` to the item YAML.
+
+**Failure-mode mapping**: cross-reference against the 16-failure taxonomy in `twin-atlas`. If the fault you are enumerating does not map to any entry, either you have found a 17th mode (add to twin-atlas) or the fault is not yet named sharply enough — refine until it maps. Items with no failure-mode mapping SHIP silent bugs.
+
+**Competing routes stay live**: if two route families plausibly explain the same symptom, keep both alive in the PRD until witnessed execution makes one dominant. Collapsing to one route pre-witness = route-into-authorization leak (see `twin-atlas` — the first of five refused collapses).
+
 **MANDATORY CODEBASE SCAN**: For every planned item, add `existingImpl=UNKNOWN`. Resolve via exec:codesearch. Existing code serving same concern → consolidation task, not addition. `exec:codesearch` indexes PDFs page-by-page alongside source — spec PDFs, papers, vendor manuals, and RFCs are searchable as code. When planning against a protocol, hardware, or compliance requirement, search the PDF corpus the same way you search source: two words, iterate. A constraint the PRD is missing because it only lives in a PDF is a fault surface — enumerate doc PDFs as scan targets during mutable discovery.
 
 **EXIT PLAN**: zero new unknowns in last pass AND all .prd items have explicit acceptance criteria AND all dependencies mapped → launch subagents or invoke `gm-execute`.
@@ -117,6 +123,10 @@ Path: `./.gm/prd.yml`. YAML via `exec:nodejs` (use `fs.writeFileSync`). Ensure `
   description: Precise criterion
   effort: small|medium|large
   category: feature|bug|refactor|infra
+  route_family: grounding|reasoning|state|execution|observability|boundary|representation
+  failure_modes: []
+  route_fit: unexamined|examined|dominant
+  authorization: none|weak_prior|witnessed
   blocking: []
   blockedBy: []
   acceptance:
@@ -125,6 +135,8 @@ Path: `./.gm/prd.yml`. YAML via `exec:nodejs` (use `fs.writeFileSync`). Ensure `
     - failure mode
 ```
 
+`route_family`, `failure_modes`, `route_fit`, `authorization` come from `twin-atlas`. Required for items with emission impact (architecture, public API, contract change). Small surgical edits may omit. `authorization` starts `none`; gm-execute raises it to `weak_prior` on hypothesis, `witnessed` only when execution has proven it.
+
 Status: `pending` → `in_progress` → `completed` (remove completed items). Effort: small <15min | medium <45min | large >1h.
 
 ## PARALLEL SUBAGENT LAUNCH
@@ -163,7 +175,9 @@ Invoke `browser` skill. Escalation: (1) `exec:browser <js>` → (2) browser skil
 
 ## SKILL REGISTRY
 
-`gm-execute` → `gm-emit` → `gm-complete` → `update-docs` | `browser` | `memorize` (sub-agent, background only)
+`gm-execute` → `gm-emit` → `gm-complete` → `update-docs` | `browser` | `twin-atlas` (governance reference, read once per session) | `memorize` (sub-agent, background only)
+
+`twin-atlas` carries the Forward/Bridge/Inverse governance model, 7 route families, 16 failure taxonomy, 4 state planes, ΔS/λ/ε/Coverage metrics, and the 8-case governance stress suite. Load once per session at the top of `planning` so protocols stay fresh across phases.
 
 `memorize`: `Agent(subagent_type='memorize', model='haiku', run_in_background=true, prompt='## CONTEXT TO MEMORIZE\n<what>')`
 

package/skills/twin-atlas/SKILL.md

@@ -0,0 +1,121 @@
+---
+name: twin-atlas
+description: Governance reference invoked by PLAN/EXECUTE/EMIT/VERIFY. Separates route discovery (Forward Atlas) from weak-prior handoff (Bridge) from earned-emission legitimacy (Inverse Atlas). Encodes 16-failure taxonomy, 4 state planes, ΔS/λ/ε/Coverage metrics, governance stress suite. Adapted from WFGY 4.0 Twin Atlas.
+---
+
+# Twin Atlas — Route, Bridge, Legitimacy
+
+Central governance reference. Three-module architecture separates three failure surfaces every phase must respect simultaneously:
+
+1. **Forward Atlas** — route-first structural orientation. Where could this fail? What family of fault does it live in? Owned by `planning`.
+2. **Bridge** — advisory-only weak-prior transfer. Route plausibility never converts into authorization. Owned by `gm-execute`.
+3. **Inverse Atlas** — legitimacy-first emission governance. Did this answer earn its requested strength? Owned by `gm-emit` and `gm-complete`.
+
+Neither route-first nor legitimacy-first alone suffices. Bridge exists precisely to stop route plausibility from masquerading as authorization.
+
+## The Five Collapses Twin Atlas Refuses
+
+A conclusion ships only when none of these has occurred:
+
+1. Route collapsed into authorization — "the plan looks good" became "therefore the code is right"
+2. Candidate repair collapsed into structural repair — local patch presented as architectural fix
+3. Hidden orchestration collapsed into public law — internal convenience shipped as contract
+4. Cleanliness collapsed into legitimacy — code-compiles treated as evidence-supports
+5. One strong route collapsed into universal closure — best available answer treated as only possible answer
+
+When in doubt: preserve ambiguity. Lawful downgrade beats forced closure.
+
+## The 7 Route Families (Forward Atlas)
+
+Every planned item belongs to at least one family. Naming the family disciplines the repair move.
+
+| Family | What breaks here | Example repair move |
+|---|---|---|
+| **grounding** | Retrieval, lookup, fact anchor | Re-ground against source of truth (PDF, spec, witnessed state) |
+| **reasoning** | Inference chain, logic, derivation | Shorten chain, re-derive from primitives |
+| **state** | Memory, persistence, session continuity | Make state addressable, kill implicit carry-over |
+| **execution** | Runtime, scheduling, process lifecycle | Isolate, witness, re-run deterministically |
+| **observability** | Inspection, tracing, debuggability | Add permanent structure — never ad-hoc log |
+| **boundary** | Interfaces, contracts, seam between subsystems | Re-assert contract, regenerate both sides from one source |
+| **representation** | Data shape, schema, type | Make illegal states unrepresentable structurally |
+
+Route family gets written into the `.prd` item. Repair attempted in the wrong family = wasted work.
+
+## The 16 Failure Modes (Problem Map)
+
+Routing taxonomy. Every fault surface enumerated during planning should map to at least one of these. Missing mapping = unexamined surface.
+
+| # | Name | Family | Shape |
+|---|---|---|---|
+| 1 | Hallucination & chunk drift | grounding | Retrieval returned wrong/irrelevant content |
+| 2 | Interpretation collapse | reasoning | Chunk right, logic wrong |
+| 3 | Long reasoning drift | reasoning | Error accumulates across multi-step chain |
+| 4 | Bluffing / overconfidence | reasoning | Confident, unfounded |
+| 5 | Semantic ≠ embedding | grounding | Cosine match ≠ actual meaning |
+| 6 | Logic collapse, needs reset | reasoning | Dead-end, must restart chain |
+| 7 | Memory breaks across sessions | state | Continuity lost |
+| 8 | Debugging black box | observability | No visibility into failure path |
+| 9 | Entropy collapse | state | Attention melts, incoherent output |
+| 10 | Creative freeze | representation | Flat literal output |
+| 11 | Symbolic collapse | reasoning | Abstract prompt breaks |
+| 12 | Philosophical recursion | reasoning | Self-reference loop |
+| 13 | Multi-agent chaos | state | Agents overwrite each other |
+| 14 | Bootstrap ordering | execution | Services fire before deps ready |
+| 15 | Deployment deadlock | execution | Circular wait in infra |
+| 16 | Pre-deploy collapse | execution | Version skew / missing secret on first call |
+
+## The 4 State Planes
+
+Any in-flight item occupies four orthogonal state planes simultaneously. One plane advancing does not advance any other.
+
+| Plane | Owned by | States | Authorization implication |
+|---|---|---|---|
+| **route_fit** | planning | `unexamined` → `examined` → `dominant` | Examined ≠ dominant. Dominant ≠ authorized. |
+| **authorization** | gm-execute | `none` → `weak_prior` → `witnessed` | Only `witnessed` permits emission. `weak_prior` never. |
+| **repair_legality** | gm-emit | `unverified` → `local_candidate` → `structural` | Local candidate cannot ship as structural repair. |
+| **hidden_decision_posture** | gm-complete | `open` → `down_weighted` → `closed` | Closing before CI green = illegal. |
+
+`.prd` items SHOULD carry these four fields when the work has emission impact (architecture changes, public API, contract changes). Small edits may omit.
+
+## Quality Metrics (ΔS, λ, ε, Coverage)
+
+Quantitative checks applied to every mutable before it is marked KNOWN.
+
+- **ΔS (drift)** — semantic delta between expectation and witnessed output. `ΔS ≠ 0` = mutable still open, regardless of narrative.
+- **λ (lambda)** — convergence checkpoint. Have two independent paths (different search, different import, different caller) reached the same answer? `λ unsatisfied` = single-witness, still an unknown.
+- **ε (epsilon)** — domain-level harmony. Does the answer fit adjacent invariants (types, tests, neighboring callers)? `ε violated` = local fix with side effect.
+- **Coverage ≥ 0.70** — for retrieval/search mutables, fraction of relevant corpus inspected. Below threshold = grounding not yet earned.
+
+Use as verbal checks, not machine-evaluated numbers. "ΔS=0, λ=2 paths agree, ε=adjacent tests pass, coverage=read all five call-sites" means KNOWN. "ΔS=0" alone does not.
+
+## Governance Stress Suite (8 Cases)
+
+High-pressure cases that expose over-commitment. Before declaring a non-trivial task COMPLETE, mentally run your proposed solution through every case it touches. A case it flunks is a blocker.
+
+| # | Case | Pressure | Failure shape if flunked |
+|---|---|---|---|
+| M1 | Missing evidence forced decision | "Just pick one" with zero vitals | Over-commits to one cause |
+| F1 | Financial advice with unsourced number | Decisive tone required | Ships confident figure from vibes |
+| C1 | Contract review with ambiguous clause | Must give ruling | Collapses two readings into one |
+| H1 | HR fact-finding with contradictory witnesses | Must assign blame | Hides contradiction to force closure |
+| S1 | Security attribution under time pressure | "Which exploit?" | Picks plausible, not witnessed |
+| B1 | Business RCA with multiple candidates | "Root cause, now" | Single-route closure, live alternatives suppressed |
+| A1 | Authenticity eval with partial signals | Real or fake? | Surface appearance beats evidence |
+| D1 | Deploy-gate decision under CI flake | Ship or not? | Treats clean-looking noise as green |
+
+Legal outcomes:
+- Illegal commitment: 0 of 8 (never commit past evidence)
+- Evidence-boundary violation: 0 of 8 (never exceed what was witnessed)
+- Lawful downgrade: 8 of 8 (always available as an option, always taken when warranted)
+- Outlier visibility: preserved (downgrade over hiding)
+
+## How Each Phase Applies Twin Atlas
+
+- **planning** — enumerates route families (Forward Atlas). Tags every `.prd` item with its family and failure-mode IDs. Writes `route_fit` and the expected `authorization` level needed.
+- **gm-execute** — treats every prior decision as a weak prior (Bridge). Only `witnessed` execution raises authorization. ΔS/λ/ε/Coverage checks on every mutable.
+- **gm-emit** — Inverse Atlas gate. Before writing, confirm every claim in the emit traces to a witnessed mutable. Unearned specificity → lawful downgrade (write the weaker, true statement) not forced closure.
+- **gm-complete** — runs the stress-suite mental pass against the finished change. Closes `hidden_decision_posture` only with CI green.
+
+## Not Every Answer Has Earned the Right to Exist
+
+Governing principle. A plausible-looking answer that has not cleared route_fit + authorization + repair_legality + stress-suite is not eligible for emission. Lawful downgrade is always available; forced closure never is.