@mindrian_os/cli 2.0.0-beta.3 → 2.0.0-beta.31

package/data/harness-manifest.json

@@ -0,0 +1,74 @@
+{
+  "ontology_ref": "data/command-registry.json + data/connector-registry.json + data/brain-orchestration-projection.json",
+  "generated_note": "GENERATED by scripts/build-harness-manifest.cjs - do not edit by hand",
+  "methodology_tier": "mindrian-operation",
+  "version": 2,
+  "maps": [
+    {
+      "role": "posture",
+      "path": "data/command-registry.json",
+      "digest": "b542137a9db903196a4713a6078f76b1a10d43ec161dc22ef46eb2f6ce903b4a",
+      "source_count": 113
+    },
+    {
+      "role": "wiring",
+      "path": "data/connector-registry.json",
+      "digest": "55b998bdbb59b6c367e1fbf52cfaba8b130eb2ccf9dc3f621122af1d80852735",
+      "source_count": 209
+    },
+    {
+      "role": "ranked_next_reach",
+      "path": "data/brain-orchestration-projection.json",
+      "digest": "71e496f203e45bf934634e1801ca85758145a072d3967af452c4efea9573b82f",
+      "source_count": 384
+    }
+  ],
+  "runtime_surfaces": [
+    {
+      "role": "orchestration_spine",
+      "path": "lib/core/chain-executor.cjs",
+      "digest": "ed0593828237f70c1c24f39a84ce04c9bf83c4c34b4216df56b273550ce018a6"
+    },
+    {
+      "role": "decide_engine",
+      "path": "lib/core/navigation-engine.cjs",
+      "digest": "5c00a2662f234534ef983bb558cf292f313feffe3e6ac8404851bb7f42a5ec5b"
+    },
+    {
+      "role": "cockpit",
+      "path": "lib/statusline/cockpit-renderer.cjs",
+      "digest": "f7d6cef6a8c4828dc5e172b9b64e92fa46caf32c57bf9e6e613c9c24be2891ca"
+    },
+    {
+      "role": "brain_orchestration",
+      "path": "lib/core/directive-envelope.cjs",
+      "digest": "ed2451cb75df1ac7b09d5f71b10fc6e5426f182c9339713f49f759fe3a04c81e"
+    }
+  ],
+  "policies": {
+    "path": "data/harness-policies",
+    "digest": "10960f4c9e8f0cdb9f7e322ae3ddadb58c75e86c0c056dd67c5e29d97944c17f",
+    "count": 13
+  },
+  "larry_surfaces": [
+    {
+      "role": "cli_agent",
+      "path": "agents/larry-extended.md",
+      "digest": "902007f12c36bdfd83da2167c07034a44c5af47e3fd77f2599f9b8410c909d9e"
+    },
+    {
+      "role": "personality_skill",
+      "path": "skills/larry-personality/SKILL.md",
+      "digest": "096044c41ad58b4e62129391aac8af6b8eef4f8c1f120ade4b1fe461ba4de04c"
+    },
+    {
+      "role": "desktop_cowork_wire",
+      "path": "lib/mcp/runtime-instructions.cjs",
+      "digest": "f4a1366d21e16d1fcaf5313980cca0d53007724e6f3f7394340b9d36ec9c3f09"
+    }
+  ],
+  "fixture_ref": {
+    "path": "data/harness-fixtures/converged-room",
+    "digest": "6f2635a32872d7e066964da4dacd6f67a0cf7afc9bfde281fe4fb6f2035e0401"
+  }
+}

package/data/harness-policies/CONTEXT.md

@@ -0,0 +1,55 @@
+# data/harness-policies/ - Layer 2 Contract
+
+Phase 298 (SEED-032, Harness-as-Code). This directory declares which policies exist and at
+what enforcement rung. It is a Layer 2 Contract in the ICM sense: a rung is a statement about
+enforcement, not a runtime decision, and it must be diffable in one line.
+
+## Reads
+
+Nothing. This directory is read BY its consumers; it never reads anything itself.
+
+## Does
+
+Declares policies. One JSON file per gate, voice rule, memory-write policy, or contract-parity
+set, each validated against `_schema.json`'s closed vocabulary. A policy file states its `id`,
+`kind`, `runner` (or an honest `null` ghost), `args`, `rung` (`declared | logged | blocking`),
+`evidence_log`, a `promotion_rule` decided before any log is read, `owner`, `pinned_by`,
+`applies_to`, and `notes` explaining the rung choice.
+
+## Writes
+
+Nothing. A policy file is hand-authored and hand-edited; nothing in this directory is
+generated, and nothing here writes to disk. `scripts/build-harness-manifest.cjs` digests the
+directory into `data/harness-manifest.json`'s `policies` key, but the digest is a fingerprint,
+never a copy of contents, and it lives outside this directory.
+
+## Human check
+
+A rung change (declared -> logged, logged -> blocking, or a demotion the other way) is a
+one-line human edit to one policy file's `rung` key, made after reading
+`evaluatePromotion(policy, lines)`'s verdict against the policy's own `promotion_rule`, and
+reviewed like any other diff. No script edits `rung`. No script edits `promotion_rule` either;
+that value is decided once, at authoring time, specifically so the counting rule cannot be
+loosened after the fact to make a verdict pass.
+
+## Change-impact table
+
+Every consumer below must be re-run when a policy file (or `_schema.json`) changes, because
+each one either validates against, digests, or spawns from this directory.
+
+| Consumer | What it does with this directory |
+|---|---|
+| `scripts/build-harness-manifest.cjs --check` | Validates every policy file against `_schema.json` before digesting; an invalid file fails `--check` naming the file and the failing key. Also runs the `contract-*` policies inline (pure file reads: every phrase present in each named surface, the byte budget within limit). |
+| The pre-commit drift guard (`scripts/hooks/pre-commit-room-minto-guard.sh` and its byte-identical twin `scripts/hooks/pre-commit`) | Fires `build-harness-manifest.cjs --check` when a staged path matches its trigger regex, so a policy edit that goes stale in the manifest is caught before commit. |
+| `scripts/run-harness.cjs` | Loads every policy file, executes each one at its declared rung (`declared` = report only, `logged` = spawn and append evidence, never fail; `blocking` = spawn and fail the tier on non-zero exit), and reports a ghost for any policy whose `runner` is `null` or whose named file is missing. |
+| The `harness-policies` doctor acceptance point (`scripts/doctor.cjs`, wired after slice 1 is green) | Spawns `scripts/run-harness.cjs --check --json` and treats a non-zero exit as a release-blocking finding. |
+
+## Migration slices
+
+This phase (298) lands slice 1 only: `gate-worktree-hygiene`, `gate-shape-declaration`,
+`gate-tool-honesty`, `gate-card-fire`, `gate-graph-derive-health`, the four voice policies,
+`memory-write-policy`, and `contract-parity-larry` - eleven files. Slice 2 (the remaining
+`check-*.cjs` gates `verify-release` already runs), slice 3 (the 21 doctor modules, each as a
+`gate-*` policy), and slice 4 (the remaining gates plus the promotion review) are explicitly
+out of scope here and land in later phases, per the migration order in the design doc
+(`docs/superpowers/specs/2026-09-07-harness-manifest-v2-design.md` section 12).

package/data/harness-policies/_schema.json

@@ -0,0 +1,42 @@
+{
+  "_doc": {
+    "purpose": "Phase 298 (SEED-032, Harness-as-Code): the closed policy vocabulary every file in data/harness-policies/ must validate against. Registry-is-the-table: this file IS the schema (the vocabulary enums inline), mirroring data/hitl-shape-declaration-schema.json's own idiom. Read by scripts/build-harness-manifest.cjs at --check time (an invalid policy file fails --check naming the file and the failing key, and the manifest is not regenerated) and by scripts/run-harness.cjs when it loads the policy directory to decide, per policy, whether to report-only, spawn-and-log, or spawn-and-fail-the-tier. This schema DECLARES the vocabulary; it does not execute a policy and does not re-implement the runner.",
+    "rung_vocabulary": ["declared", "logged", "blocking"],
+    "rung_meaning": [
+      { "rung": "declared", "meaning": "listed in the report only. Nothing is spawned. This is the honest default for a policy with no evidence log yet, or a policy that is a declared ghost (runner is null)." },
+      { "rung": "logged", "meaning": "the runner spawns the policy's runner and appends one JSONL line to evidence_log; the finding never fails the tier, regardless of exit code." },
+      { "rung": "blocking", "meaning": "the runner spawns the policy's runner; a non-zero exit fails the tier the policy applies_to." }
+    ],
+    "kind_vocabulary": ["gate", "voice", "memory", "contract"],
+    "kind_meaning": [
+      { "kind": "gate", "meaning": "a check script or doctor module: a deterministic pass/fail scanner over repo state." },
+      { "kind": "voice", "meaning": "a rule over Larry's own turn output, evaluated at the Stop hook." },
+      { "kind": "memory", "meaning": "a write policy: what channel, what verb, what confirmation gate a memory operation must pass through." },
+      { "kind": "contract", "meaning": "a parity set: phrases or invariants that must be present across named surfaces, byte budgets included." }
+    ],
+    "applies_to_vocabulary": ["pre-flight", "pre-tag", "full", "stop-hook", "room"],
+    "policy_fields": [
+      { "field": "id", "type": "string", "meaning": "unique, kebab-case, matches the filename without the .json extension." },
+      { "field": "kind", "type": "string, member of kind_vocabulary", "meaning": "what class of check this policy declares." },
+      { "field": "runner", "type": "string or null", "meaning": "the script the runner spawns, or null for an honest declared ghost. See runner_path_rule." },
+      { "field": "args", "type": "array of strings", "meaning": "passed to the runner verbatim, no interpolation. Empty array when the runner takes none, or when runner is null." },
+      { "field": "rung", "type": "string, member of rung_vocabulary", "meaning": "the enforcement level this policy runs at TODAY, never an aspiration." },
+      { "field": "evidence_log", "type": "string or null", "meaning": "a path under MINDRIAN_HOME the policy's evidence accumulates in, or null when no log exists yet." },
+      { "field": "promotion_rule", "type": "object, see promotion_rule_doc", "meaning": "decided before any log is read; the runner reports whether it is met and never edits rung itself." },
+      { "field": "owner", "type": "string", "meaning": "the exclusive owner surface for this policy's behavior (the 9-property architecture ownership column)." },
+      { "field": "pinned_by", "type": "array of strings", "meaning": "test files that already pin this policy's underlying behavior." },
+      { "field": "applies_to", "type": "array of strings, members of applies_to_vocabulary", "meaning": "which tier(s) the runner considers this policy for." },
+      { "field": "notes", "type": "string", "meaning": "one paragraph, hyphens only, stating which rung the policy is on and why." }
+    ],
+    "runner_path_rule": "runner is either null (an honest declared ghost, counted as a ghost, never counted as passing) or a repo-relative path that begins with the literal prefix scripts/, contains no parent-directory segment (no ..), is not an absolute path, and resolves inside the repository root when joined to it. A path that escapes the root, an absolute path, or a path containing a parent-directory segment is rejected at --check. The runner is always spawned with an argv array (spawnSync('node', [resolved, ...args])), never a shell string, and args is passed verbatim with no interpolation, so args can never smuggle a second command.",
+    "promotion_rule_doc": {
+      "shape": { "window_runs": "integer, the evidence window the rate is computed over", "max_false_positive_rate": "number between 0 and 1", "min_true_positives": "integer, the minimum count of human-labeled true-positive rows required before a promotion verdict can be MET" },
+      "counting_rule": "max_false_positive_rate is computed as false_positives / (true_positives + false_positives), over HUMAN-LABELED rows only. A JSONL row whose result is still the default 'fire' (unlabeled) counts as NEITHER a true positive NOR a false positive. Counting unlabeled rows as an implicit true positive, or excluding them from the denominator in a way that makes an unreviewed log read as a zero-percent false-positive rate, auto-satisfies any threshold and manufactures a false success (the T-233-04 / T-217-01 defect class). A promotion verdict of MET on a log nobody has read is always wrong.",
+      "decided_before_the_log_is_read": "Every promotion_rule is authored at declaration time, before any evidence exists. The runner and the doctor module both call ONE exported evaluatePromotion(policy, lines) and report MET or NOT MET against the rule as written; neither computes a different rule and neither may loosen it to make a verdict pass.",
+      "promotion_is_a_human_edit": "Promotion is a one-line human edit to rung in the policy's own file, reviewed in a diff, made by a human after reading evaluatePromotion's verdict. The runner never promotes a policy itself, and never demotes one either; both directions are the same one-line human edit."
+    },
+    "gate_count_principle": "The gate set this schema's policies draw from is enumerated from disk at check time (scripts/check-*.cjs minus its *.test.cjs companions) and is NEVER a frozen literal in this file or in the generator, mirroring data/hitl-shape-declaration-schema.json's own surface_count_principle. The count is whatever the enumeration finds; migrating a gate into a policy file never requires editing a count anywhere.",
+    "validation_rule": "An unknown top-level key on a policy file, an unknown rung, an unknown kind, an unknown applies_to member, a missing required field from policy_fields, or a runner that fails runner_path_rule is a hard --check failure naming the file and the failing key. memory-write-policy.json and contract-parity-larry.json each carry a small set of additional keys beyond the eleven common fields (channels/verbs/basket_fires_at/claim_review_status/toggled_off_writes/narrated for the former; surfaces/byte_budget for the latter); those two files' extra keys are explicitly permitted and are not unknown-key violations.",
+    "default_on_miss": "reject (fail closed): a policy file that fails validation is reported by name and skipped from digesting, and --check still fails overall. The generator never regenerates the manifest against an invalid policy directory."
+  }
+}

package/data/harness-policies/contract-parity-larry.json

@@ -0,0 +1,63 @@
+{
+  "id": "contract-parity-larry",
+  "kind": "contract",
+  "runner": null,
+  "args": [],
+  "rung": "blocking",
+  "evidence_log": null,
+  "promotion_rule": { "window_runs": 50, "max_false_positive_rate": 0.05, "min_true_positives": 10 },
+  "owner": "persona",
+  "pinned_by": [
+    "tests/test-298-contract-parity.cjs",
+    "lib/mcp/no-instructions.test.cjs",
+    "tests/test-205-elevation-doctrine-floor.cjs",
+    "tests/test-gate-native-fire-w1.cjs",
+    "tests/test-canon-entry-38-sourced-claims-floor.cjs",
+    "tests/test-chain-executor-part8-leak.cjs",
+    "tests/test-143.2-doctrine-presence.cjs",
+    "tests/test-larry-handoff-seam.cjs"
+  ],
+  "applies_to": ["pre-flight", "pre-tag", "full"],
+  "surfaces": [
+    {
+      "role": "agent",
+      "path": "agents/larry-extended.md",
+      "phrases": [
+        "Operating the machinery",
+        "ZERO user-content egress",
+        "chain-executor",
+        "runChain",
+        "no card, no picture (SEED-021)",
+        "## Decision Gates",
+        "AskUserQuestion",
+        "## Elevation (Part 12",
+        "vertical",
+        "horizontal",
+        "lateral",
+        "larry-personality skill",
+        "## If asked about Theo by name"
+      ]
+    },
+    {
+      "role": "wire",
+      "path": "lib/mcp/runtime-instructions.cjs",
+      "phrases": [
+        "BOUNDARIES (Canon Part 8, absolute): room content NEVER enters brain_* calls - only generic methodology handles (framework names, problem types). Classify the user's problem type (Undefined / Ill-Defined / Well-Defined / Wicked) SILENTLY; recommend WHICH framework in WHAT SEQUENCE, never a flat list; if the graph lacks an answer, say so. Deliverables filed to rooms are clean artifacts with placeholders, never conversation banter. Heavy pipeline work belongs in Claude Code - say so when asked for it here.",
+        "THEO: Theo is MindrianOS's own teaching-graph backend behind every brain_* call. Asked by name, say so briefly and honestly; never volunteer it."
+      ]
+    },
+    {
+      "role": "skill",
+      "path": "skills/larry-personality/SKILL.md",
+      "phrases": [
+        "## The Three Directions of Elevation (Part 12)"
+      ]
+    }
+  ],
+  "byte_budget": {
+    "path": "lib/mcp/runtime-instructions.cjs",
+    "constant": "RUNTIME_INSTRUCTIONS",
+    "limit": 1950
+  },
+  "notes": "runner is null because the generator (scripts/build-harness-manifest.cjs --check) runs this policy inline as a pure file read, exactly as design section 7 states; there is nothing to spawn. Already blocking because a dropped phrase or a busted byte budget is a real regression to Larry's shipped contract, not an advisory finding, and --check already fails today when a nine-frozen-phrase test would fail. Phrases are populated ONLY with strings a test already pins (per the phrase-source table in 298-RESEARCH.md), plus the two genuine unpinned beta.27 gaps: the Theo-by-name heading in agents/larry-extended.md and the THEO clause in RUNTIME_INSTRUCTIONS. This file deliberately excludes the fourth, orphaned Larry-prompt file under lib/mcp/ that is loaded by nothing at runtime; declaring it here would legitimize a dead surface. Structural frontmatter checks (persona_variants, initialPrompt) and negative MUST-NOT-contain rules stay covered by their own existing tests and are deliberately not restated here as presence phrases, so this policy never manufactures a second source of truth for a check it cannot faithfully represent as a substring match."
+}

package/data/harness-policies/gate-card-fire.json

@@ -0,0 +1,13 @@
+{
+  "id": "gate-card-fire",
+  "kind": "gate",
+  "runner": "scripts/check-card-fire.cjs",
+  "args": [],
+  "rung": "logged",
+  "evidence_log": "$MINDRIAN_HOME/card-fire-intercepts.log",
+  "promotion_rule": { "window_runs": 200, "max_false_positive_rate": 0.15, "min_true_positives": 20 },
+  "owner": "hooks",
+  "pinned_by": ["tests/test-209-card-fire-gate.cjs", "tests/test-doctor-card-fire-health.cjs"],
+  "applies_to": ["stop-hook"],
+  "notes": "This is the D8 precedent: the only one of the 35 check-*.cjs gates that already writes a real evidence log (card-fire-intercepts.log under MINDRIAN_HOME, TTL-pruned at 24 hours today). Its own 86 percent false-positive history, recorded before this phase, is exactly why the declared/logged/blocking ladder and the labeled-rows-only promotion rule exist: a raw fire count alone would have looked like a healthy gate. It runs with no flags (stdin envelope, not argv), and enters this declaration at rung logged, matching what it already runs at today, not blocking."
+}

package/data/harness-policies/gate-graph-derive-health.json

@@ -0,0 +1,13 @@
+{
+  "id": "gate-graph-derive-health",
+  "kind": "gate",
+  "runner": "scripts/check-graph-derive-health.cjs",
+  "args": ["--all", "--json"],
+  "rung": "logged",
+  "evidence_log": null,
+  "promotion_rule": { "window_runs": 30, "max_false_positive_rate": 0.1, "min_true_positives": 10 },
+  "owner": "derive-gate",
+  "pinned_by": ["tests/test-233-graph-derive-health.cjs", "tests/test-298-derive-health.cjs"],
+  "applies_to": ["room", "full"],
+  "notes": "This gate wraps lib/core/doctor/graph-derive-health-module.cjs's detectRoomHealth (Canon Part 7 reuse before build: the SEED-037 4d derivation-health signal already ships, with two existing consumers, and this is a third consumer, never a second copy). The committed converged-room fixture has no room.db, so a run against it reports status skip, and skip is the correct green case, not a failure. Declared at rung logged for one release per requirement 5, promoted to blocking only by the D8 rule once a log exists and promotion_rule is met; it is the runner's first declared policy because SEED-037 established that Larry must never write into a graph nothing downstream derives."
+}

package/data/harness-policies/gate-shape-declaration.json

@@ -0,0 +1,13 @@
+{
+  "id": "gate-shape-declaration",
+  "kind": "gate",
+  "runner": "scripts/check-shape-declaration.cjs",
+  "args": ["--check"],
+  "rung": "logged",
+  "evidence_log": null,
+  "promotion_rule": { "window_runs": 50, "max_false_positive_rate": 0.1, "min_true_positives": 10 },
+  "owner": "doctor",
+  "pinned_by": ["tests/test-canon-entry-36-shape-declaration-floor.cjs", "tests/test-shape-declaration-advisory.cjs"],
+  "applies_to": ["pre-tag", "full"],
+  "notes": "This gate has been advisory WARN since Phase 210 (the codebase's own prior decision, not new here): --check exits 0 on a violating tree and enumerates every violation instead of failing, and --strict restores the pre-210 hard-fail. Declaring it at rung logged (not blocking) matches what it already does today; declaring it blocking would silently turn an advisory gate into a release blocker, which this phase explicitly avoids. No evidence log exists for this gate today (34 of 35 gates lack one), so it enters at logged and cannot promote to blocking until a log exists and promotion_rule is met, per D8."
+}

package/data/harness-policies/gate-tool-honesty.json

@@ -0,0 +1,13 @@
+{
+  "id": "gate-tool-honesty",
+  "kind": "gate",
+  "runner": "scripts/check-tool-honesty.cjs",
+  "args": ["--check"],
+  "rung": "logged",
+  "evidence_log": null,
+  "promotion_rule": { "window_runs": 50, "max_false_positive_rate": 0.1, "min_true_positives": 10 },
+  "owner": "doctor",
+  "pinned_by": ["tests/test-276-tool-honesty-findings-closed.cjs", "tests/test-276-tool-honesty-switch-branches.cjs", "tests/test-ljj-tool-honesty.cjs"],
+  "applies_to": ["pre-tag", "full"],
+  "notes": "This gate is advisory in release.sh today (invoked as check-tool-honesty.cjs --check with the failure swallowed by || true) and rides the same advisory-by-default posture as gate-shape-declaration in scripts/doctor.cjs. Declaring it at rung logged (not blocking) matches what it already does today. No evidence log exists for this gate today (34 of 35 gates lack one), so it enters at logged and cannot promote to blocking until a log exists and promotion_rule is met, per D8."
+}

package/data/harness-policies/gate-worktree-hygiene.json

@@ -0,0 +1,13 @@
+{
+  "id": "gate-worktree-hygiene",
+  "kind": "gate",
+  "runner": "scripts/check-worktree-hygiene.cjs",
+  "args": ["--check", "--json"],
+  "rung": "blocking",
+  "evidence_log": null,
+  "promotion_rule": { "window_runs": 50, "max_false_positive_rate": 0.05, "min_true_positives": 5 },
+  "owner": "doctor",
+  "pinned_by": ["tests/test-260906-t3s-worktree-hygiene.cjs"],
+  "applies_to": ["pre-tag", "full"],
+  "notes": "Already blocking: this gate is an existing scripts/doctor.cjs acceptance point (id worktree-hygiene, severity blocker) invoked with --check --json exactly as declared here, so declaring it at rung blocking is a transcription, not a behavior change. No evidence log exists for it today (34 of 35 gates lack one), so promotion_rule is decided now, before any log exists, and is retained for schema completeness and for a future demotion-then-repromotion review, never for an upcoming promotion since blocking is already the ceiling rung."
+}

package/data/harness-policies/memory-write-policy.json

@@ -0,0 +1,28 @@
+{
+  "id": "memory-write-policy",
+  "kind": "memory",
+  "runner": null,
+  "args": [],
+  "rung": "declared",
+  "evidence_log": null,
+  "promotion_rule": { "window_runs": 50, "max_false_positive_rate": 0.1, "min_true_positives": 10 },
+  "owner": "persona",
+  "pinned_by": ["tests/test-189-governance-candidates.cjs", "tests/test-189-f8-governance-gate.cjs"],
+  "applies_to": ["room"],
+  "channels": {
+    "Belief": "the room graph, reached only through navigation.cjs",
+    "Progress": "STATE.md",
+    "Experience": ".planning/seeds/ plus memory_event"
+  },
+  "verbs": {
+    "track": "context_assemble",
+    "commit": ["graph_write", "claim_write"],
+    "recall": ["graph_query", "graph_reason"],
+    "note": "memory_event"
+  },
+  "basket_fires_at": 2,
+  "claim_review_status": "proposed",
+  "toggled_off_writes": "NOT_REMEMBERED_BECAUSE",
+  "narrated": false,
+  "notes": "This is a declaration, not an executable check: runner is null by design (a memory write policy has no scanner to spawn), never a ghost awaiting a scanner. It states D3 (operating by intent, routing reads and writes to the right component), D4 (memory_event fires silently on every substantive turn; claim_write and graph_write land only as proposed, only after an F.8 governance basket that fires at two or more candidate writes; a toggled-off candidate writes NOT_REMEMBERED_BECAUSE), and D5 (the Belief, Progress, Experience channel mapping) exactly as stated in prose in skills/larry-personality/SKILL.md. narrated is false per Canon Part 12 (pedagogy and invisibility): the governance basket is the only visible surface, and Larry never narrates a memory operation."
+}

package/data/harness-policies/registry-drift.json

@@ -0,0 +1,13 @@
+{
+  "id": "registry-drift",
+  "kind": "gate",
+  "runner": "scripts/check-registry-drift.cjs",
+  "args": ["--check"],
+  "rung": "logged",
+  "evidence_log": "$MINDRIAN_HOME/registry-drift.jsonl",
+  "promotion_rule": { "window_runs": 20, "max_false_positive_rate": 0.1, "min_true_positives": 3 },
+  "owner": "doctor",
+  "pinned_by": ["tests/test-341-registry-drift.cjs"],
+  "applies_to": ["pre-tag", "full"],
+  "notes": "On logged today because no evidence exists yet. D-08 requires exactly one release at logged before a human promotion. Promotion is a one-line human edit to rung, reviewed in a diff, never an automatic change. The runner compares the current data/command-registry.json against the previous release's copy (git tag, or the committed tests/fixtures/341-registry-drift-baseline.json fallback on a fresh clone with no tags) and reports commands that vanished without an explicit deprecation entry - the folded 2026-07-03 registry-drift todo."
+}

package/data/harness-policies/release-payload-ceiling.json

@@ -0,0 +1,13 @@
+{
+  "id": "release-payload-ceiling",
+  "kind": "gate",
+  "runner": "scripts/check-release-payload-ceiling.cjs",
+  "args": ["--check"],
+  "rung": "blocking",
+  "evidence_log": null,
+  "promotion_rule": { "window_runs": 50, "max_false_positive_rate": 0.1, "min_true_positives": 10 },
+  "owner": "doctor",
+  "pinned_by": ["tests/test-341-payload-ceiling.cjs"],
+  "applies_to": ["pre-tag", "full"],
+  "notes": "Enters at blocking, not logged, because the platform does NOT enforce the 256 MiB / 20000-entry ceiling on the npm source path - measured against Claude Code 2.1.266, where that check is reachable only from the command source. This policy is the only enforcement that exists for this artifact, so a logged rung would be an honest-looking no-op. The runner is offline: npm pack --dry-run --json makes no registry call. It asserts entryCount and unpackedSize, never packed size, because compression ratio is not a property the platform reads."
+}

package/data/harness-policies/voice-backend-noun-free.json

@@ -0,0 +1,13 @@
+{
+  "id": "voice-backend-noun-free",
+  "kind": "voice",
+  "runner": null,
+  "args": [],
+  "rung": "declared",
+  "evidence_log": null,
+  "promotion_rule": { "window_runs": 100, "max_false_positive_rate": 0.1, "min_true_positives": 50 },
+  "owner": "voice",
+  "pinned_by": [],
+  "applies_to": ["stop-hook"],
+  "notes": "Larry never names backend implementation nouns (graph internals, table names, model names) in a turn to the user. No scanner exists for this rule yet, so runner is null: this is an honest declared ghost, and the runner must never count it as passing. min_true_positives is grounded in the same 50 to 100 human-reviewed-traces baseline from Vanishing Gradients Ep. 57 (Shreya Shankar) used by the other voice policies, so the promotion_rule is ready the moment a scanner lands and the rung can move to logged."
+}

package/data/harness-policies/voice-fork-as-card.json

@@ -0,0 +1,13 @@
+{
+  "id": "voice-fork-as-card",
+  "kind": "voice",
+  "runner": "scripts/check-card-fire.cjs",
+  "args": [],
+  "rung": "logged",
+  "evidence_log": "$MINDRIAN_HOME/card-fire-intercepts.log",
+  "promotion_rule": { "window_runs": 100, "max_false_positive_rate": 0.1, "min_true_positives": 50 },
+  "owner": "voice",
+  "pinned_by": ["tests/test-209-card-fire-gate.cjs"],
+  "applies_to": ["stop-hook"],
+  "notes": "A genuine fork (a real Decision Gate) always renders as a card via AskUserQuestion, never as prose describing the options. card-fire already logs this rule mechanically today (rendered_prose_instead_of_card is one of its four hard invariants), so declaring it at rung logged is a transcription of what already runs, not a behavior change. min_true_positives is grounded in the same 50 to 100 human-reviewed-traces baseline from Vanishing Gradients Ep. 57 (Shreya Shankar) as the other voice policies."
+}

package/data/harness-policies/voice-glyph-present.json

@@ -0,0 +1,13 @@
+{
+  "id": "voice-glyph-present",
+  "kind": "voice",
+  "runner": "scripts/check-voice-style.cjs",
+  "args": [],
+  "rung": "declared",
+  "evidence_log": "$MINDRIAN_HOME/voice-style.jsonl",
+  "promotion_rule": { "window_runs": 100, "max_false_positive_rate": 0.1, "min_true_positives": 50 },
+  "owner": "voice",
+  "pinned_by": ["tests/test-298-voice-log.cjs"],
+  "applies_to": ["stop-hook"],
+  "notes": "Every Larry turn opens with exactly one De Stijl glyph naming the move (detectVoiceMark's valid contract in lib/hmi/voice-color-mark.cjs). This rule already holds in practice today; declared here (not logged) because no window of evidence has been reviewed by a human yet, even though the same Stop hook and voice-style.jsonl log this phase shipped mechanically evaluates it. min_true_positives is grounded in the same sourced figure as voice-hyphens-only: the 50 to 100 human-reviewed-traces baseline from Vanishing Gradients Ep. 57 (Shreya Shankar), low end adopted as the promotion floor."
+}

package/data/harness-policies/voice-hyphens-only.json

@@ -0,0 +1,13 @@
+{
+  "id": "voice-hyphens-only",
+  "kind": "voice",
+  "runner": "scripts/check-voice-style.cjs",
+  "args": [],
+  "rung": "declared",
+  "evidence_log": "$MINDRIAN_HOME/voice-style.jsonl",
+  "promotion_rule": { "window_runs": 100, "max_false_positive_rate": 0.1, "min_true_positives": 50 },
+  "owner": "voice",
+  "pinned_by": ["tests/test-298-voice-log.cjs"],
+  "applies_to": ["stop-hook"],
+  "notes": "No em-dashes or en-dashes in Larry's own turn output. Rung 1 (the rule stated in prose) has been declared three times already by quick 260907-qta; the machinery to log it mechanically (rung 2, the Stop hook plus lib/hmi/voice-style-log.cjs) shipped earlier in this phase, but this declaration conservatively enters at declared pending a human's own review of the first window of evidence. min_true_positives is grounded in the only sourced figure available: the 50 to 100 human-reviewed-traces baseline from Vanishing Gradients Ep. 57 (Shreya Shankar); this policy adopts the low end of that range as its promotion floor."
+}