@herjarsa/omo-meta-governor 0.17.1 → 0.17.3

package/README.md

@@ -376,6 +376,144 @@ No user action required. The audit detection now correctly fires when the agent
 - **If your agent previously wrote `@ts-ignore` without being flagged**: it will now be flagged with `[GRAVE] no-type-suppression: ...` injected as a synthetic user message.
 - **If you want to disable the audit**: set `protocolEnforcement.auditToolCalls: false` (already supported).
 
+
+
+## v0.17.2 — Fix escalation dead code + 4 audit gaps
+
+v0.17.2 closes 4 gaps discovered during live verification of v0.17.0/v0.17.1. The most important: F5.1 (escalate → Oracle) was effectively dead in production due to two compounding bugs.
+
+### Highlights
+
+#### Gap C (CRITICAL) — Escalation now actually fires
+
+The score formula's `noProgress` and `deviations` inputs were hardcoded as `false` and `[]` in the plugin. This meant the `no-progress-detector` (weight 0.20) and `deviation-detector` (weight 0.20) signals always contributed 0. Combined with default thresholds, the maximum possible score was -0.55 — never reaching `escalateThreshold: 0.6` or `stopThreshold: 0.8`.
+
+**Fix:**
+1. **Derive `noProgress`** from the recent tool call window. If the last 5 tool calls contain no `write`/`edit`/`task` (i.e. the agent is only reading/grepping without producing artifacts), `noProgress = true`.
+2. **Derive `deviations`** from accumulated protocol violations. The audit hook now stores violations in `state.accumulatedDeviations` (capped at 5 per session); the orchestrator input reads them.
+3. **Lower default thresholds** to match the new worst-case math:
+   - `escalateThreshold`: 0.6 → 0.45
+   - `stopThreshold`: 0.8 → 0.55
+
+Now worst-case state (no oracle, no progress, 2 grave deviations, iteration at limit, stop-advice lessons) produces score ≈ -0.55 → `stop` action fires.
+
+#### Gap Q (HIGH) — File paths threaded through pipeline
+
+`orchestrator.ts` was hardcoding `filesChanged: []` instead of `input.filePaths`. This meant lesson extraction never saw the actual changed files, so F7.5's file-basename FTS indexing was empty.
+
+**Fix:**
+1. Track `recentWriteFilePaths` in AuditState (alongside existing `recentWriteContents`).
+2. Capture `filePath` from `toolInput.args` on write/edit tool calls.
+3. New `MetaGovernorInput.filePaths?: readonly string[]` passed through to `observeAndLearn`.
+
+#### Gap D (HIGH) — Three config fields now actually do something
+
+Three fields were in the schema and config projection but NEVER consulted by the logic:
+
+- `closedLoop.saveLessons` — parallel to `saveDecisions`. When `false`, lessons are skipped (decision records still save).
+- `intervention.includeDecisionHistory` — when `true`, `messages.transform` prepends recent intervention texts (capped at `maxHistoryMessages`) so the LLM sees its history of decisions.
+- `intervention.maxHistoryMessages` — limit for the above (default 5).
+
+**Fix:** All three fields now control behavior. Track `recentInterventionTexts` in AuditState, format them into the injection text.
+
+#### Bonus — iteration-budget signal wired (Oracle finding)
+
+Oracle flagged a pre-existing gap alongside Gap C: `iteration` was hardcoded `0` in the orchestrator input, making the `iteration-budget` signal (weight 0.15) effectively dead.
+
+Fix:
+- Added `iteration: number` to `AuditState`, incremented per tool call.
+- Threaded `iteration: sessionState?.iteration ?? 0` into `MetaGovernorInput`.
+- `maxIterations` now reads from config instead of being hardcoded.
+
+Worst-case score math updated: with iteration at 100% (-0.12), all signals bad, no oracle → score = -0.65 → `stop` action fires.
+
+#### Gap I (MEDIUM) — `verifyDelivery` return type includes "expired"
+
+The TypeScript signature was `Promise<"delivered" | "pending">` but the registry could return `"expired"`. The expired case leaked through as `"pending"` silently.
+
+**Fix:** Signature updated to `Promise<"delivered" | "pending" | "expired">`. Bridge tools now distinguish: `"delivered"` (verified), `"pending"` (still polling), `"expired"` (TTL elapsed).
+
+### Test & build status
+
+- **521/521 tests pass** (up from 518 — 3 new tests for the v0.17.2 fixes).
+- `bun run typecheck` clean.
+- `bun build.ts` clean (0.34 MB dist).
+- `npm pack --dry-run` validated.
+
+### Migration
+
+No user action required. Two behavior changes:
+
+1. **Escalation now fires more aggressively.** If your agent has been producing violations and not making progress, expect to see escalate → Oracle prompts more often. This is the intended behavior; v0.17.0 was incorrectly silent.
+2. **`includeDecisionHistory` and `maxHistoryMessages` are now functional.** If you set them in v0.17.0 expecting them to work, they will now actually take effect.
+
+### Audit roadmap (status as of v0.17.2)
+
+| Release | Status | Scope |
+|---------|--------|-------|
+| v0.15.1 (F0) | ✅ | Hotfix self-dep |
+| v0.16.0 (F1-F7) | ✅ | Memory hygiene, dead code, tool coverage, CI |
+| v0.17.0 | ✅ | F5.1 escalate, F5.4 cap, F3.6 delivery verify |
+| v0.17.1 | ✅ | Audit args fix |
+| v0.17.2 | ✅ | Gap C (escalation live), Q (file paths), D (config fields), I (delivery expired) |
+
+
+
+## v0.17.3 — Fix Gap I properly (patch)
+
+v0.17.3 is a single-bug patch. During live verification of v0.17.2, Gap I was found to be incompletely fixed.
+
+### The bug (v0.17.2 cosmetic fix)
+
+The `verifyDelivery` export signature was widened to include `"expired"` in v0.17.2, and the bridge tools' title/output text was updated to handle it. **BUT the underlying `pollForDelivery` helper was still collapsing `"expired"` → `"pending"` silently:**
+
+```ts
+// v0.17.2 (BUG):
+return status === "delivered" ? "delivered" : "pending"
+```
+
+So bridge tools could never report `"expired"` to the user, even though the registry correctly tracked it. Live verification confirmed: `deliveryStatus` always showed `"pending"`.
+
+### The fix (v0.17.3)
+
+```ts
+// v0.17.3:
+return await pendingRegistryRef.awaitDelivery({ sessionID, mcpTool, timeoutMs })
+```
+
+Now the actual status from the registry propagates through. `"expired"` flows end-to-end to the bridge tool's `metadata.deliveryStatus` and title.
+
+### Tests
+
+Added 3 RED tests in `src/custom-tools.test.ts`:
+- Returns `"expired"` when registry entry exists past timeout (real registry instance)
+- Returns `"delivered"` when `markDelivered` fires before timeout
+- Returns `"pending"` when no registry is configured
+
+### Test & build status
+
+- **525/525 tests pass** (up from 522 in v0.17.2 — 3 new tests for pollForDelivery).
+- `bun run typecheck` clean.
+- `bun build.ts` clean (0.34 MB dist).
+
+### Migration
+
+No user action required. Bridge tools will now correctly distinguish all three delivery states:
+- `"delivered"` — LLM's MCP tool call was observed within 1.5s
+- `"expired"` — TTL elapsed without delivery (entry expires after 10s, but bridge tool sees this immediately as "expired" when polling times out at 1.5s)
+- `"pending"` — no registry configured (graceful degradation for tests/mocks)
+
+### Audit roadmap (status as of v0.17.3)
+
+| Release | Status | Scope |
+|---------|--------|-------|
+| v0.15.1 → v0.17.2 | ✅ | All audit findings + 5 gap fixes |
+| v0.17.3 | ✅ | Gap I real fix (pollForDelivery returns "expired") |
+
+Two remaining gaps documented but require SDK support to fix:
+- `recentTurnTokens: []` — token-predictor signal dead (10% of score); needs per-turn token counts from OpenCode SDK
+- `agentName` defaults to `"unknown"` — cosmetic, no functional impact
+
 ## Auto-upgrade (v0.12.0)
 
 On plugin load, queries npm/pip registries to check whether newer versions

package/dist/custom-tools.d.ts

@@ -46,6 +46,16 @@ declare let pendingRegistryRef: {
  * Exposed as a setter so we don't need to thread it through every tool deps.
  */
 export declare function setPendingDeliveryRegistry(registry: typeof pendingRegistryRef): void;
+/**
+ * v0.17.2 (Gap I): Updated return type to include "expired" so the bridge
+ * tools can distinguish between "LLM hasn't called yet" (pending) and
+ * "TTL elapsed without delivery" (expired).
+ *
+ * Returns "delivered" if the LLM's MCP tool call was observed within the
+ * timeout, "expired" if the pending entry's TTL elapsed without delivery,
+ * "pending" otherwise (poll still active, no result yet).
+ */
+export declare function verifyDelivery(sessionID: string, mcpTool: string): Promise<"delivered" | "pending" | "expired">;
 export interface OmoSearchDeps {
     graphRetrieval: GraphRetrieval;
     cwd: string;