comisai 1.0.45 → 1.0.46

package/node_modules/@comis/agent/dist/executor/verification-gate.js

@@ -1,13 +1,13 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * R4 verification-gate.ts — critic orchestration.
+ * verification-gate.ts — critic orchestration.
  *
  * Gate: only fires on completion-claiming responses past minResponseChars.
  * Seam: single bounded completeSimple call, fail-closed on every error.
  * Delivery: not-verified ⇒ honest first-person unmet-list (sanitized REQ tokens),
  *   never the agent-directed redirect — postExecution is terminal, no re-queue
- *   consumer exists yet (CR-01/D5; the re-queue/full-retry loop is R5/Phase 155+).
- * L5: native-reasoning profiles size maxOutputTokens to avoid verdict starvation.
+ *   consumer exists yet (a re-queue/full-retry loop may be added later).
+ * Native-reasoning profiles size maxOutputTokens to avoid verdict starvation.
  *
  * File-size: ≤350 lines (do not exceed).
  * Forbidden: Date.now(), raw setTimeout/clearTimeout, new Date().
@@ -16,17 +16,17 @@
  * @module
  */
 import { systemSetTimeout, systemClearTimeout } from "@comis/core";
-import { completeSimple, getModel } from "@earendil-works/pi-ai";
+import { completeSimple, getModel } from "@earendil-works/pi-ai/compat";
 import { isCompletionClaim, wrapReviewedOutput, detectCanaryLeakage, detectImpliedToolCall, parseCriticVerdict, buildCriticSystemPrompt, } from "./critic-isolation.js";
 import { buildSafetySection } from "../bootstrap/sections/core-sections.js";
 // ---------------------------------------------------------------------------
-// L5: reasoning-budget-aware maxOutputTokens sizing
+// Reasoning-budget-aware maxOutputTokens sizing
 // ---------------------------------------------------------------------------
 const VERDICT_RESERVE_TOKENS = 512;
 /**
- * CRITIC-ONLY verdict budget — NOT the main answer path (CR-01). 512 tokens
+ * CRITIC-ONLY verdict budget — NOT the main answer path. 512 tokens
  * suffices for the small verdict JSON; native-reasoning profiles get 4× so
- * reasoning_content doesn't starve the verdict (D7).
+ * reasoning_content doesn't starve the verdict.
  */
 export function resolveMaxOutputTokens(profile) {
     return profile.reasoningStyle === "native"
@@ -35,14 +35,14 @@ export function resolveMaxOutputTokens(profile) {
 }
 /**
  * Reasoning-headroom floor for the MAIN answer path (native profiles only).
- * SD4 (Phase 158): raised 4096→16384 so reasoning_content cannot starve the
- * visible answer on small native-reasoning models (observed finishReason:"length"
- * with qwen3.6:35b at maxTokens:8192 — reasoning consumed nearly all tokens).
+ * 16384 so reasoning_content cannot starve the visible answer on small
+ * native-reasoning models (observed finishReason:"length" with qwen3.6:35b
+ * at maxTokens:8192 — reasoning consumed nearly all tokens).
  * Math.max means models already reporting maxOutputTokens≥16384 are unchanged.
  */
 const NATIVE_REASONING_MAIN_PATH_FLOOR = 16_384;
 /**
- * CR-01: MAIN-PATH output budget. Returns the model's REAL maxOutputTokens so
+ * MAIN-PATH output budget. Returns the model's REAL maxOutputTokens so
  * the visible answer is never clamped to the 512-token verdict reserve.
  * Non-reasoning → full profile budget; native-reasoning → sized UP (never
  * down) so reasoning_content can't starve the answer. config.maxTokens (when
@@ -79,9 +79,9 @@ function extractResponseText(response) {
 // callCritic — one bounded completeSimple call, fail-closed
 // ---------------------------------------------------------------------------
 async function callCritic(response, plan, deps) {
-    // S2: wrap the reviewed output as untrusted BEFORE the user message is built
+    // Wrap the reviewed output as untrusted BEFORE the user message is built
     const wrapped = wrapReviewedOutput(response);
-    const safetyCore = buildSafetySection(false); // S2: always false — never []
+    const safetyCore = buildSafetySection(false); // always false — never []
     const systemPrompt = buildCriticSystemPrompt({
         checklist: plan.steps,
         canaryToken: deps.canaryToken,
@@ -122,7 +122,7 @@ async function callCritic(response, plan, deps) {
             signal: controller.signal,
         });
         const text = extractResponseText(raw);
-        // S2: check canary leak BEFORE Zod parse
+        // Check canary leak BEFORE Zod parse
         if (detectCanaryLeakage(text, deps.canaryToken)) {
             deps.eventBus.emit("critic.isolation.canary_leak", {
                 timestamp: deps.clock.now(),
@@ -136,7 +136,7 @@ async function callCritic(response, plan, deps) {
             };
         }
         const parsed = parseCriticVerdict(text); // Zod safeParse — never throws
-        // S2: check implied tool calls in verdict followUp
+        // Check implied tool calls in verdict followUp
         if (parsed.verdict !== "skipped") {
             const verdictWithProps = parsed;
             const followUpText = verdictWithProps.followUp ?? "";
@@ -172,9 +172,9 @@ async function callCritic(response, plan, deps) {
 }
 // ---------------------------------------------------------------------------
 // shouldRunCritic — gate check for the post-execution hook
-// WR-02: keyless-only (no secretManager at this layer). Mirrors KEYLESS_PROVIDER_TYPES.
+// Keyless-only (no secretManager at this layer). Mirrors KEYLESS_PROVIDER_TYPES.
 // ---------------------------------------------------------------------------
-/** Exported for parity assertion in scaffold-defaults.ts (Plan 02 cross-file test). */
+/** Exported for the cross-file parity assertion in the scaffold-defaults tests. */
 export const KEYLESS_CRITIC_PROVIDERS = new Set(["ollama", "lm-studio"]);
 export function shouldRunCritic(params) {
     const { capabilityClass, config, executionPlanRef, provider, logger, effectiveEnabled } = params;
@@ -191,7 +191,7 @@ export function shouldRunCritic(params) {
             provider,
             errorKind: "config",
             step: "verification",
-            hint: "Verification critic skipped: cloud API-key threading is deferred to Phase 155. " +
+            hint: "Verification critic skipped: cloud API-key threading is not yet wired. " +
                 "Use a keyless provider (ollama/lm-studio) to exercise the critic.",
         }, "Verification critic skipped (cloud key threading not yet wired)");
         return false;
@@ -200,16 +200,16 @@ export function shouldRunCritic(params) {
 }
 // ---------------------------------------------------------------------------
 // runVerificationCritic — entry point for the post-execution hook
-// Gate + single critic call + honest first-person delivery on not-verified (R4/D5)
+// Gate + single critic call + honest first-person delivery on not-verified
 //
-// `maxRetries` is accepted but NOT branched on in this phase: postExecution is
-// terminal (no re-queue consumer), so a not-verified verdict ALWAYS delivers the
-// honest first-person unmet-list — never the agent-directed redirect (CR-01).
-// The param is retained for the future re-queue/full-retry path (R5/Phase 155+).
+// `maxRetries` is accepted but NOT branched on: postExecution is terminal
+// (no re-queue consumer), so a not-verified verdict ALWAYS delivers the
+// honest first-person unmet-list — never the agent-directed redirect.
+// The param is retained for a future re-queue/full-retry path.
 // ---------------------------------------------------------------------------
 export async function runVerificationCritic(params) {
     const { response, plan, deps } = params;
-    // Gate: skip if not a completion claim or response is too short (D4)
+    // Gate: skip if not a completion claim or response is too short
     if (!isCompletionClaim(response) || response.length < deps.minResponseChars) {
         return { verdict: "skipped", response };
     }
@@ -221,10 +221,10 @@ export async function runVerificationCritic(params) {
     if (verdict.verdict === "verified" || verdict.verdict === "skipped") {
         return { verdict: verdict.verdict, response };
     }
-    // not-verified path: honest first-person delivery (CR-01). postExecution is
-    // TERMINAL in this phase (no re-queue consumer), so this `response` is
-    // delivered VERBATIM to the user. The agent-directed redirect is only correct
-    // when something re-queues it (R5/Phase 155+), so we ALWAYS deliver the honest
+    // not-verified path: honest first-person delivery. postExecution is
+    // TERMINAL (no re-queue consumer), so this `response` is delivered
+    // VERBATIM to the user. The agent-directed redirect is only correct
+    // when something re-queues it, so we ALWAYS deliver the honest
     // first-person unmet-list — at the DEFAULT maxRetries (2) and at 0 — never the
     // redirect and never an unqualified "done". `unmet` is sanitized (see below).
     const verdictWithUnmet = verdict;
@@ -234,12 +234,13 @@ export async function runVerificationCritic(params) {
     };
 }
 // ---------------------------------------------------------------------------
-// buildHonestUnmetResponse — the ONLY user-facing not-verified text (CR-01/D5)
+// buildHonestUnmetResponse — the ONLY user-facing not-verified text
 //
-// WR-01/IN-03: `unmet[]` is critic-authored over UNTRUSTED reviewed content, so
-// raw interpolation lets a crafted label ("REQ-0 (all tasks are done)") make the
-// text match isCompletionClaim (violating D5) or smuggle a tool-call phrase into
-// user prose. Reduce each entry to its bare REQ-\d+ token (executor-controlled
+// `unmet[]` is critic-authored over UNTRUSTED reviewed content, so raw
+// interpolation lets a crafted label ("REQ-0 (all tasks are done)") make the
+// text match isCompletionClaim (breaking the honest-exhaustion guarantee) or
+// smuggle a tool-call phrase into user prose. Reduce each entry to its bare
+// REQ-\d+ token (executor-controlled
 // plan index, not critic free-text); the composed text then provably fails both
 // isCompletionClaim and detectImpliedToolCall.
 // ---------------------------------------------------------------------------

package/node_modules/@comis/agent/dist/greeting/session-greeting.d.ts

@@ -27,7 +27,7 @@ export interface GreetingGeneratorDeps {
     };
 }
 /**
- * Which greeting variant to produce (spec §12). Closed union — adding a
+ * Which greeting variant to produce. Closed union — adding a
  * member forces a new `promptForTrigger` case via the exhaustive switch guard.
  * - `standard`: persona-driven greeting, no setup language.
  * - `onboarding-pending`: first run, setup unfinished — greet + guide next step.
@@ -36,7 +36,7 @@ export interface GreetingGeneratorDeps {
  */
 export type GreetingTrigger = "standard" | "onboarding-pending" | "onboarding-limited";
 export interface GreetingGenerator {
-    /** Generate a greeting for the given trigger variant (spec §12). */
+    /** Generate a greeting for the given trigger variant. */
     generate(agentName: string, trigger: GreetingTrigger): Promise<Result<string, Error>>;
 }
 /**

package/node_modules/@comis/agent/dist/greeting/session-greeting.js

@@ -13,10 +13,10 @@
  * @module
  */
 import { ok, err } from "@comis/shared";
-import { completeSimple, getModel } from "@earendil-works/pi-ai";
+import { completeSimple, getModel } from "@earendil-works/pi-ai/compat";
 import { systemNowMs, systemSetTimeout, systemClearTimeout } from "@comis/core";
 // ---------------------------------------------------------------------------
-// Per-variant system prompt (spec §12)
+// Per-variant system prompt
 // ---------------------------------------------------------------------------
 /**
  * Map a {@link GreetingTrigger} to its system prompt. Pure; the closed-union

package/node_modules/@comis/agent/dist/index.d.ts

@@ -29,7 +29,7 @@ export type { BudgetGuard, BudgetSnapshot } from "./budget/budget-guard.js";
 export { createCostTracker } from "./budget/cost-tracker.js";
 export type { CostTracker, CostRecord, UsageInput } from "./budget/cost-tracker.js";
 export { createSpendAccumulator, SpendError } from "./budget/spend-accumulator.js";
-export type { SpendAccumulator, SpendScope } from "./budget/spend-accumulator.js";
+export type { SpendAccumulator, SpendScope, SpendLimb, SpendUnit } from "./budget/spend-accumulator.js";
 export { checkSpendCeiling } from "./budget/budget-guard.js";
 export type { SpendGateOutcome, SpendGateConfig } from "./budget/budget-guard.js";
 export { createTurnBudgetTracker } from "./budget/turn-budget-tracker.js";
@@ -181,7 +181,7 @@ export { createThinkingTagFilter } from "./response-filter/thinking-tag-filter.j
 export type { ThinkingTagFilter, ThinkingTagFilterOptions } from "./response-filter/thinking-tag-filter.js";
 export { resolveOperationModel, resolveProviderFamily } from "./model/operation-model-resolver.js";
 export type { OperationModelResolution } from "./model/operation-model-resolver.js";
-export { resolveModelProfile, capabilityClassFromProvider } from "./executor/model-profile.js";
+export { resolveModelProfile, capabilityClassFromProvider, autoRepairForClass } from "./executor/model-profile.js";
 export type { CapabilityClass } from "./executor/model-profile.js";
 export { seedDefaultDagTemplates } from "./executor/dag-templates.js";
 export { matchRawGraphToTemplate } from "./executor/dag-template-match.js";
@@ -197,6 +197,8 @@ export { createCorrectionDetectorSeam } from "./memory/index.js";
 export type { CorrectionVerdict } from "./memory/index.js";
 export { createOutcomeJudgeSeam } from "./memory/index.js";
 export type { CustomCompletionsModelSpec } from "./memory/index.js";
+export { createOrchestrateRepairSeam } from "./memory/index.js";
+export type { OrchestrateRepairSeam, OrchestrateRepairSeamDeps } from "./memory/index.js";
 export { createLlmReflectionAdapter, runReflection, classifyReflectOutcome } from "./memory/index.js";
-export { REFLECT_PROMPT, PROFILE_REFLECT_PROMPT, TOPIC_REFLECT_PROMPT } from "./memory/index.js";
+export { REFLECT_PROMPT, PROFILE_REFLECT_PROMPT, TOPIC_REFLECT_PROMPT, PROCEDURE_REFLECT_PROMPT } from "./memory/index.js";
 export type { LlmReflectionAdapterDeps, ReflectionAdapter, ReflectInput, ReflectionResult, RunReflectionDeps, RunReflectionResult, RunReflectionConfig, ReflectionSourceTrajectory, ReflectAdmissionOutcome, } from "./memory/index.js";

package/node_modules/@comis/agent/dist/index.js

@@ -20,17 +20,16 @@ export { createOverflowRecovery, createOverflowRecoveryWrapper } from "./executo
 // Budget
 export { createBudgetGuard, BudgetError } from "./budget/budget-guard.js";
 export { createCostTracker } from "./budget/cost-tracker.js";
-// Spend kill-switch (Phase 177): the surface the daemon composition root consumes
+// Spend kill-switch: the surface the daemon composition root consumes
 // cross-package — the daemon-wide accumulator factory + the instance type + the
 // scope type (setup-observability wiring + the PiExecutorDeps thread).
 // The 3-state spend GATE (`checkSpendCeiling` + its outcome/error/ceilings/config
-// types) is now consumed cross-package by the per-`rootRunId` budget meter
-// (`@comis/daemon` `autonomy/per-root-budget.ts`, Phase 213-04 BUDGET-01/02/03):
-// the meter REUSES the shipped 3-state gate VERBATIM for its $-limb (the
-// ffe11736 fail-closed semantics — never re-implemented). These were previously
-// intra-package; the daemon meter is the first cross-package consumer, so they
-// graduate to the barrel (public-export-consumers gate: a barrel export needs a
-// cross-package consumer, which now exists).
+// types) is consumed cross-package by the per-`rootRunId` budget meter
+// (`@comis/daemon` `autonomy/per-root-budget.ts`): the meter REUSES the 3-state
+// gate VERBATIM for its $-limb (same fail-closed semantics — never
+// re-implemented). Exported here because the daemon meter consumes them
+// cross-package (public-export-consumers gate: a barrel export needs a
+// cross-package consumer).
 export { createSpendAccumulator, SpendError } from "./budget/spend-accumulator.js";
 export { checkSpendCeiling } from "./budget/budget-guard.js";
 export { createTurnBudgetTracker } from "./budget/turn-budget-tracker.js";
@@ -69,16 +68,16 @@ export { createContextWindowResolver } from "./model/context-window-resolver.js"
 export { getCacheProviderInfo } from "./executor/cache-usage-helpers.js";
 // Model scanner
 export { createModelScanner } from "./model/model-scanner.js";
-// Ollama capacity probe (CWF-03: boot-time served num_ctx discovery)
+// Ollama capacity probe (boot-time served num_ctx discovery)
 // Only probeAllOllamaProviders is consumed cross-package (daemon boot);
 // the remaining probe symbols (prewarmOllamaModel et al.) are intra-package
 // only — invoked internally by probeAllOllamaProviders, never re-exported.
 export { probeAllOllamaProviders } from "./model/ollama-capacity-probe.js";
-// Served-window comparator (KNOB-01: boot-time served<configured WARN).
+// Served-window comparator (boot-time served<configured WARN).
 // Only compareServedWindowForProvider + the comparison type are consumed
 // cross-package (daemon setup-agents wiring); the latch reset is test-only.
 export { compareServedWindowForProvider, resetServedWindowWarnForTest } from "./model/served-window-comparator.js";
-// Viable floor (FLOOR-01: boot-time minViable WARN). collectAgentBootWindowInfo +
+// Viable floor (boot-time minViable WARN). collectAgentBootWindowInfo +
 // evaluateViableFloorForAgent are consumed cross-package (daemon boot wiring);
 // computeMinViableEquation + the drift-pin surface are intra-package/test only.
 export { collectAgentBootWindowInfo, evaluateViableFloorForAgent } from "./context-engine/viable-floor.js";
@@ -94,7 +93,7 @@ export { createAuthUsageTracker } from "./model/auth-usage-tracker.js";
 export { createLastKnownModelTracker } from "./model/last-known-model.js";
 // Routing symbols (router factory, resolver, router type) live in
 // @comis/orchestrator. Consumers import from @comis/orchestrator.
-// Session lifecycle (renamed from session-manager.ts)
+// Session lifecycle
 export { createSessionLifecycle } from "./session/session-lifecycle.js";
 // Session label store (human-readable session names via metadata.label)
 export { createSessionLabelStore } from "./session/session-label-store.js";
@@ -116,15 +115,12 @@ export { createIdentityUpdater } from "./identity/identity-updater.js";
 export { createGreetingGenerator } from "./greeting/session-greeting.js";
 // Memory review (periodic session history extraction)
 export { runMemoryReview } from "./memory/memory-review-job.js";
-// (The offline triple-extraction job — runMemoryTripleExtraction + its
-// MemoryTripleExtractionDeps/Config/Stats + TripleCandidate types — was DELETED in Phase 226
-// SIMPLIFY-03 alongside its dormant __MEMORY_TRIPLE_EXTRACTION__ cron (the `extract` returned
-// [] — no triples were ever written). The TripleStorePort + its sqlite adapter + the
-// graphSpread recall lane (recall-graph-spread-lane.ts) survive — only the extraction JOB went.)
-// (The offline usefulness-judge seam — createUsefulnessJudgeSeam — was DELETED in Phase 226
-// SIMPLIFY-03 alongside its dormant __USEFULNESS_JUDGE__ cron. The keyless citation-marker
-// attribution stays the usefulness signal; the FORGET-02 recordUsage reward write lives in
-// the daemon's setup-learning.ts (a separate seam), not this factory.)
+// (There is no offline triple-extraction job: the TripleStorePort + its sqlite
+// adapter + the graphSpread recall lane (recall-graph-spread-lane.ts) exist
+// without a batch extraction writer.)
+// (There is no offline usefulness-judge seam: the keyless citation-marker
+// attribution is the usefulness signal; the recordUsage reward write lives in
+// the daemon's setup-learning.ts, a separate seam.)
 // Query-time dialectic synthesis seam (the ONE allowed query-time LLM
 // surface). The factory the daemon `memory.ask` handler calls to BUILD the
 // synthesize() seam from a cheap resolved model (the daemon injects it), keeping DIALECTIC_PROMPT
@@ -141,13 +137,6 @@ export { createDialecticSeam } from "./memory/memory-dialectic-seam.js";
 // shapes (AssembledSynthesis/CitationChain/ParsedSynthesis) are inferred at the consumer and
 // kept module-internal (no unconsumed type surface on the public barrel).
 export { orderByTrust, assembleSynthesis, citationChains } from "./memory/memory-dialectic-synthesis.js";
-// (The offline directional relationship builder (runRelationshipBuild), its cheap-model
-//  seam factory (createRelationshipSeam), and the directional builder prompt + parser
-//  (RELATIONSHIP_PROMPT / buildRelationshipPrompt / parseRelationshipOutput) — the
-//  WRITE path of the social-modeling subsystem — were DELETED in Phase 226 SIMPLIFY-03
-//  with the rest of that subsystem (the __SOCIAL_MODELING__ cron, the RelationshipStore
-//  port + adapter, the `relationship` table, the relationship-block prompt injection).
-//  No alias, I1.)
 // RAG (Retrieval-Augmented Generation)
 export { formatMemorySection } from "./rag/rag-retriever.js";
 // Queue symbols live in @comis/orchestrator:
@@ -281,28 +270,28 @@ export { sanitizeAssistantResponse, setSanitizeLogger, extractFinalTagContent }
 // Response filter utilities (used by ThinkingTagFilter)
 export { stripReasoningTagsFromText } from "./response-filter/reasoning-tags.js";
 export { findCodeRegions, isInsideCode } from "./response-filter/code-regions.js";
-// Thinking tag filter (moved from @comis/channels)
+// Thinking tag filter
 export { createThinkingTagFilter } from "./response-filter/thinking-tag-filter.js";
 // Operation model resolver
 export { resolveOperationModel, resolveProviderFamily } from "./model/operation-model-resolver.js";
 // ModelProfile resolver — the immutable capability/capacity profile. Exported so
-// the daemon wiring can derive the memory-job capabilityClass (R6) the SAME way
-// pi-executor does per-execution (CR-01: R6 was never reaching production). Only
+// the daemon wiring derives the memory-job capabilityClass the SAME way
+// pi-executor does per-execution (any parallel derivation silently drifts from
+// what actually runs). Only
 // the resolver + CapabilityClass type cross the package boundary; the memory jobs
 // (also in @comis/agent) own the resolveMemoryOpsStrategy call internally.
-export { resolveModelProfile, capabilityClassFromProvider } from "./executor/model-profile.js";
-// O2 (WR-02): canonical DAG template seeding. seedDefaultDagTemplates is wired
+export { resolveModelProfile, capabilityClassFromProvider, autoRepairForClass } from "./executor/model-profile.js";
+// Canonical DAG template seeding. seedDefaultDagTemplates is wired
 // into daemon bootstrap (idempotent INSERT-OR-IGNORE) so the four canonical
 // small-model templates exist in the named-graph store at startup.
 export { seedDefaultDagTemplates } from "./executor/dag-templates.js";
-// AUTHOR-01 (Phase 174-03, the former "Phase 157"): the deterministic,
-// conservative repair matcher. INJECTED into the daemon's buildGraphInput via
+// The deterministic, conservative repair matcher. INJECTED into the daemon's buildGraphInput via
 // deps.repairMatch (rpc-dispatch composition site) — the daemon never imports
 // dag-templates directly. fillDagTemplate / CANONICAL_DAG_TEMPLATES stay
 // package-internal (the matcher is the only weak-model repair consumer; it fills
 // internally), so only the matcher fn + its result types cross the boundary.
 export { matchRawGraphToTemplate } from "./executor/dag-template-match.js";
-// AUTHOR-02 (Phase 174-04): the deterministic intent → ExecutionGraph
+// The deterministic intent → ExecutionGraph
 // synthesizer. Imported by the pipeline tool's from_intent action (@comis/skills)
 // — it RETURNS a validated graph, never executes one; the tool dispatches it
 // through the existing graph.execute path so governance applies automatically.
@@ -313,34 +302,41 @@ export { resolveCompactionModel } from "./model/compaction-model-resolver.js";
 export { createSessionLatch } from "./executor/session-latch.js";
 // Background task infrastructure
 export * from "./background/index.js";
-// Correction-detector seam (Verified Learning WS1, Phase 199 P0.5). The cost-gated
+// Correction-detector seam. The cost-gated
 // `fast`-tier detector the daemon constructs (setup-learning-reactions
 // buildReactionWiringDeps) ONLY when `learningOutcome.correction.enabled` is opted
 // in, then runs over a follow-up user turn → a `corrected`/`correction` soft-
-// failure verdict (CORRECT-01). Re-exported from the memory sub-barrel beside its
-// daemon consumer (this plan) so the public-export-consumers gate never sees an
+// failure verdict. Re-exported from the memory sub-barrel beside its
+// daemon consumer so the public-export-consumers gate never sees an
 // orphan. The prompt + triple-bound (wrap + reward-cap + strip-parse) stay
 // agent-internal.
 export { createCorrectionDetectorSeam } from "./memory/index.js";
-// Verified Learning WS1 (OUTCOME-04, Phase 202). The cost-gated LLM outcome-judge
+// The cost-gated LLM outcome-judge
 // seam the daemon constructs on the `outcomeJudge` fast tier as the FALLBACK source
 // for a CONVERSATIONAL turn (an `unknown` deterministic resolve). Re-exported beside
 // its daemon consumer (setup-learning-reactions wiring) so the public-export-consumers
 // gate never sees an orphan. The prompt + triple-bound + the reward cap stay
 // agent-internal (the daemon `observe()`s the seam's already-capped `cappedConfidence`).
 export { createOutcomeJudgeSeam } from "./memory/index.js";
-// v2.31 Reflection engine (Phase 223 Plan 04, REFLECT-01/03/04/05/06). The
+// The one-shot orchestrate repair seam the daemon mints (in buildAutonomyToolWiring)
+// on a repair-eligible capability class and injects into the orchestrate runner:
+// ONE bounded utility-model re-prompt regenerates a failed script for exactly one
+// re-run. Built on the SAME resolveJudgeModel+completeSimple machinery as the
+// outcome-judge seam so @comis/skills consumes only the injected closure and never
+// imports the model layer — resolveJudgeModel stays package-internal; the FACTORY
+// (+ its closure/deps types) is the export.
+export { createOrchestrateRepairSeam } from "./memory/index.js";
+// Reflection engine. The
 // reflection JOB (runReflection) + the cheap-model reflect adapter + the
-// prompt/parser the daemon invokes from the __REFLECT__ cron (Plan 05): SELECT
+// prompt/parser the daemon invokes from the __REFLECT__ cron: SELECT
 // trusted-origin success → group-by-topicKey → ≥2-distinct corroboration gate →
 // delta-ops reflect → validateLearnedDocBody guard → admit at candidate/learned.
 // Consumes @comis/core PORT TYPES + the static guard + the pure delta-ops only
 // (the agent↛memory / agent↛skills cut); the daemon injects the store + adapter.
-// Re-exported from the memory sub-barrel beside its daemon consumer (Plan 05) so
-// the public-export-consumers gate never sees an orphan. This REPLACED the dead
-// embedding-clustering synthesis pipeline (job/adapter/prompt), deleted in Plan 06.
+// Re-exported from the memory sub-barrel beside its daemon consumer so
+// the public-export-consumers gate never sees an orphan.
 export { createLlmReflectionAdapter, runReflection, classifyReflectOutcome } from "./memory/index.js";
 // The per-kind reflect prompts the daemon `__REFLECT__` cron injects as the adapter
-// `systemPrompt` (Plan 04): REFLECT_PROMPT (skill default), PROFILE_REFLECT_PROMPT
-// (Plan 02), TOPIC_REFLECT_PROMPT (Plan 03). One engine, varied per-kind prompt.
-export { REFLECT_PROMPT, PROFILE_REFLECT_PROMPT, TOPIC_REFLECT_PROMPT } from "./memory/index.js";
+// `systemPrompt`: REFLECT_PROMPT (skill default), PROFILE_REFLECT_PROMPT,
+// TOPIC_REFLECT_PROMPT, PROCEDURE_REFLECT_PROMPT. One engine, varied per-kind prompt.
+export { REFLECT_PROMPT, PROFILE_REFLECT_PROMPT, TOPIC_REFLECT_PROMPT, PROCEDURE_REFLECT_PROMPT } from "./memory/index.js";

package/node_modules/@comis/agent/dist/memory/__fixtures__/recall-eval-fixtures.d.ts

@@ -121,14 +121,14 @@ export declare const TEMPORAL_EVAL_FIXTURES: EvalQuery[];
  * Trust-first contradiction fixtures — the `"temporal"` group's TRUST
  * case, scored against the EXISTING trust lever (`score()` with `trustAlpha>0`;
  * score.ts UNCHANGED). Kept in a SEPARATE exported array from {@link TEMPORAL_EVAL_FIXTURES}
- * (Pitfall 1) so the temporal-group lift/neutral assertions over T1/T2 stay
+ * so the temporal-group lift/neutral assertions over T1/T2 stay
  * byte-untouched and green (the documented zero-regression discipline).
  *
  * THE TRUST LEAPFROG (vs Hindsight's latest-mentioned-wins). Each query pairs a NEWER
  * LOW-trust claim carrying the HIGHER fusion `score` (it would win on recency/lexical alone)
  * against an OLDER HIGHER-trust fact carrying the LOWER fusion `score`. Because single-lane
  * `fuse()` is order-preserving, the fusion-only baseline ranks the newer low-trust claim at
- * rank 1 and MISSES the higher-trust id at recall@1 (Pitfall 2 — the distractor is given the
+ * rank 1 and MISSES the higher-trust id at recall@1 (the distractor is given the
  * headroom so the trust lever has work to do). Ranking through `score(..., { trustAlpha:0.5 })`
  * (system 1.0 / learned 0.5 / external 0.0; recencyAlpha 0 so `createdAt` does NOT move the
  * order) lifts the higher-trust fact to rank 1: a newer low-trust claim does NOT supersede an
@@ -298,9 +298,9 @@ export declare function usefulnessByIdFor(q: EvalQuery): ReadonlyMap<string, Use
  * scenario scored against the LIVE proof lever (`score()` with `proofAlpha>0`; the proof
  * log curve × confidence half-life — score.ts:166-198 UNCHANGED, proofAlpha is already live).
  * Kept in a SEPARATE exported array from the prior groups' fixtures so their assertions
- * stay untouched and green (the documented zero-regression discipline, Pitfall 1).
+ * stay untouched and green (the documented zero-regression discipline).
  *
- * THE PROOF-ACCRUAL PAYOFF (HINDSIGHT_VS_COMIS.md N2 PARITY). A fact corroborated across MULTIPLE
+ * THE PROOF-ACCRUAL PAYOFF. A fact corroborated across MULTIPLE
  * consolidation runs accrues proof and OUT-RANKS a one-off mention. The fold path
  * grows `proof_count` via the UNIONed source-set cardinality + refreshes `occurredAt`/`confidence`;
  * this group proves the read side rewards that accrual. Each query pairs a ONE-OFF mention (a raw,

package/node_modules/@comis/agent/dist/memory/__fixtures__/recall-eval-fixtures.js

@@ -227,14 +227,14 @@ export const TEMPORAL_EVAL_FIXTURES = [
  * Trust-first contradiction fixtures — the `"temporal"` group's TRUST
  * case, scored against the EXISTING trust lever (`score()` with `trustAlpha>0`;
  * score.ts UNCHANGED). Kept in a SEPARATE exported array from {@link TEMPORAL_EVAL_FIXTURES}
- * (Pitfall 1) so the temporal-group lift/neutral assertions over T1/T2 stay
+ * so the temporal-group lift/neutral assertions over T1/T2 stay
  * byte-untouched and green (the documented zero-regression discipline).
  *
  * THE TRUST LEAPFROG (vs Hindsight's latest-mentioned-wins). Each query pairs a NEWER
  * LOW-trust claim carrying the HIGHER fusion `score` (it would win on recency/lexical alone)
  * against an OLDER HIGHER-trust fact carrying the LOWER fusion `score`. Because single-lane
  * `fuse()` is order-preserving, the fusion-only baseline ranks the newer low-trust claim at
- * rank 1 and MISSES the higher-trust id at recall@1 (Pitfall 2 — the distractor is given the
+ * rank 1 and MISSES the higher-trust id at recall@1 (the distractor is given the
  * headroom so the trust lever has work to do). Ranking through `score(..., { trustAlpha:0.5 })`
  * (system 1.0 / learned 0.5 / external 0.0; recencyAlpha 0 so `createdAt` does NOT move the
  * order) lifts the higher-trust fact to rank 1: a newer low-trust claim does NOT supersede an
@@ -538,9 +538,9 @@ export function usefulnessByIdFor(q) {
  * scenario scored against the LIVE proof lever (`score()` with `proofAlpha>0`; the proof
  * log curve × confidence half-life — score.ts:166-198 UNCHANGED, proofAlpha is already live).
  * Kept in a SEPARATE exported array from the prior groups' fixtures so their assertions
- * stay untouched and green (the documented zero-regression discipline, Pitfall 1).
+ * stay untouched and green (the documented zero-regression discipline).
  *
- * THE PROOF-ACCRUAL PAYOFF (HINDSIGHT_VS_COMIS.md N2 PARITY). A fact corroborated across MULTIPLE
+ * THE PROOF-ACCRUAL PAYOFF. A fact corroborated across MULTIPLE
  * consolidation runs accrues proof and OUT-RANKS a one-off mention. The fold path
  * grows `proof_count` via the UNIONed source-set cardinality + refreshes `occurredAt`/`confidence`;
  * this group proves the read side rewards that accrual. Each query pairs a ONE-OFF mention (a raw,

package/node_modules/@comis/agent/dist/memory/benchmark/__fixtures__/qa-judge-stub.d.ts

@@ -4,9 +4,9 @@
  * answer-prompt -> judge-prompt -> verdict-parse -> aggregate) WITHOUT a real
  * provider or the `COMIS_BENCH` gate.
  *
- * This is the ONLY genuinely new shape here: the repo has no shared
- * fake-`completeSimple` stub — LLM call sites are
- * mocked ad hoc per test. The return shape is dictated by pi-ai's
+ * The repo has no shared fake-`completeSimple` stub — LLM call sites are
+ * mocked ad hoc per test — so this fixture provides the one reusable
+ * stub. The return shape is dictated by pi-ai's
  * `AssistantMessage.content: (TextContent | ...)[]` with
  * `TextContent = { type: "text"; text: string }`
  * (node_modules/@earendil-works/pi-ai/dist/types.d.ts:143,189-191), so the value

package/node_modules/@comis/agent/dist/memory/benchmark/__fixtures__/qa-judge-stub.js

@@ -5,9 +5,9 @@
  * answer-prompt -> judge-prompt -> verdict-parse -> aggregate) WITHOUT a real
  * provider or the `COMIS_BENCH` gate.
  *
- * This is the ONLY genuinely new shape here: the repo has no shared
- * fake-`completeSimple` stub — LLM call sites are
- * mocked ad hoc per test. The return shape is dictated by pi-ai's
+ * The repo has no shared fake-`completeSimple` stub — LLM call sites are
+ * mocked ad hoc per test — so this fixture provides the one reusable
+ * stub. The return shape is dictated by pi-ai's
  * `AssistantMessage.content: (TextContent | ...)[]` with
  * `TextContent = { type: "text"; text: string }`
  * (node_modules/@earendil-works/pi-ai/dist/types.d.ts:143,189-191), so the value

package/node_modules/@comis/agent/dist/memory/benchmark/ablation-sweep.d.ts

@@ -5,8 +5,8 @@
  * (turn it OFF, re-run, read the delta). The keyless head-to-head harness drives
  * `sweepCells` at $0.
  *
- * THE VERIFIED KNOB LEAVES (re-grepped against recall-types.ts:142-182 +
- * schema-memory-reasoning.ts:41 -- an earlier draft was WRONG on three of them):
+ * THE VERIFIED KNOB LEAVES (verified against recall-types.ts:142-182 +
+ * schema-memory-reasoning.ts:41 -- three of them have plausible-but-wrong spellings):
  *   - kg-graph-spread    -> `lanes.graphSpread.enabled`        (recall-types.ts:157)
  *   - iq-mmr             -> `mmr.enabled`         (recall-types.ts:178; NOT `rag.mmr.enabled`)
  *   - iq-intent          -> `queryUnderstanding.intentReweight` (recall-types.ts:182; the leaf
@@ -69,8 +69,8 @@ export interface AblationFactor {
 export declare const REASON_WRITE_SIDE_FACTOR: "reason-observations";
 /**
  * The shipped ablation factors, each pinned to its VERIFIED knob leaf. The
- * grep acceptance criteria + the off=byte-identity test guard these paths against
- * the stale `rag.`-prefixed / `.intent` draft leaves.
+ * off=byte-identity test guards these paths against the plausible-but-wrong
+ * `rag.`-prefixed / `.intent` spellings (which are not real config leaves).
  */
 export declare const V28_ABLATION_FACTORS: readonly AblationFactor[];
 /**

package/node_modules/@comis/agent/dist/memory/benchmark/ablation-sweep.js

@@ -6,8 +6,8 @@
  * (turn it OFF, re-run, read the delta). The keyless head-to-head harness drives
  * `sweepCells` at $0.
  *
- * THE VERIFIED KNOB LEAVES (re-grepped against recall-types.ts:142-182 +
- * schema-memory-reasoning.ts:41 -- an earlier draft was WRONG on three of them):
+ * THE VERIFIED KNOB LEAVES (verified against recall-types.ts:142-182 +
+ * schema-memory-reasoning.ts:41 -- three of them have plausible-but-wrong spellings):
  *   - kg-graph-spread    -> `lanes.graphSpread.enabled`        (recall-types.ts:157)
  *   - iq-mmr             -> `mmr.enabled`         (recall-types.ts:178; NOT `rag.mmr.enabled`)
  *   - iq-intent          -> `queryUnderstanding.intentReweight` (recall-types.ts:182; the leaf
@@ -47,8 +47,8 @@
 export const REASON_WRITE_SIDE_FACTOR = "reason-observations";
 /**
  * The shipped ablation factors, each pinned to its VERIFIED knob leaf. The
- * grep acceptance criteria + the off=byte-identity test guard these paths against
- * the stale `rag.`-prefixed / `.intent` draft leaves.
+ * off=byte-identity test guards these paths against the plausible-but-wrong
+ * `rag.`-prefixed / `.intent` spellings (which are not real config leaves).
  */
 export const V28_ABLATION_FACTORS = [
     { factor: "kg-graph-spread", knobPath: "lanes.graphSpread.enabled", on: true, off: false },

package/node_modules/@comis/agent/dist/memory/benchmark/competitor-adapter.d.ts

@@ -14,7 +14,7 @@
  * — a shape that carries NO score field. {@link AdapterResult} is a DISCRIMINATED
  * UNION whose `ran:false` arm has no `accuracy`/`overall`/`score`/`manifestRef`,
  * so it is structurally IMPOSSIBLE for an absent system to fabricate a number
- * (CONTEXT.md decision 4; the anti-fabrication threat). The keyless CI
+ * (the anti-fabrication threat). The keyless CI
  * always hits the skip branch (no env, no install) — that IS the wiring proof; the
  * operator-costed run (keys + competitor installs + LLM spend) fills the real
  * numbers in a costed re-run, honestly deferred.

package/node_modules/@comis/agent/dist/memory/benchmark/competitor-adapter.js

@@ -15,7 +15,7 @@
  * — a shape that carries NO score field. {@link AdapterResult} is a DISCRIMINATED
  * UNION whose `ran:false` arm has no `accuracy`/`overall`/`score`/`manifestRef`,
  * so it is structurally IMPOSSIBLE for an absent system to fabricate a number
- * (CONTEXT.md decision 4; the anti-fabrication threat). The keyless CI
+ * (the anti-fabrication threat). The keyless CI
  * always hits the skip branch (no env, no install) — that IS the wiring proof; the
  * operator-costed run (keys + competitor installs + LLM spend) fills the real
  * numbers in a costed re-run, honestly deferred.

package/node_modules/@comis/agent/dist/memory/benchmark/cross-judge-spread.d.ts

@@ -5,15 +5,13 @@
  *
  * THE MOAT: credibility = cross-judge
  * >=2 + a published spread. A result is trusted ONLY if it survives -- i.e. the
- * two judges agree within tolerance on that category. This module replaces the
- * hand-authored arithmetic in
- * `benchmarks/results/2026-05-31-j1-baseline/cross-judge-spread.md` with a
+ * two judges agree within tolerance on that category. This module is a
  * tested, reproducible fold over >=2 committed judge manifests.
  *
- * THE SURVIVAL TOLERANCE (from the j1-baseline reference math): a category
+ * THE SURVIVAL TOLERANCE: a category
  * SURVIVES if |A-B| <= 5.0 points. Per-category n=20 gives a binomial SE of
- * ~10-11pt, so a >5pt inter-judge gap is judge-noise, not signal -- e.g.
- * single-session-preference did NOT survive in the j1 baseline (30 vs 45 = 15pt)
+ * ~10-11pt, so a >5pt inter-judge gap is judge-noise, not signal -- e.g. a
+ * single-session-preference reading of 30 vs 45 (= 15pt) does NOT survive
  * and must NOT headline.
  *
  * PURE (no Result, no throws, no I/O, no clock, no env -- AGENTS.md 2.1 pure-fn
@@ -60,7 +58,7 @@ import type { AccuracyResult } from "./qa-accuracy.js";
  * The survival tolerance in accuracy points: a category SURVIVES (is stable
  * across the two judges) when its inter-judge spread |A-B| <= this value.
  *
- * 5.0pt is the j1-baseline protocol value. Rationale: per-category n=20 yields a
+ * 5.0pt is the protocol value. Rationale: per-category n=20 yields a
  * binomial standard error of ~10-11pt, so a <=5pt inter-judge gap is well within
  * noise (the judges agree), whereas a larger gap signals the number is too
  * judge-dependent to headline.