comisai 1.0.44 → 1.0.46

package/node_modules/@comis/agent/dist/memory/topic-key.js

@@ -1,10 +1,9 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * `normalizeOpeningRequest` (Phase 223, D-01 / REFLECT-02) — the deterministic,
- * keyless, **content-light** `topicKey` that REPLACES embedding-cosine clustering.
+ * `normalizeOpeningRequest` — the deterministic, keyless, **content-light**
+ * `topicKey` (deliberately embedding-free).
  *
- * This is the milestone's concentrated risk (the failure mode the dead embedding-cosine
- * clustering had, re-lived from the other direction): the reflection job groups successful
+ * This is the keyless approach's concentrated risk: the reflection job groups successful
  * outcomes by this key, so two genuinely same-topic sessions worded DIFFERENTLY MUST collide
  * on one key — else corroboration never reaches the >=2 distinct (session,sender) gate and
  * `admitted:0` persists forever. The collision-maximizing decision is the order-insensitive
@@ -21,10 +20,10 @@
  *  3. TOKENIZE on whitespace; drop {@link STOPWORDS} and tokens of length <= 1.
  *  4. DE-DUPLICATE into a Set, then SORT — order-insensitive. "deploy the app" and
  *     "app deploy please" carry the same {app,deploy} token set and collide; a
- *     token-SEQUENCE/string hash would NOT (RESEARCH Alternative + Pitfall 3 + A1).
+ *     token-SEQUENCE/string hash would NOT.
  *  5. HASH the sorted-unique join with sha256 (the repo's deterministic-id convention,
  *     mirroring `sqlite-mental-model-store.ts:mentalModelId`) and return the hex —
- *     content-light, NEVER the raw transcript (INV-6). An empty token list returns ""
+ *     content-light, NEVER the raw transcript. An empty token list returns ""
  *     (the reflection job treats "" as ungroupable — a singleton that never corroborates).
  *
  * Pure: no IO, no embedder, no clock/random. The sole consumer is the reflection job
@@ -80,11 +79,10 @@ const STOPWORDS = new Set([
 ]);
 /**
  * Recover the RAW user request from a stored inbound message by stripping the
- * executor's injected envelope. LIFTED verbatim (Phase 223 D-07) from
- * `setup-channels-skill-synthesis-deps.ts` (`stripUserSystemContext`) — that deps
- * file's embedding wiring is deleted in Plan 05, but this envelope-strip is the
- * correct PRE-STEP for the topicKey (Pitfall 3), so it is copied here as a private
- * helper. The executor wraps every inbound turn as
+ * executor's injected envelope. Copied verbatim from
+ * `setup-channels-skill-synthesis-deps.ts` (`stripUserSystemContext`) — this
+ * envelope-strip is the correct PRE-STEP for the topicKey, so it is kept here
+ * as a private helper. The executor wraps every inbound turn as
  * `[System context]\n<preamble incl. a VOLATILE timestamp>\n[End system context]\n\n[<channel>] <id> (<time>):\n<actual message>`
  * (envelope-wrapper.ts). Both the system-context preamble AND the channel header
  * carry a per-turn timestamp, so the stored "user message" of two IDENTICAL requests
@@ -122,7 +120,7 @@ export function normalizeOpeningRequest(signature) {
     // An empty token list is ungroupable — return "" (the reflection job treats it as a never-corroborating singleton).
     if (uniqueSorted.length === 0)
         return "";
-    // 5. Hash the canonical join (content-light — never the raw text; INV-6).
+    // 5. Hash the canonical join (content-light — never the raw text).
     return createHash("sha256").update(uniqueSorted.join(" ")).digest("hex");
 }
 /**
@@ -133,7 +131,7 @@ export function normalizeOpeningRequest(signature) {
  * DIFFERENTLY (e.g. a "fire" vs a "medical" dispatch on the same evening-rush cross-river
  * route) land on SEPARATE topicKeys and never reach the ≥2 corroboration gate. The
  * reflection job merges groups whose token sets are highly similar ({@link jaccardSimilarity}
- * ≥ threshold) — keyless, deterministic, NO embeddings (the v2.31 collapse deleted those on purpose).
+ * ≥ threshold) — keyless, deterministic, NO embeddings (the embedding-free invariant is deliberate).
  *
  * Pure: same pipeline as `normalizeOpeningRequest` steps 1-4, minus the hash.
  *
@@ -141,29 +139,29 @@ export function normalizeOpeningRequest(signature) {
  * @returns the sorted, de-duplicated content tokens (possibly empty).
  */
 export function openingRequestTokens(signature) {
-    // 1. Strip the volatile per-turn envelope FIRST (Pitfall 3).
+    // 1. Strip the volatile per-turn envelope FIRST (its timestamps would defeat collision).
     const stripped = stripUserSystemContext(signature);
     // 2. Lowercase; collapse non-alphanumeric runs to a single space; trim.
     const cleaned = stripped
         .toLowerCase()
         .replace(/[^a-z0-9]+/g, " ")
         .trim();
-    // 3. Tokenize; drop stopwords and tokens of length <= 1; STEM each survivor (REFLECT-02b: collapse
+    // 3. Tokenize; drop stopwords and tokens of length <= 1; STEM each survivor (collapse
     //    morphological variants so two genuinely-same-task openings worded differently — "deliver"/"delivered"/
     //    "delivering", "package"/"packages", "report"/"reports" — share tokens and reach the corroboration /
-    //    reuse-credit overlap they otherwise miss). Keyless + deterministic + pure (no embeddings — the v2.31
-    //    invariant holds). Stem AFTER the stopword check (stopwords are base forms).
+    //    reuse-credit overlap they otherwise miss). Keyless + deterministic + pure (no embeddings —
+    //    the keyless invariant holds). Stem AFTER the stopword check (stopwords are base forms).
     const tokens = cleaned.length === 0
         ? []
         : cleaned
             .split(/\s+/)
             .filter((t) => t.length > 1 && !STOPWORDS.has(t))
             .map(stemToken);
-    // 4. De-duplicate into a Set, then SORT — order-insensitive (the collision-maximizing decision, A1).
+    // 4. De-duplicate into a Set, then SORT — order-insensitive (the collision-maximizing decision).
     return [...new Set(tokens)].sort();
 }
 /**
- * A deliberately CONSERVATIVE inflectional stemmer (REFLECT-02b) — collapses the common,
+ * A deliberately CONSERVATIVE inflectional stemmer — collapses the common,
  * low-risk English inflections so morphological variants of the same word land on ONE token,
  * widening the corroboration / reuse-credit overlap beyond byte-identical phrasings WITHOUT
  * re-introducing embeddings (keyless + deterministic + pure). It strips ONLY regular inflections
@@ -176,8 +174,8 @@ export function openingRequestTokens(signature) {
  * It is intentionally imperfect (base-vs-inflected like "navigate"/"navigated" may not fully
  * reconcile) — applied UNIFORMLY at admit-core and reuse-turn time, so inflected↔inflected and
  * base↔base variants match; the win is monotone (more true merges, guarded against false ones).
- * NOTE: changing the token shape changes the topicKey hash — skills learned on the PRE-stemming
- * build store an un-stemmed core and re-accrue after an upgrade (a learning system re-learns; not a
+ * NOTE: changing the token shape changes the topicKey hash — skills learned under an older
+ * token shape store a different core and simply re-accrue (a learning system re-learns; not a
  * code regression). Pure: no IO/clock/random.
  */
 export function stemToken(token) {
@@ -256,7 +254,7 @@ export function tokenSetCoverage(core, turn) {
  *  0.5 credits real reuse without false-crediting an unrelated/adjacent turn. An earlier 0.6 (set
  *  without data) MISSED genuine reuse → the learned skill never promoted on a real behavioral
  *  instance. (HEAVY synonym variation below ~0.4 is the lexical-match ceiling — a semantic /
- *  LLM-topic-tag grouping would be the escalation, deliberately out of the keyless v2.31 scope.) */
+ *  LLM-topic-tag grouping would be the escalation, deliberately out of the keyless design's scope.) */
 export const DEFAULT_TOPIC_MATCH_THRESHOLD = 0.5;
 /** Absolute floor: a turn that shares at least this many of a skill's CORE tokens is credited
  *  as a reuse EVEN IF the coverage FRACTION is below {@link DEFAULT_TOPIC_MATCH_THRESHOLD}. This
@@ -269,7 +267,7 @@ export const DEFAULT_TOPIC_MATCH_THRESHOLD = 0.5;
 export const MIN_ABSOLUTE_CORE_MATCH = 8;
 /**
  * The deterministic, keyless TOPIC-MATCH reuse attribution. Skill attribution otherwise
- * requires the agent to explicitly `read` the SKILL.md file (ATTR-01 in pi-event-bridge.ts) —
+ * requires the agent to explicitly `read` the SKILL.md file (the read-attribution path in pi-event-bridge.ts) —
  * a skill the agent APPLIED from the surfaced `<available_skills>` summary / from recall,
  * without opening the file, was never credited, so its `proof_count` never bumped and it
  * never promoted (the reuse leg of the learning loop was LLM-fragile). This credits a SURFACED skill whose
@@ -293,7 +291,7 @@ export function topicMatchedSkillNames(turnSignature, surfaced, threshold = DEFA
     // The CREDITED subset of topicMatchScores (de-duplicated, input order) — the carrier the
     // pi-event-bridge merges into the turn's usedSkillIds. Output is byte-identical to the
     // pre-refactor loop; the scoring now lives in topicMatchScores so the NEGATIVE path
-    // (surfaced-but-uncredited) is observable too (finding A).
+    // (surfaced-but-uncredited) is observable too.
     const matched = new Set();
     for (const s of topicMatchScores(turnSignature, surfaced, threshold))
         if (s.credited)

package/node_modules/@comis/agent/dist/model/compaction-model-resolver.d.ts

@@ -4,19 +4,18 @@
  * is the empty string.
  *
  * Why this exists:
- *   The `compactionModel` field used to default to a hardcoded Anthropic
- *   literal ("anthropic:claude-haiku-4-5-20250929"). When operators
- *   switched their primary provider to OpenRouter, Google, etc., compaction
- *   still routed to Claude Haiku, defeating the cost-tiering intent and
- *   causing cross-provider auth confusion (no Anthropic API key configured).
+ *   A hardcoded provider-specific default (e.g. an Anthropic Haiku literal)
+ *   would keep routing compaction to that provider even when an operator's
+ *   primary provider is OpenRouter, Google, etc. — defeating the
+ *   cost-tiering intent and causing cross-provider auth confusion (no
+ *   Anthropic API key configured).
  *
  *   The schema default is "" and the value is resolved at runtime: pick the
  *   fast-tier model from `resolveOperationDefaults(primaryProvider)`, with
  *   `getModels(primaryProvider)[0]` as the catalog-fallback.
  *
- *   Note: explicit `compactionModel` values from existing YAML configs win
- *   unchanged (length > 0 short-circuits the resolver). No backward-compat
- *   shim per `feedback_no_backward_compat.md`.
+ *   Note: explicit `compactionModel` values from YAML configs win unchanged
+ *   (length > 0 short-circuits the resolver).
  *
  * @module
  */

package/node_modules/@comis/agent/dist/model/compaction-model-resolver.js

@@ -5,23 +5,22 @@
  * is the empty string.
  *
  * Why this exists:
- *   The `compactionModel` field used to default to a hardcoded Anthropic
- *   literal ("anthropic:claude-haiku-4-5-20250929"). When operators
- *   switched their primary provider to OpenRouter, Google, etc., compaction
- *   still routed to Claude Haiku, defeating the cost-tiering intent and
- *   causing cross-provider auth confusion (no Anthropic API key configured).
+ *   A hardcoded provider-specific default (e.g. an Anthropic Haiku literal)
+ *   would keep routing compaction to that provider even when an operator's
+ *   primary provider is OpenRouter, Google, etc. — defeating the
+ *   cost-tiering intent and causing cross-provider auth confusion (no
+ *   Anthropic API key configured).
  *
  *   The schema default is "" and the value is resolved at runtime: pick the
  *   fast-tier model from `resolveOperationDefaults(primaryProvider)`, with
  *   `getModels(primaryProvider)[0]` as the catalog-fallback.
  *
- *   Note: explicit `compactionModel` values from existing YAML configs win
- *   unchanged (length > 0 short-circuits the resolver). No backward-compat
- *   shim per `feedback_no_backward_compat.md`.
+ *   Note: explicit `compactionModel` values from YAML configs win unchanged
+ *   (length > 0 short-circuits the resolver).
  *
  * @module
  */
-import { getModels } from "@earendil-works/pi-ai";
+import { getModels } from "@earendil-works/pi-ai/compat";
 import { resolveOperationDefaults } from "./operation-model-defaults.js";
 /**
  * Resolve the effective compaction model id, in "provider:modelId" format.

package/node_modules/@comis/agent/dist/model/fresh-tail-clamp.d.ts

@@ -5,11 +5,11 @@
  * Rules:
  * - If effectiveWindow is not finite (frontier/mid models), returns
  *   configuredTurns UNCHANGED — byte-identical guarantee.
- * - Conservative 30% budget floor (§7 open measurement question):
+ * - Conservative 30% budget floor:
  *   affordable = floor(effectiveWindow * 0.3 / avgTokPerStep)
  * - Returns min(configuredTurns, affordable), minimum 1.
- * - The 30% fraction is a placeholder until the Phase-171 harness measures
- *   actual recall impact. Do not treat it as a calibrated constant.
+ * - The 30% fraction is a conservative placeholder pending real recall-impact
+ *   measurement. Do not treat it as a calibrated constant.
  *
  * Pure function — no side effects, no async, no DI. All context passed
  * as parameters. Mirrors the effective-context-window.ts pattern.
@@ -27,11 +27,12 @@
  *   If omitted, estimated as `max(1, floor(effectiveWindow / 20))` (5% floor).
  * @returns The clamped turn count — always ≥ 1, always ≤ configuredTurns.
  *
- * NB (ISSUE #1, 2026-06-22): this turn-count clamp is the COARSE upper bound only.
+ * NB: this turn-count clamp is the COARSE upper bound only.
  * The PRECISE residual enforcement for the protected fresh tail (so it always fits a
  * small window where the system prompt dominates) lives in the lcd-assembler as a
- * post-B-8 TOTAL token bound (boundFreshTailTotalToResidual) — a turn-count clamp is
- * unreliable there because one oversized message (which B-8 bounds anyway) skews the
- * per-step estimate.
+ * TOTAL token bound applied after the per-message char cap
+ * (boundFreshTailTotalToResidual) — a turn-count clamp is unreliable there
+ * because one oversized message (which the per-message cap bounds anyway) skews
+ * the per-step estimate.
  */
 export declare function resolveClampedFreshTailTurns(effectiveWindow: number, configuredTurns: number, avgTokensPerStep?: number): number;

package/node_modules/@comis/agent/dist/model/fresh-tail-clamp.js

@@ -6,11 +6,11 @@
  * Rules:
  * - If effectiveWindow is not finite (frontier/mid models), returns
  *   configuredTurns UNCHANGED — byte-identical guarantee.
- * - Conservative 30% budget floor (§7 open measurement question):
+ * - Conservative 30% budget floor:
  *   affordable = floor(effectiveWindow * 0.3 / avgTokPerStep)
  * - Returns min(configuredTurns, affordable), minimum 1.
- * - The 30% fraction is a placeholder until the Phase-171 harness measures
- *   actual recall impact. Do not treat it as a calibrated constant.
+ * - The 30% fraction is a conservative placeholder pending real recall-impact
+ *   measurement. Do not treat it as a calibrated constant.
  *
  * Pure function — no side effects, no async, no DI. All context passed
  * as parameters. Mirrors the effective-context-window.ts pattern.
@@ -28,12 +28,13 @@
  *   If omitted, estimated as `max(1, floor(effectiveWindow / 20))` (5% floor).
  * @returns The clamped turn count — always ≥ 1, always ≤ configuredTurns.
  *
- * NB (ISSUE #1, 2026-06-22): this turn-count clamp is the COARSE upper bound only.
+ * NB: this turn-count clamp is the COARSE upper bound only.
  * The PRECISE residual enforcement for the protected fresh tail (so it always fits a
  * small window where the system prompt dominates) lives in the lcd-assembler as a
- * post-B-8 TOTAL token bound (boundFreshTailTotalToResidual) — a turn-count clamp is
- * unreliable there because one oversized message (which B-8 bounds anyway) skews the
- * per-step estimate.
+ * TOTAL token bound applied after the per-message char cap
+ * (boundFreshTailTotalToResidual) — a turn-count clamp is unreliable there
+ * because one oversized message (which the per-message cap bounds anyway) skews
+ * the per-step estimate.
  */
 export function resolveClampedFreshTailTurns(effectiveWindow, configuredTurns, avgTokensPerStep) {
     if (!isFinite(effectiveWindow))

package/node_modules/@comis/agent/dist/model/model-allowlist.d.ts

@@ -3,7 +3,7 @@
  * is permitted to use.
  *
  * When the allowlist is empty (no entries), all models are allowed
- * (backward-compatible default). When populated, only models matching
+ * (the permissive default). When populated, only models matching
  * an entry are permitted.
  *
  * Supports two matching formats:
@@ -45,7 +45,7 @@ export interface ModelAllowlist {
  * - "provider/modelId" for exact provider+model matching
  * - "modelId" for provider-agnostic matching
  *
- * An empty array means "allow all models" (backward compatible).
+ * An empty array means "allow all models" (the permissive default).
  *
  * @param allowedModels - List of permitted model strings
  */

package/node_modules/@comis/agent/dist/model/model-allowlist.js

@@ -4,7 +4,7 @@
  * is permitted to use.
  *
  * When the allowlist is empty (no entries), all models are allowed
- * (backward-compatible default). When populated, only models matching
+ * (the permissive default). When populated, only models matching
  * an entry are permitted.
  *
  * Supports two matching formats:
@@ -23,7 +23,7 @@
  * - "provider/modelId" for exact provider+model matching
  * - "modelId" for provider-agnostic matching
  *
- * An empty array means "allow all models" (backward compatible).
+ * An empty array means "allow all models" (the permissive default).
  *
  * @param allowedModels - List of permitted model strings
  */

package/node_modules/@comis/agent/dist/model/model-registry-adapter.d.ts

@@ -17,7 +17,7 @@ import type { ModelAllowlist } from "./model-allowlist.js";
 /**
  * Result of initial model resolution.
  *
- * SA6a: This interface mirrors the shape of `InitialModelResult` in
+ * This interface mirrors the shape of `InitialModelResult` in
  * `@earendil-works/pi-coding-agent` (dist/core/model-resolver.d.ts).
  * The SDK type is structurally identical but is NOT re-exported from the
  * package root and `./core/model-resolver` is absent from the package's

package/node_modules/@comis/agent/dist/model/model-registry-adapter.js

@@ -10,7 +10,7 @@
  * @module
  */
 import { ModelRegistry } from "@earendil-works/pi-coding-agent";
-import { getModels, getProviders } from "@earendil-works/pi-ai";
+import { getModels, getProviders } from "@earendil-works/pi-ai/compat";
 import { KEYLESS_PROVIDER_TYPES, KEYLESS_API_KEY_SENTINEL } from "@comis/core";
 /**
  * Create a ModelRegistry from an AuthStorage instance.

package/node_modules/@comis/agent/dist/model/model-scanner.js

@@ -12,7 +12,7 @@
  *
  * @module
  */
-import { getModels, getProviders } from "@earendil-works/pi-ai";
+import { getModels, getProviders } from "@earendil-works/pi-ai/compat";
 import { systemNowMs, systemSetTimeout, systemClearTimeout } from "@comis/core";
 // ---------------------------------------------------------------------------
 // Catalog-driven endpoint resolution

package/node_modules/@comis/agent/dist/model/oauth-token-manager.js

@@ -297,7 +297,7 @@ function buildBootstrapProfile(provider, seed) {
  */
 export function createOAuthTokenManager(deps) {
     const { secretManager, eventBus, credentialStore, logger, dataDir, fileLock, keyPrefix = "OAUTH_", } = deps;
-    // Hot-path read cache (Discretion item — invalidated on persisted writes).
+    // Hot-path read cache — invalidated on persisted writes.
     // Keyed by canonical profileId.
     const cache = new Map();
     // De-dup sets — fire WARN/INFO once per (provider, process).
@@ -311,8 +311,7 @@ export function createOAuthTokenManager(deps) {
     // Logger de-dup: fire the configured-profile-resolved INFO once per
     // (provider, configured-profile, process) when the configured profile is
     // first used. withDedup (@comis/core — NOT @comis/infra; agent↛infra)
-    // replaces the prior hand-rolled loggedConfiguredProviders Set with the
-    // shared decorator; same process-lifetime semantics (default TTL = unbounded).
+    // is the shared decorator; process-lifetime semantics (default TTL = unbounded).
     const dedupLogger = withDedup(logger);
     // -------------------------------------------------------------------------
     // chokidar watcher on auth-profiles.json (file adapter only). chokidar's
@@ -898,7 +897,7 @@ export function createOAuthTokenManager(deps) {
                 }
                 // Once-per-(provider, configured-profile, process) INFO log when the
                 // configured profile is first used. withDedup collapses repeats by the
-                // dedupKey field (process-lifetime), replacing the prior Set guard.
+                // dedupKey field (process-lifetime).
                 dedupLogger.info({
                     dedupKey: `${providerId}::${configured}`,
                     provider: providerId,
@@ -966,7 +965,7 @@ export function createOAuthTokenManager(deps) {
                 const firstProfile = tierCList.value[0];
                 if (firstProfile) {
                     // Conflict detection on the picked profile — env var may diverge
-                    // from stored refresh (R7c silent vs WARN path).
+                    // from stored refresh (maybeWarnEnvConflict decides silent vs WARN).
                     const envRawForC = secretManager.get(toSecretKey(providerId, keyPrefix));
                     const envSeedForC = parseEnvCredentials(envRawForC);
                     maybeWarnEnvConflict(providerId, firstProfile, envSeedForC);
@@ -1010,7 +1009,7 @@ export function createOAuthTokenManager(deps) {
                     expiresAt,
                     secsUntilExpiry,
                 }, "Profile loaded from store");
-                // Conflict detection (R7c) — env var diverges from stored refresh.
+                // Conflict detection — env var may diverge from stored refresh.
                 maybeWarnEnvConflict(providerId, profile, envSeed);
             }
             else {
@@ -1062,7 +1061,7 @@ export function createOAuthTokenManager(deps) {
             // `comis auth login`, persisted in the store) reads as ABSENT through
             // hasCredentials — which froze a Codex agent's image generation
             // unavailable for the daemon's life even though the SAME OAuth credential
-            // answered its text completions (the CRED-01/CDX-01 follow-main regression).
+            // answered its text completions.
             if (Array.from(cache.values()).some((p) => p.provider === providerId))
                 return true;
             if (secretManager.has(toSecretKey(providerId, keyPrefix)))

package/node_modules/@comis/agent/dist/model/ollama-capacity-probe.d.ts

@@ -87,7 +87,7 @@ export interface ProbeAllOllamaProvidersParams {
     /** Timeout per provider probe in ms. */
     timeoutMs: number;
     /**
-     * IMP-2a (package-delivery-20260628): also fire a fire-and-forget LOAD-ONLY warm-up
+     * When true, also fire a fire-and-forget LOAD-ONLY warm-up
      * (`prewarmOllamaModel`) per ollama provider so the model is resident before the first user
      * turn — a cold model's first inference can exceed the per-inference stall budget and abort the
      * first turn after a (re)start "request took too long". Default false (byte-identical when unset).
@@ -100,9 +100,9 @@ export interface ProbeAllOllamaProvidersParams {
     };
 }
 /**
- * Single source for "which model did the probe use" (KNOB-01's served-window
- * comparator shares it — the 17fdd1e5 bug class was two sites deriving this
- * expression differently). ProviderEntrySchema has NO `defaultModel`; the
+ * Single source for "which model did the probe use" (the served-window
+ * comparator shares it — two sites deriving this expression independently
+ * is how they once disagreed). ProviderEntrySchema has NO `defaultModel`; the
  * model lives under `models[].id`. Falling back to the first configured model
  * lets the /api/show cold-start path (boot, before any inference, when
  * /api/ps is empty) send a real model name instead of "" (which Ollama
@@ -118,7 +118,7 @@ export declare function resolveProbedModelId(entry: {
  * Fire-and-forget LOAD-ONLY warm-up: POST {nativeBaseUrl}/api/generate with the model + keep_alive and an
  * EMPTY prompt, which loads the model into memory WITHOUT generating (the Ollama preload idiom).
  *
- * Why (IMP-2a, package-delivery-20260628, local qwen3.6:35b): a cold local model's FIRST inference —
+ * Why (observed live with a local qwen3.6:35b): a cold local model's FIRST inference —
  * prompt-processing the full tool-corpus prompt — emits no tokens for a long time and can exceed the
  * per-inference stall budget (`agents.<id>.promptTimeout.promptTimeoutMs`, default 180s), so the FIRST
  * user turn after a daemon (re)start aborts "request took too long" BEFORE any tool call. Warming at boot

package/node_modules/@comis/agent/dist/model/ollama-capacity-probe.js

@@ -33,12 +33,12 @@ export function deriveOllamaNativeBase(configuredBaseUrl) {
 /**
  * Minimum plausible served context window. Anything smaller is treated as a
  * bogus third-party value (e.g. a bad Modelfile `PARAMETER num_ctx`) and
- * rejected so it cannot shrink every turn's budget (IN-02, Phase 176 review).
+ * rejected so it cannot shrink every turn's budget.
  * No model Ollama serves runs below 512 tokens.
  */
 const MIN_PLAUSIBLE_SERVED_WINDOW = 512;
 /**
- * IN-02 input hardening: validate + clamp a raw `context_length` value from
+ * Input hardening: validate + clamp a raw `context_length` value from
  * an Ollama API response before it drives the budget reconcile.
  * Returns the FLOORED integer when the value is a finite number within the
  * sane range; undefined otherwise (caller falls through to the /api/show
@@ -112,7 +112,7 @@ export async function probeOllamaServedWindow(nativeBaseUrl, modelId, deps) {
     const matchingEntry = psBody.models.find((entry) => !modelId ||
         entry.model === modelId ||
         entry.name === modelId);
-    // IN-05: distinguish a PRESENT-but-rejected value from an absent field so
+    // Distinguish a PRESENT-but-rejected value from an absent field so
     // the final err can name the right lever (a bogus Modelfile num_ctx vs a
     // genuinely missing context_length).
     let rejectedImplausible = false;
@@ -170,8 +170,8 @@ export async function probeOllamaServedWindow(nativeBaseUrl, modelId, deps) {
     }
     if (rawDetailsContextLength !== undefined)
         rejectedImplausible = true;
-    // Both endpoints exhausted. IN-05 (Phase 176 review): branch presence vs
-    // absence — a value Ollama DID return but the IN-02 sanitization rejected
+    // Both endpoints exhausted. Branch presence vs absence — a value Ollama
+    // DID return but the sanitization rejected
     // (e.g. a typo'd Modelfile PARAMETER num_ctx) is bad input ("validation"),
     // and reporting it as "not found" would send the operator to restart a
     // server that is up and answering. A genuinely absent field keeps the
@@ -189,9 +189,9 @@ export async function probeOllamaServedWindow(nativeBaseUrl, modelId, deps) {
     });
 }
 /**
- * Single source for "which model did the probe use" (KNOB-01's served-window
- * comparator shares it — the 17fdd1e5 bug class was two sites deriving this
- * expression differently). ProviderEntrySchema has NO `defaultModel`; the
+ * Single source for "which model did the probe use" (the served-window
+ * comparator shares it — two sites deriving this expression independently
+ * is how they once disagreed). ProviderEntrySchema has NO `defaultModel`; the
  * model lives under `models[].id`. Falling back to the first configured model
  * lets the /api/show cold-start path (boot, before any inference, when
  * /api/ps is empty) send a real model name instead of "" (which Ollama
@@ -222,7 +222,7 @@ const PREWARM_TIMEOUT_MS = 300_000;
  * Fire-and-forget LOAD-ONLY warm-up: POST {nativeBaseUrl}/api/generate with the model + keep_alive and an
  * EMPTY prompt, which loads the model into memory WITHOUT generating (the Ollama preload idiom).
  *
- * Why (IMP-2a, package-delivery-20260628, local qwen3.6:35b): a cold local model's FIRST inference —
+ * Why (observed live with a local qwen3.6:35b): a cold local model's FIRST inference —
  * prompt-processing the full tool-corpus prompt — emits no tokens for a long time and can exceed the
  * per-inference stall budget (`agents.<id>.promptTimeout.promptTimeoutMs`, default 180s), so the FIRST
  * user turn after a daemon (re)start aborts "request took too long" BEFORE any tool call. Warming at boot
@@ -275,7 +275,7 @@ export async function probeAllOllamaProviders(params) {
             continue;
         const nativeBase = deriveOllamaNativeBase(entry.baseUrl ?? "http://localhost:11434");
         const modelId = resolveProbedModelId(entry);
-        // IMP-2a: fire-and-forget LOAD-ONLY warm-up (NOT awaited — must not block boot; the model loads in
+        // Fire-and-forget LOAD-ONLY warm-up (NOT awaited — must not block boot; the model loads in
         // the background so the first user turn runs warm). Non-fatal; detached from the probe `tasks`.
         if (prewarm) {
             // prewarmOllamaModel never throws; suppressError satisfies the no-empty-catch rule for the void promise.
@@ -293,11 +293,11 @@ export async function probeAllOllamaProviders(params) {
                 }, "Ollama served context window discovered");
             }
             else {
-                // W12 (obs-llm-troubleshooting): branch the hint by failure class. An
+                // Branch the hint by failure class. An
                 // HTTP status means Ollama responded — "start Ollama" points the
                 // operator away from the real cause (live: HTTP 400 from /api/show
                 // while the server was up; the model name/payload was the suspect).
-                // IN-05: errorKind "validation" means Ollama responded WITH a value
+                // errorKind "validation" means Ollama responded WITH a value
                 // that the sanitizer rejected as implausible — the lever is the
                 // Modelfile num_ctx, not the server, and not the probe opt-out.
                 const isHttpStatusFailure = result.error.message.startsWith("HTTP ");

package/node_modules/@comis/agent/dist/model/operation-model-defaults.d.ts

@@ -1,15 +1,15 @@
 /**
  * Operation model defaults: catalog-derived per-provider model tiering.
  *
- * Replaces the previous hardcoded `OPERATION_MODEL_DEFAULTS` table (which
- * pinned `mid`/`fast` model IDs per provider family) with a pure function
- * that reads the live pi-ai catalog at call time. Two design properties:
+ * A pure function reads the live pi-ai catalog at call time rather than
+ * consulting a hardcoded table of `mid`/`fast` model IDs per provider
+ * family. Two design properties:
  *
  * 1. Pi-ai SDK upgrades automatically light up new providers/models — no
- *    per-release source edits to bump `OPERATION_MODEL_DEFAULTS` literals.
- * 2. Closes the latent bug where switching primary to a non-Anthropic
- *    provider left cron/heartbeat/compaction routed to Claude Sonnet
- *    (because the old map was Anthropic/OpenAI/Google only).
+ *    per-release source edits to bump literal model IDs.
+ * 2. Every provider is covered uniformly: a hardcoded family map covering
+ *    only a few providers would leave cron/heartbeat/compaction routed to
+ *    Claude Sonnet when the primary provider is anything else.
  *
  * Tier picking: filter to text-capable models with non-zero cost, sort
  * ascending by total cost (input + output), pick top-of-cohort (lex-greatest

package/node_modules/@comis/agent/dist/model/operation-model-defaults.js

@@ -2,15 +2,15 @@
 /**
  * Operation model defaults: catalog-derived per-provider model tiering.
  *
- * Replaces the previous hardcoded `OPERATION_MODEL_DEFAULTS` table (which
- * pinned `mid`/`fast` model IDs per provider family) with a pure function
- * that reads the live pi-ai catalog at call time. Two design properties:
+ * A pure function reads the live pi-ai catalog at call time rather than
+ * consulting a hardcoded table of `mid`/`fast` model IDs per provider
+ * family. Two design properties:
  *
  * 1. Pi-ai SDK upgrades automatically light up new providers/models — no
- *    per-release source edits to bump `OPERATION_MODEL_DEFAULTS` literals.
- * 2. Closes the latent bug where switching primary to a non-Anthropic
- *    provider left cron/heartbeat/compaction routed to Claude Sonnet
- *    (because the old map was Anthropic/OpenAI/Google only).
+ *    per-release source edits to bump literal model IDs.
+ * 2. Every provider is covered uniformly: a hardcoded family map covering
+ *    only a few providers would leave cron/heartbeat/compaction routed to
+ *    Claude Sonnet when the primary provider is anything else.
  *
  * Tier picking: filter to text-capable models with non-zero cost, sort
  * ascending by total cost (input + output), pick top-of-cohort (lex-greatest
@@ -20,7 +20,7 @@
  *
  * @module
  */
-import { getModels, getProviders } from "@earendil-works/pi-ai";
+import { getModels, getProviders } from "@earendil-works/pi-ai/compat";
 // ---------------------------------------------------------------------------
 // Tier resolution
 // ---------------------------------------------------------------------------
@@ -40,10 +40,10 @@ function totalCost(m) {
  * claude-sonnet-4-5) the lex order matches recency.
  *
  * Why not just take `sortedAsc[idx]`? JavaScript's stable sort preserves
- * original-array order within a cost-tied block, so the previous algorithm
- * picked whatever the catalog happened to enumerate first. With 9 priced
- * Anthropic Sonnets all at $18/MTok, that was `claude-sonnet-4-5` — picked
- * by accident, not by quality signal.
+ * original-array order within a cost-tied block, so a bare index pick
+ * returns whatever the catalog happened to enumerate first. With 9 priced
+ * Anthropic Sonnets all at $18/MTok, that would be an arbitrary Sonnet —
+ * picked by accident, not by quality signal.
  *
  * Module-internal — not exported.
  */
@@ -133,10 +133,10 @@ export const OPERATION_TIER_MAP = {
     compaction: "fast",
     taskExtraction: "fast",
     condensation: "fast",
-    verification: "primary", // R4: local-only self-check; configured: cheap model via Level 2
-    planning: "primary", // R5: same resolution path; deferrable on M2
-    outcomeJudge: "fast", // OUTCOME-04: optional cost-gated outcome judge — a fast classification op
-    skillSynthesis: "mid", // SKILL-09: offline procedural synthesis — a mid-tier synthesis op (NOT a fast classify)
+    verification: "primary", // local-only self-check; configured: cheap model via Level 2
+    planning: "primary", // same resolution path as verification
+    outcomeJudge: "fast", // optional cost-gated outcome judge — a fast classification op
+    skillSynthesis: "mid", // offline procedural synthesis — a mid-tier synthesis op (NOT a fast classify)
 };
 /**
  * Default timeout per operation type in milliseconds.
@@ -152,9 +152,9 @@ export const OPERATION_TIMEOUT_DEFAULTS = {
     compaction: 60_000,
     taskExtraction: 30_000,
     condensation: 30_000,
-    verification: 120_000, // R4: same ceiling as the critic LLM_TIMEOUT_MS (Phase 154)
-    outcomeJudge: 30_000, // OUTCOME-04: mirrors condensation — a fast classification op
-    skillSynthesis: 150_000, // SKILL-09: mirrors cron — an offline batch op (Phase 201)
+    verification: 120_000, // same ceiling as the critic LLM_TIMEOUT_MS
+    outcomeJudge: 30_000, // mirrors condensation — a fast classification op
+    skillSynthesis: 150_000, // mirrors cron — an offline batch op
 };
 /**
  * Default cache retention hint per operation type.
@@ -168,8 +168,8 @@ export const OPERATION_CACHE_DEFAULTS = {
     taskExtraction: "none",
     condensation: "short",
     cron: "short", // 5m TTL: covers within-execution multi-step reuse, avoids 1h write premium across hourly runs.
-    verification: "none", // R4: critic responses consume the reviewed response and must not be cached (Phase 154)
-    planning: "none", // R5: planner responses are request-specific; caching wastes storage (Phase 154)
-    outcomeJudge: "none", // OUTCOME-04: judge responses are request-specific (like verification)
-    skillSynthesis: "short", // SKILL-09: offline reuse within a run, like cron (Phase 201)
+    verification: "none", // critic responses consume the reviewed response and must not be cached
+    planning: "none", // planner responses are request-specific; caching wastes storage
+    outcomeJudge: "none", // judge responses are request-specific (like verification)
+    skillSynthesis: "short", // offline reuse within a run, like cron
 };