comisai 1.0.45 → 1.0.46

package/node_modules/@comis/agent/dist/executor/executor-context-engine-setup.js

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: Apache-2.0
-// @allow-throw: the R1 (132-05) degrade-observability wrapper RE-THROWS the gate's
+// @allow-throw: the degrade-observability wrapper RE-THROWS the gate's
 // caught SummarizerDegradeError (after emitting a content-free WARN + dag_degraded)
 // so the leaf/condense ladder's existing catch floors to truncation-only — the
 // throw is a boundary re-raise that preserves the ladder contract, not control flow.
@@ -19,7 +19,7 @@
 import { ContextEngineConfigSchema, safePath, } from "@comis/core";
 import { createContextEngine } from "../context-engine/index.js";
 import { CHARS_PER_TOKEN_RATIO } from "../context-engine/constants.js";
-// Phase 129 (C1): the leaf-summarizer deps the afterTurn trigger consumes. The
+// The leaf-summarizer deps the afterTurn trigger consumes. The
 // production `summarize` seam wraps the SDK generateSummary; the model + key are
 // resolved via the SAME getCompactionDeps chain (no duplicate resolveProviderApiKey).
 import { buildLeafSummarizeFn, wrapSummarizerWithFailover, } from "../context-engine/lcd-leaf-summarizer.js";
@@ -32,16 +32,16 @@ import { getBreakpointIndex, getBreakpointIndexMapSize, getSessionLatches, } fro
 import { shouldDropSignedFields } from "./replay-drift-detector.js";
 import { readFileSync } from "node:fs";
 /**
- * R1 (132-05): wrap a GATED {@link LeafSummarizer} so a spend/breaker DEGRADE
+ * Wrap a GATED {@link LeafSummarizer} so a spend/breaker DEGRADE
  * bypass is OBSERVABLE at the wiring boundary (the breaker module itself is
  * content-free + log-free by design — observability lives where the `ComisLogger`
  * + `eventBus` are injected). On a {@link SummarizerDegradeError} we emit a
  * content-free WARN (`errorKind` `dependency` for breaker_open, `resource` for
- * spend_cap) + the `context:dag_degraded` event (reason `breaker_open`/`spend_cap`,
- * from 132-04), then RE-THROW so the leaf/condense ladder floors to truncation-only.
+ * spend_cap) + the `context:dag_degraded` event (reason `breaker_open`/`spend_cap`),
+ * then RE-THROW so the leaf/condense ladder floors to truncation-only.
  * Any non-degrade throw (an inner LLM failure the gate already recorded) and every
  * success pass straight through untouched. NEVER logs summary/message content —
- * ids / reason / durationMs only (AGENTS.md §2.2; T-132-05-04).
+ * ids / reason / durationMs only (AGENTS.md §2.2).
  */
 function wrapSummarizerWithDegradeObservability(gated, ctx) {
     return async (messages, opts) => {
@@ -95,7 +95,7 @@ function wrapSummarizerWithDegradeObservability(gated, ctx) {
  */
 export function setupContextEngine(params) {
     const { config, deps, formattedKey, tenantId, msg, sm, session, executionOverrides, cacheBreakDetector, contextEngineRef, getCachedSystemTokensEstimate, getCachedFreshTailPreambleTokens, getTokenAnchor, onAnchorReset, currentDiscoveryTracker, modelProfile, windowProvenance, } = params;
-    // DAG-CRIT-1: prefer the caller-supplied turn agentId (the positional
+    // Prefer the caller-supplied turn agentId (the positional
     // execute() arg threaded as params.agentId) over deps.agentId — on the
     // executeAgent path deps.agentId (= frozenDeps.agentId) is undefined, so this
     // is what makes the dag read scope == the ingest write scope (the assembler no
@@ -103,14 +103,15 @@ export function setupContextEngine(params) {
     // it directly. This `agentId` flows into the createContextEngine deps (the LCD
     // read scope) and the getSummarizerDeps wiring below.
     const agentId = params.agentId ?? deps.agentId;
-    // contextEngineOverrides removed from ExecutionOverrides -- compaction model resolved via operationModels chain
+    // ExecutionOverrides carries no per-execution context-engine override -- the
+    // compaction model is resolved via the operationModels chain below.
     const contextEngineConfig = config.contextEngine ?? ContextEngineConfigSchema.parse({});
-    // RETR-02/03 (Phase 173): resolve the relevance-first policy ONCE (the capability +
+    // Resolve the relevance-first policy ONCE (the capability +
     // supportsPromptCache gate; explicit > capability-default > off) and thread it to the
     // dag assembler, which CONSUMES the boolean (it does NOT recompute the gate). Resolved
     // here (mirrors prompt-assembly.ts's recall-side resolution) only when modelProfile is
     // present; absent ⇒ undefined ⇒ the assembler takes the verbatim recency path (frontier/
-    // mid byte-identical). The small/nano DEFAULT-ON flip is measurement-gated (VALIDATION.md).
+    // mid byte-identical).
     const relevanceFirst = modelProfile
         ? resolveScaffoldDefaults(modelProfile, config).relevanceFirst
         : undefined;
@@ -118,9 +119,8 @@ export function setupContextEngine(params) {
     // Memoized per-execute() so all pipeline runs in a single execute() see a
     // consistent decision (cleaner + scrubber must agree). The closure reads
     // the latest model identity each time (handles cycleModel mid-execute).
-    // Returns the identity/idle drift only — the kvl tool-defs dimension was
-    // removed in favor of the unconditional latest-message
-    // preserving scrub in signature-replay-scrubber.
+    // Returns the identity/idle drift only — tool-defs drift is covered by the
+    // unconditional latest-message-preserving scrub in signature-replay-scrubber.
     let memoizedDrift;
     const computeDriftIfNeeded = () => {
         if (memoizedDrift !== undefined)
@@ -166,13 +166,13 @@ export function setupContextEngine(params) {
     let signatureScrubsToolCallsAffected = 0;
     // Shared compaction-model resolution (getModel / getApiKey / overrideModel) —
     // the SINGLE source for both the pipeline `getCompactionDeps` (Layer 8) and the
-    // Phase-129 leaf `getSummarizerDeps`. Both compaction surfaces resolve the
+    // leaf `getSummarizerDeps`. Both compaction surfaces resolve the
     // SAME 5-level operation-model chain + route getApiKey through resolveProviderApiKey
-    // (no duplicate resolver — CLAUDE.md DRY / Plan 129-06 step 1). Returns the
+    // (no duplicate resolver). Returns the
     // getters + the optional override model+key.
     const resolveCompactionModelChain = (
-    // WR-04: when present, the chain uses this CAPTURED model snapshot verbatim
-    // instead of reading `session.agent.state.model`. The deferred (C4) compaction
+    // When present, the chain uses this CAPTURED model snapshot verbatim
+    // instead of reading `session.agent.state.model`. The deferred compaction
     // path captures it at the afterTurn boundary (before session.dispose()) so a
     // detached pass — which resolves the chain when it RUNS, possibly post-dispose
     // — never touches a torn-down session. Absent ⇒ the live per-call read (honors
@@ -189,12 +189,12 @@ export function setupContextEngine(params) {
                 reasoning: model?.reasoning ?? false,
             };
         },
-        // B-5: the REAL pi-ai Model<any> for the PRIMARY leaf/condense summarizer path
+        // The REAL pi-ai Model<any> for the PRIMARY leaf/condense summarizer path
         // (generateSummary needs a real Model, not the 4-field snapshot getModel
         // returns). It is the executor-resolved model (pi-executor.ts resolvedModel),
         // threaded in as params.resolvedModel. Captured at setup time (a closure over
-        // the setup-time value, resolved BEFORE session.dispose()), so the WR-04
-        // DEFERRED (C4) pass — which runs post-dispose — still returns a real Model
+        // the setup-time value, resolved BEFORE session.dispose()), so the
+        // DEFERRED pass — which runs post-dispose — still returns a real Model
         // without reading session.agent.state. Kept `unknown` end-to-end; the single
         // sanctioned `as any` cast lives at the generateSummary call site.
         getRealModel: () => params.resolvedModel,
@@ -206,12 +206,12 @@ export function setupContextEngine(params) {
             oauthManager: deps.oauthManager,
             agentConfig: config,
         }),
-        // INT-W1: the served window binding the PRIMARY (getRealModel) summarizer.
-        // windowProvenance.served is the Phase-176 pair ALREADY provider-gated by
+        // The served window binding the PRIMARY (getRealModel) summarizer.
+        // windowProvenance.served is the boot-probe pair ALREADY provider-gated by
         // the executor reconcile (`servedContextWindow.providerKey ===
-        // resolvedProviderKey` — pi-executor WR-02), so it binds exactly the model
+        // resolvedProviderKey` in pi-executor), so it binds exactly the model
         // params.resolvedModel holds. A plain setup-time number — dispose-safe on
-        // the WR-04 deferred path. Absent ⇒ no served truth (byte-identical).
+        // the deferred path. Absent ⇒ no served truth (byte-identical).
         ...(windowProvenance?.served !== undefined && {
             primaryServedWindow: windowProvenance.served,
         }),
@@ -230,10 +230,10 @@ export function setupContextEngine(params) {
                 try {
                     const compactionModel = deps.modelRegistry.find(compactionResolution.provider, compactionResolution.modelId);
                     if (compactionModel) {
-                        // INT-W1: gate the Phase-176 served pair for THIS override in
+                        // Gate the boot-probe served pair for THIS override in
                         // CONFIG space — the pair's providerKey and the operation-model
-                        // resolution's provider are both `providers.entries` keys (WR-02:
-                        // never compare model.provider, which the registry alias fallback
+                        // resolution's provider are both `providers.entries` keys (never
+                        // compare model.provider, which the registry alias fallback
                         // can rename). A cross-provider override gets NO served bound.
                         const overrideServedWindow = deps.servedContextWindow !== undefined &&
                             deps.servedContextWindow.providerKey === compactionResolution.provider
@@ -261,31 +261,31 @@ export function setupContextEngine(params) {
             return {};
         })(),
     });
-    // The Phase-129 leaf-summarizer deps getter: the shared model chain + the
-    // production summarizer seam (buildLeafSummarizeFn wraps the SDK generateSummary;
-    // Phase 132 swaps it for a spend-governed variant) + the injected logger. The
+    // The leaf-summarizer deps getter: the shared model chain + the
+    // production summarizer seam (buildLeafSummarizeFn wraps the SDK generateSummary,
+    // gated by the spend breaker below) + the injected logger. The
     // afterTurn trigger (executor-post-execution.ts → runLeafPassAfterTurn) calls
     // this; when wired the leaf pass fires live over threshold. Resolved fresh per
     // call so model cycling mid-session is honored (same as getCompactionDeps).
     //
-    // WR-04: a `modelSnapshot` (when supplied by the DEFERRED compaction path)
+    // A `modelSnapshot` (when supplied by the DEFERRED compaction path)
     // flows into the chain so BOTH `getModel` AND the `buildLeafSummarizeFn`-internal
     // model read use the captured value — a deferred pass resolving this AFTER
     // `session.dispose()` then never reads `session.agent.state`.
     const getSummarizerDeps = (modelSnapshot) => {
         const chain = resolveCompactionModelChain(modelSnapshot);
         const inner = buildLeafSummarizeFn(chain);
-        // SUM-03: wrap the primary summarizer with the ordered failover list before the
+        // Wrap the primary summarizer with the ordered failover list before the
         // spend-breaker. The failover list tries each provider in sequence; only when ALL
         // providers are exhausted does it throw — at which point the outer breaker records
-        // exactly ONE failure (Pitfall 4: throw-last, not throw-per-provider). Empty list
+        // exactly ONE failure (throw-last, not throw-per-provider). Empty list
         // (default) = zero behavioral change for existing deployments.
         const fallbackProviders = contextEngineConfig.compaction?.summarizerFallbackProviders ?? [];
         const fallbackSummarizers = fallbackProviders.map((providerModel) => {
             // Each entry is "provider:modelId". Build a LeafSummarizer using the same
             // model resolution pattern as strongerSummarizerModel (resolveCompactionModelChain
             // with an override), falling back to the primary chain for unknown providers.
-            // For Phase 171, all fallback entries resolve via the existing registry lookup;
+            // All fallback entries resolve via the existing registry lookup;
             // if the model is not found in the registry, the fallback silently uses the
             // primary chain (the outer failover wrapper will still catch any resulting error).
             const [provider, modelId] = providerModel.split(":", 2);
@@ -310,13 +310,14 @@ export function setupContextEngine(params) {
             return buildLeafSummarizeFn(chain);
         });
         const innerWithFailover = wrapSummarizerWithFailover(inner, fallbackSummarizers, deps.logger);
-        // R1 (132-05): wrap the leaf summarizer seam with the daemon-owned per-tenant
+        // Wrap the leaf summarizer seam with the daemon-owned per-tenant
         // spend+breaker gate keyed on the LIVE tenantId (the SAME tenant the afterTurn
         // ingest scope uses). On open-breaker / over-cap the gate THROWS the degrade
         // signal → the leaf/condense ladder catches it (lcd-leaf-summarizer tryLevel)
-        // → deterministic Level-3 floor → truncation-only assembly → R2's Task-1
+        // → deterministic Level-3 floor → truncation-only assembly → the
         // fallback marker fires on the resulting fallback:true summary. NO retry of
-        // the inner LLM (the RESEARCH anti-pattern). The same getter feeds BOTH the
+        // the inner LLM (retrying a spend-gated call would defeat the gate). The
+        // same getter feeds BOTH the
         // leaf and the condense pass, so condense degrades identically. Absent breaker
         // ⇒ the raw seam (non-daemon callers / tests).
         const summarize = deps.summarizerSpendBreaker
@@ -327,11 +328,11 @@ export function setupContextEngine(params) {
             logger: deps.logger,
             summarize,
             getModel: chain.getModel,
-            // B-5: the REAL primary Model<any> (params.resolvedModel) buildLeafSummarizeFn
+            // The REAL primary Model<any> (params.resolvedModel) buildLeafSummarizeFn
             // hands generateSummary — not the 4-field snapshot getModel returns.
             getRealModel: chain.getRealModel,
             getApiKey: chain.getApiKey,
-            // INT-W1: overrideModel carries its own provider-gated servedWindow;
+            // overrideModel carries its own provider-gated servedWindow;
             // primaryServedWindow binds the getRealModel candidate. Both feed
             // resolveSummarizerWindowTokens' candidate-bound served selection.
             overrideModel: chain.overrideModel,
@@ -343,8 +344,8 @@ export function setupContextEngine(params) {
         eventBus: deps.eventBus,
         agentId,
         sessionKey: formattedKey,
-        // R4 (132-03): the dag assembler builds an agent + tenant scoped LCD read
-        // from agentId + tenantId + conversationId (the WR-02 close).
+        // The dag assembler builds an agent + tenant scoped LCD read
+        // from agentId + tenantId + conversationId.
         tenantId,
         getModel: () => {
             // Lazy model getter handles model cycling mid-session
@@ -365,7 +366,7 @@ export function setupContextEngine(params) {
         objective: executionOverrides?.spawnPacket?.objective, // Objective reinforcement
         getSystemTokensEstimate: getCachedSystemTokensEstimate,
         getFreshTailPreambleTokensEstimate: getCachedFreshTailPreambleTokens,
-        // G-09: Notify cache break detector when observation masking modifies content
+        // Notify cache break detector when observation masking modifies content
         onContentModified: () => cacheBreakDetector.notifyContentModification(formattedKey),
         // Accumulate signature-replay scrub counts per-execute. Only
         // counts emissions that actually scrubbed something (zero-touch turns
@@ -397,7 +398,7 @@ export function setupContextEngine(params) {
             return undefined; // Use default keepTurns
         },
         // LLM compaction deps (Layer 8). The getModel / getApiKey / overrideModel
-        // resolution is shared with the Phase-129 leaf getSummarizerDeps via
+        // resolution is shared with the leaf getSummarizerDeps via
         // resolveCompactionModelChain() above — one source for both compaction
         // surfaces (the 5-level operation-model chain + the shared resolveProviderApiKey
         // dispatch; no duplicate resolver). The override is set only when the
@@ -504,14 +505,14 @@ export function setupContextEngine(params) {
                 deps.eventBus?.emit("context:overflow", {
                     agentId: agentId ?? config.name,
                     sessionKey: formattedKey,
-                    contextTokens: Math.ceil(stats.contextChars / CHARS_PER_TOKEN_RATIO), // flat-by-design: aggregate stats logging (TOK-01)
-                    budgetTokens: Math.ceil(stats.budgetChars / CHARS_PER_TOKEN_RATIO), // flat-by-design: aggregate stats logging (TOK-01)
+                    contextTokens: Math.ceil(stats.contextChars / CHARS_PER_TOKEN_RATIO), // flat-by-design: aggregate stats logging
+                    budgetTokens: Math.ceil(stats.budgetChars / CHARS_PER_TOKEN_RATIO), // flat-by-design: aggregate stats logging
                     recoveryAction: stats.recoveryAction,
                     timestamp: deps.clock.now(),
                 });
             },
         }),
-        // Phase 128 dag-mode: thread the injected LCD store + the conversation id
+        // dag-mode: thread the injected LCD store + the conversation id
         // (= formattedKey) so context-engine's `dag` branch returns the LCD
         // assembler, which reconstructs history from what the afterTurn ingest
         // wrote. Absent ⇒ the branch WARN-falls-through to the pipeline.
@@ -520,30 +521,30 @@ export function setupContextEngine(params) {
         // The dag assembler stamps assembly duration + synthesized-tool-result
         // timestamps via this injected clock (production never calls Date.now()).
         clock: deps.clock,
-        // C1 (Phase 165): the resolved ModelProfile for budget-aware eviction cap.
+        // The resolved ModelProfile for budget-aware eviction cap.
         // Absent ⇒ lcd-assembler applies the fail-closed nano cap + WARN.
         modelProfile,
-        // KNOB-02: served-window provenance for the budget's rawContextWindowTokens.
+        // Served-window provenance for the budget's rawContextWindowTokens.
         windowProvenance,
-        // RETR-02/03 (Phase 173): the resolved relevance-first policy + the shared scorer for
+        // The resolved relevance-first policy + the shared scorer for
         // the margin arbiter. relevanceFirst undefined/false ⇒ the assembler takes the verbatim
-        // recency path (frontier/mid byte-identical, LOCKED #2).
+        // recency path (frontier/mid byte-identical).
         relevanceFirst,
-        // RETR-02 / DEPTH-01: the injected shared relevance scorer. As of Phase 174 it has a LIVE
-        // caller — the DEPTH-01 within-history middle-band relevance pass (rankMiddleBandByRelevance,
-        // injected as marginArbitrate's middleBandRanker in lcd-arbiter-seam) calls it over the
+        // The injected shared relevance scorer. Its live
+        // caller is the within-history middle-band relevance pass (rankMiddleBandByRelevance,
+        // injected as marginArbitrate's middleBandRanker in lcd-arbiter-seam), which calls it over the
         // FTS-the-band lane to re-rank the evictable middle band cache-safely. It is injected here
-        // (executor/ may import rag/) so the context-engine never imports rag/ (the I2 cut). The
-        // cross-tier LTM/KG lanes remain EMPTY on the C2 assembly path, so the scorer's live use is
+        // (executor/ may import rag/) so the context-engine never imports rag/ (the import cut). The
+        // cross-tier LTM/KG lanes remain EMPTY on the assembly path, so the scorer's live use is
         // the middle-band pass, not the cross-session tiers.
         relevanceScorer: scoreRelevance,
-        // Phase 166 T-S4: security-pin markers so the dag eviction never drops security context.
+        // Security-pin markers so the dag eviction never drops security context.
         securityPinMarkers: params.securityPinMarkers,
         onAssembledInputTokens: params.onAssembledInputTokens,
         onEffectiveWindow: params.onEffectiveWindow,
         onThinkingDownshifted: params.onThinkingDownshifted,
         getThinkingLevel: params.getThinkingLevel,
-        // WR-02 (Phase 166): thread the operator-configurable floor so the schema field
+        // Thread the operator-configurable floor so the schema field
         // contextEngine.budget.minVisibleOutputTokens has live runtime effect.
         // Absent ⇒ pre-flight uses the compile-time constant (768) — byte-identical.
         minVisibleOutputTokens: contextEngineConfig.budget?.minVisibleOutputTokens,
@@ -576,7 +577,7 @@ export function setupContextEngine(params) {
             signatureScrubs,
             signatureScrubsToolCallsAffected,
         }),
-        // Expose the leaf-summarizer deps getter (Phase 129) so pi-executor can
+        // Expose the leaf-summarizer deps getter so pi-executor can
         // thread it into postExecution's deps.getSummarizerDeps — wiring the
         // afterTurn leaf pass live.
         getSummarizerDeps,

package/node_modules/@comis/agent/dist/executor/executor-input-guard.d.ts

@@ -13,7 +13,7 @@
 import { type SessionKey, type NormalizedMessage, type TypedEventBus, type InputValidationResult, type InputSecurityGuard, type InjectionRateLimiter, type ClockPort } from "@comis/core";
 import type { ComisLogger } from "@comis/core";
 /**
- * GIANT-INPUT-WEDGE guard (30uc-20260624 UC-09): the default maximum inbound user-message
+ * GIANT-INPUT-WEDGE guard: the default maximum inbound user-message
  * size, in characters, processed by a single turn. A message larger than this is rejected
  * honestly BEFORE the jailbreak scan + the downstream tokenize / LCD-ingest / embed path —
  * a multi-MB message otherwise drives O(words) synchronous work (the scan's typoglycemia

package/node_modules/@comis/agent/dist/executor/executor-input-guard.js

@@ -13,7 +13,7 @@
  */
 import { formatSessionKey, } from "@comis/core";
 /**
- * GIANT-INPUT-WEDGE guard (30uc-20260624 UC-09): the default maximum inbound user-message
+ * GIANT-INPUT-WEDGE guard: the default maximum inbound user-message
  * size, in characters, processed by a single turn. A message larger than this is rejected
  * honestly BEFORE the jailbreak scan + the downstream tokenize / LCD-ingest / embed path —
  * a multi-MB message otherwise drives O(words) synchronous work (the scan's typoglycemia
@@ -35,7 +35,7 @@ export function validateInput(params) {
     const { msg, sessionKey, agentId, inputValidator, inputGuard, rateLimiter, eventBus, logger, clock } = params;
     const maxInputChars = params.maxInputChars ?? DEFAULT_MAX_INPUT_CHARS;
     let safetyReinforcement;
-    // GIANT-INPUT-WEDGE (30uc UC-09): reject an over-cap message BEFORE the jailbreak scan AND the
+    // GIANT-INPUT-WEDGE: reject an over-cap message BEFORE the jailbreak scan AND the
     // downstream tokenize/LCD-ingest/embed. A multi-MB message drives O(words) sync work that blocks
     // the Node event loop for minutes (the whole daemon freezes). This early-return keeps the gateway
     // responsive and the failure HONEST + reason-coded; the context engine handles within-cap inputs.
@@ -76,7 +76,7 @@ export function validateInput(params) {
             // Original msg preserved (immutable NormalizedMessage), sanitized used for display/audit only
         }
     }
-    // : Jailbreak scoring
+    // Jailbreak scoring
     if (inputGuard) {
         const guardResult = inputGuard.scan(msg.text);
         // Block if configured and high risk