comisai 1.0.44 → 1.0.46

package/node_modules/@comis/agent/dist/executor/scaffold-defaults.js

@@ -1,10 +1,11 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * Capability-gated reliability defaults resolver (Phase 158).
+ * Capability-gated reliability defaults resolver.
  *
- * SD1/SD2/SD3: for small/nano models, resolves capability defaults
+ * For small/nano models, resolves capability defaults
  * for GoalAnchor, rag.baseFloor, and verification critic.
- * frontier/mid: returns all OFF (behavior-neutral, byte-identical to v2.14).
+ * frontier/mid: returns all OFF (behavior-neutral — identical to running
+ * without this resolver).
  *
  * Precedence (load-bearing): explicit per-agent config > capability default > off.
  * Implementation: `resolved = explicit ?? capabilityDefault(class)`.
@@ -18,9 +19,9 @@ import { resolveOperationModel, resolveProviderFamily } from "../model/operation
 // Named constants
 // ---------------------------------------------------------------------------
 /**
- * Conservative base-score floor for small/nano models (SD2).
+ * Conservative base-score floor for small/nano models.
  *
- * Proven by Phase-153 poison fixture: blocks recall entries with base≤0.12
+ * Proven by the poison-recall fixture: blocks recall entries with base≤0.12
  * (CASE A in memory-recall-floor.test.ts: base=0.12 < 0.15 → dropped by
  * passesBaseFloor gate). Keeps entries with base≥0.15 (legit recall).
  *
@@ -31,18 +32,18 @@ import { resolveOperationModel, resolveProviderFamily } from "../model/operation
  */
 const SMALL_NANO_DEFAULT_BASE_FLOOR = 0.15;
 /**
- * Keyless providers where the verification critic can auto-enable (SD3 cost-gate).
+ * Keyless providers where the verification critic can auto-enable (cost-gate).
  *
  * Mirrors KEYLESS_CRITIC_PROVIDERS in verification-gate.ts — kept separate to
- * avoid a circular import between scaffold-defaults and verification-gate.
- * // Parity with verification-gate.ts KEYLESS_CRITIC_PROVIDERS is asserted in Plan 02 (cross-file).
+ * avoid a circular import between scaffold-defaults and verification-gate;
+ * parity between the two sets is asserted by a cross-file test.
  *
- * WR-02: a cloud-keyed provider (anthropic, openai) can never trigger the
- * auto-critic default — the KEYLESS_CRITIC_PROVIDERS check is the primary barrier.
- * SD3 Test 11 proves frontier→false regardless.
+ * A cloud-keyed provider (anthropic, openai) can never trigger the
+ * auto-critic default — the KEYLESS_CRITIC_PROVIDERS check is the primary barrier,
+ * and the capability gate independently forces frontier→false.
  */
 const KEYLESS_CRITIC_PROVIDERS = new Set(["ollama", "lm-studio"]);
-/** SD6: per-file bootstrap.maxChars default for small/nano.
+/** Per-file bootstrap.maxChars default for small/nano.
  *  At 3_500, AGENTS.md (6,780 chars) truncates to head(2,450)+tail(700);
  *  all other workspace files (≤2,844 chars) are included whole.
  *  Rationale: tames the observed 495% bootstrap-budget warning without truncating
@@ -53,27 +54,27 @@ const SMALL_NANO_DEFAULT_BOOTSTRAP_MAX_CHARS = 3_500;
  *  receives exactly 20_000 from the parsed PerAgentConfig.
  *  Detection: configuredMaxChars === BOOTSTRAP_MAX_CHARS_SENTINEL → apply capability default.
  *  NOTE: an operator who explicitly sets bootstrap.maxChars:20000 also triggers the sentinel;
- *  document in D2 that to force 20_000 on a small model, use capabilityClassOverride:"frontier". */
+ *  to force 20_000 on a small model, use capabilityClassOverride:"frontier". */
 const BOOTSTRAP_MAX_CHARS_SENTINEL = 20_000;
-/** SD7/F1: active-tool ceiling for the `small` capability class.
+/** Active-tool ceiling for the `small` capability class.
  *  Keeps CORE_TOOLS (~15) + recently-used head-room + warm tools active;
  *  defers cold long-tail behind discover_tools (no capability removal).
- *  Retuned F1 from 40→24: 15 CORE_TOOLS + 9 discretionary slots.
+ *  24 = 15 CORE_TOOLS + 9 discretionary slots.
  *  At 24 tools × ~300 chars avg ≈ 7K chars (vs 12K at 40) — ~750–1250 tokens saved. */
 const SMALL_DEFAULT_ACTIVE_TOOL_CEILING = 24;
-/** SD9/F2: total bootstrap budget for small/nano. Caps the SUM of all bootstrap file chars.
+/** Total bootstrap budget for small/nano. Caps the SUM of all bootstrap file chars.
  *  Per-file cap (3_500) still applies as an upper bound per file; this cap is the aggregate.
  *  Rationale: only AGENTS.md (6780) exceeds 3500 so per-file trimming only trims that one
  *  file — total bootstrap stays ~12.9K. A total cap of 5_000 forces proportional reduction. */
 const SMALL_NANO_DEFAULT_BOOTSTRAP_TOTAL_CHARS = 5_000;
-/** v2.19: per-tool-result char cap for the `small` class.
+/** Per-tool-result char cap for the `small` class.
  *  The schema default (50_000 chars ≈ 12.5K tokens) lets ONE tool result eat ~40% of a
- *  32K-token window — two exhaust it. Live, NVDA analysts hit context_exhausted at
+ *  32K-token window — two exhaust it. Observed live: sessions hit context_exhausted at
  *  assembled ~33-35K because web_search/read results were 20-35K chars each. 12_000 chars
  *  (~3K tokens) lets a small model hold ~7 results in its working space; the offload
  *  guard keeps the full content on disk so nothing is permanently lost. */
 const SMALL_DEFAULT_MAX_TOOL_RESULT_CHARS = 12_000;
-/** v2.19: tighter per-tool-result char cap for `nano` (16K window) — ~2K tokens. */
+/** Tighter per-tool-result char cap for `nano` (16K window) — ~2K tokens. */
 const NANO_DEFAULT_MAX_TOOL_RESULT_CHARS = 8_000;
 /** The schema .default() for maxToolResultChars (50_000). Doubles as the "unset" sentinel,
  *  exactly like BOOTSTRAP_MAX_CHARS_SENTINEL: a small model that explicitly sets 50_000 still
@@ -87,32 +88,32 @@ const MAX_TOOL_RESULT_CHARS_SENTINEL = 50_000;
  *
  * Pure — no I/O, no side effects, deterministic for equal inputs.
  *
- * @param modelProfile - Immutable profile resolved once per execution (Phase 151).
+ * @param modelProfile - Immutable profile resolved once per execution.
  * @param config - Per-agent config (post top-level parse; sub-blocks may be undefined).
- * @param criticContext - Optional cost-gate inputs for SD3 (provider, agentModel, operationModels).
+ * @param criticContext - Optional cost-gate inputs for the verification critic (provider, agentModel, operationModels).
  *   Pass undefined when the critic wiring is not available at the call site.
  */
 export function resolveScaffoldDefaults(modelProfile, config, criticContext) {
     // Determine the capability tier — scaffoldLevel==="max" iff capabilityClass ∈ {small, nano}
     const isSmallNano = modelProfile.scaffoldLevel === "max";
     // -------------------------------------------------------------------------
-    // SD1: GoalAnchor — explicit ?? capability default
+    // GoalAnchor — explicit ?? capability default
     //
     // config.goalAnchor?.enabled is `boolean | undefined` (the block is .optional()
     // at agent level — an unconfigured agent yields config.goalAnchor === undefined).
     // Do NOT re-parse through GoalAnchorConfigSchema — that collapses undefined→false
-    // (Pitfall 1: the schema .default(false) fires and kills the ?? gate).
+    // (the schema-re-parse trap: the schema .default(false) fires and kills the ?? gate).
     // The ?? expression fires only when the left side is `undefined` — explicit `false`
     // is preserved (the load-bearing force-off operator path).
     // -------------------------------------------------------------------------
     const goalAnchorEnabled = config.goalAnchor?.enabled ?? isSmallNano;
     // -------------------------------------------------------------------------
-    // SD2: rag.baseFloor — explicit (non-zero) wins; 0 is the "unset" sentinel
+    // rag.baseFloor — explicit (non-zero) wins; 0 is the "unset" sentinel
     //
     // config.rag is always a fully-defaulted RagConfig (schema .default({})),
     // so config.rag.baseFloor is always `number`. The schema default is 0.
     // `> 0` treats both the schema default (0) and an explicit `rag.baseFloor: 0`
-    // as "unset" — documented in D1.
+    // as "unset".
     // Defensive: tests may pass a cast PerAgentConfig without a fully-parsed rag block.
     // In production the top-level parse always provides config.rag (schema .default({})).
     // -------------------------------------------------------------------------
@@ -121,22 +122,22 @@ export function resolveScaffoldDefaults(modelProfile, config, criticContext) {
         ? configuredBaseFloor
         : isSmallNano
             ? SMALL_NANO_DEFAULT_BASE_FLOOR // LOAD-BEARING under the arbiter (relevanceFirst): the
-            : 0; // recall gate enforces THIS floor, not a silent 0 (WR-02).
+            : 0; // recall gate enforces THIS floor, not a silent 0.
     // -------------------------------------------------------------------------
-    // RETR-03 / RETR-04 / WR-02 (Phase 173): relevanceFirst — the full policy gate
+    // relevanceFirst — the full policy gate
     // (the unified-arbiter-active signal AND the relevance-first vs recency-first policy).
     //
-    // Copies the SD1 capability-gate shape EXACTLY: explicit config wins (both ways),
-    // else the capability default. The gate axis is supportsPromptCache (design §14.1),
+    // Copies the GoalAnchor capability-gate shape EXACTLY: explicit config wins (both ways),
+    // else the capability default. The gate axis is supportsPromptCache,
     // NOT a provider-string predicate: small/nano on a non-caching model (typical Ollama)
     // reorders relevance-first for free; a caching model pays a cache-break so it stays
     // recency-first (false) below the fence — frontier (Anthropic, caches) → false; mid
-    // (caching) → false. Frontier/mid recency-first is BYTE-IDENTICAL (LOCKED #2): the
-    // arbiter does not run for them. The small/nano DEFAULT-ON flip stays MEASUREMENT-GATED
-    // (the Phase-171 harness) — this resolves the mechanism; the live flip is an operator step.
+    // (caching) → false. Frontier/mid recency-first is BYTE-IDENTICAL to running without
+    // this resolver: the arbiter does not run for them. The small/nano default-ON flip
+    // stays measurement-gated — this resolves the mechanism; the live flip is an operator step.
     //
-    // Pitfall 1 (the schema re-parse trap): read the relevance.firstByDefault field DIRECTLY
-    // via the optional chain below — the schema field (173-03) is OPTIONAL with NO `.default()`,
+    // The schema-re-parse trap: read the relevance.firstByDefault field DIRECTLY
+    // via the optional chain below — the schema field is OPTIONAL with NO `.default()`,
     // so an omitted field stays `undefined` and the `??` falls through to the capability gate.
     // Do NOT re-parse the sub-block through its schema (a `.default()` would collapse
     // undefined→false and silently kill the capability gate). An explicit boolean (true OR
@@ -145,15 +146,15 @@ export function resolveScaffoldDefaults(modelProfile, config, criticContext) {
     const relevanceFirst = config.contextEngine?.relevance?.firstByDefault ??
         (isSmallNano && !modelProfile.supportsPromptCache);
     // -------------------------------------------------------------------------
-    // SD3: verification — explicit wins; otherwise cost-gated capability default
+    // verification — explicit wins; otherwise cost-gated capability default
     //
     // config.verification?.enabled is `boolean | undefined` (block is .optional()).
     // Do NOT re-parse through VerificationConfigSchema.
     //
-    // WR-01: resolve the verification operation model ONCE. It drives BOTH the
+    // Resolve the verification operation model ONCE. It drives BOTH the
     // cost-gate distinctness decision AND the model the critic actually runs on
     // (criticModel below) — so the critic can never auto-enable on a cheap model's
-    // existence yet silently run the primary (the bug WR-01 fixed).
+    // existence yet silently run the primary.
     // Uses the real 5-level priority chain from resolveOperationModel — no mocking.
     // -------------------------------------------------------------------------
     const verificationResolution = criticContext
@@ -186,10 +187,10 @@ export function resolveScaffoldDefaults(modelProfile, config, criticContext) {
                 verificationResolution.modelId !== criticContext.agentModel;
     }
     else {
-        // frontier/mid → always off (SD5); or no criticContext → no cost-gate check → off
+        // frontier/mid → always off; or no criticContext → no cost-gate check → off
         verificationEnabled = false;
     }
-    // WR-01: when the critic runs, it MUST run on the resolved verification model
+    // When the critic runs, it MUST run on the resolved verification model
     // (the distinct cheap model the cost-gate gated on, or the operator's configured
     // verification model in the explicit force-on path) — NEVER silently the agent
     // primary. The synthetic critic deps run keyless (apiKey:""), so expose the
@@ -202,10 +203,10 @@ export function resolveScaffoldDefaults(modelProfile, config, criticContext) {
         ? { provider: verificationResolution.provider, modelId: verificationResolution.modelId }
         : undefined;
     // -------------------------------------------------------------------------
-    // SD6: bootstrap.maxChars capability default.
+    // bootstrap.maxChars capability default.
     // bootstrap uses .default() not .optional() — config.bootstrap.maxChars is always a number.
     // The "unset" detection is === BOOTSTRAP_MAX_CHARS_SENTINEL (not ?? undefined).
-    // See RESEARCH.md Pitfall 1 and the SD2 === 0 sentinel analog in this file.
+    // Same shape as the rag.baseFloor === 0 sentinel above.
     // -------------------------------------------------------------------------
     const configuredMaxChars = config.bootstrap?.maxChars ?? BOOTSTRAP_MAX_CHARS_SENTINEL;
     const bootstrapMaxChars = configuredMaxChars !== BOOTSTRAP_MAX_CHARS_SENTINEL
@@ -214,7 +215,7 @@ export function resolveScaffoldDefaults(modelProfile, config, criticContext) {
             ? SMALL_NANO_DEFAULT_BOOTSTRAP_MAX_CHARS // small/nano capability default
             : BOOTSTRAP_MAX_CHARS_SENTINEL; // frontier/mid: 20_000 (unchanged)
     // -------------------------------------------------------------------------
-    // SD7/F1: active-tool ceiling — ONLY for small class.
+    // Active-tool ceiling — ONLY for small class.
     // nano is excluded: it already has aggressive CORE_TOOLS-only deferral at applyToolDeferral:413.
     // Adding a ceiling of 24 on top of nano's ~15-active output would be a no-op, and confusing.
     // frontier/mid: no ceiling (behavior-neutral).
@@ -223,15 +224,15 @@ export function resolveScaffoldDefaults(modelProfile, config, criticContext) {
         ? SMALL_DEFAULT_ACTIVE_TOOL_CEILING
         : undefined;
     // -------------------------------------------------------------------------
-    // SD9/F2: total bootstrap budget — small/nano only.
+    // Total bootstrap budget — small/nano only.
     // Applied in buildBootstrapContextFiles as a second pass after per-file truncation.
-    // frontier/mid: undefined (no total cap — byte-identical to v2.14).
+    // frontier/mid: undefined (no total cap — behavior-neutral).
     // -------------------------------------------------------------------------
     const bootstrapTotalMaxChars = isSmallNano
         ? SMALL_NANO_DEFAULT_BOOTSTRAP_TOTAL_CHARS
         : undefined;
     // -------------------------------------------------------------------------
-    // v2.19: per-tool-result char cap — small/nano only, when unset (=== 50_000 sentinel).
+    // Per-tool-result char cap — small/nano only, when unset (=== 50_000 sentinel).
     // Returns a number ONLY when the scaffold wants to override; undefined means "use
     // config.maxToolResultChars as-is" (frontier/mid, or an explicit operator value).
     // -------------------------------------------------------------------------

package/node_modules/@comis/agent/dist/executor/session-health-rollup.d.ts

@@ -1,13 +1,18 @@
 import type { ErrorKind } from "@comis/core";
 /**
- * F1 session-health rollup: the five health fields persisted onto
+ * Session-health rollup: the five health fields persisted onto
  * `SessionMetadata.sessionEnd` and emitted on `session:summary`.
  *
  * - `degraded`        — the run finished in a non-clean state (tool errors,
  *                       budget/breaker/provider trips, or a hard error). Derived
  *                       from the mapped `endReason` (≠ "success"), the SAME source
  *                       of truth as the co-persisted `sessionEnd.endReason`.
- * - `costUsd`         — cumulative USD cost for the session.
+ * - `costUsd`         — THIS execution's USD cost. The rollup is emitted once
+ *                       per execution, so a session's total is the SUM of its
+ *                       executions' rollups — the fleet aggregate and the
+ *                       explain assembler both sum; the last-written
+ *                       `sessionEnd` alone under-reports a multi-execution
+ *                       session.
  * - `toolStats`       — per-tool {ok, failed} counts; bounded by the distinct
  *                       tool count of one execution (small).
  * - `breakerTripCount`— how many times a tool circuit breaker opened.
@@ -28,11 +33,11 @@ export interface SessionHealthRollup {
 }
 /**
  * The narrow structural slice of `bridgeResult` this reduce consumes — NOT the
- * whole `ExecutionResult`. After Phase 152-01, `toolExecResults[i]` carries the
- * optional `errorKind`, and `breakerTripCount` is accumulated on the bridge.
+ * whole `ExecutionResult`. `toolExecResults[i]` carries the optional
+ * `errorKind`, and `breakerTripCount` is accumulated on the bridge.
  */
 interface RollupInput {
-    sessionCostUsd?: number;
+    executionCostUsd?: number;
     breakerTripCount?: number;
     toolExecResults?: ReadonlyArray<{
         toolName: string;
@@ -41,11 +46,11 @@ interface RollupInput {
     }>;
 }
 /**
- * Pure reduce over accumulated bridge state → the five F1 health fields.
+ * Pure reduce over accumulated bridge state → the five health fields.
  *
  * Same inputs always produce the same output: no reads, no writes, no time
  * source, no event emission. The fire-and-forget persist/emit guards live at the
- * post-execution call site (Plan 04).
+ * post-execution call site.
  *
  * @param bridgeResult - the narrow bridge slice (cost, breaker trips, per-tool results).
  * @param endReason - the ALREADY-MAPPED `SessionMetadata.sessionEnd.endReason`

package/node_modules/@comis/agent/dist/executor/session-health-rollup.js

@@ -5,23 +5,23 @@
  * (executor-post-execution.ts — the one authoritative table) and passes that
  * SAME mapped value here, so `degraded` and `endReason` can never disagree.
  *
- * This replaces the prior dual closed sets (DEGRADED_REASONS +
- * ERROR_CLASS_REASONS) that were hand-maintained separately from
- * `END_REASON_MAP` and silently diverged: `loop_detected` and `session_reset`
- * mapped to `endReason:"error"` via the map's fallthrough yet were in NEITHER
- * set, recording a runaway-loop / session-reset abort as `degraded:false`
- * (Phase 152 CR-01). Deriving from the mapped `endReason` removes the second
- * domain entirely, so a newly added finish reason cannot reopen the divergence.
+ * Do NOT reintroduce a separate hand-maintained set of degraded finish
+ * reasons: any set kept apart from `END_REASON_MAP` silently diverges (e.g.
+ * `loop_detected` and `session_reset` map to `endReason:"error"` via the
+ * map's fallthrough, so a set that omits them would record a runaway-loop /
+ * session-reset abort as `degraded:false`). Deriving from the mapped
+ * `endReason` removes the second domain entirely, so a newly added finish
+ * reason cannot open a divergence.
  */
 const CLEAN_END_REASONS = new Set(["success"]);
 /** How many distinct ErrorKinds the rollup keeps — hard cap, DoS-bounded. */
 const TOP_ERROR_KINDS_CAP = 3;
 /**
- * Pure reduce over accumulated bridge state → the five F1 health fields.
+ * Pure reduce over accumulated bridge state → the five health fields.
  *
  * Same inputs always produce the same output: no reads, no writes, no time
  * source, no event emission. The fire-and-forget persist/emit guards live at the
- * post-execution call site (Plan 04).
+ * post-execution call site.
  *
  * @param bridgeResult - the narrow bridge slice (cost, breaker trips, per-tool results).
  * @param endReason - the ALREADY-MAPPED `SessionMetadata.sessionEnd.endReason`
@@ -57,11 +57,11 @@ export function buildSessionHealthRollup(bridgeResult, endReason) {
         .slice(0, TOP_ERROR_KINDS_CAP));
     // Single source of truth: degraded iff the mapped endReason is not "success".
     // The chokepoint derives endReason once via END_REASON_MAP and passes it here,
-    // so this can never contradict the co-persisted sessionEnd.endReason (CR-01).
+    // so this can never contradict the co-persisted sessionEnd.endReason.
     const degraded = !CLEAN_END_REASONS.has(endReason);
     return {
         degraded,
-        costUsd: bridgeResult.sessionCostUsd ?? 0,
+        costUsd: bridgeResult.executionCostUsd ?? 0,
         toolStats: Object.fromEntries(toolStatsMap),
         breakerTripCount: bridgeResult.breakerTripCount ?? 0,
         topErrorKinds,

package/node_modules/@comis/agent/dist/executor/signed-replay-detector.d.ts

@@ -2,10 +2,8 @@
  * Provider-agnostic signed-replay error detector.
  *
  * Detects provider rejections of stored signed thinking / reasoning state on
- * the latest assistant message during replay. Triggered by a real production
- * incident on `srv1593437` 2026-04-24T20:07:34Z (trace
- * `93ba66cf-4283-4ed4-92bd-73d00b4eeb76`, request_id
- * `req_011CaPCYYKfJRpuG3w2y5s52`) where Anthropic returned
+ * the latest assistant message during replay. Motivated by a real production
+ * incident where Anthropic returned
  * `400 invalid_request_error: messages.5.content.17: 'thinking' or
  * 'redacted_thinking' blocks in the latest assistant message cannot be
  * modified` after a 74-min idle gap with multiple daemon restarts.

package/node_modules/@comis/agent/dist/executor/signed-replay-detector.js

@@ -3,10 +3,8 @@
  * Provider-agnostic signed-replay error detector.
  *
  * Detects provider rejections of stored signed thinking / reasoning state on
- * the latest assistant message during replay. Triggered by a real production
- * incident on `srv1593437` 2026-04-24T20:07:34Z (trace
- * `93ba66cf-4283-4ed4-92bd-73d00b4eeb76`, request_id
- * `req_011CaPCYYKfJRpuG3w2y5s52`) where Anthropic returned
+ * the latest assistant message during replay. Motivated by a real production
+ * incident where Anthropic returned
  * `400 invalid_request_error: messages.5.content.17: 'thinking' or
  * 'redacted_thinking' blocks in the latest assistant message cannot be
  * modified` after a 74-min idle gap with multiple daemon restarts.

package/node_modules/@comis/agent/dist/executor/stream-wrappers/config-resolver.d.ts

@@ -21,20 +21,20 @@ export interface ConfigResolverConfig {
     /** Cache retention for Anthropic provider. Default: "long".
      *  Accepts a getter function for per-execution dynamic resolution. */
     cacheRetention?: CacheRetention | (() => CacheRetention | undefined);
-    /** Phase 166 Fix 3: getter returning the assembled input token count for the current dispatch.
+    /** Getter returning the assembled input token count for the current dispatch.
      *  Set by lcd-assembler.transformContext via the onAssembledInputTokens callback.
      *  When provided and > 0, max_tokens is clamped to effectiveWindow − assembledInputTokens.
      *  When absent or returns 0 (frontier/mid or pre-assembler call), falls back to static config.maxTokens. */
     getAssembledInputTokens?: () => number | undefined;
-    /** Phase 166 Fix 3: getter for the effective context window.
+    /** Getter for the effective context window.
      *  Source: effectiveWindowRef.current set by the onEffectiveWindow callback from lcd-assembler.
      *  When absent or returns Infinity, the dynamic clamp is skipped (frontier/mid byte-identical). */
     getEffectiveWindow?: () => number | undefined;
-    /** Phase 166 Fix 3: reasoning-aware output headroom floor for dynamic max_tokens clamp.
+    /** Reasoning-aware output headroom floor for the dynamic max_tokens clamp.
      *  Computed from computeOutputHeadroom(reasoningStyle, thinkingLevel) before the dispatch. */
     getOutputHeadroom?: () => number;
 }
-/** SYS-BOUNDARY: Deterministic marker between static and dynamic system prompt blocks.
+/** Deterministic marker between static and dynamic system prompt blocks.
  *  Enables cache break diagnostics to identify whether changes are in the
  *  stable or dynamic portion. Appended to staticPrefix block. */
 export declare const SYSTEM_PROMPT_DYNAMIC_BOUNDARY = "\n\n---SYSTEM-PROMPT-DYNAMIC-BOUNDARY---\n\n";

package/node_modules/@comis/agent/dist/executor/stream-wrappers/config-resolver.js

@@ -12,7 +12,7 @@ import { MIN_VISIBLE_OUTPUT_TOKENS } from "../../context-engine/output-headroom.
 // ---------------------------------------------------------------------------
 // Internal helpers
 // ---------------------------------------------------------------------------
-/** SYS-BOUNDARY: Deterministic marker between static and dynamic system prompt blocks.
+/** Deterministic marker between static and dynamic system prompt blocks.
  *  Enables cache break diagnostics to identify whether changes are in the
  *  stable or dynamic portion. Appended to staticPrefix block. */
 export const SYSTEM_PROMPT_DYNAMIC_BOUNDARY = "\n\n---SYSTEM-PROMPT-DYNAMIC-BOUNDARY---\n\n";
@@ -28,8 +28,8 @@ export function resolveBreakpointStrategy(strategy, _provider) {
     if (strategy === "single" || strategy === "multi-zone") {
         return strategy;
     }
-    // W11: All providers use multi-zone strategy. Previous "single" for
-    // direct anthropic caused breakpoint budget exhaustion.
+    // All providers use the multi-zone strategy — "single" for direct
+    // Anthropic exhausts the breakpoint budget.
     return "multi-zone";
 }
 /**
@@ -56,7 +56,7 @@ export function createConfigResolver(config, logger) {
             // Inject maxTokens and temperature for all providers when configured.
             // Skip temperature for reasoning models (e.g. OpenAI o-series, gpt-5.4-mini)
             // -- they don't support the parameter and the API returns 400.
-            // Phase 166 Fix 3: dynamic max_tokens clamp.
+            // Dynamic max_tokens clamp.
             // clamp(configuredMax, MIN_VISIBLE_OUTPUT_TOKENS, effectiveWindow − assembledInputTokens)
             // Guard: both assembled > 0 AND effectiveWindow < Infinity must hold;
             // otherwise fall back to static config.maxTokens (frontier/mid byte-identical path).

package/node_modules/@comis/agent/dist/executor/stream-wrappers/index.js

@@ -16,8 +16,8 @@ export { createTurnResultBudgetWrapper } from "./turn-result-budget-wrapper.js";
 export { createValidationErrorFormatter } from "./validation-error-formatter.js";
 export { createConfigResolver, resolveBreakpointStrategy, SYSTEM_PROMPT_DYNAMIC_BOUNDARY } from "./config-resolver.js";
 // createCacheTraceWriter (+ parseSize, rotateIfNeeded, CacheTraceConfig)
-// were removed when cache-trace moved to
-// @comis/observability/cache-trace/*. Operators using
+// live in @comis/observability/cache-trace/* and are deliberately NOT
+// re-exported here. Operators using
 // `diagnostics.cacheTrace.enabled` get the schema-versioned artifact;
 // `agents.<name>.tracing.enabled` gates the api-payload-trace below.
 export { createApiPayloadTraceWriter } from "./api-payload-trace-writer.js";

package/node_modules/@comis/agent/dist/executor/stream-wrappers/jsonl-rotation.d.ts

@@ -1,9 +1,9 @@
 /**
  * Shared JSONL append + rotation helper.
  *
- * The legacy api-payload-trace writer relies on this rotation logic
+ * The api-payload-trace writer relies on this rotation logic
  * — kept in this module so both the api-payload trace and any future
- * legacy-style JSONL trace can share the size/rotation handling.
+ * sync-fs JSONL trace can share the size/rotation handling.
  *
  * Use `@comis/observability/cache-trace/*` for new artifacts. This helper
  * is intentionally a thin wrapper around sync fs primitives — the

package/node_modules/@comis/agent/dist/executor/stream-wrappers/jsonl-rotation.js

@@ -2,9 +2,9 @@
 /**
  * Shared JSONL append + rotation helper.
  *
- * The legacy api-payload-trace writer relies on this rotation logic
+ * The api-payload-trace writer relies on this rotation logic
  * — kept in this module so both the api-payload trace and any future
- * legacy-style JSONL trace can share the size/rotation handling.
+ * sync-fs JSONL trace can share the size/rotation handling.
  *
  * Use `@comis/observability/cache-trace/*` for new artifacts. This helper
  * is intentionally a thin wrapper around sync fs primitives — the

package/node_modules/@comis/agent/dist/executor/stream-wrappers/request-body/breakpoint-orchestration.js

@@ -108,10 +108,10 @@ export function runCacheBreakpointPhase(result, model, config, needsCacheBreakpo
     // 2 system blocks that didn't exist in the original params.
     const existingCount = countCacheBreakpoints(result);
     let slotsAvailable = 4 - existingCount;
-    // Tracks the message index where Fix E placed (or upgraded) the 1h
-    // anchor. Consumed by the recent-zone retention ternary below so the
-    // earlier breakpoint placed by placeCacheBreakpoints uses 1h instead
-    // of 5m — preventing the daemon.1.log:439 ordering violation at the
+    // Tracks the message index where the UNTRUSTED_ anchor placed (or
+    // upgraded) the 1h marker. Consumed by the recent-zone retention ternary
+    // below so the earlier breakpoint placed by placeCacheBreakpoints uses 1h
+    // instead of 5m — preventing a live-observed TTL ordering violation at the
     // source (the monotonic-ttl safety net remains as defense in depth).
     let eFixFiredAt;
     // Breakpoint budget audit -- 1 per API call.
@@ -165,7 +165,7 @@ export function runCacheBreakpointPhase(result, model, config, needsCacheBreakpo
             }
         }
     }
-    // Fix E (log-review): turns carrying a large stable untrusted block
+    // UNTRUSTED_ anchor: turns carrying a large stable untrusted block
     // (e.g., link-understanding output ~32KB) need a 1h cache anchor —
     // the default 5m TTL gets evicted under Anthropic capacity pressure
     // and the entire ~32KB re-uploads on every turn. Scan the last few
@@ -183,10 +183,10 @@ export function runCacheBreakpointPhase(result, model, config, needsCacheBreakpo
             const hasUntrusted = content.some((b) => typeof b.text === "string" && b.text.includes("<<<UNTRUSTED_"));
             if (!hasUntrusted)
                 continue;
-            // Only anchor a LARGE untrusted block — the ~32KB link-understanding case this Fix exists
-            // for. A SMALL untrusted tool_result (e.g. an `echo` result) is cheap to re-upload and the
-            // SDK's last-user marker already covers it; anchoring it would waste a breakpoint slot that
-            // the lookback gap-bridge needs (cache C-FIX-2, 2026-06-18: small tool_results were stealing
+            // Only anchor a LARGE untrusted block — the ~32KB link-understanding case this anchor
+            // exists for. A SMALL untrusted tool_result (e.g. an `echo` result) is cheap to re-upload and
+            // the SDK's last-user marker already covers it; anchoring it would waste a breakpoint slot
+            // that the lookback gap-bridge needs (observed live: small tool_results were stealing
             // the slot, leaving the mid-conversation gap unbridged on alternating tool turns). The
             // delimiter-wrap security is independent of this cost anchor, so gating on size is safe.
             const EFIX_MIN_UNTRUSTED_CHARS = 16384; // ~4K tokens — clearly "large", well below ~32KB
@@ -221,12 +221,12 @@ export function runCacheBreakpointPhase(result, model, config, needsCacheBreakpo
         const placed = placeCacheBreakpoints(result.messages, {
             minTokens,
             maxBreakpoints: inCooldown ? 1 : slotsAvailable,
-            // Recent zone defaults to 5m, but when Fix E placed a 1h anchor on the
-            // latest user message, any earlier breakpoint must also be >= 1h to
-            // satisfy Anthropic's monotonicity rule (tools->system->messages).
+            // Recent zone defaults to 5m, but when the UNTRUSTED_ anchor placed 1h
+            // on the latest user message, any earlier breakpoint must also be >= 1h
+            // to satisfy Anthropic's monotonicity rule (tools->system->messages).
             // Without this coordination, enforceMonotonicTtlOrdering catches the
-            // violation but logs a WARN — Fix 2 makes the placement intent correct
-            // at source so the safety net stays quiet in the common UNTRUSTED path.
+            // violation but logs a WARN — coordinating here makes the placement intent
+            // correct at source so the safety net stays quiet in the common UNTRUSTED path.
             retention: eFixFiredAt !== undefined ? "long" : "short",
             resolvedRetention: inCooldown ? "short" : messageRetention, // Force "short" during cooldown
             strategy: resolveBreakpointStrategy(config.cacheBreakpointStrategy, model.provider),
@@ -333,8 +333,8 @@ export function runCacheBreakpointPhase(result, model, config, needsCacheBreakpo
     }
     // Safety-net sweep: enforce Anthropic's monotonic non-increasing TTL
     // invariant across tools->system->messages payload order. No-op when
-    // layout is already monotonic; load-bearing defense even when Fix 2
-    // (Fix-E-aware retention) correctly coordinates the source placement.
+    // layout is already monotonic; load-bearing defense even when the
+    // anchor-aware retention above correctly coordinates the source placement.
     // Logs WARN with errorKind:"internal" if any upgrade fires — that
     // indicates an upstream placement bug.
     enforceMonotonicTtlOrdering(result, logger);

package/node_modules/@comis/agent/dist/executor/stream-wrappers/request-body/breakpoint-placement.d.ts

@@ -37,11 +37,11 @@ export declare function placeSingleBreakpoint(messages: Array<Record<string, unk
 /**
  * Place cache breakpoints at strategic positions within the messages array.
  *
- * Budget = 4 − system − SDK-auto(last message) = 2 Comis message markers (cache C-FIX-4):
+ * Budget = 4 − system − SDK-auto(last message) = 2 Comis message markers:
  * - Semi-stable ANCHOR: a FIXED early boundary — the compaction summary if present, else the
- *   2nd user message (cache C-FIX-10). Pinned (never moves) so it is cached once and never
- *   re-written. A cheap durable fallback; superseded the drifting 50%-token/C-FIX-2b anchor
- *   whose movement churn became the dominant waste once #C2/#C4 removed the recurring drops.
+ *   2nd user message. Pinned (never moves) so it is cached once and never
+ *   re-written. A cheap durable fallback; a fixed anchor beats a drifting one (e.g. a
+ *   50%-token split) because every anchor move re-writes the whole [0..anchor] prefix.
  * - RECENT: the second-to-last user message, adjacent to the SDK's last-message marker; the
  *   load-bearing marker whose fresh write caches the whole growing prefix.
  *