comisai 1.0.33 → 1.0.36

package/node_modules/@comis/agent/dist/bridge/thinking-block-hash-invariant.js

@@ -4,9 +4,8 @@
  *
  * Observed problem: Anthropic 400 `messages.N.content.M: thinking/redacted_thinking
  * blocks cannot be modified` errors keep firing in production even after the
- * surrogate-guard, drift-scrubber, and signed-replay-detector layers shipped
- * (260425-rvm), and even after the immutable-section redirect (260425-t40).
- * Trace `c7b91328-9dc5-4618-9ae8-ca207b4b93df` on 2026-04-28 hit a 400 ~2.2s
+ * surrogate-guard, drift-scrubber, and signed-replay-detector layers shipped,
+ * and even after the immutable-section redirect. A trace hit a 400 ~2.2s
  * after `turn_end` -- meaning *some other layer* mutates a signed thinking
  * block between the assistant turn and the next replay. We don't know which.
  *
@@ -24,12 +23,12 @@
  * - NEVER mutates inputs. Pure read; only output is the structured log.
  * - NEVER alters request flow. The mismatch is observable signal only --
  *   Anthropic's 400 still surfaces through the existing error path
- *   (signed-replay-detector -> executor-prompt-runner). Bug A behavior fix
- *   is a separate quick task gated on what this diagnostic reveals.
+ *   (signed-replay-detector -> executor-prompt-runner). The behavior fix
+ *   for the underlying bug is gated on what this diagnostic reveals.
  *
  * Logging surface follows CLAUDE.md canonical Pino fields:
  * - object-first signature: `error({...fields}, "msg")`
- * - `module: "agent.bridge.hash-invariant"`
+ * - `submodule: "bridge.hash-invariant"` (parent logger binds `module: "agent"`)
  * - `errorKind: "internal"` (classification per AGENTS.md §2.1)
  * - `hint`: actionable next step for the on-call diagnoser
  * - `responseId`, `blockIndex`, `oldHash`, `newHash`,
@@ -51,7 +50,7 @@ import { readFile as fsReadFile } from "node:fs/promises";
 const HINT = "Locate the context-engine layer that touched this block " +
     "(likely between turn_end and the next pi-ai serialize step). " +
     "Compare oldText.firstChars vs newText.firstChars to identify mutation type.";
-const MODULE_FIELD = "agent.bridge.hash-invariant";
+const SUBMODULE_FIELD = "bridge.hash-invariant";
 const ERROR_KIND = "internal";
 const TEXT_PREFIX_LEN = 32;
 /** Format the four-field hash payload for one block. */
@@ -148,7 +147,7 @@ export function assertThinkingBlocksUnchanged(prior, current, responseId, deps)
         if (!now) {
             mismatchesLogged++;
             safeLog(deps, {
-                module: MODULE_FIELD,
+                submodule: SUBMODULE_FIELD,
                 responseId,
                 blockIndex: old.blockIndex,
                 oldHash: old.hash,
@@ -166,7 +165,7 @@ export function assertThinkingBlocksUnchanged(prior, current, responseId, deps)
         if (now.hash !== old.hash) {
             mismatchesLogged++;
             safeLog(deps, {
-                module: MODULE_FIELD,
+                submodule: SUBMODULE_FIELD,
                 responseId,
                 blockIndex: old.blockIndex,
                 oldHash: old.hash,
@@ -187,14 +186,14 @@ export function assertThinkingBlocksUnchanged(prior, current, responseId, deps)
     };
 }
 // ---------------------------------------------------------------------------
-// 260428-hoy: Canonical thinking-block restoration
+// Canonical thinking-block restoration
 //
 // Heals cross-turn mutation of signed thinking blocks before pi-ai serializes
 // the next API request. Pure / idempotent / never-throws. Runs AFTER
 // `assertThinkingBlocksUnchanged` so the diagnostic ERROR log captures every
 // mutation before the heal overwrites it.
 // ---------------------------------------------------------------------------
-const RESTORE_MODULE_FIELD = "agent.bridge.canonical-restore";
+const RESTORE_SUBMODULE_FIELD = "bridge.canonical-restore";
 const RESTORE_WARN_HINT = "Canonical restore aborted on malformed input; in-memory messages " +
     "returned unchanged. Inspect prior context-engine layers for shape drift.";
 /** Best-effort logger.info / logger.warn -- swallows logger errors. */
@@ -232,7 +231,7 @@ function safeRestoreLog(deps, level, payload, msg) {
  * Never throws. On any unexpected error during the walk (e.g. malformed
  * canonical entry whose getter throws), the entire result is `{ messages:
  * <input ref>, restoredCount: 0, affectedResponseIds: [] }` and ONE WARN log
- * fires with `module: RESTORE_MODULE_FIELD, errorKind: "internal"`.
+ * fires with `submodule: RESTORE_SUBMODULE_FIELD, errorKind: "internal"`.
  */
 export function restoreCanonicalThinkingBlocks(messages, canonicalStore, deps) {
     if (!Array.isArray(messages)) {
@@ -274,7 +273,7 @@ export function restoreCanonicalThinkingBlocks(messages, canonicalStore, deps) {
             };
         }
         safeRestoreLog(deps, "info", {
-            module: RESTORE_MODULE_FIELD,
+            submodule: RESTORE_SUBMODULE_FIELD,
             restoredCount,
             affectedResponseIds,
         }, "Restored canonical thinking blocks before resend");
@@ -284,7 +283,7 @@ export function restoreCanonicalThinkingBlocks(messages, canonicalStore, deps) {
         // Defensive last-resort: malformed canonical entry or any other thrown
         // error during the walk. Return the in-memory shape unchanged + WARN log.
         safeRestoreLog(deps, "warn", {
-            module: RESTORE_MODULE_FIELD,
+            submodule: RESTORE_SUBMODULE_FIELD,
             errorKind: "internal",
             hint: RESTORE_WARN_HINT,
         }, "Canonical restore aborted on malformed input");
@@ -380,16 +379,16 @@ function blockHash(block) {
     return createHash("sha256").update(input).digest("hex");
 }
 // ---------------------------------------------------------------------------
-// 260428-iag: Wire-edge diagnostic — diff in-memory thinking blocks against
-// persisted JSONL canonical.
+// Wire-edge diagnostic — diff in-memory thinking blocks against persisted
+// JSONL canonical.
 //
 // Fires from the pi-event-bridge LLM-error path when Anthropic returns 400
 // with a "thinking blocks ... cannot be modified" signature, even after the
-// 260428-hoy canonical-restore layer ran pre-serialize. The persisted JSONL
-// is the only truly immutable record of the assistant message — written
-// byte-for-byte from Anthropic's stream at receipt time. A divergence between
-// in-memory content and persisted canonical at this point implies the
-// mutation occurred AFTER the bridge restoration hook (likely inside pi-ai's
+// canonical-restore layer ran pre-serialize. The persisted JSONL is the only
+// truly immutable record of the assistant message — written byte-for-byte
+// from Anthropic's stream at receipt time. A divergence between in-memory
+// content and persisted canonical at this point implies the mutation
+// occurred AFTER the bridge restoration hook (likely inside pi-ai's
 // `sanitizeSurrogates` during request serialization).
 //
 // Behavior contract (mirrors the rest of this module):
@@ -398,7 +397,7 @@ function blockHash(block) {
 //   and an empty result. The diagnostic must NEVER abort agent flow.
 // - Caller passes a resolved `jsonlPath`; this helper does not compose paths.
 // ---------------------------------------------------------------------------
-const WIRE_DIFF_MODULE_FIELD = "agent.bridge.wire-diff";
+const WIRE_DIFF_SUBMODULE_FIELD = "bridge.wire-diff";
 export const WIRE_DIFF_HINT_FILE_MISSING = "JSONL session file unreadable; wire-edge diff skipped. " +
     "Confirm session path resolution and filesystem permissions.";
 export const WIRE_DIFF_HINT_NOT_FOUND = "responseId not present in persisted JSONL session file; " +
@@ -462,7 +461,7 @@ export async function diffThinkingBlocksAgainstPersisted(inMemoryContent, respon
         }
         catch (readErr) {
             safeWireDiffLog(deps, "warn", {
-                module: WIRE_DIFF_MODULE_FIELD,
+                submodule: WIRE_DIFF_SUBMODULE_FIELD,
                 errorKind: ERROR_KIND,
                 hint: WIRE_DIFF_HINT_FILE_MISSING,
                 jsonlPath,
@@ -506,7 +505,7 @@ export async function diffThinkingBlocksAgainstPersisted(inMemoryContent, respon
         }
         if (persistedContent === null) {
             safeWireDiffLog(deps, "warn", {
-                module: WIRE_DIFF_MODULE_FIELD,
+                submodule: WIRE_DIFF_SUBMODULE_FIELD,
                 errorKind: ERROR_KIND,
                 hint: WIRE_DIFF_HINT_NOT_FOUND,
                 jsonlPath,
@@ -554,7 +553,7 @@ export async function diffThinkingBlocksAgainstPersisted(inMemoryContent, respon
         // computeThinkingBlockHashes against malformed input, etc.), degrade
         // to WARN + empty result.
         safeWireDiffLog(deps, "warn", {
-            module: WIRE_DIFF_MODULE_FIELD,
+            submodule: WIRE_DIFF_SUBMODULE_FIELD,
             errorKind: ERROR_KIND,
             hint: WIRE_DIFF_HINT_INTERNAL,
             jsonlPath,

package/node_modules/@comis/agent/dist/budget/cost-tracker.d.ts

@@ -3,7 +3,7 @@ import type { ModelOperationType } from "@comis/core";
  * Per-agent/channel/execution cost record.
  *
  * Captures token counts and cost breakdown for attribution
- * and observability. Persistence via event bus comes in Phase 6.
+ * and observability. Persistence via event bus is planned.
  */
 export interface CostRecord {
     timestamp: number;

package/node_modules/@comis/agent/dist/context-engine/constants.d.ts

@@ -84,7 +84,7 @@ export declare const BOOTSTRAP_BUDGET_WARN_PERCENT = 85;
  * Minimum cacheable token thresholds by model family prefix.
  * Used by: cache breakpoint placement.
  *
- * Values are Anthropic's official API minimums verified 2026-03-30.
+ * Values are Anthropic's official API minimums.
  * Setting lower causes silent no-ops (breakpoints ignored, tokens not cached).
  */
 export declare const MIN_CACHEABLE_TOKENS: Record<string, number>;
@@ -131,11 +131,11 @@ export declare const MAX_REHYDRATION_FILES = 5;
 export declare const MAX_REHYDRATION_FILE_CHARS = 8000;
 /** Maximum total chars for all rehydration content (safety cap). Used by: rehydration overflow check. */
 export declare const MAX_REHYDRATION_TOTAL_CHARS = 30000;
-/** Maximum chars per individual skill in rehydration restoration (POST-COMPACT-BUDGET).
+/** Maximum chars per individual skill in rehydration restoration.
  *  Skills exceeding this limit are truncated at the boundary with closing tag repair.
  *  Used by: rehydration layer buildSkillsContent(). */
 export declare const MAX_REHYDRATION_CHARS_PER_SKILL = 5000;
-/** Maximum total chars for all rehydration content combined (POST-COMPACT-BUDGET).
+/** Maximum total chars for all rehydration content combined.
  *  This is the token-budgeted restoration cap (50K chars ~ 12,500 tokens at 4:1 ratio).
  *  Used by: rehydration layer overflow check. */
 export declare const MAX_REHYDRATION_TOKEN_BUDGET_CHARS = 50000;
@@ -169,10 +169,10 @@ export declare const XML_WRAPPER_OVERHEAD_TOKENS = 40;
 export declare const RECALL_GUIDANCE = "You have access to a context DAG containing your full conversation history. Summaries marked with <context_summary> tags contain compressed versions of earlier exchanges. To view full details of any summarized content, use the ctx_inspect tool with the summary ID. For broader recall across your history, use ctx_search with a text query.";
 /** Relative threshold for cache break detection. Break detected when cacheRead
  *  drops by more than this fraction of the previous baseline.
- *  Used by: cache break detector Phase 2 check. */
+ *  Used by: cache break detector. */
 export declare const CACHE_BREAK_RELATIVE_THRESHOLD = 0.05;
 /** Absolute threshold for cache break detection (tokens). Break detected when
  *  cacheRead drops by more than this many tokens from the previous baseline.
  *  Both relative AND absolute thresholds must be exceeded to trigger detection.
- *  Used by: cache break detector Phase 2 check. */
+ *  Used by: cache break detector. */
 export declare const CACHE_BREAK_ABSOLUTE_THRESHOLD = 2000;

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

@@ -29,12 +29,12 @@ export const CONTEXT_ROT_BUFFER_PERCENT = 25;
 /** Consecutive layer failures before circuit breaker disables the layer for the session. Used by: layer pipeline runner. */
 export const LAYER_CIRCUIT_BREAKER_THRESHOLD = 3;
 // ---------------------------------------------------------------------------
-// Thinking Block Cleaner (Layer 4: )
+// Thinking Block Cleaner (Layer 4)
 // ---------------------------------------------------------------------------
 /** Default number of recent assistant turns that retain thinking blocks. Used by: thinking block cleaner (Layer 4). */
 export const DEFAULT_KEEP_WINDOW_TURNS = 10;
 // ---------------------------------------------------------------------------
-// Microcompaction Guard (Layer 2: )
+// Microcompaction Guard (Layer 2)
 // ---------------------------------------------------------------------------
 /** Default inline threshold for tool result microcompaction (chars). Used by: microcompaction guard. */
 export const MAX_INLINE_TOOL_RESULT_CHARS = 8_000;
@@ -59,7 +59,7 @@ export const MAX_REHYDRATION_SKILL_CHARS = 15_000;
 /** Maximum number of individual skills to restore after compaction. Used by: rehydration skill restoration. */
 export const MAX_REHYDRATION_SKILLS = 10;
 // ---------------------------------------------------------------------------
-// Observation Masker (Layer 3: )
+// Observation Masker (Layer 3)
 // ---------------------------------------------------------------------------
 /** Default observation masking keep window (most recent N tool uses retained). Used by: observation masker. */
 export const DEFAULT_OBSERVATION_KEEP_WINDOW = 15;
@@ -121,7 +121,7 @@ export const CHARS_PER_TOKEN_RATIO = 3.5;
  *  over the flat 4:1 ratio. Used by: content-aware token estimation. */
 export const CHARS_PER_TOKEN_RATIO_STRUCTURED = 3;
 // ---------------------------------------------------------------------------
-// Cache Optimization (Layer: )
+// Cache Optimization
 // ---------------------------------------------------------------------------
 /** Length of truncated SHA-256 digest for system prompt hash comparison. Used by: prompt-assembly hash validation. */
 export const SYSTEM_PROMPT_HASH_LENGTH = 16;
@@ -131,7 +131,7 @@ export const BOOTSTRAP_BUDGET_WARN_PERCENT = 85;
  * Minimum cacheable token thresholds by model family prefix.
  * Used by: cache breakpoint placement.
  *
- * Values are Anthropic's official API minimums verified 2026-03-30.
+ * Values are Anthropic's official API minimums.
  * Setting lower causes silent no-ops (breakpoints ignored, tokens not cached).
  */
 export const MIN_CACHEABLE_TOKENS = {
@@ -167,7 +167,7 @@ export const CACHE_LOOKBACK_WINDOW = 20;
  *  Used by: llm-compaction layer cache-aware trigger. */
 export const CACHE_AWARE_COMPACTION_BLOCK_THRESHOLD = 60;
 // ---------------------------------------------------------------------------
-// LLM Compaction (Layer 5: )
+// LLM Compaction (Layer 5)
 // ---------------------------------------------------------------------------
 /** Context utilization percentage that triggers LLM compaction. Used by: llm-compaction layer. */
 export const COMPACTION_TRIGGER_PERCENT = 85;
@@ -192,7 +192,7 @@ export const DEFAULT_COMPACTION_PREFIX_ANCHOR_TURNS = 2;
  *  Used by: llm-compaction layer. */
 export const MIN_MIDDLE_MESSAGES_FOR_COMPACTION = 3;
 // ---------------------------------------------------------------------------
-// Post-Compaction Rehydration (Layer 6: )
+// Post-Compaction Rehydration (Layer 6)
 // ---------------------------------------------------------------------------
 /** Maximum number of recently-accessed files to re-inject after compaction. Used by: rehydration layer. */
 export const MAX_REHYDRATION_FILES = 5;
@@ -200,16 +200,16 @@ export const MAX_REHYDRATION_FILES = 5;
 export const MAX_REHYDRATION_FILE_CHARS = 8_000;
 /** Maximum total chars for all rehydration content (safety cap). Used by: rehydration overflow check. */
 export const MAX_REHYDRATION_TOTAL_CHARS = 30_000;
-/** Maximum chars per individual skill in rehydration restoration (POST-COMPACT-BUDGET).
+/** Maximum chars per individual skill in rehydration restoration.
  *  Skills exceeding this limit are truncated at the boundary with closing tag repair.
  *  Used by: rehydration layer buildSkillsContent(). */
 export const MAX_REHYDRATION_CHARS_PER_SKILL = 5_000;
-/** Maximum total chars for all rehydration content combined (POST-COMPACT-BUDGET).
+/** Maximum total chars for all rehydration content combined.
  *  This is the token-budgeted restoration cap (50K chars ~ 12,500 tokens at 4:1 ratio).
  *  Used by: rehydration layer overflow check. */
 export const MAX_REHYDRATION_TOKEN_BUDGET_CHARS = 50_000;
 // ---------------------------------------------------------------------------
-// Dead Content Evictor (Layer: )
+// Dead Content Evictor
 // ---------------------------------------------------------------------------
 /**
  * Default minimum age (in tool result positions) before content is eligible
@@ -294,10 +294,10 @@ export const RECALL_GUIDANCE = `You have access to a context DAG containing your
 // ---------------------------------------------------------------------------
 /** Relative threshold for cache break detection. Break detected when cacheRead
  *  drops by more than this fraction of the previous baseline.
- *  Used by: cache break detector Phase 2 check. */
+ *  Used by: cache break detector. */
 export const CACHE_BREAK_RELATIVE_THRESHOLD = 0.05;
 /** Absolute threshold for cache break detection (tokens). Break detected when
  *  cacheRead drops by more than this many tokens from the previous baseline.
  *  Both relative AND absolute thresholds must be exceeded to trigger detection.
- *  Used by: cache break detector Phase 2 check. */
+ *  Used by: cache break detector. */
 export const CACHE_BREAK_ABSOLUTE_THRESHOLD = 2_000;

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

@@ -201,13 +201,22 @@ export function createContextEngine(config, deps) {
     // most recent one. Always-on — no drift-mode gate. The latest
     // assistant message keeps its signatures so Anthropic's
     // extended-thinking continuation can validate the immediate next
-    // call's prefix. (Replaces 260428-kvl detection-based approach;
-    // see .planning/quick/260428-lm6 for rationale.)
+    // call's prefix.
     if (deps.getReplayDriftMode) {
         const getReplayDriftMode = deps.getReplayDriftMode;
         layers.push(createSignatureReplayScrubber({
             getReplayDriftMode,
-            onScrubbed: (stats) => { callbackState.signatureReplayScrubber = stats; },
+            onScrubbed: (stats) => {
+                // Snapshot the latest stats for the existing post-pipeline DEBUG
+                // summary at lines ~720-727 (preserves legacy aliases dropped /
+                // signaturesStripped / reason).
+                callbackState.signatureReplayScrubber = stats;
+                // Also forward stats to the optional per-execute accumulator wired
+                // in by executor-context-engine-setup.ts. The two callbacks are
+                // deliberately separate: the snapshot consumer overwrites per-call,
+                // the accumulator sums across calls.
+                deps.onSignatureReplayScrubbed?.(stats);
+            },
             logger: deps.logger,
         }));
     }
@@ -241,7 +250,7 @@ export function createContextEngine(config, deps) {
     // Observation masker: masks old tool results beyond the keep window.
     // Always active when context engine is enabled. The threshold check inside the
     // masker handles short-session bypass.
-    // Phase 8: Three-tier masking (protected/standard/ephemeral) with per-tier counters.
+    // Three-tier masking (protected/standard/ephemeral) with per-tier counters.
     const observationKeepWindow = config.observationKeepWindow ?? 25;
     const observationTriggerChars = config.observationTriggerChars ?? 120_000;
     const ephemeralKeepWindow = config.ephemeralKeepWindow;

package/node_modules/@comis/agent/dist/context-engine/dag-annotator.d.ts

@@ -15,8 +15,7 @@
  * - Already-annotated and already-masked results are skipped
  * - Recent tool results within the keep window are preserved
  *
- * DAG Assembly & Annotation.
- * Phase 8: Tier-aware annotation.
+ * DAG Assembly & Annotation. Tier-aware annotation.
  *
  * @module
  */

package/node_modules/@comis/agent/dist/context-engine/dag-annotator.js

@@ -16,8 +16,7 @@
  * - Already-annotated and already-masked results are skipped
  * - Recent tool results within the keep window are preserved
  *
- * DAG Assembly & Annotation.
- * Phase 8: Tier-aware annotation.
+ * DAG Assembly & Annotation. Tier-aware annotation.
  *
  * @module
  */

package/node_modules/@comis/agent/dist/context-engine/llm-compaction.js

@@ -279,21 +279,7 @@ export function createLlmCompactionLayer(config, deps) {
                         return messages;
                     }
                 }
-                // Step 4: Unified log (conditional spread keeps JSON shape clean).
-                deps.logger.warn({
-                    messageCount,
-                    ...(blockCountExceeded
-                        ? { blockThreshold, trigger: "block_count" }
-                        : { contextTokens, thresholdTokens, trigger: "token_threshold" }),
-                    windowTokens: budget.windowTokens,
-                    errorKind: "resource",
-                    hint: blockCountExceeded
-                        ? "Message count approaching breakpoint lookback limit; compacting to prevent cache fragmentation"
-                        : "Context approaching capacity; LLM compaction will summarize older messages to free space",
-                }, blockCountExceeded
-                    ? "LLM compaction triggered: message count exceeds cache lookback threshold"
-                    : "LLM compaction triggered: context exceeds 85% threshold");
-                // Step 5: Resolve model
+                // Step 4: Resolve model
                 /* eslint-disable @typescript-eslint/no-explicit-any */
                 let model;
                 let apiKey;
@@ -365,10 +351,28 @@ export function createLlmCompactionLayer(config, deps) {
                 /* eslint-enable security/detect-object-injection */
                 // Zone 2: Middle (to be summarized)
                 const middleMessages = messages.slice(headEndIndex, tailStartIndex);
-                // Skip if middle is empty or too small to warrant summarization
+                // Skip if middle is empty or too small to warrant summarization.
+                // Reset the cooldown counter so we don't re-evaluate (and re-warn)
+                // on every subsequent turn when the conversation shape stays this way.
                 if (middleMessages.length < MIN_MIDDLE_MESSAGES_FOR_COMPACTION) {
+                    turnsSinceLastCompaction = 0;
                     return messages;
                 }
+                // Trigger log fires only when compaction will actually run, so log
+                // volume reflects real compaction work (not infeasibility re-checks).
+                deps.logger.warn({
+                    messageCount,
+                    ...(blockCountExceeded
+                        ? { blockThreshold, trigger: "block_count" }
+                        : { contextTokens, thresholdTokens, trigger: "token_threshold" }),
+                    windowTokens: budget.windowTokens,
+                    errorKind: "resource",
+                    hint: blockCountExceeded
+                        ? "Message count approaching breakpoint lookback limit; compacting to prevent cache fragmentation"
+                        : "Context approaching capacity; LLM compaction will summarize older messages to free space",
+                }, blockCountExceeded
+                    ? "LLM compaction triggered: message count exceeds cache lookback threshold"
+                    : "LLM compaction triggered: context exceeds 85% threshold");
                 // Step 7: Summarize ONLY the middle zone (do NOT pass head or tail to generateSummary)
                 const compactionResult = await compactWithFallback(middleMessages, model, apiKey, budget.outputReserveTokens, deps.logger);
                 // Build compaction summary message matching SDK format

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

@@ -96,7 +96,7 @@ function buildSkillsContent(deps) {
         return "";
     const closingTag = "</skill>";
     const skillOpenRegex = /<skill[\s>]/g;
-    // Phase 1: Extract individual skill blocks and apply per-skill truncation (POST-COMPACT-BUDGET)
+    // Extract individual skill blocks and apply per-skill truncation
     const skillBlocks = [];
     let match;
     while ((match = skillOpenRegex.exec(xml)) !== null) {
@@ -105,7 +105,7 @@ function buildSkillsContent(deps) {
         if (closeIdx === -1)
             break; // malformed -- stop
         let skillBlock = xml.slice(startIdx, closeIdx + closingTag.length);
-        // POST-COMPACT-BUDGET: Per-skill truncation with closing tag repair
+        // Per-skill truncation with closing tag repair
         if (skillBlock.length > MAX_REHYDRATION_CHARS_PER_SKILL) {
             // Reserve space for repair suffix: "\n...</skill>" = 12 chars
             const repairSuffix = "\n..." + closingTag;
@@ -120,14 +120,14 @@ function buildSkillsContent(deps) {
     }
     if (skillBlocks.length === 0)
         return "";
-    // Phase 2: Enforce MAX_REHYDRATION_SKILLS count limit
+    // Enforce MAX_REHYDRATION_SKILLS count limit
     const limitedBlocks = skillBlocks.slice(0, MAX_REHYDRATION_SKILLS);
-    // Phase 3: Reassemble XML with wrapper
+    // Reassemble XML with wrapper
     // Preserve opening tag from original XML
     const wrapperOpenEnd = xml.indexOf(">") + 1;
     const wrapperOpen = xml.slice(0, wrapperOpenEnd);
     let truncatedXml = wrapperOpen + "\n" + limitedBlocks.join("\n") + "\n</available_skills>";
-    // Phase 4: Enforce char budget (15K)
+    // Enforce char budget (15K)
     if (truncatedXml.length > MAX_REHYDRATION_SKILL_CHARS) {
         // Re-truncate at skill boundaries within char budget
         const withinBudget = truncatedXml.slice(0, MAX_REHYDRATION_SKILL_CHARS);
@@ -236,7 +236,7 @@ export function createRehydrationLayer(deps) {
             // Position 1 (stable content -- changes rarely, good for KV-cache prefix):
             const position1Parts = [agentsMdSection, fileSection, skillsSection].filter(Boolean);
             let position1Text = position1Parts.join("\n\n");
-            // POST-COMPACT-BUDGET: Enforce total rehydration token budget
+            // Enforce total rehydration token budget
             if (position1Text.length > MAX_REHYDRATION_TOKEN_BUDGET_CHARS) {
                 deps.logger.warn({
                     totalChars: position1Text.length,

package/node_modules/@comis/agent/dist/context-engine/signature-replay-scrubber.d.ts

@@ -7,20 +7,20 @@
  * blocks are never modified, anywhere.
  *
  * Rationale: Anthropic's signed-thinking validation operates on the full
- * (system + tools + history) prefix. After 8 quick tasks of progressively
- * narrower drift detection (gj6 → kvl) we proved targeted detection is
- * intractable; trace 679c8927 had stable tools (49138 bytes across 4 turns)
- * but the system prompt grew +1824 bytes and the 400 fired anyway.
+ * (system + tools + history) prefix. After progressive refinement of
+ * narrower drift detection we proved targeted detection is intractable;
+ * traces with stable tools but a system prompt that grew by ~1.8KB still
+ * fired the 400.
  *
- * 260428-lm6 introduced an unconditional drop that preserved the LATEST
- * assistant message's signatures, on the theory that the immediate-next
- * continuation could still validate them. 260428-nzp's repro proved that
- * carve-out doesn't work: cross-turn signature validation covers the whole
- * request body (system + tools + history) and comis's dynamic context
- * guarantees the surrounding context changes turn-to-turn. So the latest's
- * signatures get invalidated too. Drop them all.
+ * An earlier unconditional drop preserved the LATEST assistant message's
+ * signatures, on the theory that the immediate-next continuation could
+ * still validate them. Repro proved that carve-out doesn't work:
+ * cross-turn signature validation covers the whole request body (system +
+ * tools + history) and comis's dynamic context guarantees the surrounding
+ * context changes turn-to-turn. So the latest's signatures get
+ * invalidated too. Drop them all.
  *
- * 260430-anthropic-400-thinking-block: the prior cache-fence skip
+ * The prior cache-fence skip
  * (`if (i <= budget.cacheFenceIndex) preserve`) caused a per-execution
  * regression. In iteration 1 of an execution the fence is -1 so all signed
  * thinking blocks are stripped and the wire body establishes a cached

package/node_modules/@comis/agent/dist/context-engine/signature-replay-scrubber.js

@@ -8,20 +8,20 @@
  * blocks are never modified, anywhere.
  *
  * Rationale: Anthropic's signed-thinking validation operates on the full
- * (system + tools + history) prefix. After 8 quick tasks of progressively
- * narrower drift detection (gj6 → kvl) we proved targeted detection is
- * intractable; trace 679c8927 had stable tools (49138 bytes across 4 turns)
- * but the system prompt grew +1824 bytes and the 400 fired anyway.
+ * (system + tools + history) prefix. After progressive refinement of
+ * narrower drift detection we proved targeted detection is intractable;
+ * traces with stable tools but a system prompt that grew by ~1.8KB still
+ * fired the 400.
  *
- * 260428-lm6 introduced an unconditional drop that preserved the LATEST
- * assistant message's signatures, on the theory that the immediate-next
- * continuation could still validate them. 260428-nzp's repro proved that
- * carve-out doesn't work: cross-turn signature validation covers the whole
- * request body (system + tools + history) and comis's dynamic context
- * guarantees the surrounding context changes turn-to-turn. So the latest's
- * signatures get invalidated too. Drop them all.
+ * An earlier unconditional drop preserved the LATEST assistant message's
+ * signatures, on the theory that the immediate-next continuation could
+ * still validate them. Repro proved that carve-out doesn't work:
+ * cross-turn signature validation covers the whole request body (system +
+ * tools + history) and comis's dynamic context guarantees the surrounding
+ * context changes turn-to-turn. So the latest's signatures get
+ * invalidated too. Drop them all.
  *
- * 260430-anthropic-400-thinking-block: the prior cache-fence skip
+ * The prior cache-fence skip
  * (`if (i <= budget.cacheFenceIndex) preserve`) caused a per-execution
  * regression. In iteration 1 of an execution the fence is -1 so all signed
  * thinking blocks are stripped and the wire body establishes a cached
@@ -85,11 +85,11 @@ export function createSignatureReplayScrubber(deps) {
             for (let i = 0; i < messages.length; i++) {
                 // eslint-disable-next-line security/detect-object-injection -- numeric index
                 const original = messages[i];
-                // 260430-anthropic-400-thinking-block: cacheFenceIndex is intentionally
-                // NOT consulted here. Stripping uniformly across the array keeps the
-                // scrubbed prefix identical across iterations of the same execution,
-                // which is what Anthropic's prompt-cache validator requires. See
-                // module docstring for the full rationale.
+                // cacheFenceIndex is intentionally NOT consulted here. Stripping
+                // uniformly across the array keeps the scrubbed prefix identical
+                // across iterations of the same execution, which is what Anthropic's
+                // prompt-cache validator requires. See module docstring for the full
+                // rationale.
                 const msg = original;
                 if (msg.role !== "assistant" || !Array.isArray(msg.content)) {
                     // eslint-disable-next-line security/detect-object-injection -- numeric index
@@ -183,16 +183,30 @@ export function createSignatureReplayScrubber(deps) {
                 signaturesStripped: toolCallsAffected,
                 reason: undefined,
             });
-            // Emit INFO once per execute() when at least one assistant message
-            // was actually scrubbed. Pino object-first; no string interp.
+            // Emit once per execute() when at least one assistant message was
+            // actually scrubbed. Pino object-first; no string interp.
+            //
+            // Conditional log level — DEBUG when toolCallsAffected===0
+            // (pure-thinking-block scrub, the routine high-frequency case per
+            // CLAUDE.md "N times per request -> DEBUG"); INFO when >0
+            // (post-incident-visibility case where a tool_call's thoughtSignature
+            // was stripped, the originally documented diagnostic value). The outer
+            // `scrubbedAssistantMessages > 0` gate is unchanged — empty scrubs
+            // never log at this site.
             if (scrubbedAssistantMessages > 0) {
-                deps.logger.info({
-                    module: "agent.context-engine.signature-replay-scrub",
+                const payload = {
+                    submodule: "context-engine.signature-replay-scrub",
                     scrubbedAssistantMessages,
                     blocksAffected,
                     toolCallsAffected,
                     latestAssistantIdx: latestIdx,
-                }, "Dropped thinking signatures from all assistant messages (cross-turn replay)");
+                };
+                if (toolCallsAffected > 0) {
+                    deps.logger.info(payload, "Dropped thinking signatures from all assistant messages (cross-turn replay)");
+                }
+                else {
+                    deps.logger.debug(payload, "Dropped thinking signatures from all assistant messages (cross-turn replay)");
+                }
             }
             // Zero-allocation early return when nothing was actually changed.
             if (!anyChanged)

package/node_modules/@comis/agent/dist/context-engine/types-core.d.ts

@@ -131,6 +131,21 @@ export interface ContextEngineDeps {
     /** Called when observation masking modifies content (maskedCount > 0).
      *  Used by cache break detector to suppress false-positive CacheBreakEvents. */
     onContentModified?: () => void;
+    /** Optional sink for the signature-replay scrubber's per-apply stats.
+     *  Receives the SAME shape the scrubber emits to its `onScrubbed` callback,
+     *  so callers can accumulate per-execute totals without the scrubber owning
+     *  the accumulator. Invoked alongside the existing snapshot wiring in
+     *  context-engine.ts (the snapshot consumer at lines ~720-727 still works
+     *  because field names and shapes are unchanged). */
+    onSignatureReplayScrubbed?: (stats: {
+        scrubbedAssistantMessages: number;
+        blocksAffected: number;
+        toolCallsAffected: number;
+        latestAssistantIdx: number;
+        dropped: number;
+        signaturesStripped: number;
+        reason?: string;
+    }) => void;
     /** Optional getter for the API-grounded token anchor.
      *  Returns the last API response's input_tokens and message count.
      *  When absent or returning null, estimation falls back to char-based heuristics. */

package/node_modules/@comis/agent/dist/executor/cache-break-detection.d.ts

@@ -1,10 +1,10 @@
 /**
  * Cache break detection module: two-phase detection with Anthropic adapter.
  *
- * Phase 1 (pre-call): Records prompt state via provider-specific adapter
+ * Pre-call: Records prompt state via provider-specific adapter
  * (system hash, per-tool schema hashes, cache_control metadata hash).
  *
- * Phase 2 (post-call): Compares cacheRead tokens against baseline using
+ * Post-call: Compares cacheRead tokens against baseline using
  * AND-based dual threshold (both >5% relative AND >2000 absolute must exceed).
  * Provider-agnostic.
  *
@@ -202,15 +202,15 @@ export interface CacheBreakDetectorOptions {
  */
 export declare function createCacheBreakDetector(logger: DetectorLogger, options?: CacheBreakDetectorOptions): CacheBreakDetector;
 /**
- * Extract prompt state from Anthropic API payload for Phase 1 recording.
+ * Extract prompt state from Anthropic API payload for pre-call recording.
  *
  * CRITICAL: Does NOT mutate the original params object. Creates shallow copies
- * for hashing with cache_control stripped (D-08). Per-tool hashing uses
- * input_schema (D-09).
+ * for hashing with cache_control stripped. Per-tool hashing uses
+ * input_schema.
  */
 export declare function extractAnthropicPromptState(params: Record<string, unknown>, modelId: string, retention: string | undefined, sessionKey: string, agentId: string, headers?: Record<string, string>): RecordPromptStateInput;
 /**
- * Extract Gemini-native prompt state for Phase 1 cache break detection.
+ * Extract Gemini-native prompt state for pre-call cache break detection.
  *
  * Gemini payload structure differs from Anthropic:
  * - System prompt: config.systemInstruction (string, not array of blocks)