comisai 1.0.45 → 1.0.46

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

@@ -6,8 +6,6 @@
  * when compaction has occurred -- avoids token waste on normal turns where
  * the objective is already present in the system prompt.
  *
- * - Objective survives compaction via transformContext hook
- *
  * @module
  */
 import type { ContextLayer } from "./types.js";

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

@@ -7,8 +7,6 @@
  * when compaction has occurred -- avoids token waste on normal turns where
  * the objective is already present in the system prompt.
  *
- * - Objective survives compaction via transformContext hook
- *
  * @module
  */
 import { systemNowMs } from "@comis/core";
@@ -39,7 +37,8 @@ function getMessageTextContent(msg) {
 /**
  * Check if a message is a compaction summary by inspecting both the
  * `compactionSummary` flag and the content text for the "[Compaction Summary]"
- * marker. Different versions of the compaction layer use different markers.
+ * marker. Both checks are needed: a summary may carry the runtime flag, the
+ * text marker, or both, depending on which compaction path produced it.
  */
 function isCompactionMessage(msg) {
     // Flag-based detection (typed as any since compactionSummary is a runtime extension)

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

@@ -1,11 +1,10 @@
 /**
- * Reasoning-aware output headroom primitives (Fix 3 / Phase 166 CWF-02).
+ * Reasoning-aware output headroom primitives.
  *
  * Pure functions; no runtime deps, no I/O.
  * Consumers: lcd-assembler.ts (pre-flight fit check),
  *            config-resolver.ts (dynamic max_tokens clamp),
  *            executor-stream-setup.ts (thinking-effort governor).
- * Design ref: design/small-model-context-fidelity.md §4 Fix 3 + §3 corollaries.
  */
 import { MIN_VISIBLE_OUTPUT_TOKENS } from "./constants.js";
 export { MIN_VISIBLE_OUTPUT_TOKENS };
@@ -17,7 +16,7 @@ export { MIN_VISIBLE_OUTPUT_TOKENS };
  * "native" style = empirically sized per qwen3.6:35b observed blocks.
  *
  * Sizing rationale — sized to qwen3.6:35b observed high-thinking block
- * (the NVDA incident: in 32477 + out 291 = 32768, stopReason 'length',
+ * (observed live: in 32477 + out 291 = 32768, stopReason 'length',
  * 0 tool calls — high-effort thinking consumed essentially the entire 32K window).
  * native/high=8192 is a defensible middle ground: large enough that a high-thinking
  * block is not silently truncated on a 32K window (leaves ~23K for history),
@@ -37,7 +36,7 @@ export declare const THINKING_RESERVE_TOKENS: Readonly<Record<"none" | "native",
  * @param reasoningStyle      - Model's reasoning style: "none" (standard) or "native" (thinking block).
  * @param thinkingLevel       - Configured thinking effort level.
  * @param visibleOutputFloor  - Minimum visible output tokens (defaults to MIN_VISIBLE_OUTPUT_TOKENS).
- *                              WR-02: pass contextEngine.budget.minVisibleOutputTokens here when
+ *                              Pass contextEngine.budget.minVisibleOutputTokens here when
  *                              the config value differs from the compile-time constant.
  * @returns Total output headroom tokens to reserve from the context window.
  */
@@ -45,7 +44,7 @@ export declare function computeOutputHeadroom(reasoningStyle: "none" | "native",
 /**
  * The FLOOR output headroom — the minimum the thinking-effort governor can reach.
  *
- * ISSUE #3b (2026-06-22): the SINGLE source for "how much headroom is reserved at the
+ * The SINGLE source for "how much headroom is reserved at the
  * floor", used by BOTH the pre-flight (the bound it ultimately throws against, after the
  * governor down-shifts as far as it can) AND the fresh-tail residual
  * (boundFreshTailTotalToResidual via lcd-assembler) so the two can NEVER diverge.
@@ -54,8 +53,9 @@ export declare function computeOutputHeadroom(reasoningStyle: "none" | "native",
  * …→medium→low→undefined) and at "off" for a non-reasoning model. So the floor reserve is
  * the native "low" thinking reserve (1024) for native, 0 for none — PLUS the visible
  * output floor. A native model that omits this reserve under-counts by ~1024+ and ships a
- * fresh tail the pre-flight then exhausts on (the live gpt-5-nano t6: Fix C used the none
- * floor 768, the pre-flight reserved the native floor 1792 → an ~873 gap → overflow).
+ * fresh tail the pre-flight then exhausts on (observed live on gpt-5-nano: the fresh-tail
+ * bound used the none floor 768, the pre-flight reserved the native floor 1792 → an ~873
+ * gap → overflow).
  *
  * @param reasoningStyle - the model's ACTUAL reasoning style ("none" | "native").
  * @param visibleOutputFloor - minVisibleOutputTokens (defaults to MIN_VISIBLE_OUTPUT_TOKENS).

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

@@ -1,12 +1,11 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * Reasoning-aware output headroom primitives (Fix 3 / Phase 166 CWF-02).
+ * Reasoning-aware output headroom primitives.
  *
  * Pure functions; no runtime deps, no I/O.
  * Consumers: lcd-assembler.ts (pre-flight fit check),
  *            config-resolver.ts (dynamic max_tokens clamp),
  *            executor-stream-setup.ts (thinking-effort governor).
- * Design ref: design/small-model-context-fidelity.md §4 Fix 3 + §3 corollaries.
  */
 import { MIN_VISIBLE_OUTPUT_TOKENS } from "./constants.js";
 export { MIN_VISIBLE_OUTPUT_TOKENS };
@@ -18,7 +17,7 @@ export { MIN_VISIBLE_OUTPUT_TOKENS };
  * "native" style = empirically sized per qwen3.6:35b observed blocks.
  *
  * Sizing rationale — sized to qwen3.6:35b observed high-thinking block
- * (the NVDA incident: in 32477 + out 291 = 32768, stopReason 'length',
+ * (observed live: in 32477 + out 291 = 32768, stopReason 'length',
  * 0 tool calls — high-effort thinking consumed essentially the entire 32K window).
  * native/high=8192 is a defensible middle ground: large enough that a high-thinking
  * block is not silently truncated on a 32K window (leaves ~23K for history),
@@ -41,7 +40,7 @@ export const THINKING_RESERVE_TOKENS = {
  * @param reasoningStyle      - Model's reasoning style: "none" (standard) or "native" (thinking block).
  * @param thinkingLevel       - Configured thinking effort level.
  * @param visibleOutputFloor  - Minimum visible output tokens (defaults to MIN_VISIBLE_OUTPUT_TOKENS).
- *                              WR-02: pass contextEngine.budget.minVisibleOutputTokens here when
+ *                              Pass contextEngine.budget.minVisibleOutputTokens here when
  *                              the config value differs from the compile-time constant.
  * @returns Total output headroom tokens to reserve from the context window.
  */
@@ -51,7 +50,7 @@ export function computeOutputHeadroom(reasoningStyle, thinkingLevel, visibleOutp
 /**
  * The FLOOR output headroom — the minimum the thinking-effort governor can reach.
  *
- * ISSUE #3b (2026-06-22): the SINGLE source for "how much headroom is reserved at the
+ * The SINGLE source for "how much headroom is reserved at the
  * floor", used by BOTH the pre-flight (the bound it ultimately throws against, after the
  * governor down-shifts as far as it can) AND the fresh-tail residual
  * (boundFreshTailTotalToResidual via lcd-assembler) so the two can NEVER diverge.
@@ -60,8 +59,9 @@ export function computeOutputHeadroom(reasoningStyle, thinkingLevel, visibleOutp
  * …→medium→low→undefined) and at "off" for a non-reasoning model. So the floor reserve is
  * the native "low" thinking reserve (1024) for native, 0 for none — PLUS the visible
  * output floor. A native model that omits this reserve under-counts by ~1024+ and ships a
- * fresh tail the pre-flight then exhausts on (the live gpt-5-nano t6: Fix C used the none
- * floor 768, the pre-flight reserved the native floor 1792 → an ~873 gap → overflow).
+ * fresh tail the pre-flight then exhausts on (observed live on gpt-5-nano: the fresh-tail
+ * bound used the none floor 768, the pre-flight reserved the native floor 1792 → an ~873
+ * gap → overflow).
  *
  * @param reasoningStyle - the model's ACTUAL reasoning style ("none" | "native").
  * @param visibleOutputFloor - minVisibleOutputTokens (defaults to MIN_VISIBLE_OUTPUT_TOKENS).

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

@@ -1,5 +1,5 @@
 /**
- * DEPTH-01 — cache-stable relevance eviction of the evictable middle band.
+ * Cache-stable relevance eviction of the evictable middle band.
  *
  * The relevance-ranked replacement for the pure-recency middle-band fill at the
  * `margin-arbiter.ts:296` seam (`evictHistoryUnderBudget(middleBand, remainingPool)`). Where
@@ -9,7 +9,7 @@
  * SELECTION only — never output order (the downstream assembler assumes chronological order:
  * transcript repair + breakpoint math).
  *
- * Cache stability (the LOCKED #1 / lossless-claw §6.3 mistake to avoid — CONSERVATIVE gate):
+ * Cache stability (CONSERVATIVE gate — reordering a cached prefix is the costly mistake to avoid):
  *   - `supportsPromptCache === true` → DO NOT reorder. Return the recency fill
  *     (`evictHistoryUnderBudget`) so the cached prefix is byte-stable. A caching profile keeps
  *     recency above the cache fence; the "re-rank below the fence" sub-case is NOT built (the
@@ -20,9 +20,9 @@
  *   - `supportsPromptCache === false` (the typical local small/nano relevance-first
  *     population) → re-rank the full evictable band freely.
  *
- * Pinned survival (DEPTH-01 success criterion #2 / RETR-05): the middle band ALREADY excludes
- * S4 pins by construction (`margin-arbiter.ts:235`, step-atomic via `expandFloorsToStepIds`),
- * so the band this pass receives is pin-free. As DEFENSE-IN-DEPTH (the S4 suite extends to
+ * Pinned survival: the middle band ALREADY excludes security
+ * pins by construction (`margin-arbiter.ts:235`, step-atomic via `expandFloorsToStepIds`),
+ * so the band this pass receives is pin-free. As DEFENSE-IN-DEPTH (security pinning extends to
  * eviction), if a security-marked message nonetheless reaches the band (markers threaded), it
  * is separated out as an UNCONDITIONAL survivor (mirroring the arbiter's floor handling) and
  * is NEVER subject to the relevance budget — a relevance score can never evict a pin.
@@ -36,13 +36,13 @@
  *   - agent↛memory: it imports only `@comis/core` TYPES + in-package context-engine modules —
  *     never `@comis/memory` (forbidden by `test/architecture` "agent → memory"). The store
  *     arrives as the core `ContextStorePort` TYPE only (injected via `deps.contextStore`).
- *   - I2 (context-engine ↮ rag): the shared relevance scorer (`rag/relevance-scorer.ts`) is
+ *   - context-engine ↮ rag: the shared relevance scorer (`rag/relevance-scorer.ts`) is
  *     NOT imported — it is INJECTED as `deps.relevanceScorer` (DI at the budget boundary), and
  *     the `ArbiterRelevanceQuery`/`RelevanceScorerFn` shapes are the LOCAL structural contracts
- *     declared in `margin-arbiter.ts` (the rag types are structurally compatible). This is the
- *     exact cut that bit 173-03 — the relevance scorer is INJECTED, never imported.
+ *     declared in `margin-arbiter.ts` (the rag types are structurally compatible). The
+ *     relevance scorer is INJECTED, never imported.
  *
- * Match (WR-01, Phase 174-04): a band message is associated with its `searchLcd` hit by the
+ * Match: a band message is associated with its `searchLcd` hit by the
  * hit's STABLE `refId` (= `lcd_messages.id`), carried onto the BudgetItem as `lcdId` at
  * assembly (`lcd-assembler.resolveContextItem`). This is the robust match — a pure
  * `tool_use`/`tool_result` message (whose live block-text render is "" but whose stored FTS
@@ -55,7 +55,7 @@
  * `contextStore.searchLcd` read. Same input → same output (the FTS read is deterministic for a
  * fixed store + query). The `searchLcd` snippet is UNTRUSTED — on the id-less fallback it is
  * used ONLY transiently to associate a band message with its rank position and is NEVER logged
- * or returned (T-174-01-03); only the resulting ordinal rank drives selection.
+ * or returned; only the resulting ordinal rank drives selection.
  *
  * @module
  */
@@ -69,10 +69,10 @@ import type { ContextEngineDeps } from "./types.js";
  * pin-protecting replacement for the recency `evictHistoryUnderBudget(middleBand, poolTokens)`
  * at the margin-arbiter middle-band seam.
  *
- * @param deps - the context-engine deps (modelProfile, relevanceScorer, contextStore, R4 scope).
+ * @param deps - the context-engine deps (modelProfile, relevanceScorer, contextStore, read scope).
  * @param middleBand - the evictable (already pin-free by construction) history band + tokens.
  * @param poolTokens - the discretionary pool the middle band may consume (post-floors).
- * @param _liveMessages - CURRENTLY UNUSED (IN-02, Phase 174-04). The relevance query is built
+ * @param _liveMessages - CURRENTLY UNUSED. The relevance query is built
  *   UPSTREAM by the seam (`evictUnderArbiter` → `buildAssemblyRelevanceQuery`) and passed as
  *   `query`; this function never reads the live array. Retained for signature symmetry with the
  *   recency fill / a possible future in-function query refinement — prefixed `_` (unused).

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

@@ -1,6 +1,6 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * DEPTH-01 — cache-stable relevance eviction of the evictable middle band.
+ * Cache-stable relevance eviction of the evictable middle band.
  *
  * The relevance-ranked replacement for the pure-recency middle-band fill at the
  * `margin-arbiter.ts:296` seam (`evictHistoryUnderBudget(middleBand, remainingPool)`). Where
@@ -10,7 +10,7 @@
  * SELECTION only — never output order (the downstream assembler assumes chronological order:
  * transcript repair + breakpoint math).
  *
- * Cache stability (the LOCKED #1 / lossless-claw §6.3 mistake to avoid — CONSERVATIVE gate):
+ * Cache stability (CONSERVATIVE gate — reordering a cached prefix is the costly mistake to avoid):
  *   - `supportsPromptCache === true` → DO NOT reorder. Return the recency fill
  *     (`evictHistoryUnderBudget`) so the cached prefix is byte-stable. A caching profile keeps
  *     recency above the cache fence; the "re-rank below the fence" sub-case is NOT built (the
@@ -21,9 +21,9 @@
  *   - `supportsPromptCache === false` (the typical local small/nano relevance-first
  *     population) → re-rank the full evictable band freely.
  *
- * Pinned survival (DEPTH-01 success criterion #2 / RETR-05): the middle band ALREADY excludes
- * S4 pins by construction (`margin-arbiter.ts:235`, step-atomic via `expandFloorsToStepIds`),
- * so the band this pass receives is pin-free. As DEFENSE-IN-DEPTH (the S4 suite extends to
+ * Pinned survival: the middle band ALREADY excludes security
+ * pins by construction (`margin-arbiter.ts:235`, step-atomic via `expandFloorsToStepIds`),
+ * so the band this pass receives is pin-free. As DEFENSE-IN-DEPTH (security pinning extends to
  * eviction), if a security-marked message nonetheless reaches the band (markers threaded), it
  * is separated out as an UNCONDITIONAL survivor (mirroring the arbiter's floor handling) and
  * is NEVER subject to the relevance budget — a relevance score can never evict a pin.
@@ -37,13 +37,13 @@
  *   - agent↛memory: it imports only `@comis/core` TYPES + in-package context-engine modules —
  *     never `@comis/memory` (forbidden by `test/architecture` "agent → memory"). The store
  *     arrives as the core `ContextStorePort` TYPE only (injected via `deps.contextStore`).
- *   - I2 (context-engine ↮ rag): the shared relevance scorer (`rag/relevance-scorer.ts`) is
+ *   - context-engine ↮ rag: the shared relevance scorer (`rag/relevance-scorer.ts`) is
  *     NOT imported — it is INJECTED as `deps.relevanceScorer` (DI at the budget boundary), and
  *     the `ArbiterRelevanceQuery`/`RelevanceScorerFn` shapes are the LOCAL structural contracts
- *     declared in `margin-arbiter.ts` (the rag types are structurally compatible). This is the
- *     exact cut that bit 173-03 — the relevance scorer is INJECTED, never imported.
+ *     declared in `margin-arbiter.ts` (the rag types are structurally compatible). The
+ *     relevance scorer is INJECTED, never imported.
  *
- * Match (WR-01, Phase 174-04): a band message is associated with its `searchLcd` hit by the
+ * Match: a band message is associated with its `searchLcd` hit by the
  * hit's STABLE `refId` (= `lcd_messages.id`), carried onto the BudgetItem as `lcdId` at
  * assembly (`lcd-assembler.resolveContextItem`). This is the robust match — a pure
  * `tool_use`/`tool_result` message (whose live block-text render is "" but whose stored FTS
@@ -56,7 +56,7 @@
  * `contextStore.searchLcd` read. Same input → same output (the FTS read is deterministic for a
  * fixed store + query). The `searchLcd` snippet is UNTRUSTED — on the id-less fallback it is
  * used ONLY transiently to associate a band message with its rank position and is NEVER logged
- * or returned (T-174-01-03); only the resulting ordinal rank drives selection.
+ * or returned; only the resulting ordinal rank drives selection.
  *
  * @module
  */
@@ -68,10 +68,10 @@ import { isSecurityRelevantMessage } from "./security-context-pinner.js";
  * pin-protecting replacement for the recency `evictHistoryUnderBudget(middleBand, poolTokens)`
  * at the margin-arbiter middle-band seam.
  *
- * @param deps - the context-engine deps (modelProfile, relevanceScorer, contextStore, R4 scope).
+ * @param deps - the context-engine deps (modelProfile, relevanceScorer, contextStore, read scope).
  * @param middleBand - the evictable (already pin-free by construction) history band + tokens.
  * @param poolTokens - the discretionary pool the middle band may consume (post-floors).
- * @param _liveMessages - CURRENTLY UNUSED (IN-02, Phase 174-04). The relevance query is built
+ * @param _liveMessages - CURRENTLY UNUSED. The relevance query is built
  *   UPSTREAM by the seam (`evictUnderArbiter` → `buildAssemblyRelevanceQuery`) and passed as
  *   `query`; this function never reads the live array. Retained for signature symmetry with the
  *   recency fill / a possible future in-function query refinement — prefixed `_` (unused).
@@ -79,8 +79,8 @@ import { isSecurityRelevantMessage } from "./security-context-pinner.js";
  * @returns the kept middle-band messages in chronological order (selection reorders; output does not).
  */
 export function rankMiddleBandByRelevance(deps, middleBand, poolTokens, _liveMessages, query) {
-    // --- Cache gate (CONSERVATIVE — LOCKED #1): a caching profile keeps recency. ---
-    // Reordering above the cache fence shreds the KV cache (the lossless-claw §6.3 mistake).
+    // --- Cache gate (CONSERVATIVE): a caching profile keeps recency. ---
+    // Reordering above the cache fence shreds the KV cache.
     // Returning the recency fill keeps the cached prefix byte-stable.
     if (deps.modelProfile?.supportsPromptCache === true) {
         return evictHistoryUnderBudget(middleBand, poolTokens);
@@ -91,10 +91,10 @@ export function rankMiddleBandByRelevance(deps, middleBand, poolTokens, _liveMes
     if (middleBand.length === 0 || store === undefined || scorer === undefined || query.degraded) {
         return evictHistoryUnderBudget(middleBand, poolTokens);
     }
-    // --- Pinned survival (DEFENSE-IN-DEPTH, RETR-05): separate any security-marked item out
+    // --- Pinned survival (DEFENSE-IN-DEPTH): separate any security-marked item out
     //     of the relevance-evictable set as an UNCONDITIONAL survivor (never budget-evicted).
     //     The band is already pin-free by construction (margin-arbiter.ts:235); this is the
-    //     S4-extends-to-eviction backstop. Identity-tracked so the chronological restore re-admits
+    //     pinning-extends-to-eviction backstop. Identity-tracked so the chronological restore re-admits
     //     it regardless of the relevance budget. ---
     const markers = deps.securityPinMarkers;
     const pinnedMsgs = new Set();
@@ -109,9 +109,9 @@ export function rankMiddleBandByRelevance(deps, middleBand, poolTokens, _liveMes
         }
     }
     // --- FTS-the-band: rank the evictable band by BM25 relevance. ---
-    // The R4 scope is built from the same deps the assembler uses (WR-02 agent/tenant isolation).
+    // The read scope is built from the same deps the assembler uses (agent/tenant isolation).
     // searchLcd returns BM25-`rank`ed hits best-first; the per-item rank is the position of the
-    // hit whose STABLE `refId` (= lcd_messages.id) equals the band item's `lcdId` (WR-01 — the
+    // hit whose STABLE `refId` (= lcd_messages.id) equals the band item's `lcdId` (the
     // robust id-based match). Unmatched items rank WORST (after every matched item) — they fall
     // back to recency within the chronological restore. NB: the snippet is read ONLY for the
     // legacy id-less fallback association, never logged/returned.
@@ -124,9 +124,9 @@ export function rankMiddleBandByRelevance(deps, middleBand, poolTokens, _liveMes
     // OR-join the (already alphanumeric, stopworded) terms so ANY term match ranks a message
     // (FTS5 MATCH of space-separated terms is implicit AND — an AND of the whole query almost
     // never matches a single historical message, which would silently degrade the pass to
-    // recency, the CR-01 no-op trap). The terms come from buildAssemblyRelevanceQuery's
-    // tokenizer (quotes + non-\p{L}\p{N} already stripped) AND ASSEMBLY_STOPWORDS now drops the
-    // FTS5 operator keywords near/and/or/not (IN-04), so the bare ` OR ` is FTS5-safe — no term
+    // recency — a no-op trap). The terms come from buildAssemblyRelevanceQuery's
+    // tokenizer (quotes + non-\p{L}\p{N} already stripped) AND ASSEMBLY_STOPWORDS drops the
+    // FTS5 operator keywords near/and/or/not, so the bare ` OR ` is FTS5-safe — no term
     // can be a bare operator that alters MATCH parsing.
     const ftsQuery = query.terms.join(" OR ");
     const result = store.searchLcd(scope, ftsQuery, {
@@ -138,11 +138,11 @@ export function rankMiddleBandByRelevance(deps, middleBand, poolTokens, _liveMes
         // No FTS signal → deterministic recency floor over the FULL band (pins included).
         return evictHistoryUnderBudget(middleBand, poolTokens);
     }
-    // WR-01: build the refId → best-first POSITION map ONCE. `searchLcd` returns hits ordered
+    // Build the refId → best-first POSITION map ONCE. `searchLcd` returns hits ordered
     // by relevance (BM25 rank on the FTS path; SELECT order on the LIKE fallback), and each hit
     // carries `refId` = lcd_messages.id. Keying on the stable id (not the snippet text) is the
     // robust match: a pure tool_use/tool_result message — whose live block-text render is "" but
-    // whose stored FTS snippet is the toolName+JSON envelope — now ranks correctly, and the
+    // whose stored FTS snippet is the toolName+JSON envelope — ranks correctly, and the
     // non-FTS5 LIKE fallback (snippet = metadata JSON, no BM25 rank) ranks too. First occurrence
     // of an id wins (best position). The snippet is NEVER read for an id-carrying item.
     const positionByRefId = new Map();
@@ -152,7 +152,7 @@ export function rankMiddleBandByRelevance(deps, middleBand, poolTokens, _liveMes
             positionByRefId.set(refId, i);
     }
     // Per-item relevance ordinal: lower = more relevant (best-first hit position). An item with a
-    // store `lcdId` matches by `refId` (WR-01 robust path); an id-less item (live/synthetic — no
+    // store `lcdId` matches by `refId` (the robust path); an id-less item (live/synthetic — no
     // store row) falls back to the legacy snippet-substring association. Unmatched → a large
     // sentinel so matched items always outrank them; ties broken by chronological index so the
     // order is total + deterministic.
@@ -181,7 +181,8 @@ export function rankMiddleBandByRelevance(deps, middleBand, poolTokens, _liveMes
         const id = `band-${idx}`;
         idToItem.set(id, it);
         // content is intentionally empty — the scorer ranks the pre-ordered single FTS lane by
-        // position (single-lane identity), it does not re-tokenize content here (I2 cut clean).
+        // position (single-lane identity), it does not re-tokenize content here (keeps the
+        // context-engine ↮ rag cut clean).
         return { entry: { id, content: "" }, score: undefined };
     });
     const fused = scorer([{ results: lane, weight: 1 }], query, {
@@ -211,7 +212,7 @@ export function rankMiddleBandByRelevance(deps, middleBand, poolTokens, _liveMes
     const chronoAdmitted = wholeStepsContaining(middleBand, admittedEvictable);
     const keptEvictable = new Set(evictHistoryUnderBudget(chronoAdmitted, poolTokens));
     // --- Stitch the kept history in ORIGINAL order: pins (UNCONDITIONAL, ride on top of the
-    //     pool — a relevance score can never evict a pin, RETR-05) PLUS the kept evictable set,
+    //     pool — a relevance score can never evict a pin) PLUS the kept evictable set,
     //     preserving middleBand order. The relevance pass drove SELECTION; output is chronological. ---
     const kept = [];
     for (const it of middleBand) {
@@ -225,16 +226,16 @@ export function rankMiddleBandByRelevance(deps, middleBand, poolTokens, _liveMes
 // Internals (pure)
 // ---------------------------------------------------------------------------
 /**
- * LEGACY id-less FALLBACK (WR-01): the relevance ordinal for a band message that has NO store
+ * LEGACY id-less FALLBACK: the relevance ordinal for a band message that has NO store
  * `lcdId` (a live/synthetic item — the fresh tail, a coalesced summary message, a unit fixture
  * with no id) — the position (0-based, best-first) of the FIRST FTS hit whose snippet contains
  * the message's rendered text. Returns a large sentinel when no hit matches (so matched items
  * always outrank unmatched ones). The snippet is used here ONLY for the transient match (never
  * logged/returned). A blank rendering never matches (→ sentinel).
  *
- * KNOWN RESIDUAL (DEPTH-01 limitation): this substring fallback STILL mis-handles a pure
+ * KNOWN RESIDUAL: this substring fallback STILL mis-handles a pure
  * tool_use/tool_result message (empty block-text render) and the non-FTS5 LIKE metadata-JSON
- * snippet for an id-less item. Store-resolved message-refs now carry `lcdId` and never reach
+ * snippet for an id-less item. Store-resolved message-refs carry `lcdId` and never reach
  * here (they match by `refId`), so the residual is confined to the genuinely id-less items
  * above — which are rare on the relevance-first path and correctly fall back to recency.
  */

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

@@ -1,7 +1,7 @@
 /**
  * Security context pinning for eviction/compaction passes.
  *
- * S4: identifies stored messages that contain security-critical markers
+ * Identifies stored messages that contain security-critical markers
  * (canary token, wrapExternalContent delimiters, sender-trust prefixes,
  * safety reinforcement text) and marks them as ineligible for eviction.
  * Fail-closed: uncertain → treat as security-relevant (pin it).
@@ -20,7 +20,7 @@ export interface SecurityPinMarkers {
      * Sender-trust section prefix for messages injected by buildSenderTrustSection
      * (canonical value: "## Authorized Senders"). When set, messages containing
      * this prefix are pinned — the trust table must survive compaction so the model
-     * always knows which senders are authorized. S4: one of the four pinned categories.
+     * always knows which senders are authorized. One of the four pinned categories.
      */
     senderTrustPrefix?: string;
 }
@@ -28,7 +28,7 @@ export interface SecurityPinMarkers {
  * Returns true if the message text contains any security-critical marker.
  * Fail-closed: empty/undefined message text → true (pin it).
  *
- * S4: pinned messages are excluded from eviction/summarization chunk selection
+ * Pinned messages are excluded from eviction/summarization chunk selection
  * in BOTH the pipeline llm-compaction layer AND the LCD leaf-summarizer.
  */
 export declare function isSecurityRelevantMessage(msg: {

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

@@ -2,7 +2,7 @@
 /**
  * Security context pinning for eviction/compaction passes.
  *
- * S4: identifies stored messages that contain security-critical markers
+ * Identifies stored messages that contain security-critical markers
  * (canary token, wrapExternalContent delimiters, sender-trust prefixes,
  * safety reinforcement text) and marks them as ineligible for eviction.
  * Fail-closed: uncertain → treat as security-relevant (pin it).
@@ -13,7 +13,7 @@
  * Returns true if the message text contains any security-critical marker.
  * Fail-closed: empty/undefined message text → true (pin it).
  *
- * S4: pinned messages are excluded from eviction/summarization chunk selection
+ * Pinned messages are excluded from eviction/summarization chunk selection
  * in BOTH the pipeline llm-compaction layer AND the LCD leaf-summarizer.
  */
 export function isSecurityRelevantMessage(

package/node_modules/@comis/agent/dist/context-engine/signature-surrogate-guard.d.ts

@@ -10,9 +10,8 @@
  * Reference: pi-ai upstream behavior in
  * `node_modules/.pnpm/@mariozechner+pi-ai@0.67.68_HASH/node_modules/@earendil-works/pi-ai/dist/providers/anthropic.js`
  * around line 661 calls `sanitizeSurrogates(block.thinking)` while preserving
- * the original `signature`. A separate upstream PR is recommended (out of
- * scope for this commit); Comis is protected by this guard regardless of
- * upstream.
+ * the original `signature`. An upstream fix would remove the root cause;
+ * Comis is protected by this guard regardless of upstream.
  *
  * Strategy: scan every type:"thinking" block on every assistant message
  * for unpaired UTF-16 surrogates in `block.thinking`. When found AND the
@@ -53,7 +52,7 @@ export interface SignatureSurrogateGuardDeps {
 /**
  * Create the signature-surrogate-guard pipeline layer.
  *
- * Layer ordering: runs AFTER `signature-replay-scrubber` (Fix #2) and
+ * Layer ordering: runs AFTER `signature-replay-scrubber` and
  * BEFORE `reasoning-tag-stripper`. Always added unconditionally — the cost
  * is one walk over thinking blocks with two regex tests per block, which
  * is negligible compared to a single 400-rejection round trip.

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

@@ -11,9 +11,8 @@
  * Reference: pi-ai upstream behavior in
  * `node_modules/.pnpm/@mariozechner+pi-ai@0.67.68_HASH/node_modules/@earendil-works/pi-ai/dist/providers/anthropic.js`
  * around line 661 calls `sanitizeSurrogates(block.thinking)` while preserving
- * the original `signature`. A separate upstream PR is recommended (out of
- * scope for this commit); Comis is protected by this guard regardless of
- * upstream.
+ * the original `signature`. An upstream fix would remove the root cause;
+ * Comis is protected by this guard regardless of upstream.
  *
  * Strategy: scan every type:"thinking" block on every assistant message
  * for unpaired UTF-16 surrogates in `block.thinking`. When found AND the
@@ -50,7 +49,7 @@ const UNPAIRED_LOW_SURROGATE = /(?<![\uD800-\uDBFF])[\uDC00-\uDFFF]/;
 /**
  * Create the signature-surrogate-guard pipeline layer.
  *
- * Layer ordering: runs AFTER `signature-replay-scrubber` (Fix #2) and
+ * Layer ordering: runs AFTER `signature-replay-scrubber` and
  * BEFORE `reasoning-tag-stripper`. Always added unconditionally — the cost
  * is one walk over thinking blocks with two regex tests per block, which
  * is negligible compared to a single 400-rejection round trip.

package/node_modules/@comis/agent/dist/context-engine/summarize-prompt-style.d.ts

@@ -1,5 +1,5 @@
 /**
- * Depth-aware summarization prompt style builder (Phase 171, SUM-01).
+ * Depth-aware summarization prompt style builder.
  *
  * Pure function: no I/O, no imports from @comis/memory (agent↛memory cut).
  * NEVER logs instruction content — only structural metadata.
@@ -7,10 +7,11 @@
  * @module
  */
 /**
- * GEN-01 (I7): the single language-preservation sentence shared by the dag depth
+ * The single language-preservation sentence shared by the dag depth
  * templates AND the pipeline compaction instructions (`llm-compaction.ts`). Defined
  * ONCE here and imported by the pipeline site so the two can never silently drift.
- * Content-anchored — it sidesteps the DET-02 reply-language tag (no tag).
+ * Content-anchored — it keys the language to the source content itself rather than
+ * to a reply-language tag.
  *
  * The machine-parsed scaffolding (the dag `Files:`/`Expand for:`/`[SUPERSEDED]`
  * tokens, the pipeline's nine `## Section` headings) stays verbatim-English; only
@@ -20,7 +21,7 @@ export declare const LANGUAGE_PRESERVATION_INSTRUCTION = "Write the summary in t
 /**
  * Build the depth-keyed summarization instructions for the leaf/condense passes.
  *
- * Depth mapping (per design/lcd-v3-unified-substrate.md §6.4):
+ * Depth mapping:
  *  - d0 (leaf): extractive — concrete facts, file paths, decisions, tool outcomes
  *  - d1 (timeline): chronological — mark superseded decisions
  *  - d2 (trajectory): high-level arc — drop per-session minutiae

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

@@ -1,6 +1,6 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * Depth-aware summarization prompt style builder (Phase 171, SUM-01).
+ * Depth-aware summarization prompt style builder.
  *
  * Pure function: no I/O, no imports from @comis/memory (agent↛memory cut).
  * NEVER logs instruction content — only structural metadata.
@@ -8,10 +8,11 @@
  * @module
  */
 /**
- * GEN-01 (I7): the single language-preservation sentence shared by the dag depth
+ * The single language-preservation sentence shared by the dag depth
  * templates AND the pipeline compaction instructions (`llm-compaction.ts`). Defined
  * ONCE here and imported by the pipeline site so the two can never silently drift.
- * Content-anchored — it sidesteps the DET-02 reply-language tag (no tag).
+ * Content-anchored — it keys the language to the source content itself rather than
+ * to a reply-language tag.
  *
  * The machine-parsed scaffolding (the dag `Files:`/`Expand for:`/`[SUPERSEDED]`
  * tokens, the pipeline's nine `## Section` headings) stays verbatim-English; only
@@ -21,7 +22,7 @@ export const LANGUAGE_PRESERVATION_INSTRUCTION = "Write the summary in the domin
 /**
  * Build the depth-keyed summarization instructions for the leaf/condense passes.
  *
- * Depth mapping (per design/lcd-v3-unified-substrate.md §6.4):
+ * Depth mapping:
  *  - d0 (leaf): extractive — concrete facts, file paths, decisions, tool outcomes
  *  - d1 (timeline): chronological — mark superseded decisions
  *  - d2 (trajectory): high-level arc — drop per-session minutiae

package/node_modules/@comis/agent/dist/context-engine/summarize-tier-targets.d.ts

@@ -1,5 +1,5 @@
 /**
- * Tier-aware summary token target resolver (Phase 171, SUM-02).
+ * Tier-aware summary token target resolver.
  *
  * Provides resolveSummaryTargetTokens() keyed on capabilityClass + depth,
  * and the deterministic nano structured extractor floor. No I/O, no network,
@@ -12,7 +12,7 @@ import type { AgentMessage } from "@earendil-works/pi-agent-core";
 /**
  * Resolve the effective summary target token count for the given tier.
  *
- * Rules (SUM-02):
+ * Rules:
  *  - nano   → min(256, configuredTarget)
  *  - small  → min(400, configuredTarget)
  *  - mid    → min(800, configuredTarget)
@@ -35,8 +35,8 @@ configuredTarget: number): number;
  * Replaces the bare count-note (`LEAF_FALLBACK_SUMMARY_MARKER N messages`) with a
  * structured JSON block carrying decisions/files/entities/constraints extracted by
  * simple string heuristics (no LLM). The result MUST:
- *   1. Carry LEAF_FALLBACK_SUMMARY_MARKER so DOC-01 scans detect it.
- *   2. Pass the shrink invariant (C1): tokenCount strictly < chunkTokens.
+ *   1. Carry LEAF_FALLBACK_SUMMARY_MARKER so fallback-marker scans detect it.
+ *   2. Pass the shrink invariant: tokenCount strictly < chunkTokens.
  *
  * If the structured JSON exceeds `shrinkCeilingTokens` (from computeShrinkBounds),
  * falls back to a bare count-note that is guaranteed to be smaller.

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

@@ -1,6 +1,6 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * Tier-aware summary token target resolver (Phase 171, SUM-02).
+ * Tier-aware summary token target resolver.
  *
  * Provides resolveSummaryTargetTokens() keyed on capabilityClass + depth,
  * and the deterministic nano structured extractor floor. No I/O, no network,
@@ -17,7 +17,7 @@ import { estimateMessageTokens, estimateMessageChars } from "../safety/token-est
 /**
  * Resolve the effective summary target token count for the given tier.
  *
- * Rules (SUM-02):
+ * Rules:
  *  - nano   → min(256, configuredTarget)
  *  - small  → min(400, configuredTarget)
  *  - mid    → min(800, configuredTarget)
@@ -58,8 +58,8 @@ configuredTarget) {
  * Replaces the bare count-note (`LEAF_FALLBACK_SUMMARY_MARKER N messages`) with a
  * structured JSON block carrying decisions/files/entities/constraints extracted by
  * simple string heuristics (no LLM). The result MUST:
- *   1. Carry LEAF_FALLBACK_SUMMARY_MARKER so DOC-01 scans detect it.
- *   2. Pass the shrink invariant (C1): tokenCount strictly < chunkTokens.
+ *   1. Carry LEAF_FALLBACK_SUMMARY_MARKER so fallback-marker scans detect it.
+ *   2. Pass the shrink invariant: tokenCount strictly < chunkTokens.
  *
  * If the structured JSON exceeds `shrinkCeilingTokens` (from computeShrinkBounds),
  * falls back to a bare count-note that is guaranteed to be smaller.
@@ -122,7 +122,7 @@ export function buildNanoStructuredExtraction(messages, chunkTokens) {
     const renderedChars = messages.reduce((acc, m) => acc + estimateMessageChars(m), 0);
     const { shrinkCeilingTokens } = computeShrinkBounds(renderedChars, 256);
     const tokenCount = estimateMessageTokens({ role: "user", content: structured });
-    // C1 / shrink invariant: if structured output is not smaller than the ceiling,
+    // Shrink invariant: if structured output is not smaller than the ceiling,
     // fall back to the bare count-note (always tiny, always beats the ceiling)
     if (tokenCount >= shrinkCeilingTokens) {
         return buildBareCountNote(messages.length, chunkTokens);