comisai 1.0.44 → 1.0.46

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

@@ -1,12 +1,12 @@
 /**
- * SUMW-01 (Phase 178): THE one resolved-summarizer window read, extracted into
- * its own module from `lcd-leaf-summarizer.ts` (which sits at the 800-line
- * file-size cap). The deps shape stays {@link LeafSummarizerDeps} — this module
+ * THE one resolved-summarizer window read, kept in its own module so
+ * `lcd-leaf-summarizer.ts` stays under the 800-line file-size cap.
+ * The deps shape stays {@link LeafSummarizerDeps} — this module
  * only owns the window-resolution read; the summarizer seam, chunk selection,
  * and escalation ladder remain in `lcd-leaf-summarizer.ts`.
  *
  * Consumed by: llm-compaction (pipeline span clamp, via
- * {@link effectiveSummarizerWindow}) and the 178-03 LCD leaf/condense clamps
+ * {@link effectiveSummarizerWindow}) and the LCD leaf/condense clamps
  * (lcd-compaction-trigger / lcd-condense-trigger, via
  * {@link resolveSummarizerWindowTokens}).
  *
@@ -14,10 +14,10 @@
  */
 import type { LeafSummarizerDeps } from "./lcd-leaf-summarizer.js";
 /**
- * INT-W1: combine a candidate summarizer's CONFIGURED contextWindow with the
- * provider-SERVED window bound to that candidate (Phase 176) into the one
+ * Combine a candidate summarizer's CONFIGURED contextWindow with the
+ * provider-SERVED window bound to that candidate into the one
  * effective window the span clamps size against. Finite-positive guards on
- * both inputs (WR-05 parity — an invalid number never disables or inflates a
+ * both inputs (an invalid number never disables or inflates a
  * clamp):
  *
  *  - both valid → `min(configured, served)` — the provider truncates anything
@@ -34,10 +34,10 @@ import type { LeafSummarizerDeps } from "./lcd-leaf-summarizer.js";
  */
 export declare function effectiveSummarizerWindow(configuredWindow: unknown, servedWindow: number | undefined): number | undefined;
 /**
- * SUMW-01 (Phase 178): THE one resolved-summarizer window read. Mirrors
+ * THE one resolved-summarizer window read. Mirrors
  * `buildLeafSummarizeFn`'s PRIMARY model resolution EXACTLY
  * (`overrideModel?.model ?? getRealModel()`) so a span clamp and the primary
- * LLM call always agree about WHICH model summarizes. Pitfall 2: `getModel()`
+ * LLM call always agree about WHICH model summarizes. Pitfall: `getModel()`
  * is the session-PRIMARY snapshot — with an `operationModels.compaction`
  * override the summarizer is a DIFFERENT model; a clamp keyed to the primary
  * would pass a 131K span to an 8K summarizer. `getRealModel` is optional-called
@@ -47,17 +47,17 @@ export declare function effectiveSummarizerWindow(configuredWindow: unknown, ser
  * finite-positive guard falls back to `getModel().contextWindow` (the snapshot
  * — never silently huge).
  *
- * INT-W1 (served-window truth): the configured window is then min()'d with the
- * Phase-176 provider-SERVED window bound to the SAME candidate the `??` chain
+ * Served-window truth: the configured window is then min()'d with the
+ * provider-SERVED window bound to the SAME candidate the `??` chain
  * resolved — `overrideModel.servedWindow` when the override summarizes,
  * `primaryServedWindow` when the primary does. Both values were provider-gated
- * at the wiring site against the 176 `{providerKey, window}` pair (WR-02
- * config-key space), so a served bound can never clamp a summarizer on another
+ * at the wiring site against the probed `{providerKey, window}` pair, so a
+ * served bound can never clamp a summarizer on another
  * provider; on the flagship gap (configured 131_072, Ollama serving 8_192,
  * summarizer = primary) this resolves 8_192 and the leaf/condense/pipeline
  * clamps bind against the window the provider will actually serve.
  *
- * KNOWN LIMIT (review IN-02): this resolves the PRIMARY summarizer only. The
+ * KNOWN LIMIT: this resolves the PRIMARY summarizer only. The
  * production `summarize` seam may be failover-wrapped
  * (`wrapSummarizerWithFailover` via `summarizerFallbackProviders`): on
  * primary failure a fallback model with a possibly SMALLER window serves the
@@ -66,7 +66,7 @@ export declare function effectiveSummarizerWindow(configuredWindow: unknown, ser
  * through the escalation ladder (degraded, never data loss). Clamping to
  * min(primary, ...fallback windows) is a deliberate non-goal until a live
  * incident motivates it.
- * Consumed by: llm-compaction (pipeline span clamp) and the 178-03 LCD
+ * Consumed by: llm-compaction (pipeline span clamp) and the LCD
  * leaf/condense clamps.
  */
 export declare function resolveSummarizerWindowTokens(deps: Pick<LeafSummarizerDeps, "overrideModel" | "getRealModel" | "getModel" | "primaryServedWindow">): number;

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

@@ -1,23 +1,23 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * SUMW-01 (Phase 178): THE one resolved-summarizer window read, extracted into
- * its own module from `lcd-leaf-summarizer.ts` (which sits at the 800-line
- * file-size cap). The deps shape stays {@link LeafSummarizerDeps} — this module
+ * THE one resolved-summarizer window read, kept in its own module so
+ * `lcd-leaf-summarizer.ts` stays under the 800-line file-size cap.
+ * The deps shape stays {@link LeafSummarizerDeps} — this module
  * only owns the window-resolution read; the summarizer seam, chunk selection,
  * and escalation ladder remain in `lcd-leaf-summarizer.ts`.
  *
  * Consumed by: llm-compaction (pipeline span clamp, via
- * {@link effectiveSummarizerWindow}) and the 178-03 LCD leaf/condense clamps
+ * {@link effectiveSummarizerWindow}) and the LCD leaf/condense clamps
  * (lcd-compaction-trigger / lcd-condense-trigger, via
  * {@link resolveSummarizerWindowTokens}).
  *
  * @module
  */
 /**
- * INT-W1: combine a candidate summarizer's CONFIGURED contextWindow with the
- * provider-SERVED window bound to that candidate (Phase 176) into the one
+ * Combine a candidate summarizer's CONFIGURED contextWindow with the
+ * provider-SERVED window bound to that candidate into the one
  * effective window the span clamps size against. Finite-positive guards on
- * both inputs (WR-05 parity — an invalid number never disables or inflates a
+ * both inputs (an invalid number never disables or inflates a
  * clamp):
  *
  *  - both valid → `min(configured, served)` — the provider truncates anything
@@ -44,10 +44,10 @@ export function effectiveSummarizerWindow(configuredWindow, servedWindow) {
     return cfg ?? served;
 }
 /**
- * SUMW-01 (Phase 178): THE one resolved-summarizer window read. Mirrors
+ * THE one resolved-summarizer window read. Mirrors
  * `buildLeafSummarizeFn`'s PRIMARY model resolution EXACTLY
  * (`overrideModel?.model ?? getRealModel()`) so a span clamp and the primary
- * LLM call always agree about WHICH model summarizes. Pitfall 2: `getModel()`
+ * LLM call always agree about WHICH model summarizes. Pitfall: `getModel()`
  * is the session-PRIMARY snapshot — with an `operationModels.compaction`
  * override the summarizer is a DIFFERENT model; a clamp keyed to the primary
  * would pass a 131K span to an 8K summarizer. `getRealModel` is optional-called
@@ -57,17 +57,17 @@ export function effectiveSummarizerWindow(configuredWindow, servedWindow) {
  * finite-positive guard falls back to `getModel().contextWindow` (the snapshot
  * — never silently huge).
  *
- * INT-W1 (served-window truth): the configured window is then min()'d with the
- * Phase-176 provider-SERVED window bound to the SAME candidate the `??` chain
+ * Served-window truth: the configured window is then min()'d with the
+ * provider-SERVED window bound to the SAME candidate the `??` chain
  * resolved — `overrideModel.servedWindow` when the override summarizes,
  * `primaryServedWindow` when the primary does. Both values were provider-gated
- * at the wiring site against the 176 `{providerKey, window}` pair (WR-02
- * config-key space), so a served bound can never clamp a summarizer on another
+ * at the wiring site against the probed `{providerKey, window}` pair, so a
+ * served bound can never clamp a summarizer on another
  * provider; on the flagship gap (configured 131_072, Ollama serving 8_192,
  * summarizer = primary) this resolves 8_192 and the leaf/condense/pipeline
  * clamps bind against the window the provider will actually serve.
  *
- * KNOWN LIMIT (review IN-02): this resolves the PRIMARY summarizer only. The
+ * KNOWN LIMIT: this resolves the PRIMARY summarizer only. The
  * production `summarize` seam may be failover-wrapped
  * (`wrapSummarizerWithFailover` via `summarizerFallbackProviders`): on
  * primary failure a fallback model with a possibly SMALLER window serves the
@@ -76,7 +76,7 @@ export function effectiveSummarizerWindow(configuredWindow, servedWindow) {
  * through the escalation ladder (degraded, never data loss). Clamping to
  * min(primary, ...fallback windows) is a deliberate non-goal until a live
  * incident motivates it.
- * Consumed by: llm-compaction (pipeline span clamp) and the 178-03 LCD
+ * Consumed by: llm-compaction (pipeline span clamp) and the LCD
  * leaf/condense clamps.
  */
 export function resolveSummarizerWindowTokens(deps) {
@@ -86,7 +86,7 @@ export function resolveSummarizerWindowTokens(deps) {
     const configured = typeof win === "number" && Number.isFinite(win) && win > 0
         ? win
         : deps.getModel().contextWindow; // documented fallback: the snapshot — never silently huge
-    // INT-W1: the served truth rides the SAME candidate the ?? above resolved
+    // The served truth rides the SAME candidate the ?? above resolved
     // (`!= null` is the exact nullish test `??` applies to overrideModel.model),
     // so clamp and call can never disagree about which provider's bound applies.
     const served = overrideResolved != null ? deps.overrideModel?.servedWindow : deps.primaryServedWindow;

package/node_modules/@comis/agent/dist/context-engine/thinking-block-cleaner.d.ts

@@ -8,11 +8,11 @@
  * cacheFenceIndex is intentionally NOT consulted to gate stripping. The
  * cleaner is pure/deterministic — input messages → same cleaned output
  * every time — so iteration 1 strips, Anthropic caches the cleaned prefix,
- * iteration 2 strips identically, and the cache hits. The prior fence-skip
- * caused per-execution divergence: iter 1 stripped (fence=-1) and built a
- * thinking-free cached prefix, iter 2 preserved fence-protected messages
- * (fence>0) and re-introduced thinking blocks at positions Anthropic had
- * cached without them, which the prompt-cache validator rejected with
+ * iteration 2 strips identically, and the cache hits. Gating the strip on the
+ * fence causes per-execution divergence: iter 1 strips (fence=-1) and builds a
+ * thinking-free cached prefix, iter 2 preserves fence-protected messages
+ * (fence>0) and re-introduces thinking blocks at positions Anthropic has
+ * cached without them, which the prompt-cache validator rejects with
  * `400 ... blocks cannot be modified`. The cacheFenceIndex on the budget
  * is read for diagnostic stats only and never gates the strip decision.
  *
@@ -38,9 +38,9 @@ import type { ContextLayer } from "./types.js";
 export declare function createThinkingBlockCleaner(keepTurns: number, onCleaned?: (stats: {
     blocksRemoved: number;
     /** Cache fence index when present on the budget; reported for diagnostics
-     *  only. Stripping is no longer gated on the fence. */
+     *  only. Stripping is never gated on the fence. */
     cacheFenceIndex?: number;
-    /** Number of messages protected by the cache fence. Always undefined now
+    /** Number of messages protected by the cache fence. Always undefined
      *  because the fence does not protect any messages from stripping. */
     messagesProtected?: number;
     /** Total messages in the conversation. */

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

@@ -9,11 +9,11 @@
  * cacheFenceIndex is intentionally NOT consulted to gate stripping. The
  * cleaner is pure/deterministic — input messages → same cleaned output
  * every time — so iteration 1 strips, Anthropic caches the cleaned prefix,
- * iteration 2 strips identically, and the cache hits. The prior fence-skip
- * caused per-execution divergence: iter 1 stripped (fence=-1) and built a
- * thinking-free cached prefix, iter 2 preserved fence-protected messages
- * (fence>0) and re-introduced thinking blocks at positions Anthropic had
- * cached without them, which the prompt-cache validator rejected with
+ * iteration 2 strips identically, and the cache hits. Gating the strip on the
+ * fence causes per-execution divergence: iter 1 strips (fence=-1) and builds a
+ * thinking-free cached prefix, iter 2 preserves fence-protected messages
+ * (fence>0) and re-introduces thinking blocks at positions Anthropic has
+ * cached without them, which the prompt-cache validator rejects with
  * `400 ... blocks cannot be modified`. The cacheFenceIndex on the budget
  * is read for diagnostic stats only and never gates the strip decision.
  *
@@ -114,8 +114,8 @@ export function createThinkingBlockCleaner(keepTurns, onCleaned, getKeepTurnsOve
             if (!anyChanged)
                 return messages;
             // Report cleaning stats via callback. cacheFenceIndex is reported for
-            // diagnostic visibility but is no longer gating stripping. messagesProtected
-            // is intentionally omitted because no messages are fence-protected anymore.
+            // diagnostic visibility but never gates stripping. messagesProtected
+            // is intentionally omitted because no messages are fence-protected.
             onCleaned?.({
                 blocksRemoved,
                 ...(budget.cacheFenceIndex >= 0 && blocksRemoved > 0 && {

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

@@ -11,10 +11,10 @@
  * - M = safety margin (percentage-based with absolute floor)
  * - R = context rot buffer (percentage-based)
  * - P = fresh-tail preamble estimate (the WHOLE `dynamicPreamble` + `inlineMemory`
- *   block envelope-wrapper prepends into the latest user message; I1) — a SEPARATE
+ *   block envelope-wrapper prepends into the latest user message) — a SEPARATE
  *   subtrahend, NEVER folded into S
  *
- * WR-01 (133): `P` is the WHOLE injected fresh-tail preamble, not just the recalled
+ * `P` is the WHOLE injected fresh-tail preamble, not just the recalled
  * memory. `dynamicPreamble` accumulates the date/time lines, inbound metadata,
  * channel context, verbosity guidance, the recalled memory block, active-skill
  * content, sender-trust lines, the role block, MCP server instructions, deferred-
@@ -28,12 +28,12 @@
  * that reserves window headroom for it; measuring only the recalled bytes would
  * UNDER-reserve H and risk a fresh-tail overflow on a heavy-skills / many-MCP
  * agent. The over-reservation is conservative (it over-reserves H, never
- * under-reserves) and preserves I1's intent: recalled memory is a strict SUBSET of
+ * under-reserves) and preserves the design intent: recalled memory is a strict SUBSET of
  * the preamble, so a heavier recall block still grows `P` and compacts older
  * history harder.
  *
  * This is a pure function with zero side effects. All constants come from
- * the centralized constants module (not user config, per locked decision).
+ * the centralized constants module (deliberately not operator-configurable).
  *
  * @module
  */
@@ -41,7 +41,7 @@ import type { TokenBudget } from "./types.js";
 /**
  * Compute available history token budget.
  *
- * Per user decision: negative budget clamps to zero. The caller should
+ * Negative budget clamps to zero by design. The caller should
  * log a WARN when `availableHistoryTokens === 0` (degraded fresh-turn behavior).
  *
  * @param contextWindow - Model context window size in tokens (W)
@@ -49,8 +49,8 @@ import type { TokenBudget } from "./types.js";
  * @param cacheFenceIndex - Message index at or below which content must not be modified (-1 = no fence)
  * @param freshTailPreambleTokensEstimate - Estimated tokens for the WHOLE fresh-tail
  *   preamble block (the `dynamicPreamble` + `inlineMemory` prepended into the latest
- *   user message by envelope-wrapper; I1). NOT just recalled memory — see the module
- *   doc (WR-01): it covers skills XML, MCP instructions, deferred-tools context,
+ *   user message by envelope-wrapper). NOT just recalled memory — see the module
+ *   doc: it covers skills XML, MCP instructions, deferred-tools context,
  *   date/channel lines, recalled memory, etc., because that whole blob rides the
  *   UNCONDITIONALLY-shipped fresh tail and is reserved nowhere else. A SEPARATE
  *   subtrahend of H — never folded into S — so a heavier preamble (and recalled

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

@@ -12,10 +12,10 @@
  * - M = safety margin (percentage-based with absolute floor)
  * - R = context rot buffer (percentage-based)
  * - P = fresh-tail preamble estimate (the WHOLE `dynamicPreamble` + `inlineMemory`
- *   block envelope-wrapper prepends into the latest user message; I1) — a SEPARATE
+ *   block envelope-wrapper prepends into the latest user message) — a SEPARATE
  *   subtrahend, NEVER folded into S
  *
- * WR-01 (133): `P` is the WHOLE injected fresh-tail preamble, not just the recalled
+ * `P` is the WHOLE injected fresh-tail preamble, not just the recalled
  * memory. `dynamicPreamble` accumulates the date/time lines, inbound metadata,
  * channel context, verbosity guidance, the recalled memory block, active-skill
  * content, sender-trust lines, the role block, MCP server instructions, deferred-
@@ -29,12 +29,12 @@
  * that reserves window headroom for it; measuring only the recalled bytes would
  * UNDER-reserve H and risk a fresh-tail overflow on a heavy-skills / many-MCP
  * agent. The over-reservation is conservative (it over-reserves H, never
- * under-reserves) and preserves I1's intent: recalled memory is a strict SUBSET of
+ * under-reserves) and preserves the design intent: recalled memory is a strict SUBSET of
  * the preamble, so a heavier recall block still grows `P` and compacts older
  * history harder.
  *
  * This is a pure function with zero side effects. All constants come from
- * the centralized constants module (not user config, per locked decision).
+ * the centralized constants module (deliberately not operator-configurable).
  *
  * @module
  */
@@ -42,7 +42,7 @@ import { SAFETY_MARGIN_PERCENT, MIN_SAFETY_MARGIN_TOKENS, OUTPUT_RESERVE_TOKENS,
 /**
  * Compute available history token budget.
  *
- * Per user decision: negative budget clamps to zero. The caller should
+ * Negative budget clamps to zero by design. The caller should
  * log a WARN when `availableHistoryTokens === 0` (degraded fresh-turn behavior).
  *
  * @param contextWindow - Model context window size in tokens (W)
@@ -50,8 +50,8 @@ import { SAFETY_MARGIN_PERCENT, MIN_SAFETY_MARGIN_TOKENS, OUTPUT_RESERVE_TOKENS,
  * @param cacheFenceIndex - Message index at or below which content must not be modified (-1 = no fence)
  * @param freshTailPreambleTokensEstimate - Estimated tokens for the WHOLE fresh-tail
  *   preamble block (the `dynamicPreamble` + `inlineMemory` prepended into the latest
- *   user message by envelope-wrapper; I1). NOT just recalled memory — see the module
- *   doc (WR-01): it covers skills XML, MCP instructions, deferred-tools context,
+ *   user message by envelope-wrapper). NOT just recalled memory — see the module
+ *   doc: it covers skills XML, MCP instructions, deferred-tools context,
  *   date/channel lines, recalled memory, etc., because that whole blob rides the
  *   UNCONDITIONALLY-shipped fresh tail and is reserved nowhere else. A SEPARATE
  *   subtrahend of H — never folded into S — so a heavier preamble (and recalled
@@ -69,13 +69,13 @@ export function computeTokenBudget(contextWindow, systemTokensEstimate, cacheFen
     const M = Math.max(Math.ceil(W * SAFETY_MARGIN_PERCENT / 100), MIN_SAFETY_MARGIN_TOKENS);
     // R: context rot buffer -- percentage of window
     const R = Math.ceil(W * CONTEXT_ROT_BUFFER_PERCENT / 100);
-    // P: fresh-tail preamble estimate (I1 / WR-01) -- the WHOLE dynamicPreamble +
+    // P: fresh-tail preamble estimate -- the WHOLE dynamicPreamble +
     // inlineMemory blob that rides the unconditionally-shipped fresh tail, a SEPARATE
     // subtrahend clamped to >= 0 so a negative estimate never widens H. NOT folded
     // into S (preserves the recall-dag-budget-partition invariant) and deliberately
     // the whole preamble, not just recall (the only window reservation for it).
     const P = Math.max(0, freshTailPreambleTokensEstimate);
-    // B-1 (260605-oa7) — TOKEN-AUTHORITY DECISION (deliberate, do NOT "align"):
+    // TOKEN-AUTHORITY DECISION (deliberate, do NOT "align"):
     // The S and P this function subtracts are produced at the executor call sites
     // (executor-tool-assembly.ts:518-519 / :530-531) by dividing CHARS by
     // CHARS_PER_TOKEN_RATIO = 3.5, whereas the LCD history items the resulting H
@@ -84,9 +84,9 @@ export function computeTokenBudget(contextWindow, systemTokensEstimate, cacheFen
     // CONSERVATIVE and intentional: for the SAME text, ÷3.5 yields MORE tokens than
     // ÷4, so S and P OVER-reserve relative to how that text would be counted as
     // history → H below comes out SMALLER (the ceiling is TIGHTER), and the assembled
-    // prompt stays UNDER the real free window. The real-LLM review's "H larger than
-    // the true free budget → overflow" reads the direction BACKWARDS (the same
-    // reversal the B-4 fix corrected, 260605-ney). Moving S/P from 3.5 → 4 would make
+    // prompt stays UNDER the real free window. The intuition "H larger than
+    // the true free budget → overflow" reads the direction BACKWARDS.
+    // Moving S/P from 3.5 → 4 would make
     // them SMALLER and WIDEN H — the UNSAFE direction, and exactly what the
     // recall-dag-budget-partition invariant guards against. So the 3.5 ratio at the
     // call sites is kept on purpose. Executable proof + regression guard:
@@ -98,7 +98,7 @@ export function computeTokenBudget(contextWindow, systemTokensEstimate, cacheFen
         windowTokens: W,
         // Profile-unaware base: no capability-class cap is applied here, so the
         // window IS the raw window. computeTokenBudgetForProfile overrides these
-        // two when its class cap actually clamps W (W1 cap provenance).
+        // two when its class cap actually clamps W (cap provenance).
         rawContextWindowTokens: W,
         windowCapSource: "none",
         systemTokens: S,

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

@@ -1,5 +1,5 @@
 /**
- * Transcript repair (A2) — the `tool_use`<->`tool_result` pairing invariant.
+ * Transcript repair — the `tool_use`<->`tool_result` pairing invariant.
  *
  * A self-contained, pure transform over the canonical pi-ai `Message[]`
  * (`AgentMessage[]` at the boundary). It runs as the FINAL step of LCD
@@ -10,8 +10,8 @@
  * Providers HARD-REJECT a transcript where an assistant `tool_use` block is
  * not immediately followed by its matching `tool_result`, where a `tool_result`
  * has no preceding call, or where a turn is left with a dangling unpaired call.
- * The deleted DAG assembler had no such pass — emitting a flattened, unpaired
- * array is exactly what produced the prior tool-call loop. This module
+ * An assembler without such a pass can emit a flattened, unpaired array —
+ * exactly what produces a provider-rejection tool-call loop. This module
  * guarantees a provider-valid pairing on ANY input:
  *
  *  - REORDER  — an out-of-order `tool_result` is re-placed immediately after
@@ -21,7 +21,7 @@
  *  - DROP ORPHAN — a `tool_result` whose `toolCallId` matches no call is dropped.
  *  - DROP DUPLICATE — a second `tool_result` for an already-resolved id is dropped.
  *  - STRIP ABORTED — an assistant turn with `stopReason "error" | "aborted"` has
- *    its dangling `tool_use` blocks removed (text/thinking preserved, A5), so no
+ *    its dangling `tool_use` blocks removed (text/thinking preserved), so no
  *    unpaired call survives and no result is synthesized for the aborted call.
  *
  * Reasoning (`ThinkingContent`) is never re-introduced, reordered, or dropped:
@@ -39,10 +39,10 @@
  */
 import type { AgentMessage } from "@earendil-works/pi-agent-core";
 /**
- * Transcript repair (A2): guarantee a provider-valid `tool_use`<->`tool_result`
+ * Transcript repair: guarantee a provider-valid `tool_use`<->`tool_result`
  * pairing on the assembled array. The FINAL assembly step (`lcd-assembler.ts`).
  *
- * Three passes (RESEARCH §"Transcript Repair"):
+ * Three passes:
  *  - **Pass A** — index every assistant `toolCall.id` into `toolUseSeen`.
  *  - **Pass B** — collect `tool_result` messages by `toolCallId`; DROP an orphan
  *    (id not in `toolUseSeen`) and DROP a duplicate (id already collected).

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

@@ -1,6 +1,6 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * Transcript repair (A2) — the `tool_use`<->`tool_result` pairing invariant.
+ * Transcript repair — the `tool_use`<->`tool_result` pairing invariant.
  *
  * A self-contained, pure transform over the canonical pi-ai `Message[]`
  * (`AgentMessage[]` at the boundary). It runs as the FINAL step of LCD
@@ -11,8 +11,8 @@
  * Providers HARD-REJECT a transcript where an assistant `tool_use` block is
  * not immediately followed by its matching `tool_result`, where a `tool_result`
  * has no preceding call, or where a turn is left with a dangling unpaired call.
- * The deleted DAG assembler had no such pass — emitting a flattened, unpaired
- * array is exactly what produced the prior tool-call loop. This module
+ * An assembler without such a pass can emit a flattened, unpaired array —
+ * exactly what produces a provider-rejection tool-call loop. This module
  * guarantees a provider-valid pairing on ANY input:
  *
  *  - REORDER  — an out-of-order `tool_result` is re-placed immediately after
@@ -22,7 +22,7 @@
  *  - DROP ORPHAN — a `tool_result` whose `toolCallId` matches no call is dropped.
  *  - DROP DUPLICATE — a second `tool_result` for an already-resolved id is dropped.
  *  - STRIP ABORTED — an assistant turn with `stopReason "error" | "aborted"` has
- *    its dangling `tool_use` blocks removed (text/thinking preserved, A5), so no
+ *    its dangling `tool_use` blocks removed (text/thinking preserved), so no
  *    unpaired call survives and no result is synthesized for the aborted call.
  *
  * Reasoning (`ThinkingContent`) is never re-introduced, reordered, or dropped:
@@ -38,7 +38,7 @@
  *
  * @module
  */
-/** The literal marker carried by every synthesized placeholder result (T-128-01). */
+/** The literal marker carried by every synthesized placeholder result. */
 const SYNTHESIZED_RESULT_MARKER = "[tool result missing — synthesized placeholder]";
 const ABORTED_STOP_REASONS = new Set(["error", "aborted"]);
 /**
@@ -94,10 +94,10 @@ function isAbortedAssistant(m) {
  * matching `tool_result`. Providers reject an unpaired call, so the invariant
  * fills the gap rather than dropping the call.
  *
- * SECURITY (T-128-01): the placeholder is marked `isError: true` AND its only
+ * SECURITY: the placeholder is marked `isError: true` AND its only
  * content is the explicit literal {@link SYNTHESIZED_RESULT_MARKER}, so the
  * model can never read it as a genuine tool output. Both markers are asserted
- * by Test 2.
+ * by the synthesize-missing test.
  */
 function makeMissingToolResult(toolCallId, toolName, now) {
     return {
@@ -111,8 +111,8 @@ function makeMissingToolResult(toolCallId, toolName, now) {
 }
 /**
  * Return a NEW assistant message with all `toolCall` blocks removed from its
- * content, preserving text/thinking and every other field (A5:
- * strip-dangling-blocks-keep-text). Used for aborted/errored turns whose calls
+ * content, preserving text/thinking and every other field
+ * (strip-dangling-blocks-keep-text). Used for aborted/errored turns whose calls
  * never completed — leaving the calls would put an unpaired `tool_use` in front
  * of the provider. Never mutates the input.
  */
@@ -124,10 +124,10 @@ function stripDanglingToolUseBlocks(m, a) {
 // Public API
 // ---------------------------------------------------------------------------
 /**
- * Transcript repair (A2): guarantee a provider-valid `tool_use`<->`tool_result`
+ * Transcript repair: guarantee a provider-valid `tool_use`<->`tool_result`
  * pairing on the assembled array. The FINAL assembly step (`lcd-assembler.ts`).
  *
- * Three passes (RESEARCH §"Transcript Repair"):
+ * Three passes:
  *  - **Pass A** — index every assistant `toolCall.id` into `toolUseSeen`.
  *  - **Pass B** — collect `tool_result` messages by `toolCallId`; DROP an orphan
  *    (id not in `toolUseSeen`) and DROP a duplicate (id already collected).
@@ -148,8 +148,8 @@ export function sanitizeToolUseResultPairing(messages, now) {
     // A non-array input degrades to a PASS-THROUGH, never an empty array: the SDK
     // transformContext contract is "return the original messages or another safe
     // fallback" (context-engine.ts's "never no-op the whole context" stance), so
-    // nuking a malformed shape to [] would silently drop the entire conversation
-    // (WR-02). An EMPTY well-formed array is a genuine no-op and returns [].
+    // nuking a malformed shape to [] would silently drop the entire conversation.
+    // An EMPTY well-formed array is a genuine no-op and returns [].
     if (!Array.isArray(messages))
         return messages;
     if (messages.length === 0)

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

@@ -25,21 +25,21 @@ export interface CompactionLayerDeps {
     getApiKey: () => Promise<string>;
     /** Optional: resolved override model + apiKey for cheaper compaction.
      *  When provided, compaction uses this model instead of the session model.
-     *  INT-W1: `servedWindow` is the provider-served window binding THIS override
+     *  `servedWindow` is the provider-served window binding THIS override
      *  model, provider-gated at the wiring site (executor-context-engine-setup)
-     *  against the Phase-176 `servedContextWindow` pair — see the identical
+     *  against the probed `servedContextWindow` pair — see the identical
      *  fields on `LeafSummarizerDeps` (lcd-leaf-summarizer.ts). */
     overrideModel?: {
         model: unknown;
         getApiKey: () => Promise<string>;
         servedWindow?: number;
     };
-    /** INT-W1: the provider-served window binding the PRIMARY (session) model —
+    /** The provider-served window binding the PRIMARY (session) model —
      *  `windowProvenance.served`, already provider-gated by the executor
-     *  reconcile (WR-02). Consumed by the Step-4 model read in llm-compaction:
+     *  reconcile. Consumed by the Step-4 model read in llm-compaction:
      *  the served value is selected in the SAME branch that selects the
      *  summarizer model, so clamp and call always agree. Absent ⇒ configured
-     *  window governs (byte-identical pre-INT-W1 behavior). */
+     *  window governs. */
     primaryServedWindow?: number;
     /** Optional callback for reporting compaction stats. */
     onCompacted?: (stats: {
@@ -54,7 +54,7 @@ export interface CompactionLayerDeps {
     /** Optional getter for the API-grounded token anchor.
      *  When provided, compaction threshold check uses anchor + delta instead of pure char estimation. */
     getTokenAnchor?: () => TokenAnchor | null;
-    /** Optional event bus for emitting context:compaction_routed (C4/S4). */
+    /** Optional event bus for emitting context:compaction_routed. */
     eventBus?: TypedEventBus;
     /** Optional agent id for context:compaction_routed event payload. */
     agentId?: string;