@humwick/runner 0.130.4 → 0.130.6

package/src/adapters/claude-token-usage-capture.ts

@@ -1,9 +1,9 @@
 import { readdir } from "node:fs/promises";
 import { basename, dirname, join } from "node:path";
-// #2482 — per-turn Claude token-usage capture: the runner-side half of the token-usage
+// per-turn Claude token-usage capture: the runner-side half of the token-usage
 // spine's first capture seam. On every Stop-hook session turn (runner-core's
 // publishSessionTurn) the transcript's APPENDED tail is read and every NEW root-session
-// assistant entry's API-reported usage is extracted (the #2482 sibling parser beside
+// assistant entry's API-reported usage is extracted (the sibling parser beside
 // extractLatestContextUsage in adapters/claude-transcript.ts) and one spine record is
 // POSTed per entry, keyed dedupeKey=entry-uuid so the spine's dedupe_key upsert makes
 // replays and retries idempotent by construction.
@@ -33,7 +33,7 @@ import { claudeSessionIdFromTranscriptPath, TranscriptTokenUsageExtractor, type
 import { incrementalFileNeedsReset, readIncrementalFileRange, splitIncrementalFileLines, statIncrementalFile, type IncrementalFileInfo, type IncrementalFileLineBuffer, type IncrementalFileLinePolicy, type IncrementalFilePosition } from "./incremental-file";
 import type { ProviderTokenUsageRecord } from "../adapter";
 
-/** The spine's turn-less report shape (POST /api/agents/:id/token-usage, #2485). */
+/** The spine's turn-less report shape. */
 export type TranscriptTokenUsageSpineRecord = ProviderTokenUsageRecord & { provider: "claude" };
 
 // Claude assistant transcript entries carry both the full model output and usage. Eight
@@ -142,13 +142,13 @@ export class ClaudeTranscriptTokenUsageCapture {
 
   private async settleTranscript(transcriptPath: string, options: { pollIntervalMs: number; timeoutMs: number; now?: () => number; sleep?: (ms: number) => Promise<unknown> }): Promise<number> {
     // now/sleep are injectable so tests can drive the settle window on a virtual
-    // clock instead of racing real timers against host load (#2714 flake).
+    // clock instead of racing real timers against host load.
     const now = options.now ?? Date.now;
     const sleep = options.sleep ?? Bun.sleep;
     const deadline = now() + options.timeoutMs;
     const pollIntervalMs = Math.max(1, options.pollIntervalMs);
     const quietWindowMs = Math.min(options.timeoutMs, Math.max(pollIntervalMs * 2, 25));
-    // #2491 review round 3 (reverts round 2's watermark bail, which was WRONG — it exited
+    // review round 3 (reverts round 2's watermark bail, which was WRONG — it exited
     // exactly when the watermark hadn't moved yet, which is precisely the state a still-
     // flushing final entry looks like before it lands). Seed the quiet-window baseline at
     // settle START rather than leaving it undefined until something is first found: without

package/src/adapters/claude-tool-policy.ts

@@ -5,13 +5,13 @@ import type { AgentProfileProviderOptions } from "@humwick/sdk";
 import { isRecord } from "@humwick/sdk/types/guards";
 import { claudeDisallowedWebTools } from "../web-tools";
 
-// ── #2453 — eager "System tools" context trim for the coding/review profiles ─────────────────
+// ── — eager "System tools" context trim for the coding/review profiles ─────────────────
 // A BARE tool name on Claude Code's --disallowedTools removes the tool's schema from the
 // model's eager context (the /context "System tools" bucket) WITHOUT touching MCP (mcp__*),
 // skills, or core coding tools — which an allowlist (--tools) would put at risk. The Artifact
 // family + Workflow carry the heaviest eager docstrings and coding/review agents never publish
 // artifacts or run workflows; SendFeedback is misused by autonomous agents (also env-suppressed
-// fleet-wide since T3182 — this is the per-profile context-trim path); ScheduleWakeup is
+// fleet-wide since — this is the per-profile context-trim path); ScheduleWakeup is
 // loop-only; ReportFindings duplicates relay report-up.
 export const CLAUDE_CONTEXT_TRIM_DISALLOWED_TOOLS = ["Artifact", "ArtifactComments", "ArtifactData", "ArtifactCheck", "Workflow", "SendFeedback", "ScheduleWakeup", "ReportFindings"];
 
@@ -23,7 +23,7 @@ const CLAUDE_CONTEXT_TRIM_DEFAULT_PROFILES = new Set(["worker", "debugger", "rev
 
 const CLAUDE_DISALLOWED_TOOLS_ARG = /^--(?:disallowedTools|disallowed-tools)(?:=(.*))?$/;
 
-/** Effective #2453 denylist for this launch: a profile-configured array REPLACES the default,
+/** Effective denylist for this launch: a profile-configured array REPLACES the default,
  * `null` opts the profile out entirely, absent falls back to the default — and only for the
  * coding/review profiles. Resolves the profile via the established
  * `agentProfile?.name || profile` pattern. */
@@ -41,12 +41,12 @@ export function claudeProfileDisallowedTools(config: Pick<RunnerSpawnConfig, "pr
   }
   const profileName = config.agentProfile?.name || config.profile || "";
   if (!CLAUDE_CONTEXT_TRIM_DEFAULT_PROFILES.has(profileName)) return [];
-  // A profile that switches the Artifact or Workflow tool back on (#3171) gets it.
+  // A profile that switches the Artifact or Workflow tool back on gets it.
   const harness = config.agentProfile ? claudeHarnessOptions(config) : CLAUDE_HARNESS_DEFAULTS;
   return CLAUDE_CONTEXT_TRIM_DISALLOWED_TOOLS.filter((tool) => tool === "Workflow" ? !harness.workflows : !tool.startsWith("Artifact") || !harness.artifacts);
 }
 
-// ── #3171 — harness context trim: Claude Code built-ins Relay agents never use ──────────────────
+// ── — harness context trim: Claude Code built-ins Relay agents never use ──────────────────
 export type ClaudeAgentTypes = NonNullable<NonNullable<AgentProfileProviderOptions["claude"]>["agentTypes"]>;
 const CLAUDE_AGENT_TYPES: readonly ClaudeAgentTypes[] = ["all", "general-purpose", "none"];
 export interface ClaudeHarnessOptions {
@@ -146,7 +146,7 @@ function splitDisallowedToolsList(value: string): string[] {
   return value.split(",").map((name) => name.trim()).filter(Boolean);
 }
 
-/** Appends the profile's #2453 denylist to the launch args, merging with any inherited
+/** Appends the profile's denylist to the launch args, merging with any inherited
  * --disallowedTools entries into ONE canonical trailing occurrence (union, never clobber).
  * No-op (args returned as-is) when the profile resolves no denylist. The single launch seam
  * for this trim is ClaudeAdapter.buildSpawnArgs.
@@ -156,7 +156,7 @@ function splitDisallowedToolsList(value: string): string[] {
  * launch would eat the prompt as a deny value (silent no-prompt launch). The `=` form
  * terminates the flag so following positionals survive. */
 export function claudeProfileDisallowedToolsArgs(args: string[], config: Pick<RunnerSpawnConfig, "harnessSubagents" | "profile" | "agentProfile">): string[] {
-  // #3024 — a profile's switched-off web tools are hidden whatever its denylist says.
+  // a profile's switched-off web tools are hidden whatever its denylist says.
   const deny = [...(harnessSubagentsEnabled(config) ? [] : ["Agent", "Task", "TeamCreate", "TeamDelete", "SendMessage"]), ...claudeProfileDisallowedTools(config), ...claudeHarnessTrim(config).deny, ...claudeDisallowedWebTools(config)];
   if (!deny.length) return args;
   const { kept, inherited } = partitionClaudeDisallowedToolsArgs(args);

package/src/adapters/claude-transcript.ts

@@ -32,11 +32,11 @@ export interface TurnStep {
   label?: string;
   category?: ToolCallCategory;
   occurredAt?: number;
-  /** #2437 — redacted, size-capped tool input, only set for type "tool". */
+  /** redacted, size-capped tool input, only set for type "tool". */
   params?: unknown;
   /** The native tool_use id, only set for type "tool"; pairs the call with its tool_result. */
   toolUseId?: string;
-  /** #3411 — absolute file paths the tool reads or edits, only set for type "tool". */
+  /** absolute file paths the tool reads or edits, only set for type "tool". */
   paths?: string[];
 }
 
@@ -47,7 +47,7 @@ export interface TranscriptMessage {
   // Assistant entries carry the model id and the API-reported token usage for
   // the turn. `usage.{input,cache_creation_input,cache_read_input}_tokens` sum to
   // the same context-size figure the status-line probe reports — read here so the
-  // runner can keep context stats advancing every turn without the probe (#1512).
+  // runner can keep context stats advancing every turn without the probe.
   model?: string;
   usage?: TranscriptUsage;
 }
@@ -59,7 +59,7 @@ export interface TranscriptUsage {
   output_tokens?: number;
 }
 
-/** Fresh, per-turn context usage read straight from the transcript (#1512). */
+/** Fresh, per-turn context usage read straight from the transcript. */
 export interface TranscriptContextUsage {
   /** input + cache-creation + cache-read tokens — the context size sent this turn. */
   tokensUsed: number;
@@ -80,7 +80,7 @@ interface TranscriptEntry {
   // sidechain entries to keep a subagent's reasoning/tools/responses from
   // leaking into the parent agent's chat. Insights parsers (collectClaudeSession-
   // Events/countSubstantiveTurns) intentionally do NOT filter — changing them
-  // would shift the #184/#185 baselines, a separate concern.
+  // would shift the baselines, a separate concern.
   isSidechain?: boolean;
   // Every entry carries its own `uuid` and its `parentUuid` (the entry it was
   // appended after), forming a causal chain for the whole session. A `user`-type
@@ -113,7 +113,7 @@ function isRealUserPrompt(entry: TranscriptEntry): boolean {
   return blocks(entry.message).some((b) => b.type === "text" && Boolean(b.text?.trim()));
 }
 
-// #2840 — `closingOnly` keeps just the text after the turn's last tool call (the closing message,
+// `closingOnly` keeps just the text after the turn's last tool call (the closing message,
 // without narration the reasoning tail already mirrored); a turn ending on a tool keeps its last text.
 class AssistantTurnText {
   private texts: string[] = [];
@@ -282,7 +282,7 @@ interface QueueOperationEntry extends TranscriptEntry {
 }
 
 /**
- * #2919 — prompts left in Claude's input queue at the end of `jsonl`: a mid-turn prompt is `enqueue`d,
+ * prompts left in Claude's input queue at the end of `jsonl`: a mid-turn prompt is `enqueue`d,
  * then `remove`d when a tool boundary absorbs it, or `dequeue`d (every queued prompt at once) to run as
  * the next turn after the Stop. So a dequeue or a new promptId empties the queue.
  */
@@ -323,13 +323,13 @@ function canonicalJson(value: unknown): string {
 
 /**
  * True when the transcript already contains the exact assistant tool_use block that
- * triggered a PreToolUse hook — the settle-poll anchor (#1116). The hook payload's own
+ * triggered a PreToolUse hook — the settle-poll anchor. The hook payload's own
  * `tool_name`/`tool_input` are the precise call that fired it, so matching them against
  * the transcript's tool_use blocks (name + structural input equality) is exact, unlike
  * guessing from transcript shape alone. Claude's JSONL flush is async/bursty and can
  * lag well behind the hook firing, so the entry backing the hook's own call may still be
  * missing from disk the instant the hook runs — this is what the settle-poll waits for.
- * #1086 should adopt this same anchor for its own turn-final settle marker rather than
+ * should adopt this same anchor for its own turn-final settle marker rather than
  * inventing a second one.
  */
 export function transcriptHasToolUseAnchor(jsonl: string, toolName: string, toolInput: unknown): boolean {
@@ -357,7 +357,7 @@ export function transcriptHasToolUseAnchor(jsonl: string, toolName: string, tool
  * transcriptHasToolUseAnchor above). Used to stamp a resolved permission prompt with the
  * tool call's real event time — a hook's own wall-clock receipt time can end up earlier than
  * the entry the provider eventually flushes for that same call, which otherwise sorts the
- * resolved prompt ahead of the narration that led up to it (#1185).
+ * resolved prompt ahead of the narration that led up to it.
  */
 export function transcriptToolUseOccurredAt(jsonl: string, toolName: string, toolInput: unknown): number | undefined {
   const wantInput = canonicalJson(toolInput ?? {});
@@ -461,7 +461,7 @@ export class LatestTurnStepAccumulator {
   }
 }
 
-/** Narration, reasoning and tool steps of one assistant message, shared by the transcript tail and the headless stream (#2771). */
+/** Narration, reasoning and tool steps of one assistant message, shared by the transcript tail and the headless stream. */
 export function claudeAssistantSteps(message: TranscriptMessage | undefined, occurredAt?: number): TurnStep[] {
   const steps: TurnStep[] = [];
   for (const b of blocks(message)) {
@@ -477,7 +477,7 @@ export function claudeAssistantSteps(message: TranscriptMessage | undefined, occ
   return steps;
 }
 
-// #3411 — Claude's file tools and the input field naming their file. Keyed on the tool name, not the
+// Claude's file tools and the input field naming their file. Keyed on the tool name, not the
 // category: MultiEdit is "generic". Claude's file tools take absolute paths only.
 const CLAUDE_TOOL_PATH_FIELDS: Record<string, string> = {
   Read: "file_path",
@@ -493,7 +493,7 @@ export function claudeToolPaths(name: string, input: Record<string, unknown> | u
   return field && input ? absoluteToolPaths([input[field]]) : undefined;
 }
 
-/** The raw PostToolUse payload uses the same path contract as transcript tools (#3646). */
+/** The raw PostToolUse payload uses the same path contract as transcript tools. */
 export function claudeToolHookPaths(input: unknown): string[] | undefined {
   return isRecord(input) && typeof input.tool_name === "string"
     ? claudeToolPaths(input.tool_name, isRecord(input.tool_input) ? input.tool_input : undefined) : undefined;
@@ -529,7 +529,7 @@ function transcriptEntryOccurredAt(entry: TranscriptEntry): number | undefined {
  * Stable dedup keys for a turn's steps, in order. Each key is salted with how many
  * identical (type,label,text) steps preceded it in the same window — so running the
  * same tool twice with identical input within a turn yields two distinct keys and
- * both show in the activity trace (#265). Keying on occurrence-within-window rather
+ * both show in the activity trace. Keying on occurrence-within-window rather
  * than raw transcript index keeps the reasoning tailer idempotent when the "latest
  * turn" window shrinks/resets mid-poll: a surviving step recomputes to the same or a
  * lower occurrence, so an already-emitted step never re-fires.
@@ -544,7 +544,7 @@ export function stepDedupKeys(steps: TurnStep[]): string[] {
   });
 }
 
-// #1498: Claude's native tool names → the provider-independent category taxonomy. Tools with
+// Claude's native tool names → the provider-independent category taxonomy. Tools with
 // no dedicated category (Grep, Glob, LS, WebFetch, WebSearch, ...) fall to "generic" rather than
 // being force-fit into a category that doesn't really describe them — the fallback exists
 // precisely so nothing is hidden from the dashboard's activity trace.
@@ -560,7 +560,7 @@ const CLAUDE_TOOL_CATEGORIES: Record<string, ToolCallCategory> = {
   ToolSearch: "mcp-tool-search",
 };
 
-/** Map a Claude tool_use name onto the provider-independent tool-call category (#1498). */
+/** Map a Claude tool_use name onto the provider-independent tool-call category. */
 export function claudeToolCategory(name: string): ToolCallCategory {
   if (name.startsWith("mcp__")) return "mcp-tool-call";
   return CLAUDE_TOOL_CATEGORIES[name] ?? "generic";
@@ -575,7 +575,7 @@ export function summarizeToolUse(name: string, input: Record<string, unknown> |
   return clipToolSummary(summary);
 }
 
-// --- Insights #184: context-gathering ratio (epic #183, docs/self-improvement.md) ---
+// --- Insights: context-gathering ratio ---
 //
 // Computed mechanically from the whole-session transcript at session end — no model
 // involvement, so it costs zero agent tokens and the agent can't game it. The ratio is
@@ -584,7 +584,7 @@ export function summarizeToolUse(name: string, input: Record<string, unknown> |
 
 /**
  * Normalize a full Claude transcript into the provider-agnostic `SessionEvent` stream
- * (#183/#184). The classifier and ratio math live in `session-insights.ts` and are shared
+ * The classifier and ratio math live in `session-insights.ts` and are shared
  * across providers; this only knows the Claude JSONL shape. Events are emitted in
  * transcript order so `leadingGather` is meaningful.
  */
@@ -628,7 +628,7 @@ export function analyzeSession(jsonl: string): SessionAnalysis | null {
   return computeContextRatio(collectClaudeSessionEvents(jsonl));
 }
 
-/** Count substantive assistant turns — used by the #185 introspection gate. */
+/** Count substantive assistant turns — used by the introspection gate. */
 export function countSubstantiveTurns(jsonl: string): number {
   let turns = 0;
   for (const line of jsonl.split("\n")) {
@@ -650,7 +650,7 @@ export function countSubstantiveTurns(jsonl: string): number {
 }
 
 // Walk the transcript from the end and return the most recent root-session
-// assistant turn's token usage (#1512). This is the SAME input+cache token sum
+// assistant turn's token usage. This is the SAME input+cache token sum
 // the Claude status-line probe reports, but read directly from the transcript —
 // which the Stop hook hands the runner every turn — so context stats keep
 // advancing across compact/clear/native-resume even when the status-line probe
@@ -697,7 +697,7 @@ export function sumUsageTokens(usage: TranscriptUsage): number | undefined {
   return seen ? total : undefined;
 }
 
-// #2482 — the token-usage spine's capture parser. Where extractLatestContextUsage
+// the token-usage spine's capture parser. Where extractLatestContextUsage
 // collapses a turn's usage into one context-size sum, this sibling keeps the RAW
 // per-entry breakdown the spine records, one record per assistant transcript entry:
 //   dedupeKey = the entry `uuid` (entries never move once written, so the uuid is the
@@ -716,7 +716,7 @@ export interface TranscriptTokenUsageEntry {
   cacheReadTokens: number;
   cacheWriteTokens: number;
   occurredAt: number;
-  /** #3351 — an API assistant entry without a usage block: spend happened, counts are unknown. */
+  /** an API assistant entry without a usage block: spend happened, counts are unknown. */
   usageStatus?: "unknown";
 }
 
@@ -758,7 +758,7 @@ function usageEntryFromTranscriptEntry(entry: TranscriptEntry, byUuid: ReadonlyM
 }
 
 /**
- * Incremental #2482 extractor: feeds transcript lines one at a time so a caller can
+ * Incremental extractor: feeds transcript lines one at a time so a caller can
  * consume only newly-appended data without re-walking the file. promptId is resolved
  * against the entries THIS instance has seen (the uuid→entry map) — feed a whole file
  * for whole-transcript resolution, or feed only an appended tail and accept that an
@@ -794,7 +794,7 @@ function usageCount(value: number | undefined): number {
   return typeof value === "number" && Number.isSafeInteger(value) && value >= 0 ? value : 0;
 }
 
-/** Claude names its transcript file `<sessionId>.jsonl` (#1235), so the provider's own
+/** Claude names its transcript file `<sessionId>.jsonl`, so the provider's own
  *  session uuid is recoverable from the path even when the status probe is silent. */
 export function claudeSessionIdFromTranscriptPath(transcriptPath: string): string {
   const file = transcriptPath.split(/[\\/]/).pop() ?? "";