@sema-agent/core 5.40.0 → 5.42.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -4669,6 +4669,16 @@ export interface EngineNotice {
4669
4669
  * value IS in force the prepare refuses with the same code as `TaskResult.errorCode` (one
4670
4670
  * fact, one code, two loudness dialects); `detail: { raw }`.
4671
4671
  *
4672
+ * - `"task.user_steer_undrained"` / `"task.user_followup_undrained"` (#259) — user steers /
4673
+ * follow-ups whose receipts said "queued" were still in their queue at agent_end: the run
4674
+ * ended before any turn could drain them. One notice PER family (a consumer routing on `code`
4675
+ * alone must not mistake a stranded follow-up for a stranded steer — same split as the settled
4676
+ * frame's two count keys). They are NOT redelivered (a steer aimed at a finished run must not
4677
+ * fire at the next one — unlike ENGINE notes, which pend per session); the loud half of the
4678
+ * #257 contract's "accepted = enqueued, not consumed" sentence;
4679
+ * `detail: { steer, taskId? }` / `{ followUp, taskId? }`. Per-run, at most once per family
4680
+ * (the terminal sweep is a single site).
4681
+ *
4672
4682
  * Deliberately NOT a notice family: brain retry/reconnect liveness (a rate limit, a 5xx, a
4673
4683
  * transient network failure being retried). Those are per-attempt liveness frames with their own
4674
4684
  * frequency semantics and ride the wire `status` channel ({@link BrainStatus}), whose sink the
@@ -4702,6 +4712,16 @@ export declare function __resetMalformedNoticeSeatAnnouncement(): void;
4702
4712
  * · an ABSENT seat prints the historic `console.warn` line verbatim (byte-compat loudness).
4703
4713
  */
4704
4714
  export declare function deliverEngineNotice(onNotice: ((notice: EngineNotice) => void) | undefined, notice: EngineNotice): void;
4715
+ /**
4716
+ * #259 loud half, one notice PER FAMILY (`task.user_steer_undrained` / `task.user_followup_undrained`):
4717
+ * a consumer routing on `code` alone must never mistake a stranded follow-up for a stranded steer, so
4718
+ * the code carries exactly the semantics its name claims — the same two-key split the settled frame
4719
+ * uses. Pure (the terminal sweep race window is not constructible deterministically; this seam is).
4720
+ */
4721
+ export declare function undrainedUserInputNotices(counts: {
4722
+ steer: number;
4723
+ followUp: number;
4724
+ }, taskId?: string): EngineNotice[];
4705
4725
  /** Runtime dependencies shared across tasks. */
4706
4726
  export interface RunnerDeps {
4707
4727
  brain: Brain;
@@ -49,3 +49,23 @@ export function deliverEngineNotice(onNotice, notice) {
49
49
  }
50
50
  console.warn(notice.message);
51
51
  }
52
+ export function undrainedUserInputNotices(counts, taskId) {
53
+ const tid = taskId !== undefined ? { taskId } : {};
54
+ const tail = `accepted as "queued" were never consumed — the run ended first. They are NOT redelivered; re-send against a live run if still wanted.`;
55
+ const out = [];
56
+ if (counts.steer > 0) {
57
+ out.push({
58
+ code: "task.user_steer_undrained",
59
+ message: `${counts.steer} user steer(s) ${tail}`,
60
+ detail: { steer: counts.steer, ...tid },
61
+ });
62
+ }
63
+ if (counts.followUp > 0) {
64
+ out.push({
65
+ code: "task.user_followup_undrained",
66
+ message: `${counts.followUp} user follow-up(s) ${tail}`,
67
+ detail: { followUp: counts.followUp, ...tid },
68
+ });
69
+ }
70
+ return out;
71
+ }
@@ -80,6 +80,16 @@ export declare class AgentHarness<TSkill extends Skill = Skill, TPromptTemplate
80
80
  private thinkingLevel;
81
81
  /** RB-30 terminal fix — runner-set sink for engine-note payloads left undrained at agent_end. */
82
82
  onUndrainedEngineNotes?: (payloads: unknown[]) => void;
83
+ /** backlog #259 — USER-authored queue remnants at agent_end: steers/follow-ups whose receipts said
84
+ * "queued" but that no turn will ever drain (the run ended first). The engine-note sweep above
85
+ * hands ENGINE payloads back for redelivery; user inputs have no redelivery semantics (a steer
86
+ * aimed at a finished run must not silently fire at the next one), so their loss is ANNOUNCED
87
+ * instead — the runner surfaces it as an operator notice, closing the "accepted then silently
88
+ * dropped" window the #257 contract could only document. */
89
+ onUndrainedUserInputs?: (counts: {
90
+ steer: number;
91
+ followUp: number;
92
+ }) => void;
83
93
  /** design/176 — runner-set sink fired at the CONSUMPTION boundary, once per engine-note payload,
84
94
  * in consumption order (steer/followUp drain and the turn-open nextTurn splice — the two points
85
95
  * where a queued frame actually enters the model's input). The runner uses it to record the
@@ -99,6 +109,13 @@ export declare class AgentHarness<TSkill extends Skill = Skill, TPromptTemplate
99
109
  * so a double call is a no-op. */
100
110
  recoverUndrainedEngineNotes(): void;
101
111
  private sweepUndrainedEngineNotes;
112
+ /** backlog #259 — what remains in the steer/follow-up queues AFTER the engine-note sweep is USER
113
+ * input that was accepted ("queued") and will never be consumed: the run reached agent_end first.
114
+ * User inputs have no redelivery semantics (unlike engine notes — a steer aimed at a finished run
115
+ * must not fire at the next one), so the loss is ANNOUNCED, never silent. The window is a narrow
116
+ * race (an injection landing after the loop's final queue check), which is exactly why it needs a
117
+ * loud terminal account rather than an e2e reproduction. Returns the counts for the settled frame. */
118
+ private announceUndrainedUserInputs;
102
119
  private systemPrompt;
103
120
  /** S4: physical system blocks (static per leg, additive — see AgentHarnessOptions.systemBlocks). */
104
121
  private systemBlocks;
@@ -158,6 +158,7 @@ export class AgentHarness {
158
158
  model;
159
159
  thinkingLevel;
160
160
  onUndrainedEngineNotes;
161
+ onUndrainedUserInputs;
161
162
  onEngineNoteConsumed;
162
163
  recoverUndrainedEngineNotes() {
163
164
  this.sweepUndrainedEngineNotes([this.nextTurnQueue, this.steerQueue, this.followUpQueue]);
@@ -182,6 +183,17 @@ export class AgentHarness {
182
183
  }
183
184
  }
184
185
  }
186
+ announceUndrainedUserInputs() {
187
+ const counts = { steer: this.steerQueue.length, followUp: this.followUpQueue.length };
188
+ if ((counts.steer > 0 || counts.followUp > 0) && this.onUndrainedUserInputs) {
189
+ try {
190
+ this.onUndrainedUserInputs(counts);
191
+ }
192
+ catch {
193
+ }
194
+ }
195
+ return counts;
196
+ }
185
197
  systemPrompt;
186
198
  systemBlocks;
187
199
  streamOptions;
@@ -620,9 +632,15 @@ export class AgentHarness {
620
632
  if (event.type === "agent_end") {
621
633
  await this.flushPendingSessionWrites();
622
634
  this.sweepUndrainedEngineNotes([this.nextTurnQueue, this.steerQueue, this.followUpQueue]);
635
+ const undrainedUser = this.announceUndrainedUserInputs();
623
636
  this.phase = "idle";
624
637
  await this.emitAny(event, signal);
625
- await this.emitOwn({ type: "settled", nextTurnCount: this.nextTurnQueue.length }, signal);
638
+ await this.emitOwn({
639
+ type: "settled",
640
+ nextTurnCount: this.nextTurnQueue.length,
641
+ ...(undrainedUser.steer > 0 ? { undrainedSteerCount: undrainedUser.steer } : {}),
642
+ ...(undrainedUser.followUp > 0 ? { undrainedFollowUpCount: undrainedUser.followUp } : {}),
643
+ }, signal);
626
644
  return;
627
645
  }
628
646
  await this.emitAny(event, signal);
@@ -815,6 +815,11 @@ export interface AbortEvent {
815
815
  export interface SettledEvent {
816
816
  type: "settled";
817
817
  nextTurnCount: number;
818
+ /** backlog #259 (additive) — USER steers accepted ("queued") but never drained before agent_end.
819
+ * Present only when > 0; their loss is announced through `onUndrainedUserInputs` too. */
820
+ undrainedSteerCount?: number;
821
+ /** backlog #259 (additive) — same for the follow-up queue. */
822
+ undrainedFollowUpCount?: number;
818
823
  }
819
824
  export interface BeforeAgentStartEvent<TSkill extends Skill = Skill, TPromptTemplate extends PromptTemplate = PromptTemplate> {
820
825
  type: "before_agent_start";
package/dist/index.d.ts CHANGED
@@ -93,7 +93,7 @@ export { READ_FACE_DEFAULT_DENY_ENTRIES, READ_FACE_BUILTIN_DENY_TABLE, READ_DENY
93
93
  export { deploymentReadFaceClampNotice, resolveReadFace, type ReadFace, type ReadFaceInputs } from "./tools/fs/index.js";
94
94
  export { WRITE_PROTECTED_DEFAULT_TABLE, resolveWriteProtectedTable, compileWriteProtection, type WriteProtectedEntry, type WriteProtectedRow, type WriteProtectedKind, type WriteProtectedHit, type WriteProtectionMatcher, } from "./core/write-protect.js";
95
95
  export { InMemoryToolResultStore, OFFLOAD_TOOL_NAME, DEFAULT_TOOL_RESULT_THRESHOLD_CHARS, assertSafeToolResultRef, buildToolResultRef, toolResultContentSegment, MAX_MINTED_TOOL_RESULT_REF_CHARS, type ToolResultProvenance, assertToolResultProvenanceMatch, normalizeToolResultProvenance, toolResultProvenanceOf, ToolResultRefConflictError, TOOL_RESULT_REF_CONFLICT_CODE, type ToolResultStore, type ToolResultSlice, type ToolResultDeletionReport, } from "./core/tool-result-store.js";
96
- export { InMemoryCheckpointStore, CheckpointError, mintCheckpointToken, checkpointVersionOf, CURRENT_CHECKPOINT_VERSION, MAX_SUPPORTED_CHECKPOINT_VERSION, ORG_ADMISSION_CHECKPOINT_VERSION, F012_CHECKPOINT_VERSION, REAL_APPROVAL_CHECKPOINT_VERSION, RESOURCE_CHECKPOINT_VERSION, TOKEN_CHECKPOINT_VERSION, debitLedger, remainingBudgetMicroUsd, remainingTokens, winnerFromOutcome, validatePendingSteer, readPendingSteerQueue, appendPendingSteer, MAX_PENDING_STEER_CHARS, MAX_PENDING_STEER_ENTRIES, PENDING_STEER_QUEUE_BYTE_BUDGET_BYTES, PENDING_STEER_FROZEN_FIELDS, ACTOR_ASSERTION_FROZEN_FIELDS, MAX_ACTOR_FIELD_CHARS, MAX_STEER_INPUT_ID_CHARS, LEGACY_PENDING_STEER_INPUT_ID, type ActorAssertion, type PendingSteerEntry, type PendingSteerInput, riskSeverity, buildRiskDescriptor, summarizeCheckpoint, type RiskDescriptor, type ProbeCause, type ProbeCauseOperands, type CheckpointStore, type CheckpointSummary, type Checkpoint, type CheckpointToken, type CheckpointGate, type CheckpointState, type SerializedCheckpointState, type CheckpointFaultMode, type PendingAction, type ResumeOutcome, type ResolvedOutcome, type ReopenReason, type ResolveExpectation, type SafetyAxis, type RealApprovalGateBit, type ResourceLedger, type ResourceLimitReason, type PlatformLimitReason, } from "./core/checkpoint-store.js";
96
+ export { InMemoryCheckpointStore, CheckpointError, mintCheckpointId, mintCheckpointToken, checkpointVersionOf, CURRENT_CHECKPOINT_VERSION, MAX_SUPPORTED_CHECKPOINT_VERSION, ORG_ADMISSION_CHECKPOINT_VERSION, F012_CHECKPOINT_VERSION, REAL_APPROVAL_CHECKPOINT_VERSION, RESOURCE_CHECKPOINT_VERSION, TOKEN_CHECKPOINT_VERSION, debitLedger, remainingBudgetMicroUsd, remainingTokens, winnerFromOutcome, validatePendingSteer, readPendingSteerQueue, appendPendingSteer, MAX_PENDING_STEER_CHARS, MAX_PENDING_STEER_ENTRIES, PENDING_STEER_QUEUE_BYTE_BUDGET_BYTES, PENDING_STEER_FROZEN_FIELDS, ACTOR_ASSERTION_FROZEN_FIELDS, MAX_ACTOR_FIELD_CHARS, MAX_STEER_INPUT_ID_CHARS, LEGACY_PENDING_STEER_INPUT_ID, type ActorAssertion, type PendingSteerEntry, type PendingSteerInput, riskSeverity, buildRiskDescriptor, summarizeCheckpoint, type RiskDescriptor, type ProbeCause, type ProbeCauseOperands, type CheckpointStore, type CheckpointSummary, type Checkpoint, type CheckpointToken, type CheckpointGate, type CheckpointState, type SerializedCheckpointState, type CheckpointFaultMode, type PendingAction, type ResumeOutcome, type ResolvedOutcome, type ReopenReason, type ResolveExpectation, type SafetyAxis, type RealApprovalGateBit, type ResourceLedger, type ResourceLimitReason, type PlatformLimitReason, } from "./core/checkpoint-store.js";
97
97
  export { InMemoryUsageWindowStore, GLOBAL_USAGE_KEY, EMPTY_USAGE_WINDOW_RECORD, chargeUsageRecord, readUsageRecord, usageRetryAfterMs, resolveUsageWindows, type UsageWindow, type UsageWindowStore, type UsageWindowReading, type UsageWindowRecord, type UsageSlot, type UsageBucketRow, } from "./core/usage-window-store.js";
98
98
  export { FileUsageWindowStore } from "./stores/file/usage-window-store.js";
99
99
  export { ENV_LIFETIME_SUSPEND_MARGIN_MS, USAGE_WINDOW_REAP_MARGIN_MS } from "./core/runner/prepare-task.js";
package/dist/index.js CHANGED
@@ -73,7 +73,7 @@ export { READ_FACE_DEFAULT_DENY_ENTRIES, READ_FACE_BUILTIN_DENY_TABLE, READ_DENY
73
73
  export { deploymentReadFaceClampNotice, resolveReadFace } from "./tools/fs/index.js";
74
74
  export { WRITE_PROTECTED_DEFAULT_TABLE, resolveWriteProtectedTable, compileWriteProtection, } from "./core/write-protect.js";
75
75
  export { InMemoryToolResultStore, OFFLOAD_TOOL_NAME, DEFAULT_TOOL_RESULT_THRESHOLD_CHARS, assertSafeToolResultRef, buildToolResultRef, toolResultContentSegment, MAX_MINTED_TOOL_RESULT_REF_CHARS, assertToolResultProvenanceMatch, normalizeToolResultProvenance, toolResultProvenanceOf, ToolResultRefConflictError, TOOL_RESULT_REF_CONFLICT_CODE, } from "./core/tool-result-store.js";
76
- export { InMemoryCheckpointStore, CheckpointError, mintCheckpointToken, checkpointVersionOf, CURRENT_CHECKPOINT_VERSION, MAX_SUPPORTED_CHECKPOINT_VERSION, ORG_ADMISSION_CHECKPOINT_VERSION, F012_CHECKPOINT_VERSION, REAL_APPROVAL_CHECKPOINT_VERSION, RESOURCE_CHECKPOINT_VERSION, TOKEN_CHECKPOINT_VERSION, debitLedger, remainingBudgetMicroUsd, remainingTokens, winnerFromOutcome, validatePendingSteer, readPendingSteerQueue, appendPendingSteer, MAX_PENDING_STEER_CHARS, MAX_PENDING_STEER_ENTRIES, PENDING_STEER_QUEUE_BYTE_BUDGET_BYTES, PENDING_STEER_FROZEN_FIELDS, ACTOR_ASSERTION_FROZEN_FIELDS, MAX_ACTOR_FIELD_CHARS, MAX_STEER_INPUT_ID_CHARS, LEGACY_PENDING_STEER_INPUT_ID, riskSeverity, buildRiskDescriptor, summarizeCheckpoint, } from "./core/checkpoint-store.js";
76
+ export { InMemoryCheckpointStore, CheckpointError, mintCheckpointId, mintCheckpointToken, checkpointVersionOf, CURRENT_CHECKPOINT_VERSION, MAX_SUPPORTED_CHECKPOINT_VERSION, ORG_ADMISSION_CHECKPOINT_VERSION, F012_CHECKPOINT_VERSION, REAL_APPROVAL_CHECKPOINT_VERSION, RESOURCE_CHECKPOINT_VERSION, TOKEN_CHECKPOINT_VERSION, debitLedger, remainingBudgetMicroUsd, remainingTokens, winnerFromOutcome, validatePendingSteer, readPendingSteerQueue, appendPendingSteer, MAX_PENDING_STEER_CHARS, MAX_PENDING_STEER_ENTRIES, PENDING_STEER_QUEUE_BYTE_BUDGET_BYTES, PENDING_STEER_FROZEN_FIELDS, ACTOR_ASSERTION_FROZEN_FIELDS, MAX_ACTOR_FIELD_CHARS, MAX_STEER_INPUT_ID_CHARS, LEGACY_PENDING_STEER_INPUT_ID, riskSeverity, buildRiskDescriptor, summarizeCheckpoint, } from "./core/checkpoint-store.js";
77
77
  export { InMemoryUsageWindowStore, GLOBAL_USAGE_KEY, EMPTY_USAGE_WINDOW_RECORD, chargeUsageRecord, readUsageRecord, usageRetryAfterMs, resolveUsageWindows, } from "./core/usage-window-store.js";
78
78
  export { FileUsageWindowStore } from "./stores/file/usage-window-store.js";
79
79
  export { ENV_LIFETIME_SUSPEND_MARGIN_MS, USAGE_WINDOW_REAP_MARGIN_MS } from "./core/runner/prepare-task.js";
@@ -2,6 +2,14 @@ import type { AgentTool, ExecutionEnv } from "../../internal/harness-types.js";
2
2
  import type { ToolEffect } from "../../core/types.js";
3
3
  import type { ReadDenyMatcher } from "./read-deny.js";
4
4
  import type { ReadFace } from "./read-face.js";
5
+ /**
6
+ * The Grep card's details assembly, PURE over the engine text ({mode, offset} from the request) —
7
+ * exported so the text→structured mapping is pinnable with synthetic texts (the byte-truncation and
8
+ * fenced-partial shapes are impractical to construct through a live tool call). Follow-up #313 will
9
+ * replace the text parsing with structured rows from runGrep; until then this seam is the honesty
10
+ * boundary.
11
+ */
12
+ export declare function grepDetailFields(text: string, mode: "files_with_matches" | "content" | "count", offset?: number): Record<string, unknown>;
5
13
  export declare function createGrepTool(env: ExecutionEnv, rootCanonical: string, additionalRoots?: readonly string[], readDeny?: ReadDenyMatcher, readFace?: ReadFace): AgentTool;
6
14
  export declare function createGlobTool(env: ExecutionEnv, rootCanonical: string, additionalRoots?: readonly string[], readDeny?: ReadDenyMatcher, readFace?: ReadFace): AgentTool;
7
15
  /** Static side-effect class of every hand tool, by name (design/44 §3). Used by prepare-task to (a) feed
@@ -2,6 +2,72 @@ import { Type } from "typebox";
2
2
  import { defineTool, errorResult } from "../../core/tools.js";
3
3
  import { resolveKey, violationText, violationDetails } from "./safety.js";
4
4
  import { runGrepDetailed, runGlobDetailed, splitAbsoluteGlobPattern, invalidGlobTokens } from "./search.js";
5
+ export function grepDetailFields(text, mode, offset) {
6
+ const rows = text.startsWith("No matches.")
7
+ ? []
8
+ : text
9
+ .split("\n")
10
+ .filter((l) => l.length > 0 &&
11
+ l !== "No matches." &&
12
+ !l.startsWith("…[") &&
13
+ !l.startsWith("[offset") &&
14
+ !l.startsWith("[note") &&
15
+ !l.startsWith("[!]") &&
16
+ !l.startsWith("<<<UNTRUSTED ") &&
17
+ !l.startsWith("<<<END UNTRUSTED "));
18
+ const contentPathOf = (l) => {
19
+ const m = /^(.*?):(\d+):/.exec(l);
20
+ return m ? m[1] : l;
21
+ };
22
+ const byteTruncated = text.includes("…[output truncated at ");
23
+ const truncatedFlag = byteTruncated ? { truncated: true } : {};
24
+ const capMarker = /^…\[capped at (\d+) of (\d+)\]$/m.exec(text);
25
+ const cappedAt = capMarker ? Number(capMarker[1]) : undefined;
26
+ const capTotal = capMarker ? Number(capMarker[2]) : undefined;
27
+ const appliedOffset = typeof offset === "number" && offset > 0 ? { appliedOffset: offset } : {};
28
+ const appliedLimit = cappedAt !== undefined ? { appliedLimit: cappedAt } : {};
29
+ let detailFields;
30
+ if (mode === "files_with_matches") {
31
+ detailFields = { filenames: rows, numFiles: rows.length, totalFiles: capTotal ?? rows.length, ...appliedLimit, ...appliedOffset };
32
+ }
33
+ else if (mode === "count") {
34
+ const filenames = [...new Set(rows.map((l) => (l.includes(":") ? l.slice(0, l.lastIndexOf(":")) : l)))];
35
+ let numMatches = 0;
36
+ let malformed = false;
37
+ for (const l of rows) {
38
+ const m = /:(\d+)$/.exec(l);
39
+ if (m)
40
+ numMatches += Number(m[1]);
41
+ else
42
+ malformed = true;
43
+ }
44
+ detailFields = {
45
+ filenames,
46
+ numFiles: filenames.length,
47
+ content: rows.join("\n"),
48
+ ...(malformed || byteTruncated ? {} : { numMatches }),
49
+ ...truncatedFlag,
50
+ ...appliedLimit,
51
+ ...appliedOffset,
52
+ };
53
+ }
54
+ else {
55
+ const filenames = [...new Set(rows.map(contentPathOf))];
56
+ const joined = rows.join("\n");
57
+ const GREP_CONTENT_PREVIEW_CHARS = 16_000;
58
+ detailFields = {
59
+ filenames,
60
+ numFiles: filenames.length,
61
+ content: joined.length > GREP_CONTENT_PREVIEW_CHARS ? `${joined.slice(0, GREP_CONTENT_PREVIEW_CHARS)}\n…[truncated — full text in the tool output]` : joined,
62
+ numLines: rows.length,
63
+ ...(byteTruncated ? {} : { totalLines: capTotal ?? rows.length }),
64
+ ...truncatedFlag,
65
+ ...appliedLimit,
66
+ ...appliedOffset,
67
+ };
68
+ }
69
+ return detailFields;
70
+ }
5
71
  export function createGrepTool(env, rootCanonical, additionalRoots, readDeny, readFace) {
6
72
  return defineTool({
7
73
  name: "Grep",
@@ -95,64 +161,7 @@ export function createGrepTool(env, rootCanonical, additionalRoots, readDeny, re
95
161
  if (text.startsWith("Error (grep)") || text.startsWith("Error (Grep)"))
96
162
  return errorResult(text);
97
163
  const mode = a.output_mode ?? "files_with_matches";
98
- const rows = text.startsWith("No matches.")
99
- ? []
100
- : text
101
- .split("\n")
102
- .filter((l) => l.length > 0 &&
103
- !l.startsWith("…[") &&
104
- !l.startsWith("[offset") &&
105
- !l.startsWith("[note") &&
106
- !l.startsWith("[!]") &&
107
- !l.startsWith("<<<UNTRUSTED ") &&
108
- !l.startsWith("<<<END UNTRUSTED "));
109
- const contentPathOf = (l) => {
110
- const m = /^(.*?):(\d+):/.exec(l);
111
- return m ? m[1] : l;
112
- };
113
- const capMarker = /^…\[capped at (\d+) of (\d+)\]$/m.exec(text);
114
- const cappedAt = capMarker ? Number(capMarker[1]) : undefined;
115
- const capTotal = capMarker ? Number(capMarker[2]) : undefined;
116
- const appliedOffset = typeof a.offset === "number" && a.offset > 0 ? { appliedOffset: a.offset } : {};
117
- const appliedLimit = cappedAt !== undefined ? { appliedLimit: cappedAt } : {};
118
- let detailFields;
119
- if (mode === "files_with_matches") {
120
- detailFields = { filenames: rows, numFiles: rows.length, totalFiles: capTotal ?? rows.length, ...appliedLimit, ...appliedOffset };
121
- }
122
- else if (mode === "count") {
123
- const filenames = [...new Set(rows.map((l) => (l.includes(":") ? l.slice(0, l.lastIndexOf(":")) : l)))];
124
- let numMatches = 0;
125
- let malformed = false;
126
- for (const l of rows) {
127
- const m = /:(\d+)$/.exec(l);
128
- if (m)
129
- numMatches += Number(m[1]);
130
- else
131
- malformed = true;
132
- }
133
- detailFields = {
134
- filenames,
135
- numFiles: filenames.length,
136
- content: rows.join("\n"),
137
- ...(malformed ? {} : { numMatches }),
138
- ...appliedLimit,
139
- ...appliedOffset,
140
- };
141
- }
142
- else {
143
- const filenames = [...new Set(rows.map(contentPathOf))];
144
- const joined = rows.join("\n");
145
- const GREP_CONTENT_PREVIEW_CHARS = 16_000;
146
- detailFields = {
147
- filenames,
148
- numFiles: filenames.length,
149
- content: joined.length > GREP_CONTENT_PREVIEW_CHARS ? `${joined.slice(0, GREP_CONTENT_PREVIEW_CHARS)}\n…[truncated — full text in the tool output]` : joined,
150
- numLines: rows.length,
151
- totalLines: capTotal ?? rows.length,
152
- ...appliedLimit,
153
- ...appliedOffset,
154
- };
155
- }
164
+ const detailFields = grepDetailFields(text, mode, a.offset);
156
165
  return {
157
166
  content: text,
158
167
  details: { type: "grep", mode, ...detailFields, ...(grepRun.degraded ?? {}), ...(grepRun.withheld !== undefined ? { withheld: grepRun.withheld } : {}) },
@@ -1,4 +1,9 @@
1
1
  import type { ExecutionEnv } from "../../internal/harness-types.js";
2
+ /** backlog #310 — the JS scanner's own output ceiling, in bytes of emitted content rows. Pinned to the
3
+ * ripgrep leg's figure ({@link MAX_EXEC_OUTPUT_BYTES}, the exec pipe's rolling-tail bound) so the two
4
+ * legs cannot disagree about how much one search may return; see the collection site in {@link jsGrep}
5
+ * for why this leg bounds the HEAD where the pipe bounds the tail. */
6
+ export declare const JS_GREP_OUTPUT_MAX_BYTES: number;
2
7
  /** Directories never worth crawling (design/64 §10.3) — dependency/build/cache trees, PLUS the VCS
3
8
  * metadata directories CC names explicitly (RB-200 F1, 220 @368402: `Ok_ = [".git",".svn",".hg",
4
9
  * ".bzr",".jj",".sl"]` — git/svn/mercurial/bazaar/jujutsu/sapling). `.bzr`/`.jj`/`.sl` are additive here
@@ -153,6 +158,10 @@ export interface JsGrepGuards {
153
158
  budgetMs?: number;
154
159
  /** Longest line/content (chars) a gray-zone pattern may be matched against. */
155
160
  longLineLimit?: number;
161
+ /** backlog #310 — bytes of content rows this scan may accumulate before it stops collecting and
162
+ * says so (default {@link JS_GREP_OUTPUT_MAX_BYTES}). A seam for the same reason the two above
163
+ * are: the boundary is pinnable at a testable size instead of only at an 8MB harness. */
164
+ outputMaxBytes?: number;
156
165
  }
157
166
  /** design/199 件B — what a traversal withheld under the sensitive-path read deny list, structured
158
167
  * (the prose note is the model-facing twin). Three shapes (§3.1): `pruned_count` = the JS walker
@@ -208,6 +217,8 @@ export interface GrepRunResult {
208
217
  * withheld / no deny judge in play. */
209
218
  withheld?: ReadDenyWithheld;
210
219
  }
220
+ /** The output modes the rg leg and {@link jsGrep} share. */
221
+ type GrepOutputMode = "content" | "files_with_matches" | "count";
211
222
  /**
212
223
  * backlog #303 (rescan-hardened form) — the rg leg's deny TRIPWIRE. rg's exclusion globs are spelled
213
224
  * from the pattern text and cannot express the win32 component-alias family (`.aws.` / `.aws ` —
@@ -220,21 +231,31 @@ export interface GrepRunResult {
220
231
  * ripgrep prints filenames verbatim, so a path component containing a NEWLINE splits one record
221
232
  * across physical lines whose fragments carry no `:line:` boundary — the filter kept them and the
222
233
  * deny-listed content passed (a protection hole, not a precision residual); and pruning every line
223
- * of a partial run fabricated a `No matches.` row inside the structured card. Text-splitting rg
224
- * records is unsalvageable without structured output (`--json`, backlog #306), so this function no
225
- * longer edits anything: it JUDGES. Any deny hit — or any line the format cannot account for —
234
+ * of a partial run fabricated a `No matches.` row inside the structured card. So this function no
235
+ * longer edits anything: it JUDGES. Any deny hit or any record the format cannot account for —
226
236
  * trips, and the caller abandons the rg run for the JS scanner, whose walk prunes with the
227
237
  * authoritative judge and needs no path parsing at all. rg stays the fast path for the common case
228
238
  * (no guarded entries in the result); the moment a guarded spelling is involved, the engine that
229
239
  * cannot mis-parse it owns the answer.
230
240
  *
231
- * Trip conditions by mode `files_with_matches`: every line IS a path, judged whole; `count`:
232
- * `path:N` (a line without the numeric tail is a split record ambiguous, trip); `content`:
233
- * judged at EVERY `:digits:` boundary and every `-digits-` boundary (a candidate cut inside match
234
- * text can over-trip safe: the fallback re-derives the exact answer), and a non-separator line
235
- * with NO boundary at all is a split record — trip. `--` separators and empty lines pass.
241
+ * backlog #306 what it judges is now ripgrep's OWN path field ({@link parseRgRecords}, `--null`),
242
+ * not a prefix cut at a guessed separator. Trip conditions, all three modes alike: a record whose
243
+ * path the deny judge names trips as `deny-hit`; a record with no path field at all (rg's binary-file
244
+ * notice, a no-filename row, an unterminated tail, output from an env that ignored `--null`) is
245
+ * unaccountable and trips as `ambiguous-record`same abandonment, same reason code as before.
246
+ * `dropIncompleteTail` mirrors {@link parseRgRecords}: on the partial legs the cut tail is dropped
247
+ * BEFORE judging, because judging a record the caller will never receive can only cost an
248
+ * unnecessary rescan — every record that does reach the output is still judged, which is what the
249
+ * rule is for. (A killed stream ending mid-record is an EXPECTED shape with a rule of its own; the
250
+ * unaccountable arm is for records that survive that rule.)
251
+ * `--` group separators and empty lines carry no path by construction and pass. The `:digits:` /
252
+ * `-digits-` boundary family this used to reason about is gone with the guessing: both of #303's
253
+ * declared residuals (over-tripping on a cluster inside match TEXT, and the dedup key folding two
254
+ * denied files whose own names carry a cluster) were artifacts of that decision, not of the judge.
236
255
  */
237
- export declare function rgOutputDenyTripwire(stdout: string, mode: "content" | "files_with_matches" | "count", judge: Pick<ReadDenyJudge, "matchPath">): {
256
+ export declare function rgOutputDenyTripwire(stdout: string, mode: GrepOutputMode, judge: Pick<ReadDenyJudge, "matchPath">, opts?: {
257
+ dropIncompleteTail?: boolean;
258
+ }): {
238
259
  trip: false;
239
260
  } | {
240
261
  trip: true;
@@ -285,3 +306,4 @@ export declare function runGlobDetailed(env: ExecutionEnv, root: string, pattern
285
306
  error?: string;
286
307
  withheld?: ReadDenyWithheld;
287
308
  }>;
309
+ export {};