pi-braid 0.40.3 → 0.40.6

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.
Files changed (85) hide show
  1. package/AGENTS.md +17 -18
  2. package/README.md +116 -130
  3. package/TOOLS.md +35 -35
  4. package/build-info.json +4 -4
  5. package/dist/extensions/core/agent-environment/index.js +4 -3
  6. package/dist/extensions/core/prompt-builder/index.js +12 -7
  7. package/dist/extensions/core/prompt-builder/instance-composition.d.ts +51 -10
  8. package/dist/extensions/core/prompt-builder/instance-composition.js +87 -14
  9. package/dist/extensions/core/prompt-builder/system-info-tool.d.ts +3 -3
  10. package/dist/extensions/core/prompt-builder/system-info-tool.js +3 -3
  11. package/dist/extensions/core/prompt-builder/tool-guidelines.js +1 -1
  12. package/dist/extensions/core/skills-tool/auto-inject.d.ts +178 -17
  13. package/dist/extensions/core/skills-tool/auto-inject.js +264 -23
  14. package/dist/extensions/core/skills-tool/index.d.ts +70 -2
  15. package/dist/extensions/core/skills-tool/index.js +149 -15
  16. package/dist/extensions/core/skills-tool/reach.d.ts +119 -0
  17. package/dist/extensions/core/skills-tool/reach.js +213 -0
  18. package/dist/extensions/core/skills-tool/skill-triggers.d.ts +74 -13
  19. package/dist/extensions/core/skills-tool/skill-triggers.js +105 -24
  20. package/dist/extensions/enforcement/handlers/register-session-compact.js +12 -1
  21. package/dist/extensions/enforcement/handlers/register-session-shutdown-resumption.js +15 -34
  22. package/dist/extensions/enforcement/handlers/register-session-start-discoverability.d.ts +2 -1
  23. package/dist/extensions/enforcement/handlers/register-session-start-discoverability.js +55 -28
  24. package/dist/extensions/enforcement/handlers/register-turn-end-hygiene.d.ts +14 -10
  25. package/dist/extensions/enforcement/handlers/register-turn-end-hygiene.js +88 -82
  26. package/dist/extensions/enforcement/resume-state.d.ts +76 -0
  27. package/dist/extensions/enforcement/resume-state.js +90 -0
  28. package/dist/extensions/enforcement/substate/context-hygiene-state.d.ts +9 -0
  29. package/dist/extensions/enforcement/substate/context-hygiene-state.js +8 -0
  30. package/dist/extensions/enforcement/whisper.d.ts +12 -0
  31. package/dist/extensions/enforcement/whisper.js +16 -1
  32. package/dist/extensions/heartbeat/heartbeat-scheduler/index.js +14 -6
  33. package/dist/extensions/heartbeat/heartbeat-scheduler/policies.d.ts +7 -0
  34. package/dist/extensions/heartbeat/heartbeat-scheduler/policies.js +372 -0
  35. package/dist/extensions/memory/context-manager/handlers/tools/context-compress.d.ts +8 -2
  36. package/dist/extensions/memory/context-manager/handlers/tools/context-compress.js +54 -21
  37. package/dist/lib/config/braid-paths.d.ts +14 -0
  38. package/dist/lib/config/braid-paths.js +18 -0
  39. package/dist/lib/config/braid-settings.d.ts +21 -0
  40. package/dist/lib/config/braid-settings.js +16 -0
  41. package/dist/lib/detection/context-tiers.d.ts +47 -3
  42. package/dist/lib/detection/context-tiers.js +122 -13
  43. package/dist/lib/events/turn-boundary.d.ts +78 -0
  44. package/dist/lib/events/turn-boundary.js +91 -0
  45. package/dist/lib/orchestrator/workers.d.ts +3 -2
  46. package/dist/lib/orchestrator/workers.js +22 -4
  47. package/dist/lib/state/global-state.d.ts +1 -1
  48. package/docs/CONTEXT-POLICY.md +8 -0
  49. package/docs/LOCAL_DEV.md +0 -1
  50. package/docs/ROADMAP.md +5 -5
  51. package/docs/SETUP.md +19 -18
  52. package/package.json +2 -1
  53. package/skills/herdr-topology/SKILL.md +6 -1
  54. package/skills/research/SKILL.md +32 -0
  55. package/skills/subagent-orchestration/SKILL.md +2 -2
  56. package/src/extensions/core/agent-environment/index.ts +4 -3
  57. package/src/extensions/core/prompt-builder/AGENTS.md +8 -4
  58. package/src/extensions/core/prompt-builder/index.ts +12 -7
  59. package/src/extensions/core/prompt-builder/instance-composition.ts +98 -14
  60. package/src/extensions/core/prompt-builder/system-info-tool.ts +3 -3
  61. package/src/extensions/core/prompt-builder/tool-guidelines.ts +1 -1
  62. package/src/extensions/core/skills-tool/auto-inject.ts +350 -27
  63. package/src/extensions/core/skills-tool/index.ts +217 -15
  64. package/src/extensions/core/skills-tool/reach.ts +256 -0
  65. package/src/extensions/core/skills-tool/skill-triggers.ts +130 -23
  66. package/src/extensions/enforcement/handlers/register-session-compact.ts +12 -1
  67. package/src/extensions/enforcement/handlers/register-session-shutdown-resumption.ts +23 -36
  68. package/src/extensions/enforcement/handlers/register-session-start-discoverability.ts +59 -29
  69. package/src/extensions/enforcement/handlers/register-turn-end-hygiene.ts +101 -81
  70. package/src/extensions/enforcement/resume-state.ts +127 -0
  71. package/src/extensions/enforcement/substate/context-hygiene-state.ts +17 -0
  72. package/src/extensions/enforcement/whisper.ts +21 -1
  73. package/src/extensions/heartbeat/heartbeat-scheduler/index.ts +20 -5
  74. package/src/extensions/heartbeat/heartbeat-scheduler/policies.ts +370 -0
  75. package/src/extensions/memory/AGENTS.md +25 -1
  76. package/src/extensions/memory/context-manager/handlers/tools/context-compress.ts +80 -27
  77. package/src/lib/config/braid-paths.ts +20 -0
  78. package/src/lib/config/braid-settings.ts +46 -0
  79. package/src/lib/detection/context-tiers.ts +186 -13
  80. package/src/lib/events/AGENTS.md +1 -0
  81. package/src/lib/events/turn-boundary.ts +115 -0
  82. package/src/lib/orchestrator/workers.ts +33 -12
  83. package/src/lib/state/global-state.ts +1 -0
  84. package/src/lib/utils/AGENTS.md +1 -1
  85. package/src/types/extension-context.d.ts +22 -79
@@ -44,6 +44,376 @@ export const WEB_TOOLS = new Set(["fetch_content", "browser_navigate", "browser_
44
44
 
45
45
  export const CODE_TOOLS = new Set(["bash", "write", "edit"]);
46
46
 
47
+ // ═══════════════════════════════════════════════════════════════════════════
48
+ // VERIFICATION-SHAPE RECOGNITION FOR BASH
49
+ // ═══════════════════════════════════════════════════════════════════════════
50
+ // VERIFICATION_TOOLS is a tool-NAME set, and the same work routinely runs
51
+ // through bash: `grep -rn "needle" src`, `git diff main...HEAD`, `node -e
52
+ // 'console.log(m.x)'`. The set cannot see it, so the verification counter
53
+ // stayed at zero while the agent verified and the hint below told the agent it
54
+ // had not verified in ten turns. Counting every bash call would swing the
55
+ // defect the other way — an edit, a build or a publish script is not
56
+ // verification, and a counter that accepts them silences the hint it exists to
57
+ // raise.
58
+ //
59
+ // The rule, in one sentence: a command counts when it puts an inspecting
60
+ // program at a program position and no mutating program anywhere.
61
+ // - program positions are the first word of each `;` `&&` `||` `|` `&`
62
+ // segment, tokenized quote-aware (`grep -n "a|b" f` is one segment);
63
+ // - inspecting: search/locate, read, compare/extract programs, git read
64
+ // subcommands, and an inline interpreter probe (`node -e`, `python -c`)
65
+ // whose script shows no write primitive;
66
+ // - mutating: rm/mv/cp/mkdir/tee/dd/chmod…, `sed -i`, `perl -i`, git write
67
+ // subcommands, and any unquoted redirect to a file other than /dev/null.
68
+ // Both directions matter: the two named holes are an interpreter script that
69
+ // writes through a primitive outside INLINE_WRITE_RE, and a mutating program
70
+ // behind `eval` or a heredoc body (both are read as inspection when the same
71
+ // command also inspects — they are listed here rather than guessed at).
72
+
73
+ /** Programs that inspect existing state: search, locate, read, compare, extract. */
74
+ const INSPECT_PROGRAMS = new Set([
75
+ "grep",
76
+ "egrep",
77
+ "fgrep",
78
+ "rg",
79
+ "ag",
80
+ "ack",
81
+ "ack-grep",
82
+ "ugrep",
83
+ "pcregrep",
84
+ "find",
85
+ "fd",
86
+ "fdfind",
87
+ "locate",
88
+ "cat",
89
+ "tac",
90
+ "head",
91
+ "tail",
92
+ "bat",
93
+ "less",
94
+ "more",
95
+ "wc",
96
+ "nl",
97
+ "od",
98
+ "xxd",
99
+ "strings",
100
+ "file",
101
+ "stat",
102
+ "ls",
103
+ "tree",
104
+ "realpath",
105
+ "readlink",
106
+ "jq",
107
+ "yq",
108
+ "diff",
109
+ "diff3",
110
+ "cmp",
111
+ "comm",
112
+ "cut",
113
+ "uniq",
114
+ "tr",
115
+ "column",
116
+ "awk",
117
+ "gawk",
118
+ "sed",
119
+ ]);
120
+
121
+ /** Programs that mutate the tree. Any one of them vetoes the call. */
122
+ const MUTATE_PROGRAMS = new Set([
123
+ "rm",
124
+ "rmdir",
125
+ "mv",
126
+ "cp",
127
+ "ln",
128
+ "touch",
129
+ "mkdir",
130
+ "install",
131
+ "truncate",
132
+ "dd",
133
+ "shred",
134
+ "tee",
135
+ "patch",
136
+ "chmod",
137
+ "chown",
138
+ "chgrp",
139
+ ]);
140
+
141
+ /**
142
+ * `git` subcommands that only read (the inspection side of the verbs the
143
+ * destructive gate covers separately).
144
+ */
145
+ const GIT_READ_SUBS = new Set([
146
+ "diff",
147
+ "diff-tree",
148
+ "diff-index",
149
+ "diff-files",
150
+ "show",
151
+ "log",
152
+ "status",
153
+ "grep",
154
+ "blame",
155
+ "shortlog",
156
+ "whatchanged",
157
+ "ls-files",
158
+ "ls-tree",
159
+ "cat-file",
160
+ "rev-parse",
161
+ "rev-list",
162
+ "show-ref",
163
+ "for-each-ref",
164
+ "describe",
165
+ "merge-base",
166
+ "name-rev",
167
+ "count-objects",
168
+ "fsck",
169
+ ]);
170
+
171
+ /** `git` subcommands that write. Conservative: a remote/config/branch/tag
172
+ * verb may mutate, so it vetoes rather than counts. */
173
+ const GIT_WRITE_SUBS = new Set([
174
+ "add",
175
+ "commit",
176
+ "push",
177
+ "pull",
178
+ "fetch",
179
+ "checkout",
180
+ "switch",
181
+ "restore",
182
+ "reset",
183
+ "revert",
184
+ "merge",
185
+ "rebase",
186
+ "cherry-pick",
187
+ "stash",
188
+ "apply",
189
+ "am",
190
+ "clean",
191
+ "rm",
192
+ "mv",
193
+ "tag",
194
+ "branch",
195
+ "init",
196
+ "clone",
197
+ "config",
198
+ "remote",
199
+ "submodule",
200
+ "worktree",
201
+ "update-index",
202
+ "update-ref",
203
+ "symbolic-ref",
204
+ "notes",
205
+ "gc",
206
+ "prune",
207
+ "reflog",
208
+ ]);
209
+
210
+ /** Interpreters whose inline-eval form can be a probe. */
211
+ const INLINE_INTERPRETERS = new Set(["node", "deno", "bun", "python", "python3", "ruby"]);
212
+
213
+ /** Write primitives inside an inline script (both JS/TS and Python spellings). */
214
+ const INLINE_WRITE_RE =
215
+ /writeFileSync|writeFile\(|appendFileSync|createWriteStream|rmSync|unlinkSync|renameSync|mkdirSync|copyFileSync|truncateSync|chmodSync|openSync\([^)]*["'`][wa]|os\.remove|os\.rename|shutil\.|open\([^)]*["'`][wa]/;
216
+
217
+ /** Wrappers whose presence does not decide the kind of the segment. */
218
+ const WRAPPER_PROGRAMS = new Set(["sudo", "env", "time", "nice", "command", "nohup"]);
219
+
220
+ /** Tokens that end one command segment. */
221
+ const SEGMENT_SEPARATORS = new Set([";", "&&", "||", "|", "&", "\n"]);
222
+
223
+ /** The program name without its path (`/usr/bin/grep` → `grep`). */
224
+ function baseName(token: string): string {
225
+ const parts = token.split("/");
226
+ return parts[parts.length - 1] ?? token;
227
+ }
228
+
229
+ /**
230
+ * Quote-aware tokenizer that keeps shell operators as their own tokens
231
+ * (`;` `&&` `||` `|` `&` newline `>` `>>`, and `&1`/`&2` as one token) and
232
+ * strips quote characters. An escaped character stays literal, so `find …
233
+ * -exec grep … \;` yields the word `;` rather than a separator.
234
+ */
235
+ function tokenizeCommand(command: string): string[] {
236
+ const tokens: string[] = [];
237
+ let cur = "";
238
+ let quote: '"' | "'" | null = null;
239
+ const flush = (): void => {
240
+ if (cur) {
241
+ tokens.push(cur);
242
+ cur = "";
243
+ }
244
+ };
245
+ for (let i = 0; i < command.length; i++) {
246
+ const ch = command[i] ?? "";
247
+ if (quote) {
248
+ if (ch === "\\" && quote === '"') {
249
+ cur += command[i + 1] ?? "";
250
+ i++;
251
+ continue;
252
+ }
253
+ if (ch === quote) {
254
+ quote = null;
255
+ continue;
256
+ }
257
+ cur += ch;
258
+ continue;
259
+ }
260
+ if (ch === "\\") {
261
+ cur += command[i + 1] ?? "";
262
+ i++;
263
+ continue;
264
+ }
265
+ if (ch === '"' || ch === "'") {
266
+ quote = ch;
267
+ continue;
268
+ }
269
+ if (/\s/.test(ch)) {
270
+ flush();
271
+ if (ch === "\n") tokens.push("\n");
272
+ continue;
273
+ }
274
+ if (ch === ";" || ch === "&") {
275
+ flush();
276
+ const next = command[i + 1] ?? "";
277
+ if (ch === "&" && /\d/.test(next)) {
278
+ tokens.push(`&${next}`);
279
+ i++;
280
+ continue;
281
+ }
282
+ if (ch === "&" && next === "&") {
283
+ tokens.push("&&");
284
+ i++;
285
+ continue;
286
+ }
287
+ tokens.push(ch);
288
+ continue;
289
+ }
290
+ if (ch === "|") {
291
+ flush();
292
+ if ((command[i + 1] ?? "") === "|") {
293
+ tokens.push("||");
294
+ i++;
295
+ continue;
296
+ }
297
+ tokens.push("|");
298
+ continue;
299
+ }
300
+ if (ch === ">") {
301
+ flush();
302
+ if ((command[i + 1] ?? "") === ">") {
303
+ tokens.push(">>");
304
+ i++;
305
+ continue;
306
+ }
307
+ tokens.push(">");
308
+ continue;
309
+ }
310
+ cur += ch;
311
+ }
312
+ flush();
313
+ return tokens;
314
+ }
315
+
316
+ /** Split tokenized commands into segments at the separators. */
317
+ function splitSegments(tokens: string[]): string[][] {
318
+ const out: string[][] = [];
319
+ let cur: string[] = [];
320
+ for (const token of tokens) {
321
+ if (SEGMENT_SEPARATORS.has(token)) {
322
+ if (cur.length > 0) out.push(cur);
323
+ cur = [];
324
+ continue;
325
+ }
326
+ cur.push(token);
327
+ }
328
+ if (cur.length > 0) out.push(cur);
329
+ return out;
330
+ }
331
+
332
+ /**
333
+ * A redirect that writes a real file. `/dev/null` and fd duplications
334
+ * (`2>&1`) are the ordinary shapes of a silent inspection and stay benign — a
335
+ * redirect into a path is a write, whatever program it follows.
336
+ */
337
+ function writesFile(tokens: string[]): boolean {
338
+ for (let i = 0; i < tokens.length; i++) {
339
+ const token = tokens[i];
340
+ if (token !== ">" && token !== ">>") continue;
341
+ const target = tokens[i + 1];
342
+ if (!target || SEGMENT_SEPARATORS.has(target)) continue;
343
+ if (/^\/dev\/(?:null|stdout|stderr)$/.test(target)) continue;
344
+ if (/^&\d$/.test(target)) continue;
345
+ return true;
346
+ }
347
+ return false;
348
+ }
349
+
350
+ /** The first argument that is not a flag (`xargs -I{} rm` → `rm`). */
351
+ function firstArgument(args: string[]): string | null {
352
+ for (const arg of args) if (!arg.startsWith("-")) return arg;
353
+ return null;
354
+ }
355
+
356
+ /** Classify one segment by the program in its program position. */
357
+ function classifySegment(tokens: string[]): "inspect" | "mutate" | "neutral" {
358
+ let i = 0;
359
+ while (i < tokens.length && /^[A-Za-z_][A-Za-z0-9_]*=/.test(tokens[i] ?? "")) i++;
360
+ while (i < tokens.length && WRAPPER_PROGRAMS.has(baseName(tokens[i] ?? ""))) i++;
361
+ const program = tokens[i] ? baseName(tokens[i] as string) : "";
362
+ if (!program) return "neutral";
363
+ const args = tokens.slice(i + 1);
364
+
365
+ if (MUTATE_PROGRAMS.has(program)) return "mutate";
366
+ // xargs runs its first non-flag argument: classify the thing it runs.
367
+ if (program === "xargs") {
368
+ const target = firstArgument(args);
369
+ return target ? classifySegment([target]) : "neutral";
370
+ }
371
+ if (program === "git") {
372
+ const sub = args.find((a) => GIT_READ_SUBS.has(a) || GIT_WRITE_SUBS.has(a));
373
+ if (!sub) return "neutral";
374
+ return GIT_WRITE_SUBS.has(sub) ? "mutate" : "inspect";
375
+ }
376
+ const inPlace = args.some(
377
+ (a) => a === "--in-place" || /^-[A-Za-z]*i/.test(a) || /^--in-place=/.test(a),
378
+ );
379
+ if (program === "sed" || program === "perl") {
380
+ if (inPlace) return "mutate";
381
+ return "inspect";
382
+ }
383
+ if (INLINE_INTERPRETERS.has(program)) {
384
+ const inline = args.some((a) => a === "-e" || a === "-p" || a === "--eval" || a === "-c");
385
+ if (!inline) return "neutral";
386
+ return INLINE_WRITE_RE.test(args.join(" ")) ? "mutate" : "inspect";
387
+ }
388
+ if (INSPECT_PROGRAMS.has(program)) return "inspect";
389
+ return "neutral";
390
+ }
391
+
392
+ /** True when the command is inspection-shaped (see the rule above). */
393
+ export function bashCommandIsVerification(command: string): boolean {
394
+ if (!command.trim()) return false;
395
+ const tokens = tokenizeCommand(command);
396
+ if (writesFile(tokens)) return false;
397
+ let inspected = false;
398
+ for (const segment of splitSegments(tokens)) {
399
+ const kind = classifySegment(segment);
400
+ if (kind === "mutate") return false;
401
+ if (kind === "inspect") inspected = true;
402
+ }
403
+ return inspected;
404
+ }
405
+
406
+ /**
407
+ * Does this tool call count as verification activity? The name set first, then
408
+ * bash by the command's shape. Every other tool is not verification.
409
+ */
410
+ export function isVerificationToolCall(toolName: string, input: unknown): boolean {
411
+ if (VERIFICATION_TOOLS.has(toolName)) return true;
412
+ if (toolName !== "bash") return false;
413
+ const command = (input as { command?: unknown } | undefined)?.command;
414
+ return typeof command === "string" && bashCommandIsVerification(command);
415
+ }
416
+
47
417
  export function createDefaultPolicies(
48
418
  getToolTracking: () => ToolTracking,
49
419
  onSignal: (type: string, data: Record<string, unknown>) => void,
@@ -49,7 +49,8 @@ The `cancelTimersByTag` helper from `extensions/scheduling/timer/index.ts` is th
49
49
  `context_compress` returns its result immediately and defers the actual
50
50
  `ctx.compact()` to the next safe boundary (batch-settle scheduler in
51
51
  `context-manager/handlers/tools/context-compress.ts`): the first `turn_end`
52
- whose persisted `toolResults` account for every active tool call, or
52
+ whose persisted `toolResults` account for every active tool call AND whose
53
+ assistant message reads as loop-done under the shared boundary rule, or
53
54
  `agent_settled`, whichever comes first. Active calls are tracked by
54
55
  `toolCallId`, registered at `tool_execution_start` and released by its own
55
56
  `tool_execution_end` or by a boundary event whose `toolResults` carry that
@@ -71,6 +72,29 @@ deferred `ctx.compact()` still aborts the run at the settle point, so
71
72
  compaction remains a turn-boundary action, not a mid-batch one. The
72
73
  mechanism is loss-preserving; the boundary discipline stands.
73
74
 
75
+ ### The turn-boundary rule
76
+
77
+ Which boundaries may abort a run is one rule in one home,
78
+ `lib/events/turn-boundary.ts`, read by the drain above and by the enforcement
79
+ hygiene act (`enforcement/handlers/register-turn-end-hygiene.ts`). It is an
80
+ allow-list over `message.stopReason`: the reasons that end a turn's work
81
+ (`stop`, `length`, `error`, `aborted`) fire, `toolUse` parks, and so does any
82
+ boundary it cannot read or does not recognize (`pending`, `deferred`, a
83
+ missing or misspelled reason, an empty object). A caller holding no boundary
84
+ at all (`undefined`) keeps the prior contract and fires, because the tool path
85
+ is parked by its own tool call id instead. The 0.90W hard backstop cuts a live
86
+ loop anyway.
87
+
88
+ The rule is narrower than the safety property it stands for. The loop still
89
+ dispatches when a tool call rides a non-`toolUse` stop reason, and when a steer
90
+ is queued at any boundary (pi polls steering after the `turn_end` emit), so a
91
+ boundary this rule accepts can still reach the abort it exists to avoid.
92
+ Reading "is anything left to dispatch" from the loop itself would close that
93
+ gap; the rule does not, and the drain inherits the residual. Its truth table is
94
+ pinned in `test/runtime/memory/test-compress-drain-boundaries.mjs`, and the
95
+ heartbeat's forwarding of the boundary into the arm path end to end in
96
+ `test/runtime/heartbeat/test-turn-boundary-arm-wiring.mjs`.
97
+
74
98
  ## Whisper wake-up policy in this module
75
99
 
76
100
  Whispers from compaction-retry, navigation, and memory hooks use `triggerTurn: true` for messages that demand the agent's action:
@@ -26,6 +26,10 @@ import { analyzeSession } from "../../analysis.js";
26
26
  import * as timerMod from "../../../../scheduling/timer/index.js";
27
27
  import { getState } from "../../../../enforcement/state.js";
28
28
  import { CONTEXT_COMPACT_REFUSE_PERCENT } from "../../../../../lib/detection/context-thresholds.js";
29
+ import {
30
+ turnBoundaryMayAbortRun,
31
+ type TurnBoundary,
32
+ } from "../../../../../lib/events/turn-boundary.js";
29
33
  import { getCmState, isCompactionActive, log, recordNavigationAction } from "../shared.js";
30
34
  import { snapBoundaryOutsideToolBatch } from "../tool-batch-boundary.js";
31
35
  import { formatTokensK } from "../../../../../lib/detection/context-usage.js";
@@ -91,16 +95,30 @@ function optionalNumber(v: unknown): number | undefined {
91
95
  // Batch-settle compaction scheduler.
92
96
  // Contract: the tool returns its result FIRST (execute() resolves before a
93
97
  // compaction is armed), but ctx.compact() — which aborts the turn — fires only
94
- // at a safe boundary: agent_settled (the run fully settled, abort a clean
95
- // no-op) or turn_end inside a live run (the turn's messages and tool results
96
- // are already persisted; the abort lands between turns, where a queued steer
97
- // is appended before the loop polls the abort signal and its turn runs on the
98
- // post-compaction resume). Firing while a tool call is still active aborts a
99
- // still-live run (the agent is streaming its response to the context_compress
100
- // result), surfacing a spurious "This operation was aborted" error in the TUI
101
- // and tearing the finishing turn. agent_settled alone never drains a run kept
102
- // alive by continuous steers, so the armed compaction stalls behind live work
103
- // while usage keeps climbing, hence the turn_end boundary.
98
+ // at a boundary where the abort cannot tear live work: agent_settled (the run
99
+ // fully settled, abort a clean no-op), or a turn_end that reads as loop-done.
100
+ // The rule that decides the second case lives in one place
101
+ // (lib/events/turn-boundary.ts) and is shared with the hygiene act, because a
102
+ // copy of it drifts silently. Read the residual there too: a loop-done stop
103
+ // reason is necessary, not sufficient. The loop still dispatches when a tool
104
+ // call rides a non-"toolUse" stop reason, and when a steer is queued at any
105
+ // boundary (pi polls steering after the turn_end emit), so a boundary this
106
+ // rule accepts can still reach the abort it exists to avoid. The 0.90W hard
107
+ // backstop is the one deliberate mid-loop cut, kept for a loop whose usage must
108
+ // not be allowed to overflow. A boundary with a tool call still active is
109
+ // unsafe for a further reason: a parallel sibling's result has not been
110
+ // delivered, so the abort drops it while the agent is still streaming its
111
+ // response to the context_compress result, surfacing a spurious "This operation
112
+ // was aborted" error in the TUI and tearing the finishing turn. agent_settled
113
+ // alone never drains a run kept alive by continuous steers, so the armed
114
+ // compaction stalls behind live work while usage keeps climbing, hence the
115
+ // turn_end boundary. The arm side asks the same question, from the one arm site
116
+ // that can see a boundary: the heartbeat's usage-driven auto-arm runs inside a
117
+ // turn_end dispatch and hands that event in, so a live-loop boundary below the
118
+ // 0.90W backstop leaves its entry queued for the next loop-done boundary (pinned
119
+ // end to end by test/runtime/heartbeat/test-turn-boundary-arm-wiring.mjs).
120
+ // Arming inside a tool call has no boundary to read, and its own call id parks
121
+ // the queue there anyway.
104
122
  //
105
123
  // Active work is tracked by TOOL CALL ID, never by a count: tool_execution_start
106
124
  // registers an id, tool_execution_end deletes its own, and a turn boundary
@@ -317,6 +335,20 @@ export function scheduleDegradedCompact(trigger: () => void): void {
317
335
  setTimeout(trigger, DEGRADED_DEFERRAL_MS);
318
336
  }
319
337
 
338
+ /**
339
+ * Whether an arm may fire its own entry. Firing a compact aborts whatever run
340
+ * is live, so an arm that happens inside a turn boundary dispatch hands that
341
+ * boundary in and is judged by the same rule the TURN_END handler applies
342
+ * (the rule lives in lib/events/turn-boundary.ts; it owns the "no boundary
343
+ * known" case, which fires). Every other arm site runs inside a tool call,
344
+ * where there is no boundary to read and the prior contract stands: the arm
345
+ * drains unless a batch is still active (the arm's own call id keeps it
346
+ * parked, so it does not fire there).
347
+ */
348
+ function armMayDrainNow(boundary: TurnBoundary | undefined, ctx: any): boolean {
349
+ return turnBoundaryMayAbortRun(boundary, () => captureUsage(ctx));
350
+ }
351
+
320
352
  /**
321
353
  * Arm a one-shot compaction at a safe boundary (next turn end or run settle).
322
354
  * Returns "pending-exists" when an LLM-class compaction is already armed: the
@@ -335,6 +367,7 @@ function scheduleCompactionAtBatchSettle(
335
367
  meta: { kind: PendingCompaction["kind"]; cutEntryId?: string; firstKeptEntryId?: string } = {
336
368
  kind: "compact",
337
369
  },
370
+ boundary?: TurnBoundary,
338
371
  ): "queued" | "pending-exists" {
339
372
  const trigger = () => {
340
373
  try {
@@ -368,7 +401,13 @@ function scheduleCompactionAtBatchSettle(
368
401
  ` (cap ${PENDING_COMPACT_QUEUE_MAX}); the first-armed entry still fires.`,
369
402
  );
370
403
  }
371
- firePendingCompactionsIfSettled(ctx, st);
404
+ // Arming fires only where that is safe. Firing a compaction aborts whatever
405
+ // run is live, and an arm site is not in a position to tell: the tool path
406
+ // arms inside its own call (whose id parks the queue anyway), and the
407
+ // heartbeat's arm runs inside a turn_end dispatch, where whether the loop
408
+ // still has work to dispatch is a fact the event carries. That caller hands
409
+ // its boundary in; every other caller keeps the prior contract.
410
+ if (armMayDrainNow(boundary, ctx)) firePendingCompactionsIfSettled(ctx, st);
372
411
  return "queued";
373
412
  } catch {
374
413
  // Degraded path: batch-settle bookkeeping is unavailable (extension state
@@ -397,6 +436,7 @@ function armSlideCompaction(
397
436
  pi: ExtensionAPI,
398
437
  ctx: any,
399
438
  keepCount: number,
439
+ boundary?: TurnBoundary,
400
440
  ): { cutEntryId: string; keptCount: number; droppedCount: number } | null {
401
441
  const sm = ctx.sessionManager;
402
442
  if (!sm || typeof sm.getBranch !== "function") return null;
@@ -453,6 +493,7 @@ function armSlideCompaction(
453
493
  ctx,
454
494
  { customInstructions: "SLIDE_MODE:true" },
455
495
  { kind: "slide", cutEntryId },
496
+ boundary,
456
497
  );
457
498
  return { cutEntryId, keptCount, droppedCount };
458
499
  }
@@ -488,11 +529,17 @@ function getAutoArmState(ctx: any): AutoArmState {
488
529
  * settle, never mid-run.
489
530
  *
490
531
  * Called from the heartbeat tick — the only cadence that runs while a session
491
- * works. Returns what happened and why, so the caller can log it.
532
+ * works, and the only arm site that runs inside a turn boundary dispatch. It
533
+ * hands that boundary in, so the arm is judged by the same rule the drain
534
+ * applies: a "toolUse" boundary below the 0.90W backstop leaves the entry
535
+ * queued for the next loop-done boundary instead of aborting the request the
536
+ * loop is about to dispatch. Returns what happened and why, so the caller can
537
+ * log it.
492
538
  */
493
539
  export function maybeArmAutoCompaction(
494
540
  pi: ExtensionAPI,
495
541
  ctx: any,
542
+ boundary?: TurnBoundary,
496
543
  ): { armed: boolean; reason: string } {
497
544
  const config: ResolvedAutoCompactConfig = getAutoCompactConfig();
498
545
  if (!config.enabled) return { armed: false, reason: "disabled" };
@@ -513,7 +560,7 @@ export function maybeArmAutoCompaction(
513
560
  }
514
561
  const branchLength = ctx?.sessionManager?.getBranch?.().length ?? 0;
515
562
  const keep = resolveAutoSlideKeep(branchLength);
516
- const armed = armSlideCompaction(pi, ctx, keep);
563
+ const armed = armSlideCompaction(pi, ctx, keep, boundary);
517
564
  st.aboveCount = 0;
518
565
  if (!armed) return { armed: false, reason: "branch-too-short" };
519
566
  log.info(
@@ -607,10 +654,10 @@ export const contextCompressSchema = Type.Object({
607
654
  export function registerContextCompress(pi: ExtensionAPI): void {
608
655
  // Batch-settle scheduling: track the ACTIVE tool calls of every running
609
656
  // batch by call id and fire the deferred compaction only at a safe boundary
610
- // (turn_end or agent_settled). Never mid-batch (would drop a parallel
611
- // sibling's result) and never at tool_execution_end (the run is still live
612
- // there — an aborting compact would surface a spurious "This operation was
613
- // aborted" error and tear the finishing turn).
657
+ // (a loop-done turn_end or agent_settled). Never mid-batch (would drop a
658
+ // parallel sibling's result) and never at tool_execution_end (the run is
659
+ // still live there — an aborting compact would surface a spurious "This
660
+ // operation was aborted" error and tear the finishing turn).
614
661
  onEvent(pi, TOOL_EXECUTION_START, (event: any, ctx: any) => {
615
662
  try {
616
663
  const st = getBatchSettleState(ctx);
@@ -637,17 +684,22 @@ export function registerContextCompress(pi: ExtensionAPI): void {
637
684
  }
638
685
  });
639
686
  onEvent(pi, TURN_END, (event: any, ctx: any) => {
640
- // Between-turns boundary: the turn's assistant message and tool results
641
- // are already persisted when turn_end fires, so the abort inside
642
- // ctx.compact() cannot tear a live turn or drop a sibling result. The
643
- // persisted results carry the toolCallIds of the calls they answer, which
644
- // releases a call whose tool_execution_end never arrived; an active id
645
- // with NO result in this event is kept (it may still be executing, and a
646
- // count could not tell which sibling completed). A run kept alive by
687
+ // Turn boundary. The turn's assistant message and tool results are already
688
+ // persisted here, but that alone does not make the run safe to abort: a
689
+ // "toolUse" stop means the loop dispatches its next request the moment the
690
+ // handlers return, so the deferred ctx.compact() tears that request
691
+ // mid-stream and the compaction replaces the work it was about to do (the
692
+ // rule is lib/events/turn-boundary.ts, and its residual section lists the
693
+ // two cases a loop-done stop reason does not rule out). Only a loop-done
694
+ // boundary drains, except at the 0.90W hard backstop, which exists
695
+ // precisely to stop a runaway live loop before it overflows.
696
+ //
697
+ // The persisted results carry the toolCallIds of the calls they answer,
698
+ // which releases a call whose tool_execution_end never arrived; an active
699
+ // id with NO result in this event is kept (it may still be executing, and
700
+ // a count could not tell which sibling completed). A run kept alive by
647
701
  // continuous steers never reaches agent_settled — this boundary keeps an
648
- // armed compaction from stalling behind live work. A steer queued at this
649
- // moment survives: the loop appends it before the aborted response stream,
650
- // and its turn runs on the post-compaction resume. Ids are never blanket-
702
+ // armed compaction from stalling behind live work. Ids are never blanket-
651
703
  // reset here (only agent_settled owns that backstop, which is what keeps a
652
704
  // missed end event from parking the queue forever).
653
705
  try {
@@ -655,6 +707,7 @@ export function registerContextCompress(pi: ExtensionAPI): void {
655
707
  releaseSettledToolCallIds(st, event?.toolResults);
656
708
  if (st.pending.length === 0 || st.activeToolCallIds.size > 0) return;
657
709
  if (isCompactionActive(ctx)) return;
710
+ if (!turnBoundaryMayAbortRun(event, () => captureUsage(ctx))) return;
658
711
  firePendingCompactionsIfSettled(ctx, st);
659
712
  } catch {
660
713
  // A failed boundary drain must not tear the turn_end event pipeline;
@@ -122,6 +122,13 @@ export function scratchpadDir(): string {
122
122
  return path.join(braidDir(), "scratchpads");
123
123
  }
124
124
 
125
+ /** Journal root (one dir per project slug), resolved on every call. Durable
126
+ * write state: the append-only record of what a seat decided, which a
127
+ * replacement seat reads before acting, so BRAID_HOME pinning must move it. */
128
+ export function journalDir(): string {
129
+ return path.join(braidDir(), "journal");
130
+ }
131
+
125
132
  /** Healthcheck scratch and report root, resolved on every call. Durable-write
126
133
  * state: the probes and the saved report must follow a pinned BRAID_HOME. */
127
134
  export function healthcheckDir(): string {
@@ -305,6 +312,19 @@ export function tagsSlugFromCwd(cwd?: string): string {
305
312
  return cwd.split("/").filter(Boolean).join("-") || "default";
306
313
  }
307
314
 
315
+ /**
316
+ * Append-only journal file for one project: `<journal root>/<slug>/journal.jsonl`.
317
+ *
318
+ * The slug is the scratchpad's `projectSlugFromCwd` form, passed in rather than
319
+ * re-derived here: the writer, the checker and the session-start pointer must
320
+ * name ONE file, and a second slug rule is how the tag file split into a writer
321
+ * and a reader that could never meet (see tagsSlugFromCwd). Resolved on every
322
+ * call, so a pinned BRAID_HOME writes and reads its own journal.
323
+ */
324
+ export function journalFilePath(projectSlug: string): string {
325
+ return path.join(journalDir(), projectSlug, "journal.jsonl");
326
+ }
327
+
308
328
  /**
309
329
  * Get the lossless flush file path for one offloaded tool result.
310
330
  * Layout: <offload dir>/<sessionId>/<toolName>-<toolCallId>.txt.