@coreplane/switchboard 1.249.1 → 1.251.0

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.
Files changed (57) hide show
  1. package/dist/assets/config/config.example.yaml +46 -4
  2. package/dist/assets/deploy/cloudflare-memory/worker.ts +260 -1
  3. package/dist/assets/deploy/cloudflare-resident/Dockerfile +12 -3
  4. package/dist/assets/deploy/cloudflare-sandbox/Dockerfile +7 -0
  5. package/dist/assets/package-lock.json +3 -3
  6. package/dist/assets/package.json +3 -2
  7. package/dist/assets/project.json +4 -0
  8. package/dist/assets/source.json +3 -3
  9. package/dist/assets/src/agents/registry.ts +1 -1
  10. package/dist/assets/src/core/authz/policy.ts +10 -0
  11. package/dist/assets/src/core/authz/resource.ts +6 -2
  12. package/dist/assets/src/core/authz/types.ts +2 -0
  13. package/dist/assets/src/core/budgets.ts +13 -0
  14. package/dist/assets/src/core/coordinator/contract.ts +120 -0
  15. package/dist/assets/src/core/coordinator/driver.ts +17 -3
  16. package/dist/assets/src/core/modelCard.ts +19 -3
  17. package/dist/assets/src/core/prDescriptionTypes.ts +35 -0
  18. package/dist/assets/src/core/provider.ts +57 -3
  19. package/dist/assets/src/core/refusal.ts +14 -0
  20. package/dist/assets/src/core/runEvents.ts +49 -2
  21. package/dist/assets/src/core/runFriction.ts +2 -0
  22. package/dist/assets/src/core/runLedger/decisions.ts +9 -1
  23. package/dist/assets/src/core/runLedger/sessionLog.ts +10 -0
  24. package/dist/assets/src/core/runLedger/transcript.ts +13 -3
  25. package/dist/assets/src/core/runLedger/types.ts +72 -6
  26. package/dist/assets/src/core/runRecord.ts +34 -0
  27. package/dist/assets/src/core/ship/contract.ts +17 -4
  28. package/dist/assets/src/core/ship/coordinator.ts +70 -4
  29. package/dist/assets/src/core/trace/attrs.ts +13 -2
  30. package/dist/assets/src/core/trace/workerTrace.ts +3 -2
  31. package/dist/assets/src/core/types.ts +3 -0
  32. package/dist/assets/web/dist/.vite/manifest.json +58 -52
  33. package/dist/assets/web/dist/assets/DeliveryPage-DUXd-Sl-.js +1 -0
  34. package/dist/assets/web/dist/assets/HomePage-mSiqEEcN.js +2 -0
  35. package/dist/assets/web/dist/assets/{PendingTurnRow-BuRre8it.js → PendingTurnRow-DDhMhrI7.js} +1 -1
  36. package/dist/assets/web/dist/assets/{ResidentDetailPage-BIUXyz6K.js → ResidentDetailPage-BnEoOnGQ.js} +1 -1
  37. package/dist/assets/web/dist/assets/{ResidentsIndexPage-BZymgSAb.js → ResidentsIndexPage-Dxpgf-l-.js} +1 -1
  38. package/dist/assets/web/dist/assets/RunFoldRow-CSo4-vld.js +1 -0
  39. package/dist/assets/web/dist/assets/{RunRoutePage-bgkjkA0p.js → RunRoutePage-9klVWhSF.js} +6 -6
  40. package/dist/assets/web/dist/assets/{RunsIndexPage-8S944AzB.js → RunsIndexPage-BplMIgaw.js} +1 -1
  41. package/dist/assets/web/dist/assets/{ScheduledPage-8bBtG9y3.js → ScheduledPage-B_GgeJrb.js} +1 -1
  42. package/dist/assets/web/dist/assets/{SettingsPage-DQeNvfaV.js → SettingsPage-BXX4R113.js} +1 -1
  43. package/dist/assets/web/dist/assets/{StatusDot-BPE5syBa.js → StatusDot-BOaw8le9.js} +1 -1
  44. package/dist/assets/web/dist/assets/{Tooltip-DkoeZfTs.js → Tooltip-DYZZ4l4V.js} +1 -1
  45. package/dist/assets/web/dist/assets/{UnitRoutePage-BUzw--Ii.js → UnitRoutePage-BaSW5Odq.js} +1 -1
  46. package/dist/assets/web/dist/assets/budgets-BvWYKPsY.js +1 -0
  47. package/dist/assets/web/dist/assets/{dist-D11y9ZJ4.js → dist-BCVXeBJ9.js} +1 -1
  48. package/dist/assets/web/dist/assets/indexRow-Bde9OZxG.js +1 -0
  49. package/dist/assets/web/dist/assets/{main-B6LcgNM6.js → main-Dkcbtu3u.js} +2 -2
  50. package/dist/assets/web/dist/assets/sseReplay-DPwdsaok.js +9 -0
  51. package/dist/cli.js +4594 -2743
  52. package/package.json +1 -1
  53. package/dist/assets/web/dist/assets/DeliveryPage-DF4aQypG.js +0 -1
  54. package/dist/assets/web/dist/assets/HomePage-BpQRky8B.js +0 -2
  55. package/dist/assets/web/dist/assets/RunFoldRow-3m4CPRI4.js +0 -1
  56. package/dist/assets/web/dist/assets/indexRow-BmK74Vp1.js +0 -1
  57. package/dist/assets/web/dist/assets/sseReplay-DXC7kGbN.js +0 -9
@@ -114,6 +114,8 @@ export type Resource =
114
114
  readonly visibility?: ChannelVisibility;
115
115
  }
116
116
  | { readonly type: "config-scope"; readonly kind: "user"; readonly id: string }
117
+ /** A thread's runtime scope (`config set thread`, routing-and-config item 27); `id` is the thread key. */
118
+ | { readonly type: "config-scope"; readonly kind: "thread"; readonly id: string }
117
119
  | { readonly type: "config-scope"; readonly kind: "org" }
118
120
  | { readonly type: "agent"; readonly name: string }
119
121
  /** List-shaped actions with no single resource (`runs.list`, `friction.report`). */
@@ -85,6 +85,19 @@ export const DRAIN = {
85
85
  defaultMinutes: 60,
86
86
  } as const;
87
87
 
88
+ /** How long an intake receipt row is kept on the run history object
89
+ * (docs/reference/specs/run-history.md item 59; docs/decisions/0058): the larger of one day
90
+ * and the reconnect catch-up window the write named plus the drain deadline
91
+ * (`DRAIN.maxMinutes`, the longest a fleet drain may last), so a catch-up
92
+ * that runs after the longest allowed drain still reads the verdict instead
93
+ * of deciding the reply again. The window is clamped to a month so a
94
+ * misconfigured writer cannot make retention unbounded. */
95
+ export const INTAKE_WINDOW_MAX_MS = 30 * DAY_MS;
96
+ export function intakeReceiptRetentionMs(catchUpWindowMs: number): number {
97
+ const window = Math.min(Math.max(0, catchUpWindowMs), INTAKE_WINDOW_MAX_MS);
98
+ return Math.max(DAY_MS, window + minutesToMs(DRAIN.maxMinutes));
99
+ }
100
+
88
101
  /** The presets that run the tool loop, and the one pipeline preset. */
89
102
  export const LOOP_PRESETS = ["general", "coding", "review", "research", "explore", "conductor"] as const;
90
103
  export type LoopPreset = (typeof LOOP_PRESETS)[number];
@@ -81,6 +81,124 @@ export function runFinishedEventType(runId: string): string {
81
81
  return `${RUN_FINISHED_EVENT_PREFIX}${runId}`;
82
82
  }
83
83
 
84
+ /** A message into a thread an unfinished unit owns (record 0051's reply-as-event rule): one row
85
+ * of the unit's event list, appended by the dispatcher, folded into the
86
+ * unit's next coding spawn — or run as one fresh turn at the unit's end.
87
+ * `mode` is a receipt of the owner's state at append (record 0051): `steer` into a
88
+ * live run or between rounds, `wake` into an idle owner, `interrupt` into an
89
+ * owner idle after a stop — never a switch the sender fills. */
90
+ export type ThreadEventMode = "steer" | "wake" | "interrupt";
91
+
92
+ export interface ThreadEventAttachment {
93
+ mediaType: string;
94
+ data: string;
95
+ name?: string;
96
+ }
97
+
98
+ export interface ThreadEvent {
99
+ /** Assigned by the store's append, in arrival order, per unit. */
100
+ seq: number;
101
+ /** The channel's message id, when the platform gave one. */
102
+ id?: string;
103
+ /** The sender (platform-namespaced) and the display name the channel knew. */
104
+ sender: string;
105
+ senderName?: string;
106
+ text: string;
107
+ attachments?: ThreadEventAttachment[];
108
+ /** How many attachments were dropped because the event was over the cap. */
109
+ attachmentsDropped?: number;
110
+ /** How many characters were cut from the end of `text` because the row was
111
+ * still over the cap without any attachment. */
112
+ textDropped?: number;
113
+ mode: ThreadEventMode;
114
+ at: number;
115
+ /** The spawn step or run that consumed the event; absent while unconsumed. */
116
+ consumedBy?: string;
117
+ }
118
+
119
+ /** The most one durable event row may weigh, serialized — the durable inbox's
120
+ * own cap (`DURABLE_INBOX_MAX_BYTES`), restated here because this contract is
121
+ * node-free and the state Worker enforces it too. */
122
+ export const THREAD_EVENT_MAX_BYTES = 400 * 1024;
123
+
124
+ const serializedBytes = (v: unknown): number => new TextEncoder().encode(JSON.stringify(v)).length;
125
+
126
+ /** The event under the cap: attachments ride when the serialized row fits,
127
+ * else all of them are dropped and the row says how many — all or nothing,
128
+ * like the durable inbox (a partial carry would hand the model some of the
129
+ * sender's attachments as if they were all of them). A row still over the cap
130
+ * with no attachment left — a text alone past 400 KiB, possible through the
131
+ * ingress body — has its text cut from the end until the row fits, and the
132
+ * row says how many characters went: no event escapes the constant's promise.
133
+ * `TextEncoder`, not `Buffer`: both stores — the bot's and the state Worker's —
134
+ * apply it. */
135
+ export function capThreadEvent<T extends Omit<ThreadEvent, "seq"> & { seq?: number }>(
136
+ event: T,
137
+ ): T & Pick<ThreadEvent, "attachmentsDropped" | "textDropped"> {
138
+ if (serializedBytes(event) <= THREAD_EVENT_MAX_BYTES) return event;
139
+ const attachments = event.attachments;
140
+ let capped: T & Pick<ThreadEvent, "attachmentsDropped" | "textDropped"> = event;
141
+ if (attachments && attachments.length > 0) {
142
+ const { attachments: _dropped, ...rest } = event;
143
+ capped = { ...rest, attachmentsDropped: attachments.length } as typeof capped;
144
+ if (serializedBytes(capped) <= THREAD_EVENT_MAX_BYTES) return capped;
145
+ }
146
+ // Text-only overflow: the row's shape is fixed except for the text, so the
147
+ // text is cut — by characters, so a multibyte cut never splits a code point
148
+ // pair the decoder would read as garbage — and shrunk until the bytes fit.
149
+ const text = capped.text;
150
+ const overhead = serializedBytes({ ...capped, text: "", textDropped: text.length });
151
+ let keep = Math.max(0, Math.min(text.length, THREAD_EVENT_MAX_BYTES - overhead));
152
+ for (;;) {
153
+ const cut = { ...capped, text: text.slice(0, keep), textDropped: text.length - keep };
154
+ if (keep === 0 || serializedBytes(cut) <= THREAD_EVENT_MAX_BYTES) return cut;
155
+ keep = Math.floor(keep * 0.9);
156
+ }
157
+ }
158
+
159
+ const isThreadEventAttachment = (v: unknown): v is ThreadEventAttachment =>
160
+ isObject(v) &&
161
+ typeof v.mediaType === "string" &&
162
+ typeof v.data === "string" &&
163
+ (v.name === undefined || typeof v.name === "string");
164
+
165
+ const isThreadEventMode = (v: unknown): v is ThreadEventMode => v === "steer" || v === "wake" || v === "interrupt";
166
+
167
+ /** Structural check on an event row from outside the process. */
168
+ export function isThreadEvent(v: unknown): v is ThreadEvent {
169
+ if (!isObject(v)) return false;
170
+ const r = v;
171
+ if (typeof r.seq !== "number" || !Number.isInteger(r.seq) || r.seq < 1) return false;
172
+ // The fixed fields are bounded too, so the cap's text cut has a floor to
173
+ // land on: a row whose id or names alone weighed the cap could never fit.
174
+ if (!isOptionalText(r.id)) return false;
175
+ if (!isText(r.sender)) return false;
176
+ if (!isOptionalText(r.senderName)) return false;
177
+ if (typeof r.text !== "string") return false;
178
+ if (r.attachments !== undefined && (!Array.isArray(r.attachments) || !r.attachments.every(isThreadEventAttachment)))
179
+ return false;
180
+ if (r.attachmentsDropped !== undefined && !isCount(r.attachmentsDropped)) return false;
181
+ if (r.textDropped !== undefined && !isCount(r.textDropped)) return false;
182
+ if (!isThreadEventMode(r.mode)) return false;
183
+ if (!isFinite(r.at)) return false;
184
+ if (!isOptionalText(r.consumedBy)) return false;
185
+ return true;
186
+ }
187
+
188
+ /** The payload-free nudge the dispatcher sends an instance when a thread
189
+ * event lands on one of its units (record 0051's reply-as-event rule): the relay's
190
+ * alphabet — letters, digits, `_` and `-`, at most 100 characters, a colon
191
+ * refused — so the type is the two ids joined by `-`. The send addresses the
192
+ * instance by id (`workflow.get(id)`), so the type only has to name the unit
193
+ * within it: when the two ids together overflow the cap, the INSTANCE id is
194
+ * clipped and the unit is kept whole, and both ends compute the same string. */
195
+ export const UNIT_NUDGE_EVENT_PREFIX = "unit-nudge-";
196
+ export function unitNudgeEventType(key: { instanceId: string; unit: string }): string {
197
+ const suffix = `-${key.unit}`;
198
+ const room = 100 - UNIT_NUDGE_EVENT_PREFIX.length - suffix.length;
199
+ return `${UNIT_NUDGE_EVENT_PREFIX}${key.instanceId.slice(0, room)}${suffix}`;
200
+ }
201
+
84
202
  /** The event the bot's GitHub check-run intake sends a merge-waiting parent:
85
203
  * the type carries the head sha (hex — inside the platform's alphabet), so a
86
204
  * driver waiting at that head matches its own event and any other head's is
@@ -278,6 +396,8 @@ const MAX_ROUNDS = 200;
278
396
  const isText = (v: unknown, max = MAX_TEXT): v is string => typeof v === "string" && v.length > 0 && v.length <= max;
279
397
  const isOptionalText = (v: unknown): boolean => v === undefined || isText(v);
280
398
  const isFinite = (v: unknown): v is number => typeof v === "number" && Number.isFinite(v);
399
+ /** A count the writers produce: a non-negative integer, never a fraction, a negative or NaN. */
400
+ const isCount = (v: unknown): v is number => typeof v === "number" && Number.isInteger(v) && v >= 0;
281
401
  const isObject = (v: unknown): v is Record<string, unknown> => typeof v === "object" && v !== null;
282
402
  const isPr = (v: unknown): boolean => isObject(v) && isFinite(v.number) && isText(v.url, 2048);
283
403
  const isResume = (v: unknown): boolean =>
@@ -145,7 +145,7 @@ export function readBotAnswer(
145
145
 
146
146
  /** The answers the bot itself calls a passing condition — the step is asked
147
147
  * again, under the policy. Everything else, refusals included, is the machine's. */
148
- const TRANSIENT = new Set(["github_unavailable", "no_channel", "thread_failed", "unit_not_started"]);
148
+ const TRANSIENT = new Set(["github_unavailable", "no_channel", "thread_failed", "unit_not_started", "not_host"]);
149
149
  export function transientRefusal(answer: BotAnswer): string | undefined {
150
150
  const { ok, error, message } = answer.body;
151
151
  if (ok !== false || typeof error !== "string" || !TRANSIENT.has(error)) return undefined;
@@ -172,6 +172,8 @@ interface PlanFacts {
172
172
  grantSource: GrantSource;
173
173
  /** The instance's mark as the plan route answers it: a generated one-unit plan (a `plan` with no `path`). */
174
174
  generated: boolean;
175
+ /** The runs page base the bot answered: the report links a child's write-up to its run page with it. */
176
+ runPageBase?: string;
175
177
  repo: string;
176
178
  base: string;
177
179
  caps: ShipCaps;
@@ -217,6 +219,7 @@ function readPlan(a: BotAnswer): PlanFacts {
217
219
  grantSource:
218
220
  b.grantSource === "run" || b.grantSource === "user" || b.grantSource === "channel" ? b.grantSource : "org",
219
221
  generated: b.generated === true,
222
+ ...(typeof b.runPageBase === "string" && b.runPageBase.length > 0 ? { runPageBase: b.runPageBase } : {}),
220
223
  repo: b.repo,
221
224
  base: b.base,
222
225
  caps: { maxRounds: b.caps.maxRounds, maxMinutes: b.caps.maxMinutes },
@@ -318,7 +321,7 @@ function isCommitChecks(v: unknown): v is { total: number; pending: string[]; fa
318
321
  function prCheckReturn(step: string, a: BotAnswer): StepReturn {
319
322
  const { ok, state, prNumber, url, headSha, sha, mergedAt, at } = a.body;
320
323
  if (ok === true && state === "none") {
321
- const { unrecovered, aheadOfBase } = a.body;
324
+ const { unrecovered, aheadOfBase, prClosed } = a.body;
322
325
  return {
323
326
  type: "pr-check",
324
327
  step,
@@ -328,6 +331,9 @@ function prCheckReturn(step: string, a: BotAnswer): StepReturn {
328
331
  // The branch's commits over the base, when the bot could read them
329
332
  // (agent-ship item 12): zero is the `already_landed` ending's fact.
330
333
  ...(typeof aheadOfBase === "number" ? { aheadOfBase } : {}),
334
+ // The followed pull request verified closed unmerged (issue 1799):
335
+ // the machine must not brief a review round on it.
336
+ ...(prClosed === true ? { prClosed: true } : {}),
331
337
  },
332
338
  at,
333
339
  };
@@ -453,12 +459,19 @@ async function perform(
453
459
  case "pr-check":
454
460
  // `recover` rides only after a dead coding child: the bot opens the pull
455
461
  // request from the pushed branch itself instead of answering `none`.
462
+ // `pr` is the machine's adopted pull request (issue 1799): the bot
463
+ // follows it when nothing heads the unit's branch and answers its live
464
+ // state instead of `none` over a minutes-old record fact.
456
465
  return prCheckReturn(
457
466
  action.step,
458
467
  answerOf(
459
468
  "pr-check",
460
469
  await step.do(action.step, STEP_CONFIG, () =>
461
- call(bot, "pr-check", { ...tag, ...(action.recover !== undefined ? { recover: action.recover } : {}) }),
470
+ call(bot, "pr-check", {
471
+ ...tag,
472
+ ...(action.recover !== undefined ? { recover: action.recover } : {}),
473
+ ...(action.pr !== undefined ? { pr: action.pr } : {}),
474
+ }),
462
475
  ),
463
476
  ),
464
477
  );
@@ -537,6 +550,7 @@ async function runUnit(
537
550
  grant: plan.grant,
538
551
  grantSource: plan.grantSource,
539
552
  generated: plan.generated,
553
+ ...(plan.runPageBase !== undefined ? { runPageBase: plan.runPageBase } : {}),
540
554
  ...(resume !== undefined ? { resume } : {}),
541
555
  ...(lastPush !== undefined ? { lastPush } : {}),
542
556
  ...(session !== undefined ? { session } : {}),
@@ -10,7 +10,7 @@
10
10
  // catalog lives in ./modelRegistry.ts.
11
11
 
12
12
  import { EFFORT_LEVELS, type Effort } from "../effort.js";
13
- import { vendorOf, wireOf, type ProviderConfig, type Wire } from "./provider.js";
13
+ import { billerHarnessProvider, vendorOf, wireOf, type ProviderConfig, type Wire } from "./provider.js";
14
14
  import type { RegistryCard } from "./modelRegistry.js";
15
15
 
16
16
  /** Which layer named a field: the operator's block, the registry card, or the
@@ -335,13 +335,29 @@ export function decideControls(card: ModelCard, asked: AskedControls): ControlDe
335
335
  why: windowVouched ? "" : `no layer names the window; compacting at ${card.window}`,
336
336
  });
337
337
 
338
- const cacheNative = card.cache !== "unknown";
338
+ // A `markers` rule needs a harness-side write that places the markers
339
+ // (record 0052's amendment: the harness write names the biller's own
340
+ // provider). The Anthropic wire's own packages place per-block breakpoints;
341
+ // on the chat wire a biller the table names caches the aggregator's own way
342
+ // — the pinned OpenCode binary exempts the openrouter route from per-block
343
+ // placement, so the write carries OpenRouter's top-level
344
+ // `cache_control: { type: "ephemeral" }` via `settings.extraBody` (measured
345
+ // in `opencode/testing/realDriver.test.ts`), and pi's compat sends the
346
+ // per-block markers. A biller served generically degrades, never a silent
347
+ // `native`.
348
+ const markersUnplaced =
349
+ card.cache === "markers" && card.wire === "openai-chat" && billerHarnessProvider(card.block) === undefined;
350
+ const cacheNative = card.cache !== "unknown" && !markersUnplaced;
339
351
  decisions.push({
340
352
  control: "cache",
341
353
  outcome: cacheNative ? "native" : "degraded",
342
354
  applied: card.cache,
343
355
  vouched: cacheNative,
344
- why: cacheNative ? "" : `no layer names ${card.model}'s cache rule`,
356
+ why: cacheNative
357
+ ? ""
358
+ : markersUnplaced
359
+ ? `no harness-side provider vouches for the "${card.block}" biller's cache markers; the rule goes out unvouched`
360
+ : `no layer names ${card.model}'s cache rule`,
345
361
  });
346
362
 
347
363
  return decisions;
@@ -37,6 +37,41 @@ export interface RenderedPointer extends Pointer {
37
37
  anchor: RenderedPrAnchor;
38
38
  }
39
39
 
40
+ /** One reason a submitted description was refused, as the refusal and the
41
+ * `description_refused` run note carry it: the zod path (`risk`,
42
+ * `pointers.0.text`) and the schema's message; on a cap issue also the
43
+ * characters to remove (visible ones — raw for the title) and the longest
44
+ * prefix of the field that fits, cut at a word boundary (`fitToCap`), absent
45
+ * when no word of it fits. */
46
+ export interface DescriptionIssue {
47
+ path: string;
48
+ message: string;
49
+ remove?: number;
50
+ prefix?: string;
51
+ }
52
+
53
+ /** The refused object as the `description_refused` note records it: JSON,
54
+ * nested at most four objects deep, and nothing else. The record crosses the
55
+ * run store's RPC boundary, whose typing walks every field: a field of
56
+ * `unknown` types the whole record `never`, and a recursive alias is
57
+ * "excessively deep" to it, so the nesting is spelled out level by level
58
+ * (as `RouteInputValue` is). A description is three deep at most —
59
+ * `pointers[i].anchor` — so one level is to spare; `recordedJson` stores a
60
+ * deeper value as its JSON text. */
61
+ export type RecordedJsonLeaf = string | number | boolean | null;
62
+ export type RecordedJsonObject1 = { readonly [key: string]: RecordedJsonLeaf | ReadonlyArray<RecordedJsonLeaf> };
63
+ export type RecordedJson1 =
64
+ RecordedJsonLeaf | RecordedJsonObject1 | ReadonlyArray<RecordedJsonLeaf | RecordedJsonObject1>;
65
+ export type RecordedJsonObject2 = { readonly [key: string]: RecordedJson1 };
66
+ export type RecordedJson2 =
67
+ RecordedJsonLeaf | RecordedJsonObject2 | ReadonlyArray<RecordedJsonLeaf | RecordedJsonObject2>;
68
+ export type RecordedJsonObject3 = { readonly [key: string]: RecordedJson2 };
69
+ export type RecordedJson3 =
70
+ RecordedJsonLeaf | RecordedJsonObject3 | ReadonlyArray<RecordedJsonLeaf | RecordedJsonObject3>;
71
+ export type RecordedJsonObject4 = { readonly [key: string]: RecordedJson3 };
72
+ export type RecordedJson =
73
+ RecordedJsonLeaf | RecordedJsonObject4 | ReadonlyArray<RecordedJsonLeaf | RecordedJsonObject4>;
74
+
40
75
  /** The PR description as data (docs/decisions/0050): the map above the fold
41
76
  * (tldr, why, pointers, feedbackWanted, risk, verified), every field capped
42
77
  * so the map's size does not grow with the diff, and the collapsed half
@@ -95,9 +95,8 @@ export interface Provider {
95
95
 
96
96
  /** The three wire shapes a provider block may declare (record 0052):
97
97
  * Anthropic's Messages API, OpenAI's Chat Completions and OpenAI's Responses
98
- * API. `openai-responses` is a declaration only in this slice — the proxy
99
- * still serves two routes (a later slice adds the third) — and a block that names it
100
- * keeps routing through the compatible shape until then. */
98
+ * API. Each is a proxy route of its own (`PROXY_PATHS`): an `openai-responses`
99
+ * block runs on `/v1/responses`, pinned and metered like the other two. */
101
100
  export const WIRES = ["anthropic-messages", "openai-chat", "openai-responses"] as const;
102
101
  export type Wire = (typeof WIRES)[number];
103
102
 
@@ -123,6 +122,12 @@ export interface ProviderModelOverride {
123
122
  cache?: "automatic" | "markers" | "none" | "unknown";
124
123
  /** USD per million tokens, by kind. */
125
124
  price?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number };
125
+ /** The answer shapes the model can produce for a forced one-call turn (the
126
+ * router's, intake's): `tool` — a forced tool call — and `text` — the
127
+ * one-JSON-object text contract. Absent, both are assumed; an empty list
128
+ * declares neither, and a load whose intake would classify on such a card
129
+ * is refused by name (routing-and-config item 27). */
130
+ answers?: ("tool" | "text")[];
126
131
  }
127
132
 
128
133
  export interface ProviderConfig {
@@ -204,6 +209,55 @@ export function vendorOf(
204
209
  return { block, model, vendor: block, vendorId: model, vendorSource: "block" };
205
210
  }
206
211
 
212
+ /** The harness-side provider a block's biller implies (record 0052's
213
+ * amendment: the harness write names the biller's own provider, never a
214
+ * generic alias). Keyed by the biller — the block's name — for the billers
215
+ * whose protocol a harness bundles a provider for. The wires with a package
216
+ * of their own (`anthropic-messages` → `@ai-sdk/anthropic`,
217
+ * `openai-responses` → `@ai-sdk/openai`) need no entry: the wire names the
218
+ * package. A chat-wire biller not here is served generically
219
+ * (`@ai-sdk/openai-compatible`), under which a `markers` cache rule cannot be
220
+ * vouched for (`decideControls`): the generic provider places no cache
221
+ * breakpoints. */
222
+ export interface BillerHarnessProvider {
223
+ /** The AI SDK package OpenCode's configuration names for the biller
224
+ * (`openCodeProviderPackage` adds the `aisdk:` prefix). */
225
+ openCodePackage: string;
226
+ /** The compat words pi keys on the biller's identity, copied once from pi's
227
+ * own completions detection of that biller and never inferred from a URL at
228
+ * run time: through the proxy pi sees the bot's URL, so `piModelsJson` must
229
+ * say the words the biller's own base URL would have made pi detect. */
230
+ piCompat: {
231
+ /** How the wire spells reasoning: `reasoning: { effort }` under
232
+ * `"openrouter"`, never the completions shape's flat `reasoning_effort`. */
233
+ thinkingFormat: string;
234
+ /** How a session id would ride the headers, were affinity ever turned on. */
235
+ sessionAffinityFormat: string;
236
+ /** The vendor-qualified id prefixes the biller grants the developer role:
237
+ * any other id is told `supportsDeveloperRole: false`, as pi's own
238
+ * detection would say against the biller directly. */
239
+ developerRoleIdPrefixes: readonly string[];
240
+ };
241
+ }
242
+
243
+ /** The biller-to-provider table: one row per biller a harness speaks natively
244
+ * — OpenCode's package (U44) and pi's compat words (U45) side by side. */
245
+ export const BILLER_HARNESS_PROVIDERS: Readonly<Record<string, BillerHarnessProvider>> = {
246
+ openrouter: {
247
+ openCodePackage: "@openrouter/ai-sdk-provider",
248
+ piCompat: {
249
+ thinkingFormat: "openrouter",
250
+ sessionAffinityFormat: "openrouter",
251
+ developerRoleIdPrefixes: ["anthropic/", "openai/"],
252
+ },
253
+ },
254
+ };
255
+
256
+ /** The biller's harness-side provider, or undefined when it is served generically. */
257
+ export function billerHarnessProvider(biller: string | undefined): BillerHarnessProvider | undefined {
258
+ return biller === undefined ? undefined : BILLER_HARNESS_PROVIDERS[biller];
259
+ }
260
+
207
261
  /** The env var Anthropic's own SDK reads when an `anthropic` provider block names none. */
208
262
  export const ANTHROPIC_API_KEY_ENV = "ANTHROPIC_API_KEY";
209
263
 
@@ -22,6 +22,17 @@ export interface Guess {
22
22
  evidence: string;
23
23
  }
24
24
 
25
+ /** A command handler's guess hint (record 0054): the corrected chat form and
26
+ * the evidence, without the full proposal — the proposal is synthesised at
27
+ * the invocation point from the original message and this hint's `line`. The
28
+ * handler carries this lighter type on `CommandError.guess`; the adapter
29
+ * (`invokeChatCommand`, `answerCommand`) builds the `Guess` that
30
+ * `renderRefusal` expects. */
31
+ export interface CommandGuessHint {
32
+ line: string;
33
+ evidence: string;
34
+ }
35
+
25
36
  /** One cause per code — the closed table `causeOf` reads. The codes are the
26
37
  * inventory's: the dispatch gates' `refuse(code)` outcomes, the click's four
27
38
  * `confirmation_*` codes, the two `elsewhere_*` reasons that never reached a
@@ -45,6 +56,8 @@ const CAUSE_OF = {
45
56
  which_branch: "request",
46
57
  workspace_lost: "system",
47
58
  ship_budget: "request",
59
+ // one pipeline per thread (record 0060): the host key's claim answered thread-live
60
+ ship_thread_live: "request",
48
61
  setup_failed: "system",
49
62
  // the click on a confirmation (confirm.ts)
50
63
  confirmation_used: "request",
@@ -95,6 +108,7 @@ const CAUSE_OF = {
95
108
  directive_budget: "request",
96
109
  directive_severity: "request",
97
110
  directive_renewals: "request",
111
+ directive_verbosity: "request",
98
112
  provider_unknown: "request",
99
113
  // the model card's refusal (record 0052, model-proxy item 11): a control the
100
114
  // resolved card does not take, named before any card or model call
@@ -1,7 +1,7 @@
1
1
  // Types only, and from the zod-free module deliberately: this file is part of
2
2
  // the node-free contract the memory Worker and web app compile with their own
3
3
  // tsconfigs — importing prDescription.ts would drag zod into those graphs.
4
- import type { PrDescription, RenderedPointer } from "./prDescriptionTypes.js";
4
+ import type { DescriptionIssue, PrDescription, RecordedJson, RenderedPointer } from "./prDescriptionTypes.js";
5
5
  import type { HarnessScope } from "./harness/scope.js";
6
6
  import type { ModelCard } from "./modelCard.js";
7
7
 
@@ -146,6 +146,15 @@ export type RunNoteKind =
146
146
  * bounded extra model turn to submit it (docs/reference/specs/pr-description.md
147
147
  * item 5). Published by the dispatcher before that turn. */
148
148
  | "description_turn"
149
+ /** `submit_pr_description` refused the object (docs/reference/specs/pr-description.md
150
+ * item 5): the summary counts the fields over their cap and names them
151
+ * (or the issues, when none is a cap), `description` is the object as
152
+ * submitted — redacted like the accepted `pr_description` event's, and not
153
+ * necessarily a valid `PrDescription` — and `issues` the refusal's list
154
+ * with each cap's count to remove and the prefix that fits, so the record
155
+ * says what the model changed between one submit and the next. Published
156
+ * by the tool, before it answers. */
157
+ | "description_refused"
149
158
  /** A review run's loop ended on a pull request without `submit_verdict`, and
150
159
  * the same run is being given one bounded extra model turn to call it
151
160
  * (docs/reference/specs/agent-review.md item 5; verdictTurn.ts). Published by
@@ -273,7 +282,15 @@ export type RunNoteKind =
273
282
  /** The native loop's stuck-loop guard: the same tool call failed identically
274
283
  * six times in a row and the run was forced into its write-up. Written by
275
284
  * no loop since that loop's deletion; a record from before it may carry it. */
276
- | "stuck_loop";
285
+ | "stuck_loop"
286
+ /** OpenCode's reject cascade ended the execution `interrupted` after the bot
287
+ * refused one of a step's two (or more) calls — the binary declines every
288
+ * other pending ask at a reject and ends the step `session.step.failed
289
+ * {aborted}`, the execution ending `session.execution.interrupted` — and the
290
+ * model never read the refusal. The loop re-prompts with the refusal so the
291
+ * model can continue, exactly as it does after a single refusal. Published
292
+ * by the OpenCode bridge. */
293
+ | "decline_cascade";
277
294
 
278
295
  /** Every `RunNoteKind`, as a value (a reader that filters notes by kind uses
279
296
  * this; adding a kind to the union without adding it here is a type error). */
@@ -298,6 +315,7 @@ export const RUN_NOTE_KINDS = [
298
315
  "seed",
299
316
  "redispatch",
300
317
  "description_turn",
318
+ "description_refused",
301
319
  "verdict_turn",
302
320
  "cold_sandbox",
303
321
  "ledger_untracked",
@@ -316,6 +334,7 @@ export const RUN_NOTE_KINDS = [
316
334
  "directory_reached",
317
335
  "budget_salvage",
318
336
  "stuck_loop",
337
+ "decline_cascade",
319
338
  ] as const satisfies readonly RunNoteKind[];
320
339
  type _EveryKindListed = [RunNoteKind] extends [(typeof RUN_NOTE_KINDS)[number]] ? true : never;
321
340
  const _everyKindListed: _EveryKindListed = true;
@@ -566,6 +585,12 @@ export type RunEvent =
566
585
  * interval the dropped setup records covered — a `not recorded` loss. */
567
586
  from?: number;
568
587
  to?: number;
588
+ /** On a `description_refused` note only: the refused object as submitted
589
+ * (JSON, not necessarily a valid `PrDescription`; `RecordedJson` says
590
+ * why it is typed level by level), every string leaf redacted, and the
591
+ * issues the refusal named. */
592
+ description?: RecordedJson;
593
+ issues?: DescriptionIssue[];
569
594
  spanId?: string;
570
595
  seq?: number;
571
596
  at?: number;
@@ -600,6 +625,13 @@ export type RunEvent =
600
625
  * run sent rather than a person (a parent's `send_to_run`), that run's
601
626
  * id (docs/reference/specs/agent-conductor.md item 8). */
602
627
  source?: { url?: string; channel?: string; user?: string; run?: string };
628
+ /** How the turn was delivered when it consumed a unit's thread events
629
+ * (record 0051's mode-as-receipt rule): the mode read off the owner's state when
630
+ * each event arrived — a receipt, never a switch — and the sequence
631
+ * numbers consumed, so the record names the events it folded. Absent on
632
+ * every run that consumed none. */
633
+ mode?: "steer" | "wake" | "interrupt";
634
+ consumed?: number[];
603
635
  seq?: number;
604
636
  at?: number;
605
637
  }
@@ -826,6 +858,13 @@ export type RunEvent =
826
858
  * `base` is absent when the spawn knew none; the post-step then falls to
827
859
  * the coordinator store's `instance.base`. Additive: unknown → ignored. */
828
860
  | { type: "coordinator_tag"; parentInstanceId: string; unit?: string; base?: string; seq?: number; at?: number }
861
+ /** The plan runner instance a ship run's hand-off created (record 0051 R2;
862
+ * docs/reference/specs/run-history.md item 2): published by the ship branch
863
+ * after `handOffToCoordinator` succeeds, straight to the registry like
864
+ * `pr_opened`, and projected onto `RunRecord.instanceId` the way
865
+ * `coordinator_tag` is — so the thread's owner rule can find the instance
866
+ * from the page's ship run. Additive: unknown → ignored. */
867
+ | { type: "ship_handoff"; instanceId: string; seq?: number; at?: number }
829
868
  /** The review post-step's outcome when the verdict landed
830
869
  * (docs/reference/specs/agent-review.md item 18): the pull request it was
831
870
  * posted to, the head it was pinned to (the carried head after a rebase,
@@ -915,6 +954,14 @@ export type RunEvent =
915
954
  seq?: number;
916
955
  at?: number;
917
956
  }
957
+ /** A refusal the door made ([record 0054](../../docs/decisions/0054-a-refusal-the-person-caused-is-one-question-with-a-best-guess.md),
958
+ * as amended: every refusal is a run record; run-history.md item 2): the
959
+ * code, its one cause, and the sentence the person read — redacted and
960
+ * capped like a route receipt (`ROUTE_RECEIPT_CAP`). Exactly one per `door`
961
+ * record, published by `recordRefusal` beside the redacted request, so the
962
+ * door report counts every refusal — a gate refusal before any command is
963
+ * bound included — from the run store alone. Additive: unknown → ignored. */
964
+ | { type: "refusal"; code: string; cause: string; text: string; seq?: number; at?: number }
918
965
  /** The span records (docs/reference/specs/tracing.md): published, counted and stored like
919
966
  * every other event, read as timing and never as content. */
920
967
  | SpanStartEvent
@@ -422,9 +422,11 @@ export function analyzeRunFriction(events: readonly RunEvent[], opts: FrictionOp
422
422
  ev.type === "pr_description" ||
423
423
  ev.type === "pr_opened" ||
424
424
  ev.type === "coordinator_tag" ||
425
+ ev.type === "ship_handoff" ||
425
426
  ev.type === "review_posted" ||
426
427
  ev.type === "ship_round" ||
427
428
  ev.type === "route" ||
429
+ ev.type === "refusal" ||
428
430
  ev.type === "reference" ||
429
431
  ev.type === "lease" ||
430
432
  ev.type === "pushed_head"
@@ -2,7 +2,7 @@
2
2
  // Durable Object applies them inside one transaction; the in-memory ledger
3
3
  // applies them in tests; both agree because this is the only copy.
4
4
 
5
- import type { ClaimResult, FenceResult, LivePhase } from "./types.js";
5
+ import type { ClaimResult, FenceResult, IntakeReceipt, IntakeWriteResult, LivePhase } from "./types.js";
6
6
 
7
7
  /** One live run per thread. The existing row, if any, is what `live_runs` holds
8
8
  * for the thread; the same run re-claimed by its owner is idempotent (a retry
@@ -26,6 +26,14 @@ export function decideClaim(
26
26
  };
27
27
  }
28
28
 
29
+ /** First writer wins on an intake receipt (item 59): the existing row, if
30
+ * any, is what `intake_receipts` holds for the key; only when none stands
31
+ * does this write land, and every caller acts on the STORED row. */
32
+ export function decideIntakeInsert(existing: IntakeReceipt | undefined, receipt: IntakeReceipt): IntakeWriteResult {
33
+ if (existing) return { inserted: false, stored: existing };
34
+ return { inserted: true, stored: receipt };
35
+ }
36
+
29
37
  /** What an accepted claim does to the thread's row (item 42). `insert`: no row.
30
38
  * `promote`: the owner's claim WITH a prompt on its own `attaching` row — the
31
39
  * prompt, tools, card and state land and the phase goes `live`, identity and
@@ -84,6 +84,16 @@ export function roleOfStoredRow(json: string): "user" | "assistant" | undefined
84
84
  return role === "user" || role === "assistant" ? role : undefined;
85
85
  }
86
86
 
87
+ /** The platform-namespaced id of the person who authored the turn a stored
88
+ * row belongs to (e.g. `slack:U…`); absent for machine turns, compaction
89
+ * rows, unreadable rows, and rows written before record 0057. */
90
+ export function actorOfStoredRow(json: string): string | undefined {
91
+ const stored = parseStored(json);
92
+ if (!stored || "compaction" in stored) return undefined;
93
+ const actor = (stored as { actor?: unknown }).actor;
94
+ return typeof actor === "string" ? actor : undefined;
95
+ }
96
+
87
97
  /** The notepad's size (record 0035, "The notepad"): one document per session,
88
98
  * written whole, at most this many UTF-8 bytes; `notes` refuses over it naming
89
99
  * the size, and the object's write route does too. */
@@ -23,10 +23,13 @@ import {
23
23
  // ~4.2 MB on the wire.
24
24
 
25
25
  /** The stored form of one part: the turn's role rides on every row so a turn
26
- * is reconstructible from its rows alone. */
26
+ * is reconstructible from its rows alone. The actor is the platform-namespaced
27
+ * id of the person who authored the turn (e.g. `slack:U…`); absent for
28
+ * machine turns and rows written before record 0057. */
27
29
  export interface StoredPart {
28
30
  role: ChatMessage["role"];
29
31
  part: ContentPart | (ContentPart & { dataRef: string });
32
+ actor?: string;
30
33
  }
31
34
 
32
35
  /** The stored form of a compaction row: no role, no part — the entry alone. */
@@ -46,11 +49,14 @@ const isCompaction = (v: ChatMessage | StoredCompaction): v is StoredCompaction
46
49
 
47
50
  /** One turn → its rows (and any externalized attachments). Refuses a part the
48
51
  * row budget cannot hold: a transcript is never truncated. A compaction entry
49
- * is one row at its index, part 0. */
52
+ * is one row at its index, part 0. The optional `actor` is the
53
+ * platform-namespaced id of the person who authored the turn (absent for
54
+ * machine turns, compaction rows and rows written before record 0057). */
50
55
  export function turnRows(
51
56
  idx: number,
52
57
  message: ChatMessage | StoredCompaction,
53
58
  opts: { partBytes?: number; attachmentRefBytes?: number } = {},
59
+ actor?: string,
54
60
  ): { rows: TranscriptRow[]; attachments: TranscriptAttachment[] } {
55
61
  const partBytes = opts.partBytes ?? TRANSCRIPT_PART_BYTES;
56
62
  const refBytes = opts.attachmentRefBytes ?? ATTACHMENT_REF_BYTES;
@@ -73,7 +79,11 @@ export function turnRows(
73
79
  attachments.push({ ref, mediaType: part.mediaType, data: part.data });
74
80
  stored = { ...(part as ContentPart), data: "", dataRef: ref } as StoredPart["part"];
75
81
  }
76
- const json = JSON.stringify({ role: message.role, part: stored } satisfies StoredPart);
82
+ const json = JSON.stringify({
83
+ role: message.role,
84
+ part: stored,
85
+ ...(actor !== undefined ? { actor } : {}),
86
+ } satisfies StoredPart);
77
87
  const bytes = utf8ByteLength(json);
78
88
  if (bytes > partBytes) {
79
89
  throw new Error(`transcript: part ${i} of turn ${idx} is ${bytes} bytes, over the ${partBytes} row budget`);