switchroom 0.18.17 → 0.18.18

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 (54) hide show
  1. package/dist/agent-scheduler/index.js +13 -0
  2. package/dist/auth-broker/index.js +13 -0
  3. package/dist/cli/notion-write-pretool.mjs +13 -0
  4. package/dist/cli/switchroom.js +605 -479
  5. package/dist/host-control/main.js +17 -1
  6. package/dist/vault/approvals/kernel-server.js +13 -0
  7. package/dist/vault/broker/server.js +13 -0
  8. package/package.json +1 -1
  9. package/telegram-plugin/bridge/bridge.ts +7 -1
  10. package/telegram-plugin/dist/bridge/bridge.js +26 -1
  11. package/telegram-plugin/dist/gateway/gateway.js +1401 -431
  12. package/telegram-plugin/dist/server.js +26 -1
  13. package/telegram-plugin/fleet-fallback-resume.ts +26 -3
  14. package/telegram-plugin/gateway/approval-hold.ts +49 -0
  15. package/telegram-plugin/gateway/bridge-dead-watchdog.ts +61 -18
  16. package/telegram-plugin/gateway/gateway.ts +362 -71
  17. package/telegram-plugin/gateway/linear-activity.ts +20 -4
  18. package/telegram-plugin/gateway/premium-recovery-wiring.ts +122 -0
  19. package/telegram-plugin/gateway/session-model-file.ts +103 -0
  20. package/telegram-plugin/gateway/tier-downgrade-wiring.ts +121 -0
  21. package/telegram-plugin/gateway/unhandled-rejection-policy.ts +14 -1
  22. package/telegram-plugin/llm-error-present.ts +436 -0
  23. package/telegram-plugin/operator-events.ts +7 -1
  24. package/telegram-plugin/permission-title.ts +172 -10
  25. package/telegram-plugin/premium-recovery.ts +101 -0
  26. package/telegram-plugin/raw-error-scrub.ts +73 -0
  27. package/telegram-plugin/retry-api-call.ts +8 -2
  28. package/telegram-plugin/send-gate-degraded.test.ts +152 -1
  29. package/telegram-plugin/send-gate-observability.test.ts +140 -0
  30. package/telegram-plugin/send-gate-observability.ts +65 -20
  31. package/telegram-plugin/send-gate.test.ts +143 -1
  32. package/telegram-plugin/send-gate.ts +212 -19
  33. package/telegram-plugin/session-tail.ts +16 -0
  34. package/telegram-plugin/shared/local-time.ts +69 -0
  35. package/telegram-plugin/tests/approval-hold-harness.ts +6 -6
  36. package/telegram-plugin/tests/approval-hold-outcome.test.ts +10 -2
  37. package/telegram-plugin/tests/bridge-dead-watchdog.test.ts +61 -0
  38. package/telegram-plugin/tests/fleet-fallback-resume.test.ts +39 -0
  39. package/telegram-plugin/tests/flood-windows-persistence.test.ts +3 -2
  40. package/telegram-plugin/tests/linear-create-issue.test.ts +30 -2
  41. package/telegram-plugin/tests/llm-error-present.test.ts +380 -0
  42. package/telegram-plugin/tests/permission-title.test.ts +167 -4
  43. package/telegram-plugin/tests/premium-recovery-wiring.test.ts +150 -0
  44. package/telegram-plugin/tests/premium-recovery.test.ts +165 -0
  45. package/telegram-plugin/tests/reaction-gate-routing.test.ts +6 -1
  46. package/telegram-plugin/tests/retry-api-call.test.ts +21 -0
  47. package/telegram-plugin/tests/tier-downgrade-wiring.test.ts +165 -0
  48. package/telegram-plugin/tests/tier-downgrade.test.ts +141 -0
  49. package/telegram-plugin/tests/unhandled-rejection-policy.test.ts +27 -1
  50. package/telegram-plugin/tests/worker-activity-feed.test.ts +5 -2
  51. package/telegram-plugin/tests/worker-feed-coalesce.test.ts +492 -0
  52. package/telegram-plugin/tier-downgrade.ts +198 -0
  53. package/telegram-plugin/tool-activity-summary.ts +99 -0
  54. package/telegram-plugin/worker-activity-feed.ts +509 -409
@@ -0,0 +1,436 @@
1
+ /**
2
+ * llm-error-present.ts — humanized, cross-surface-deduped presentation of an
3
+ * LLM/API error, so a raw `b'{"type":"error",…}'` line never reaches a user.
4
+ *
5
+ * THE PROBLEM this closes
6
+ * -----------------------
7
+ * One Anthropic error JSONL line fans out to THREE independent render surfaces,
8
+ * none consulting each other, each historically printing the raw `detail`
9
+ * string with its `b'{…}'` byte-blob attached:
10
+ * 1. the agent turn-end "done" card (tool-activity-summary result block),
11
+ * 2. the operator-event "🚦 Rate limited" card (raw `detail`), and
12
+ * 3. the reply/answer passthrough (the synthetic-assistant error text relayed
13
+ * as the turn's answer).
14
+ *
15
+ * This module owns the ONE clean rendering + the ONE dedup authority the three
16
+ * surfaces consult, so a fan-out produces EXACTLY ONE user-facing message with
17
+ * NO raw JSON. It reuses the existing classifiers rather than re-deriving them:
18
+ * `classify429Detail` (throttle-tier.ts), `detectModelUnavailable` /
19
+ * `parseResetTime` (model-unavailable.ts), `isLitellmProxyLocal429` /
20
+ * `isTransientUpstreamSignal` (model-unavailable.ts), `classifyClaudeError`
21
+ * (operator-events.ts). `coreText` is ALWAYS built from a per-kind template —
22
+ * never from the raw `detail` — and `stripRawErrorBytes` is the belt-and-braces
23
+ * scrub so even the enrichment `reason` can carry no JSON.
24
+ *
25
+ * Pure module: no IPC, no bot, no FS. The only mutable state is the
26
+ * `ErrorPresenceGate` singleton (an in-memory dedup ledger), driven by an
27
+ * injectable `now`.
28
+ */
29
+
30
+ import {
31
+ detectModelUnavailable,
32
+ isLitellmProxyLocal429,
33
+ parseResetTime,
34
+ } from './model-unavailable.js'
35
+ import { classify429Detail } from './throttle-tier.js'
36
+ import { classifyClaudeError } from './operator-events.js'
37
+ import type { InlineKeyboardMarkup } from './operator-events.js'
38
+ import { stripRawErrorBytes, extractRequestId } from './raw-error-scrub.js'
39
+ import { fmtLocalClock, tzAbbrev } from './shared/local-time.js'
40
+
41
+ export { stripRawErrorBytes, extractRequestId } from './raw-error-scrub.js'
42
+
43
+ // ─── Types ──────────────────────────────────────────────────────────────────
44
+
45
+ export type LlmErrorKind =
46
+ | 'rate_limit'
47
+ | 'overload_529'
48
+ | 'quota_wall'
49
+ | 'auth'
50
+ | 'transient'
51
+ | 'unknown'
52
+
53
+ export type LlmErrorSource = 'anthropic' | 'litellm-local' | 'network'
54
+
55
+ export interface ParsedLlmError {
56
+ kind: LlmErrorKind
57
+ /** Human, JSON-STRIPPED one-liner built from a per-kind template. Never raw. */
58
+ coreText: string
59
+ /** Parsed reset instant, when the source carried one. */
60
+ resetAt?: Date
61
+ /** Parsed retry-after window in ms, when the source carried one. */
62
+ retryAfterMs?: number
63
+ /** Resolved model id, when the source named one. */
64
+ model?: string
65
+ /** Anthropic `request_id`, when present — the strongest dedup key. */
66
+ requestId?: string
67
+ source: LlmErrorSource
68
+ /** True when the harness is still retrying this error internally (mid-retry). */
69
+ autoRetrying: boolean
70
+ /** True when the failure is final (NOT an in-flight retry). */
71
+ terminal: boolean
72
+ }
73
+
74
+ /** Optional retry-state annotations Claude Code stamps on a retried error line. */
75
+ export interface LlmErrorRetryState {
76
+ retryAttempt: number | null
77
+ maxRetries: number | null
78
+ }
79
+
80
+ // ─── model extraction ────────────────────────────────────────────────────────
81
+
82
+ /** Pull a resolved model id (`claude-…`, `sr-…`) out of a raw string, or undefined. */
83
+ function extractModel(raw: string): string | undefined {
84
+ const m = raw.match(/["']?model["']?\s*[=:]\s*["']?((?:claude|sr)[A-Za-z0-9._-]+)/i)
85
+ return m ? m[1] : undefined
86
+ }
87
+
88
+ // ─── Parser ──────────────────────────────────────────────────────────────────
89
+
90
+ const TRANSIENT_KINDS: ReadonlySet<LlmErrorKind> = new Set<LlmErrorKind>([
91
+ 'rate_limit',
92
+ 'overload_529',
93
+ 'transient',
94
+ ])
95
+
96
+ /** The always-actionable kinds — NEVER silenced, even inside a collapse window. */
97
+ export function isActionableKind(kind: LlmErrorKind): boolean {
98
+ return kind === 'auth' || kind === 'quota_wall'
99
+ }
100
+
101
+ /**
102
+ * Parse a raw error string (a session-tail `detail`, an isApiErrorMessage
103
+ * text, or a raw stderr line) into a structured, JSON-free ParsedLlmError.
104
+ * Reuses the existing wording classifiers; maps their kinds into this union.
105
+ */
106
+ export function parseLlmError(
107
+ raw: string,
108
+ retryState?: LlmErrorRetryState,
109
+ ): ParsedLlmError {
110
+ const text = typeof raw === 'string' ? raw : ''
111
+ const requestId = extractRequestId(text)
112
+ const model = extractModel(text)
113
+
114
+ // Reset / retry-after best-effort (Anthropic + relative wordings).
115
+ const resetAt = parseResetTime(text)
116
+ const retryAfterMs =
117
+ resetAt != null ? Math.max(0, resetAt.getTime() - Date.now()) : undefined
118
+
119
+ const { kind, source } = classifyKindAndSource(text)
120
+
121
+ // Retry / terminal semantics. Only the transient family can be mid-retry;
122
+ // auth / quota_wall / model_unavailable are terminal by construction.
123
+ let autoRetrying = false
124
+ let terminal = true
125
+ if (TRANSIENT_KINDS.has(kind)) {
126
+ const { retryAttempt, maxRetries } = retryState ?? { retryAttempt: null, maxRetries: null }
127
+ if (retryAttempt != null && maxRetries != null) {
128
+ autoRetrying = retryAttempt < maxRetries
129
+ terminal = retryAttempt >= maxRetries
130
+ } else {
131
+ // No retry annotation: treat as a surfaced (terminal) failure — Claude
132
+ // writes the user-facing error shape only after its own retries are done.
133
+ autoRetrying = false
134
+ terminal = true
135
+ }
136
+ }
137
+
138
+ return {
139
+ kind,
140
+ coreText: buildCoreText(kind, source),
141
+ ...(resetAt != null ? { resetAt } : {}),
142
+ ...(retryAfterMs != null ? { retryAfterMs } : {}),
143
+ ...(model != null ? { model } : {}),
144
+ ...(requestId != null ? { requestId } : {}),
145
+ source,
146
+ autoRetrying,
147
+ terminal,
148
+ }
149
+ }
150
+
151
+ function classifyKindAndSource(text: string): { kind: LlmErrorKind; source: LlmErrorSource } {
152
+ const lower = text.toLowerCase()
153
+
154
+ // 1. Auth — always terminal, always actionable.
155
+ const claudeKind = classifyClaudeError({ message: text, type: text })
156
+ if (claudeKind === 'credentials-expired' || claudeKind === 'credentials-invalid') {
157
+ return { kind: 'auth', source: 'anthropic' }
158
+ }
159
+
160
+ // 2. LiteLLM-proxy-LOCAL 429 — the proxy's own limiter, never Anthropic.
161
+ // Checked before the generic quota/overload matchers because some LiteLLM
162
+ // bodies contain the word "limit" that a quota matcher could seize on.
163
+ if (isLitellmProxyLocal429(text)) {
164
+ return { kind: 'rate_limit', source: 'litellm-local' }
165
+ }
166
+
167
+ // 3. Model-unavailable detector owns quota-wall / overload / network wording.
168
+ const mu = detectModelUnavailable(text)
169
+ if (mu != null) {
170
+ if (mu.kind === 'quota_exhausted') return { kind: 'quota_wall', source: 'anthropic' }
171
+ if (mu.kind === 'network') return { kind: 'transient', source: 'network' }
172
+ // mu.kind === 'overload' | 'rate_limited' — split 529 overload from a 429.
173
+ if (
174
+ lower.includes('529') ||
175
+ lower.includes('overloaded')
176
+ ) {
177
+ return { kind: 'overload_529', source: 'anthropic' }
178
+ }
179
+ // A 429-family transient. Three-way classify picks the source nuance; all
180
+ // land on the calm rate_limit kind here.
181
+ const c = classify429Detail(text)
182
+ return {
183
+ kind: 'rate_limit',
184
+ source: c === 'litellm-local' ? 'litellm-local' : 'anthropic',
185
+ }
186
+ }
187
+
188
+ // 4. Credit balance is a quota-family wall (actionable).
189
+ if (claudeKind === 'credit-exhausted') {
190
+ return { kind: 'quota_wall', source: 'anthropic' }
191
+ }
192
+
193
+ // 5. Bare rate-limit classification from the operator taxonomy.
194
+ if (claudeKind === 'rate-limited') {
195
+ return { kind: 'rate_limit', source: 'anthropic' }
196
+ }
197
+ if (claudeKind === 'unknown-5xx') {
198
+ return { kind: 'overload_529', source: 'anthropic' }
199
+ }
200
+
201
+ return { kind: 'unknown', source: 'anthropic' }
202
+ }
203
+
204
+ function buildCoreText(kind: LlmErrorKind, source: LlmErrorSource): string {
205
+ switch (kind) {
206
+ case 'rate_limit':
207
+ return source === 'litellm-local'
208
+ ? 'Hit the local proxy rate limit — retrying automatically.'
209
+ : 'Rate limited by Anthropic — retrying automatically.'
210
+ case 'overload_529':
211
+ return 'Anthropic is overloaded (529) — retrying automatically.'
212
+ case 'quota_wall':
213
+ return 'Usage limit reached on this Claude subscription.'
214
+ case 'auth':
215
+ return 'Claude login needs re-authentication.'
216
+ case 'transient':
217
+ return source === 'network'
218
+ ? "Couldn't reach Anthropic (network) — retrying automatically."
219
+ : 'A temporary upstream hiccup — retrying automatically.'
220
+ case 'unknown':
221
+ return 'The model returned an error.'
222
+ }
223
+ }
224
+
225
+ // ─── Rendering ───────────────────────────────────────────────────────────────
226
+
227
+ export interface RenderedLlmError {
228
+ text: string
229
+ keyboard?: InlineKeyboardMarkup
230
+ }
231
+
232
+ /**
233
+ * Format a reset instant in the operator's local tz plus a relative tail, e.g.
234
+ * `clears ~4:52pm AEST (~in 38m)`. Returns '' when there is no reset to show.
235
+ */
236
+ export function formatResetLocal(resetAt: Date | undefined, tz: string, now: Date = new Date()): string {
237
+ if (resetAt == null) return ''
238
+ const ms = resetAt.getTime()
239
+ if (!Number.isFinite(ms)) return ''
240
+ const clock = fmtLocalClock(ms, tz)
241
+ const abbrev = tzAbbrev(ms, tz)
242
+ const rel = formatRelativeTail(ms - now.getTime())
243
+ return rel ? `clears ~${clock} ${abbrev} (${rel})` : `clears ~${clock} ${abbrev}`
244
+ }
245
+
246
+ function formatRelativeTail(deltaMs: number): string {
247
+ if (deltaMs <= 0) return '~now'
248
+ const totalMin = Math.round(deltaMs / 60_000)
249
+ if (totalMin < 1) return '~in <1m'
250
+ if (totalMin < 60) return `~in ${totalMin}m`
251
+ const hours = Math.floor(totalMin / 60)
252
+ const mins = totalMin % 60
253
+ if (hours < 24) return mins > 0 ? `~in ${hours}h ${mins}m` : `~in ${hours}h`
254
+ const days = Math.floor(hours / 24)
255
+ const remH = hours % 24
256
+ return remH > 0 ? `~in ${days}d ${remH}h` : `~in ${days}d`
257
+ }
258
+
259
+ /**
260
+ * Render ONE clean card for a parsed LLM error. Action buttons for the
261
+ * actionable kinds (auth → Reauth, quota_wall → Wait/switch). `agent` is used
262
+ * in the callback_data (URL-encoded) and the headline; `tz` localizes the reset.
263
+ */
264
+ export function renderLlmError(
265
+ parsed: ParsedLlmError,
266
+ agent: string,
267
+ tz: string,
268
+ now: Date = new Date(),
269
+ ): RenderedLlmError {
270
+ const safeAgent = escapeAgent(agent)
271
+ const emoji = kindEmoji(parsed.kind)
272
+ const lines: string[] = [`${emoji} ${parsed.coreText} (**${safeAgent}**)`]
273
+
274
+ const resetLine = formatResetLocal(parsed.resetAt, tz, now)
275
+ if (resetLine) lines.push(`_${resetLine}_`)
276
+
277
+ if (parsed.model) lines.push(`_model: ${escapeAgent(parsed.model)}_`)
278
+
279
+ const text = lines.join('\n')
280
+
281
+ switch (parsed.kind) {
282
+ case 'auth':
283
+ return {
284
+ text,
285
+ keyboard: {
286
+ inline_keyboard: [
287
+ [
288
+ { text: '🔐 Reauth now', callback_data: `op:reauth:${encodeURIComponent(agent)}` },
289
+ { text: '❌ Dismiss', callback_data: `op:dismiss:${encodeURIComponent(agent)}` },
290
+ ],
291
+ ],
292
+ },
293
+ }
294
+ case 'quota_wall':
295
+ return {
296
+ text,
297
+ keyboard: {
298
+ inline_keyboard: [
299
+ [{ text: '⏳ Wait', callback_data: `op:dismiss:${encodeURIComponent(agent)}` }],
300
+ ],
301
+ },
302
+ }
303
+ default:
304
+ return { text }
305
+ }
306
+ }
307
+
308
+ function kindEmoji(kind: LlmErrorKind): string {
309
+ switch (kind) {
310
+ case 'rate_limit':
311
+ return '🚦'
312
+ case 'overload_529':
313
+ return '🔥'
314
+ case 'quota_wall':
315
+ return '⚠️'
316
+ case 'auth':
317
+ return '🔑'
318
+ case 'transient':
319
+ return '🌐'
320
+ case 'unknown':
321
+ return '⚠️'
322
+ }
323
+ }
324
+
325
+ /** Minimal markdown-safe agent rendering (mirrors operator-events escapeMarkdown intent). */
326
+ function escapeAgent(s: string): string {
327
+ return s.replace(/([_*`\[\]])/g, '\\$1')
328
+ }
329
+
330
+ // ─── Cross-surface dedup: ErrorPresenceGate ──────────────────────────────────
331
+
332
+ /**
333
+ * How long ONE terminal error owns the "already surfaced" claim across all
334
+ * three surfaces. Distinct from — and additional to — the 5-minute per-kind
335
+ * `shouldEmitOperatorEvent` cooldown (operator-events.ts): that debounces an
336
+ * error STORM on one surface; this collapses ONE error's FAN-OUT across
337
+ * surfaces within a single turn.
338
+ */
339
+ export const ERROR_COLLAPSE_WINDOW_MS = 60_000
340
+
341
+ /**
342
+ * A tiny in-memory dedup ledger. The FIRST surface to `claim(key)` within the
343
+ * collapse window wins (returns true) and renders; every later surface loses
344
+ * (returns false) and suppresses. Key preference: the Anthropic `request_id`
345
+ * when present (exact), else a `${kind}:${agent}:${windowBucket}` coarse key so
346
+ * two distinct-but-unlabelled errors of the same kind within 60s still collapse.
347
+ */
348
+ export class ErrorPresenceGate {
349
+ private readonly claims = new Map<string, number>()
350
+
351
+ /** Build the dedup key for a parsed error + agent. */
352
+ keyFor(parsed: Pick<ParsedLlmError, 'kind' | 'requestId'>, agent: string, now: number): string {
353
+ if (parsed.requestId) return `rid:${parsed.requestId}`
354
+ const bucket = Math.floor(now / ERROR_COLLAPSE_WINDOW_MS)
355
+ return `${parsed.kind}:${agent}:${bucket}`
356
+ }
357
+
358
+ /**
359
+ * Attempt to claim ownership of `key`. Returns true for the first caller
360
+ * within the window, false for every subsequent caller. Expired claims are
361
+ * pruned lazily on each call.
362
+ */
363
+ claim(key: string, now: number = Date.now()): boolean {
364
+ this.prune(now)
365
+ const existing = this.claims.get(key)
366
+ if (existing != null && now - existing < ERROR_COLLAPSE_WINDOW_MS) {
367
+ return false
368
+ }
369
+ this.claims.set(key, now)
370
+ return true
371
+ }
372
+
373
+ /** True when `key` is currently claimed (does NOT claim). */
374
+ isClaimed(key: string, now: number = Date.now()): boolean {
375
+ const existing = this.claims.get(key)
376
+ return existing != null && now - existing < ERROR_COLLAPSE_WINDOW_MS
377
+ }
378
+
379
+ private prune(now: number): void {
380
+ for (const [k, at] of this.claims) {
381
+ if (now - at >= ERROR_COLLAPSE_WINDOW_MS) this.claims.delete(k)
382
+ }
383
+ }
384
+
385
+ /** Test-only: forget every claim. */
386
+ reset(): void {
387
+ this.claims.clear()
388
+ }
389
+ }
390
+
391
+ /** The process-wide gate the three surfaces consult. */
392
+ export const errorPresenceGate = new ErrorPresenceGate()
393
+
394
+ /**
395
+ * The single decision a surface makes before rendering a parsed LLM error.
396
+ * Returns:
397
+ * - 'render' — this surface owns the humanized card, render it.
398
+ * - 'suppress' — another surface already owns it (dedup) OR it is a
399
+ * transient error the harness is still auto-retrying — stay silent.
400
+ *
401
+ * ACTIONABLE kinds (auth / quota_wall) ALWAYS return 'render' — they are never
402
+ * deduped away and never silenced mid-retry, because the operator must always
403
+ * see (and be able to act on) a login/quota wall.
404
+ */
405
+ export function decideErrorSurface(
406
+ parsed: ParsedLlmError,
407
+ agent: string,
408
+ opts: { claim: boolean; now?: number; gate?: ErrorPresenceGate } = { claim: true },
409
+ ): 'render' | 'suppress' {
410
+ const now = opts.now ?? Date.now()
411
+ const gate = opts.gate ?? errorPresenceGate
412
+
413
+ if (isActionableKind(parsed.kind)) {
414
+ // Always render — but still record the claim so a redundant transient
415
+ // surface for the same key stays collapsed.
416
+ if (opts.claim) gate.claim(gate.keyFor(parsed, agent, now), now)
417
+ return 'render'
418
+ }
419
+
420
+ // Auto-retry silence: a transient error still being retried internally is
421
+ // NOT surfaced on ANY surface until it goes terminal. NOTE: in the current
422
+ // production wiring the operator surface reaches here only AFTER session-tail
423
+ // (readNew → `errEvent.terminal || !errEvent.transient`) has already dropped
424
+ // in-flight transients, and the gateway calls `parseLlmError(detail)` with no
425
+ // retryState (the raw retry annotations don't survive the IPC hop) — so a
426
+ // production `parsed.autoRetrying` is always false here. This branch is the
427
+ // module's own guarantee for ANY caller that DOES pass retryState (unit-tested
428
+ // as such); prod silence is owned upstream at session-tail, not re-derived here.
429
+ if (parsed.autoRetrying && !parsed.terminal) return 'suppress'
430
+
431
+ const key = gate.keyFor(parsed, agent, now)
432
+ if (opts.claim) {
433
+ return gate.claim(key, now) ? 'render' : 'suppress'
434
+ }
435
+ return gate.isClaimed(key, now) ? 'suppress' : 'render'
436
+ }
@@ -13,6 +13,7 @@
13
13
  */
14
14
 
15
15
  import { escapeMarkdown } from './format.js'
16
+ import { stripRawErrorBytes } from './raw-error-scrub.js'
16
17
 
17
18
  // ─── Taxonomy ────────────────────────────────────────────────────────────────
18
19
 
@@ -216,7 +217,12 @@ export interface RenderResult {
216
217
  */
217
218
  export function renderOperatorEvent(ev: OperatorEvent): RenderResult {
218
219
  const agent = escapeMarkdown(ev.agent)
219
- const detail = escapeMarkdown(ev.detail)
220
+ // #llm-error-surfacing: NEVER let a raw API-error byte-blob (`· b'{…}'`,
221
+ // trailing `{"type":"error"…}` JSON, `API Error:` prefix) reach a user. A
222
+ // clean human detail passes through unchanged; only smuggled raw bytes are
223
+ // stripped. This is the belt-and-braces scrub for every card kind — the
224
+ // rate-limited card in particular used to relay the raw synthetic-error text.
225
+ const detail = escapeMarkdown(stripRawErrorBytes(ev.detail))
220
226
 
221
227
  switch (ev.kind) {
222
228
  case 'credentials-expired':
@@ -20,7 +20,7 @@
20
20
  import { basename } from "node:path";
21
21
  import { escapeMarkdown } from "./card-format.js";
22
22
  import { prettyMcpServer, type ScopeOption } from "./permission-rule.js";
23
- import { redact } from "./secret-detect/redact.js";
23
+ import { redact, REDACTED_MARKER } from "./secret-detect/redact.js";
24
24
 
25
25
  const COMMAND_TITLE_MAX = 48;
26
26
  const DESCRIPTION_LINE_MAX = 240;
@@ -35,6 +35,55 @@ const ARG_SUMMARY_MAX_KEYS = 4; // how many payload keys to surface on the card
35
35
  const ARG_VALUE_MAX = 40; // per-value truncation in the arg-summary line
36
36
  const ARG_SUMMARY_LINE_MAX = 180; // total cap for the arg-summary line
37
37
 
38
+ /**
39
+ * Input keys that are routing / formatting / id noise, never operator-
40
+ * meaningful context. Stripped when synthesizing the fallback `context:`
41
+ * line (#3167) so it surfaces the substance (the reply text, the command,
42
+ * the query) and not `chat_id` / `format` / `disable_notification` chrome.
43
+ * `reason`/`why` are here too because their presence takes the `why:` path —
44
+ * they never reach the synthesizer.
45
+ */
46
+ const CONTEXT_NOISE_KEYS = new Set([
47
+ "reason",
48
+ "why",
49
+ "chat_id",
50
+ "message_id",
51
+ "message_thread_id",
52
+ "thread_id",
53
+ "origin_turn_id",
54
+ "reply_to",
55
+ "quote",
56
+ "quote_text",
57
+ "format",
58
+ "parse_mode",
59
+ "disable_web_page_preview",
60
+ "disable_notification",
61
+ "protect_content",
62
+ "single_use",
63
+ "ack_text",
64
+ "inline_keyboard",
65
+ "file_id",
66
+ ]);
67
+
68
+ /**
69
+ * Input keys whose VALUE is credential-shaped by name — hard-masked whole,
70
+ * never partially revealed (#3167 review). The token-shape detectors in
71
+ * `redact()` need a random-looking value to fire, so a low-entropy password
72
+ * or a short secret under one of these keys would otherwise slip through the
73
+ * synthesized `context:` line. Matching the key is the reliable signal.
74
+ */
75
+ const SENSITIVE_VALUE_KEY_RE =
76
+ /token|secret|password|passwd|key|auth|dsn|conn|url|credential|cookie|session/i;
77
+
78
+ /**
79
+ * Credential-bearing DSN schemes that `redactUrls()` (inside `redact()`) does
80
+ * NOT cover — it only handles http(s)/ws(s)/ftp. A `postgres://user:pass@host`
81
+ * DATABASE_URL is a real vault-key shape here, so mask the whole DSN when its
82
+ * authority carries a `user:pass@` credential (#3167 review).
83
+ */
84
+ const NON_HTTP_DSN_RE =
85
+ /\b(?:postgres(?:ql)?|mysql|mariadb|redis|rediss|mongodb(?:\+srv)?|amqp|amqps):\/\/\S*@\S+/gi;
86
+
38
87
  /**
39
88
  * Human verb-phrases for switchroom-managed MCP tools. The raw
40
89
  * `mcp__<server>__<tool>` name is operator-hostile. Phrases are written
@@ -141,15 +190,23 @@ export function formatPermissionCardBody(opts: {
141
190
  // static schema description (#2469).
142
191
  const callerReason = callerSuppliedReason(opts.inputPreview);
143
192
  const rawWhy = (callerReason ?? "").replace(/\s+/g, " ").trim();
144
- const truncatedWhy =
145
- rawWhy.length > DESCRIPTION_LINE_MAX
146
- ? rawWhy.slice(0, DESCRIPTION_LINE_MAX - 1) + "…"
147
- : rawWhy;
148
- lines.push(
149
- truncatedWhy.length > 0
150
- ? `why: _${escapeTgHtml(truncatedWhy)}_`
151
- : `why: _not provided_`,
152
- );
193
+ if (rawWhy.length > 0) {
194
+ const truncatedWhy =
195
+ rawWhy.length > DESCRIPTION_LINE_MAX
196
+ ? rawWhy.slice(0, DESCRIPTION_LINE_MAX - 1) + "…"
197
+ : rawWhy;
198
+ lines.push(`why: _${escapeTgHtml(truncatedWhy)}_`);
199
+ } else {
200
+ // No caller reason. Many tools (reply, react, edit_message, …) carry no
201
+ // `reason`/`why` argument at all, so a bare "why: not provided" gives the
202
+ // operator nothing to decide on — the post-rollout permission storm filled
203
+ // the chat with contentless cards (#3167). Fall back to an honest,
204
+ // redaction-safe `context:` line synthesized from the tool + its salient
205
+ // input, so the card always carries something meaningful. The distinct
206
+ // label keeps the agent's omission of a rationale visible (it is NOT a
207
+ // fabricated "why"), while still surfacing what the tool is about to do.
208
+ lines.push(`context: _${escapeTgHtml(synthesizeContext(opts.toolName, opts.inputPreview))}_`);
209
+ }
153
210
 
154
211
  // Third line (REST-wrapper MCP writes only): a redaction-safe summary of
155
212
  // the payload so the operator can see WHAT is being sent, not just the
@@ -597,6 +654,111 @@ function callerSuppliedReason(inputPreview: string | undefined): string | null {
597
654
  return null;
598
655
  }
599
656
 
657
+ /**
658
+ * Honest, redaction-safe fallback for the card's rationale line when the
659
+ * caller supplied no `reason`/`why` (#3167). Synthesizes context from a
660
+ * summary of the tool's salient input fields — the reply text, the command,
661
+ * the search query — so the operator always has something to decide on
662
+ * instead of a contentless "not provided". Every value passes through
663
+ * `redact()`; ids / routing / formatting keys are stripped. Falls back to
664
+ * the natural action phrase when the input exposes nothing salient.
665
+ * Deterministic (no model in the loop): same input → same line. Exported
666
+ * for unit testing.
667
+ */
668
+ export function synthesizeContext(
669
+ toolName: string,
670
+ inputPreview: string | undefined,
671
+ ): string {
672
+ const input = parseInput(inputPreview);
673
+ const summary = input ? salientInputSummary(input) : null;
674
+ // Prefer the salient input summary (the reply text, the query, the diff
675
+ // target) — the substance the title's verb-phrase does NOT already carry.
676
+ // When the input exposes nothing meaningful, fall back to the natural
677
+ // action so the line still names the tool rather than reading empty.
678
+ const base = summary ?? naturalAction(toolName, inputPreview);
679
+ // Acceptance criteria (#3167) call for "tool + summarized input +
680
+ // originating turn". When the call carries a forum-topic origin turn,
681
+ // append a compact reference so the operator can tie the card to the turn
682
+ // that spawned it. Kept short — it's a routing id, not prose.
683
+ const turnRef = input ? readString(input, "origin_turn_id") : null;
684
+ return turnRef ? `${base} · turn ${shortTurnRef(turnRef)}` : base;
685
+ }
686
+
687
+ /** Compact display form of an origin-turn id — the tail is the distinguishing
688
+ * part; a full opaque id would swamp the line. */
689
+ function shortTurnRef(turnId: string): string {
690
+ const t = turnId.trim();
691
+ return t.length <= 10 ? t : `…${t.slice(-8)}`;
692
+ }
693
+
694
+ /**
695
+ * A compact, redaction-safe summary of the operator-meaningful scalar fields
696
+ * of a tool input — used to build the synthesized `context:` line (#3167).
697
+ * Skips {@link CONTEXT_NOISE_KEYS} (ids/routing/formatting), surfaces up to
698
+ * {@link ARG_SUMMARY_MAX_KEYS} scalar `key: value` pairs (each redacted +
699
+ * truncated), and renders nested objects/arrays as the bare key name (no
700
+ * value dump — avoids leaking PII/secrets and oversized blobs). Returns null
701
+ * when nothing meaningful remains, so the caller falls back to the bare
702
+ * action phrase.
703
+ *
704
+ * Free-text fields (e.g. a reply `text`) surface as content — a conscious
705
+ * choice (#3167 review, LOW-1): the operator NEEDS to see *what* is being
706
+ * sent to judge the card. Exposure is bounded to {@link ARG_VALUE_MAX} chars
707
+ * and every value is redaction-passed via {@link redactSalientValue}, so a
708
+ * card can never leak more than the outbound reply already does in history
709
+ * (both use the same `redact()` chokepoint). We do not mask free text to a
710
+ * placeholder, which would gut the card's usefulness.
711
+ */
712
+ function salientInputSummary(input: Record<string, unknown>): string | null {
713
+ const parts: string[] = [];
714
+ for (const [key, value] of Object.entries(input)) {
715
+ if (CONTEXT_NOISE_KEYS.has(key)) continue;
716
+ if (value == null) continue;
717
+ if (parts.length >= ARG_SUMMARY_MAX_KEYS) {
718
+ parts.push("…");
719
+ break;
720
+ }
721
+ if (typeof value === "object") {
722
+ parts.push(key); // nested object/array → key name only, never dumped
723
+ continue;
724
+ }
725
+ const shown = truncate(redactSalientValue(key, String(value)), ARG_VALUE_MAX);
726
+ if (shown.length === 0) continue;
727
+ parts.push(`${key}: ${shown}`);
728
+ }
729
+ if (parts.length === 0) return null;
730
+ const joined = parts.join(", ");
731
+ return joined.length > ARG_SUMMARY_LINE_MAX
732
+ ? joined.slice(0, ARG_SUMMARY_LINE_MAX - 1) + "…"
733
+ : joined;
734
+ }
735
+
736
+ /**
737
+ * Redact a single salient value for display on the synthesized `context:`
738
+ * line, given its input KEY (#3167 review). Three layers, defense-in-depth:
739
+ *
740
+ * 1. Hard-mask whole when the key is credential-shaped
741
+ * ({@link SENSITIVE_VALUE_KEY_RE}) — the token-shape detectors need a
742
+ * random-looking value, so a low-entropy password / short secret under a
743
+ * `token`/`password`/`api_key` key would otherwise leak. The key name is
744
+ * the reliable signal.
745
+ * 2. Mask non-http DSN credentials ({@link NON_HTTP_DSN_RE}) that
746
+ * `redactUrls()` misses (postgres://user:pass@host, mysql://, redis://…).
747
+ * 3. Run `redact()` WITH the key restored as `key=value` context, so the
748
+ * contextual detectors (`kv_entropy` / `env_key_value`) that require the
749
+ * `key[:=]value` shape in one string can fire — then strip the synthetic
750
+ * `key=` prefix. Redacting the BARE value (the pre-review bug) blinded
751
+ * those detectors, since `redact()` deliberately excludes the
752
+ * low-precision `generic_high_entropy` fallback.
753
+ */
754
+ function redactSalientValue(key: string, value: string): string {
755
+ if (SENSITIVE_VALUE_KEY_RE.test(key)) return REDACTED_MARKER;
756
+ const dsnScrubbed = value.replace(NON_HTTP_DSN_RE, REDACTED_MARKER);
757
+ const scrubbed = redact(`${key}=${dsnScrubbed}`);
758
+ const prefix = `${key}=`;
759
+ return scrubbed.startsWith(prefix) ? scrubbed.slice(prefix.length) : scrubbed;
760
+ }
761
+
600
762
  /**
601
763
  * Regex-based fallback to extract a `reason` or `why` value from a raw
602
764
  * (possibly truncated / invalid-JSON) inputPreview string. Mirrors