@qmediat.io/gemini-code-context-mcp 1.23.1 → 1.25.0

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 (147) hide show
  1. package/CHANGELOG.md +161 -1
  2. package/README.md +1 -1
  3. package/dist/cache/ttl-watcher.d.ts +29 -0
  4. package/dist/cache/ttl-watcher.d.ts.map +1 -1
  5. package/dist/cache/ttl-watcher.js +104 -43
  6. package/dist/cache/ttl-watcher.js.map +1 -1
  7. package/dist/gemini/client.d.ts +7 -0
  8. package/dist/gemini/client.d.ts.map +1 -1
  9. package/dist/gemini/client.js +8 -1
  10. package/dist/gemini/client.js.map +1 -1
  11. package/dist/gemini/long-decode.d.ts +153 -0
  12. package/dist/gemini/long-decode.d.ts.map +1 -0
  13. package/dist/gemini/long-decode.js +178 -0
  14. package/dist/gemini/long-decode.js.map +1 -0
  15. package/dist/gemini/model-registry.d.ts +2 -0
  16. package/dist/gemini/model-registry.d.ts.map +1 -1
  17. package/dist/gemini/model-registry.js +5 -2
  18. package/dist/gemini/model-registry.js.map +1 -1
  19. package/dist/gemini/models.d.ts.map +1 -1
  20. package/dist/gemini/models.js +1 -0
  21. package/dist/gemini/models.js.map +1 -1
  22. package/dist/gemini/retry.d.ts +8 -0
  23. package/dist/gemini/retry.d.ts.map +1 -1
  24. package/dist/gemini/retry.js +33 -1
  25. package/dist/gemini/retry.js.map +1 -1
  26. package/dist/tools/agentic/dispatch.d.ts +68 -0
  27. package/dist/tools/agentic/dispatch.d.ts.map +1 -0
  28. package/dist/tools/agentic/dispatch.js +167 -0
  29. package/dist/tools/agentic/dispatch.js.map +1 -0
  30. package/dist/tools/agentic/input.d.ts +28 -0
  31. package/dist/tools/agentic/input.d.ts.map +1 -0
  32. package/dist/tools/agentic/input.js +66 -0
  33. package/dist/tools/agentic/input.js.map +1 -0
  34. package/dist/tools/agentic/instructions.d.ts +46 -0
  35. package/dist/tools/agentic/instructions.d.ts.map +1 -0
  36. package/dist/tools/agentic/instructions.js +170 -0
  37. package/dist/tools/agentic/instructions.js.map +1 -0
  38. package/dist/tools/agentic/iteration.d.ts +81 -0
  39. package/dist/tools/agentic/iteration.d.ts.map +1 -0
  40. package/dist/tools/agentic/iteration.js +317 -0
  41. package/dist/tools/agentic/iteration.js.map +1 -0
  42. package/dist/tools/agentic/ledger.d.ts +64 -0
  43. package/dist/tools/agentic/ledger.d.ts.map +1 -0
  44. package/dist/tools/agentic/ledger.js +110 -0
  45. package/dist/tools/agentic/ledger.js.map +1 -0
  46. package/dist/tools/agentic/loop.d.ts +10 -0
  47. package/dist/tools/agentic/loop.d.ts.map +1 -0
  48. package/dist/tools/agentic/loop.js +179 -0
  49. package/dist/tools/agentic/loop.js.map +1 -0
  50. package/dist/tools/agentic/no-progress.d.ts +90 -0
  51. package/dist/tools/agentic/no-progress.d.ts.map +1 -0
  52. package/dist/tools/agentic/no-progress.js +107 -0
  53. package/dist/tools/agentic/no-progress.js.map +1 -0
  54. package/dist/tools/agentic/pass-clock.d.ts +85 -0
  55. package/dist/tools/agentic/pass-clock.d.ts.map +1 -0
  56. package/dist/tools/agentic/pass-clock.js +162 -0
  57. package/dist/tools/agentic/pass-clock.js.map +1 -0
  58. package/dist/tools/agentic/phase-defaults.d.ts +29 -0
  59. package/dist/tools/agentic/phase-defaults.d.ts.map +1 -0
  60. package/dist/tools/agentic/phase-defaults.js +23 -0
  61. package/dist/tools/agentic/phase-defaults.js.map +1 -0
  62. package/dist/tools/agentic/rescue.d.ts +69 -0
  63. package/dist/tools/agentic/rescue.d.ts.map +1 -0
  64. package/dist/tools/agentic/rescue.js +203 -0
  65. package/dist/tools/agentic/rescue.js.map +1 -0
  66. package/dist/tools/agentic/response.d.ts +16 -0
  67. package/dist/tools/agentic/response.d.ts.map +1 -0
  68. package/dist/tools/agentic/response.js +25 -0
  69. package/dist/tools/agentic/response.js.map +1 -0
  70. package/dist/tools/agentic/results.d.ts +76 -0
  71. package/dist/tools/agentic/results.d.ts.map +1 -0
  72. package/dist/tools/agentic/results.js +228 -0
  73. package/dist/tools/agentic/results.js.map +1 -0
  74. package/dist/tools/agentic/run.d.ts +130 -0
  75. package/dist/tools/agentic/run.d.ts.map +1 -0
  76. package/dist/tools/agentic/run.js +131 -0
  77. package/dist/tools/agentic/run.js.map +1 -0
  78. package/dist/tools/agentic/sandbox.d.ts +13 -2
  79. package/dist/tools/agentic/sandbox.d.ts.map +1 -1
  80. package/dist/tools/agentic/sandbox.js +127 -111
  81. package/dist/tools/agentic/sandbox.js.map +1 -1
  82. package/dist/tools/agentic/workspace-tools.d.ts +30 -27
  83. package/dist/tools/agentic/workspace-tools.d.ts.map +1 -1
  84. package/dist/tools/agentic/workspace-tools.js +324 -385
  85. package/dist/tools/agentic/workspace-tools.js.map +1 -1
  86. package/dist/tools/ask-agentic-fallback.d.ts +33 -0
  87. package/dist/tools/ask-agentic-fallback.d.ts.map +1 -0
  88. package/dist/tools/ask-agentic-fallback.js +113 -0
  89. package/dist/tools/ask-agentic-fallback.js.map +1 -0
  90. package/dist/tools/ask-agentic.tool.d.ts +11 -44
  91. package/dist/tools/ask-agentic.tool.d.ts.map +1 -1
  92. package/dist/tools/ask-agentic.tool.js +143 -1618
  93. package/dist/tools/ask-agentic.tool.js.map +1 -1
  94. package/dist/tools/ask.tool.d.ts.map +1 -1
  95. package/dist/tools/ask.tool.js +195 -555
  96. package/dist/tools/ask.tool.js.map +1 -1
  97. package/dist/tools/code.tool.d.ts +6 -0
  98. package/dist/tools/code.tool.d.ts.map +1 -1
  99. package/dist/tools/code.tool.js +155 -306
  100. package/dist/tools/code.tool.js.map +1 -1
  101. package/dist/tools/shared/abort-timeout.d.ts +48 -4
  102. package/dist/tools/shared/abort-timeout.d.ts.map +1 -1
  103. package/dist/tools/shared/abort-timeout.js +116 -95
  104. package/dist/tools/shared/abort-timeout.js.map +1 -1
  105. package/dist/tools/shared/call-ledger.d.ts +88 -0
  106. package/dist/tools/shared/call-ledger.d.ts.map +1 -0
  107. package/dist/tools/shared/call-ledger.js +133 -0
  108. package/dist/tools/shared/call-ledger.js.map +1 -0
  109. package/dist/tools/shared/context-decode.d.ts +77 -0
  110. package/dist/tools/shared/context-decode.d.ts.map +1 -0
  111. package/dist/tools/shared/context-decode.js +125 -0
  112. package/dist/tools/shared/context-decode.js.map +1 -0
  113. package/dist/tools/shared/continuation.d.ts +124 -0
  114. package/dist/tools/shared/continuation.d.ts.map +1 -0
  115. package/dist/tools/shared/continuation.js +302 -0
  116. package/dist/tools/shared/continuation.js.map +1 -0
  117. package/dist/tools/shared/dispatch-throttle.d.ts +43 -0
  118. package/dist/tools/shared/dispatch-throttle.d.ts.map +1 -0
  119. package/dist/tools/shared/dispatch-throttle.js +65 -0
  120. package/dist/tools/shared/dispatch-throttle.js.map +1 -0
  121. package/dist/tools/shared/question-fields.d.ts +10 -0
  122. package/dist/tools/shared/question-fields.d.ts.map +1 -0
  123. package/dist/tools/shared/question-fields.js +13 -0
  124. package/dist/tools/shared/question-fields.js.map +1 -0
  125. package/dist/tools/shared/settle-call.d.ts +68 -0
  126. package/dist/tools/shared/settle-call.d.ts.map +1 -0
  127. package/dist/tools/shared/settle-call.js +59 -0
  128. package/dist/tools/shared/settle-call.js.map +1 -0
  129. package/dist/tools/shared/stream-collector.d.ts +7 -0
  130. package/dist/tools/shared/stream-collector.d.ts.map +1 -1
  131. package/dist/tools/shared/stream-collector.js +5 -1
  132. package/dist/tools/shared/stream-collector.js.map +1 -1
  133. package/dist/tools/shared/thinking.d.ts +41 -11
  134. package/dist/tools/shared/thinking.d.ts.map +1 -1
  135. package/dist/tools/shared/thinking.js +62 -9
  136. package/dist/tools/shared/thinking.js.map +1 -1
  137. package/dist/tools/shared/throttle.d.ts +7 -0
  138. package/dist/tools/shared/throttle.d.ts.map +1 -1
  139. package/dist/tools/shared/throttle.js +18 -0
  140. package/dist/tools/shared/throttle.js.map +1 -1
  141. package/dist/types.d.ts +2 -0
  142. package/dist/types.d.ts.map +1 -1
  143. package/dist/utils/cost-estimator.d.ts +17 -0
  144. package/dist/utils/cost-estimator.d.ts.map +1 -1
  145. package/dist/utils/cost-estimator.js +15 -2
  146. package/dist/utils/cost-estimator.js.map +1 -1
  147. package/package.json +4 -3
@@ -10,24 +10,27 @@ import { createHash } from 'node:crypto';
10
10
  import { resolve } from 'node:path';
11
11
  import { z } from 'zod';
12
12
  import { AttachmentError, AttachmentTokensUncountedError, INLINE_REQUEST_LIMIT_BYTES, MAX_ATTACHMENTS, MAX_ATTACHMENTS_TOTAL_BYTES, MAX_ATTACHMENT_BYTES, countAttachmentTokens, inlineRequestBytes, inspectAttachments, readAttachments, } from '../attachments.js';
13
- import { isStaleCacheError, markCacheStale, prepareContext } from '../cache/cache-manager.js';
13
+ import { prepareContext } from '../cache/cache-manager.js';
14
14
  import { CACHING_MODE_FALLBACK_DESCRIPTION, cachingModeFor, cachingRan, } from '../cache/caching-mode.js';
15
+ import { capThatRuns, preflightModeFor, resolveOutputCap, thinkingBudgetRefusal, } from '../gemini/long-decode.js';
15
16
  import { resolveModel } from '../gemini/models.js';
16
- import { abortableSleep, withNetworkRetry } from '../gemini/retry.js';
17
17
  import { countForPreflight } from '../gemini/token-counter.js';
18
18
  import { buildScanMemo, scanWorkspace } from '../indexer/workspace-scanner.js';
19
19
  import { WorkspaceValidationError, validateWorkspacePath, } from '../indexer/workspace-validation.js';
20
- import { estimateCostUsd, estimatePreCallCostUsd, toMicrosUsd } from '../utils/cost-estimator.js';
20
+ import { estimatePreCallCostUsd, toMicrosUsd } from '../utils/cost-estimator.js';
21
21
  import { logger, safeForLog } from '../utils/logger.js';
22
22
  import { createProgressEmitter } from '../utils/progress.js';
23
- import { askAgenticTool } from './ask-agentic.tool.js';
23
+ import { fallBackToAgentic } from './ask-agentic-fallback.js';
24
24
  import { errorResult, textResult } from './registry.js';
25
- import { createTimeoutController, getTimeoutKind, isTimeoutAbort } from './shared/abort-timeout.js';
25
+ import { isTimeoutAbort, timedOutCall, toolTimeoutController } from './shared/abort-timeout.js';
26
+ import { decodeOnContext } from './shared/context-decode.js';
27
+ import { CompletedRequests, LONG_DECODE_CAP_DESCRIPTION } from './shared/continuation.js';
28
+ import { DispatchThrottle } from './shared/dispatch-throttle.js';
29
+ import { PROMPT_FIELD, WORKSPACE_FIELD } from './shared/question-fields.js';
26
30
  import { SERVICE_TIER_DESCRIPTION, ServiceTierError, networkAttempts, resolveServiceTier, serviceTierConfig, statusOf, tierErrorMeta, } from './shared/service-tier.js';
27
- import { collectStream } from './shared/stream-collector.js';
31
+ import { settleAnswer, settleFailure } from './shared/settle-call.js';
28
32
  import { SYSTEM_INSTRUCTION_SAFETY_EAGER } from './shared/system-instruction-safety.js';
29
- import { THINKING_LEVEL_FIELD, THINKING_LEVEL_RESERVE, emptyAtOutputCap, } from './shared/thinking.js';
30
- import { isGemini429, parseRetryDelayMs } from './shared/throttle.js';
33
+ import { THINKING_LEVEL_FIELD, thinkingBudgetUnder, thinkingReserve } from './shared/thinking.js';
31
34
  /**
32
35
  * `ask` system instruction. Prepends the shared `SYSTEM_INSTRUCTION_SAFETY_EAGER`
33
36
  * data-vs-instruction firewall (added in v1.15.2 — pre-v1.15.2 `ask` had no
@@ -72,11 +75,8 @@ const THINKING_BUDGET_MODEL_DEFAULT = null;
72
75
  // publishes per-tier token budgets, one edit propagates to both tools.
73
76
  export const askInputSchema = z
74
77
  .object({
75
- prompt: z.string().min(1).describe('The question or analysis request.'),
76
- workspace: z
77
- .string()
78
- .optional()
79
- .describe('Absolute or cwd-relative path to the workspace. Defaults to process.cwd().'),
78
+ prompt: PROMPT_FIELD,
79
+ workspace: WORKSPACE_FIELD,
80
80
  model: z
81
81
  .string()
82
82
  .optional()
@@ -106,14 +106,17 @@ export const askInputSchema = z
106
106
  .int()
107
107
  .min(1)
108
108
  .optional()
109
- .describe("Per-call opt-in cap on response length (tokens). OMIT for the default 'auto' behaviour — Gemini uses its model-default cap (per Google docs, equal to the model's advertised `outputTokenLimit`: 65,536 for Gemini 3.x / 2.5 Pro-tier; see ai.google.dev/gemini-api/docs/models/gemini-2.5-pro). Pass a smaller value when you want a bounded response (e.g. strict summary length, tighter budget per call). Values larger than the resolved model's limit are clamped down. Budget reservation always uses the effective cap (explicit OR model limit) as worst-case, so `GEMINI_DAILY_BUDGET_USD` stays a true upper bound."),
109
+ .describe([
110
+ "Per-call opt-in cap on response length (tokens). OMIT for the default 'auto' behaviour — Gemini uses its model-default cap (per Google docs, equal to the model's advertised `outputTokenLimit`: 65,536 for Gemini 3.x / 2.5 Pro-tier; see ai.google.dev/gemini-api/docs/models/gemini-2.5-pro). Pass a smaller value when you want a bounded response (e.g. strict summary length, tighter budget per call). On any other model a value larger than the resolved model's limit is clamped down to it. Budget reservation always uses the effective cap (explicit OR model limit) as worst-case, so `GEMINI_DAILY_BUDGET_USD` bounds what a completed call books (a request aborted mid-stream is booked 0: docs/cost-model.md).",
111
+ LONG_DECODE_CAP_DESCRIPTION,
112
+ ].join(' ')),
110
113
  timeoutMs: z
111
114
  .number()
112
115
  .int()
113
116
  .min(1_000)
114
117
  .max(1_800_000)
115
118
  .optional()
116
- .describe('Per-call wall-clock TOTAL timeout in ms (1s–30min). The cost ceiling — aborts the in-flight `generateContent` request even if it is actively streaming. Gemini still finishes server-side and bills tokens for completed work. When omitted, falls back to env var `GEMINI_CODE_CONTEXT_ASK_TIMEOUT_MS`, then to disabled. v1.12.0 RECOMMENDATION: prefer `stallMs` for liveness — `timeoutMs` is the cost cap, `stallMs` is the heartbeat-aware liveness watchdog that does NOT fire while the model is actively thinking. Both can be set simultaneously. Returns `errorCode: "TIMEOUT"` with `timeoutKind: "total"` on abort.'),
119
+ .describe('Per-call wall-clock TOTAL timeout in ms (1s–30min). The cost ceiling — aborts the in-flight `generateContent` request even if it is actively streaming. A request aborted after its stream opened is booked 0; whether Google bills it is not measured (docs/cost-model.md). When omitted, falls back to env var `GEMINI_CODE_CONTEXT_ASK_TIMEOUT_MS`, then to disabled. v1.12.0 RECOMMENDATION: prefer `stallMs` for liveness — `timeoutMs` is the cost cap, `stallMs` is the heartbeat-aware liveness watchdog that does NOT fire while the model is actively thinking. Both can be set simultaneously. Returns `errorCode: "TIMEOUT"` with `timeoutKind: "total"` on abort.'),
117
120
  stallMs: z
118
121
  .number()
119
122
  .int()
@@ -141,7 +144,7 @@ export const askInputSchema = z
141
144
  onWorkspaceTooLarge: z
142
145
  .enum(['error', 'fallback-to-agentic'])
143
146
  .optional()
144
- .describe("Behaviour when the v1.5.0 preflight detects the workspace exceeds the model's `inputTokenLimit × workspaceGuardRatio` (v1.11.0+). `'error'` (default — preserves v1.5.0 behaviour) returns a structured `errorCode: 'WORKSPACE_TOO_LARGE'` and lets the caller decide how to recover. `'fallback-to-agentic'` opt-in: the server transparently re-routes through `ask_agentic` (sandboxed file-access loop — no eager upload, scales to arbitrarily large repos), wraps the agentic result, and surfaces `fallbackApplied: 'ask_agentic'` in `structuredContent` so callers can audit the swap. **Semantic divergence:** `ask_agentic` runs as a multi-iteration loop, so `timeoutMs` (ask's wall-clock cap) translates to `iterationTimeoutMs` (per-iteration cap); total wall-clock can be N × timeoutMs. Cost shape also changes (multiple smaller calls vs one big eager call). Recommended setting for orchestrators that prioritise getting an answer over a single round-trip; default stays `'error'` so existing callers see no behaviour change. **`code` does NOT support this field** — `code`'s OLD/NEW edit format is load-bearing for Claude's Edit pipeline, and `ask_agentic` returns prose only."),
147
+ .describe("Behaviour when the v1.5.0 preflight detects the workspace exceeds the model's `inputTokenLimit × workspaceGuardRatio` (v1.11.0+). `'error'` (default — preserves v1.5.0 behaviour) returns a structured `errorCode: 'WORKSPACE_TOO_LARGE'` and lets the caller decide how to recover. `'fallback-to-agentic'` opt-in: the server transparently re-routes through `ask_agentic` (sandboxed file-access loop — no eager upload, scales to arbitrarily large repos), wraps the agentic result, and surfaces `fallbackApplied: 'ask_agentic'` in `structuredContent` so callers can audit the swap. **Semantic divergence:** `ask_agentic` runs as a multi-iteration loop, so `timeoutMs` (ask's wall-clock cap, per call or `GEMINI_CODE_CONTEXT_ASK_TIMEOUT_MS`) translates to `iterationTimeoutMs` (per-iteration cap; total wall-clock can be N × timeoutMs) and `stallMs` (per call or `GEMINI_CODE_CONTEXT_ASK_STALL_MS`) to the loop's `stallMs` (a stream's silence only — a tool batch longer than `stallMs` answers); a phase neither bounds ends at `ask_agentic`'s default (600 s for a request's opening and a silent stream, 120 s without progress in an iteration's tools). The result reports `iterTimeoutMs`, `iterTimeoutSource: 'timeoutMs'` and `stallMs`, and lifts `errorCode`, `retryable`, `timeoutKind`, `timeoutPhase` and `timeoutSource` of a failed agentic call. Cost shape also changes (multiple smaller calls vs one big eager call). Recommended setting for orchestrators that prioritise getting an answer over a single round-trip; default stays `'error'` so existing callers see no behaviour change. **`code` does NOT support this field** — `code`'s OLD/NEW edit format is load-bearing for Claude's Edit pipeline, and `ask_agentic` returns prose only."),
145
148
  })
146
149
  .refine((data) => !(data.thinkingBudget !== undefined && data.thinkingLevel !== undefined), {
147
150
  message: 'Cannot specify both `thinkingBudget` and `thinkingLevel` — they are mutually exclusive. Gemini rejects the combination with 400. Choose one: `thinkingLevel` (recommended for Gemini 3) or `thinkingBudget` (required on Gemini 2.5).',
@@ -237,16 +240,15 @@ function attachmentMetadata(prepared) {
237
240
  }
238
241
  async function executeAskBody(input, ctx, workspaceRoot, model, started) {
239
242
  let reservationId = null;
240
- // `-1` signals "no reservation held" (distinct from 0 which is a valid
241
- // releaseId). Mirrors the pattern used by `reservationId`.
242
- let throttleReservationId = -1;
243
243
  // Canonical resolved-model string, captured once after `resolveModel`
244
244
  // so the outer catch can feed retry hints into the same throttle bucket
245
245
  // `reserve` used. `model` at the top of execute is the request alias
246
246
  // ("latest-pro-thinking") — different key, different bucket.
247
247
  let resolvedModelKey = null;
248
- // once a stream opened, a later transport failure is mid-stream: billed work, not to be re-sent
249
- let streamOpened = false;
248
+ // every request that completed is billed: the ledger settles from these, on an answer and on a failure alike
249
+ const completed = new CompletedRequests();
250
+ // the running decode's booking and whether a stream opened (a later failure is then billed work, not re-sent)
251
+ const tracking = { ledgerEntry: null, streamOpened: false };
250
252
  // the tier this call runs on once resolved (Vertex may force standard); before that, the requested one is reported
251
253
  let serviceTier;
252
254
  const reportedTier = () => serviceTier ?? input.serviceTier ?? ctx.config.serviceTier;
@@ -258,13 +260,15 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
258
260
  // SDK's `config.abortSignal` — whichever timer fires first aborts both
259
261
  // layers cleanly. `recordChunk()` is wired into `collectStream` to reset
260
262
  // the stall watchdog on every chunk arrival.
261
- const timeoutController = createTimeoutController({
262
- ...(input.timeoutMs !== undefined ? { totalMs: input.timeoutMs } : {}),
263
- totalEnvVar: 'GEMINI_CODE_CONTEXT_ASK_TIMEOUT_MS',
264
- ...(input.stallMs !== undefined ? { stallMs: input.stallMs } : {}),
265
- stallEnvVar: 'GEMINI_CODE_CONTEXT_ASK_STALL_MS',
266
- });
263
+ const timeoutController = toolTimeoutController('ask', input);
267
264
  const abortSignal = timeoutController.signal;
265
+ // Each request of the decode holds its own TPM reservation; a failure drops the one held.
266
+ const dispatch = new DispatchThrottle({
267
+ throttle: ctx.throttle,
268
+ enabled: ctx.config.tpmThrottleLimit > 0,
269
+ signal: abortSignal,
270
+ emit: (message) => emitter.emit(message),
271
+ });
268
272
  try {
269
273
  // Workspace path validation lives INSIDE the try so a
270
274
  // `WorkspaceValidationError` is reported as a regular tool error
@@ -325,97 +329,38 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
325
329
  manifestMemo: scanMemo,
326
330
  forceRescan: input.forceRescan === true || ctx.config.forceRescan === true,
327
331
  });
328
- // Output-cap strategy (v1.4.0): three-layer precedence.
332
+ // Output cap, three layers (`resolveOutputCap`):
329
333
  //
330
- // 1. `input.maxOutputTokens` (per-call, strongest) — user explicitly
331
- // caps this one call. Clamped to `modelOutputLimit`.
332
- // 2. `ctx.config.forceMaxOutputTokens` (env override) — when the
333
- // MCP host sets `GEMINI_CODE_CONTEXT_FORCE_MAX_OUTPUT=true`,
334
- // every call sends `maxOutputTokens = modelOutputLimit` so
335
- // code-review workloads always run at the model's full capacity
336
- // (65,536 tokens for Gemini 3.x / 2.5 Pro per Google docs).
337
- // 3. Default (no overrides) — omit `maxOutputTokens` from the
338
- // `generateContent` config. Gemini uses its model-default cap
339
- // (documented as the model's advertised `outputTokenLimit`),
340
- // letting the model size the response to the query's complexity
341
- // rather than reserving full capacity on short Q&A.
334
+ // 1. `input.maxOutputTokens` (per-call, strongest) — clamped to the
335
+ // model's per-request limit, or on a model with long decode
336
+ // (Gemini 4) to its long-decode total: a cap above the per-request
337
+ // limit continues the decode in further requests.
338
+ // 2. `ctx.config.forceMaxOutputTokens` (env override) — every call
339
+ // sends the model's per-request limit.
340
+ // 3. Default — `maxOutputTokens` is omitted: one request, the model's
341
+ // own default (its advertised `outputTokenLimit`), letting the model
342
+ // size the response to the query rather than reserving full capacity.
342
343
  //
343
- // `modelOutputLimit` is the model's advertised ceiling from
344
- // `models.list()` (or a fallback if the SDK doesn't report one).
345
- // Used as the clamp target and as the budget-reservation worst-case.
346
- const modelOutputLimit = typeof resolved.outputTokenLimit === 'number' && resolved.outputTokenLimit > 0
347
- ? resolved.outputTokenLimit
348
- : ASK_MAX_OUTPUT_TOKENS_FALLBACK;
349
- const wireMaxOutputTokens = input.maxOutputTokens !== undefined
350
- ? Math.min(input.maxOutputTokens, modelOutputLimit)
351
- : ctx.config.forceMaxOutputTokens
352
- ? modelOutputLimit
353
- : undefined;
354
- // Effective cap used for thinking-budget clamp and budget reservation.
355
- // When neither override is set, the model's internal default equals
356
- // `modelOutputLimit` per Google docs — so using the limit as
357
- // worst-case keeps `GEMINI_DAILY_BUDGET_USD` a true upper bound.
358
- const effectiveOutputCap = wireMaxOutputTokens ?? modelOutputLimit;
344
+ // The caller's `thinkingBudget` is checked against `outputCap`; the
345
+ // clamp and the reservation use the cap that runs (`runCap`).
346
+ const outputCap = resolveOutputCap({
347
+ requested: input.maxOutputTokens,
348
+ forceMax: ctx.config.forceMaxOutputTokens,
349
+ outputTokenLimit: resolved.outputTokenLimit,
350
+ longDecodeTokenLimit: resolved.longDecodeTokenLimit,
351
+ fallbackLimit: ASK_MAX_OUTPUT_TOKENS_FALLBACK,
352
+ });
359
353
  // Single source of truth for "is the caller driving reasoning via the
360
354
  // discrete-tier knob?". Three downstream call sites branch on this
361
355
  // (cost estimate, emitter, thinkingConfig build) — extracting to a
362
356
  // named const keeps them in lockstep and makes future additions hard
363
357
  // to forget.
364
358
  const usingThinkingLevel = input.thinkingLevel !== undefined;
365
- // Gemini requires `thinkingBudget < maxOutputTokens` (the thinking pool
366
- // is carved out of the candidate-output allowance). Reject the call
367
- // early when the caller's `maxOutputTokens` leaves no headroom for a
368
- // positive `thinkingBudget`, rather than silently clamping to 0 —
369
- // clamp-to-0 would make Gemini 3 Pro 400 ("thinking disabled"
370
- // rejected) and mislead the caller about the actual cause (PR #22
371
- // round-3 review finding #C). Minimum answer reserve: 1024 tokens.
372
- if (input.thinkingBudget !== undefined &&
373
- input.thinkingBudget > 0 &&
374
- input.maxOutputTokens !== undefined &&
375
- effectiveOutputCap < input.thinkingBudget + 1024) {
376
- return errorResult(`ask: thinkingBudget (${input.thinkingBudget}) + 1024-token answer reserve exceeds maxOutputTokens (${effectiveOutputCap}). Raise \`maxOutputTokens\` to at least ${input.thinkingBudget + 1024}, or lower \`thinkingBudget\`.`);
377
- }
378
- // Normalise thinkingBudget:
379
- // undefined → null (model-default path — `thinkingBudget` OMITTED on wire)
380
- // -1 → -1 (explicit legacy dynamic; kept for Gemini 2.5 and
381
- // power-users who want to force the legacy path)
382
- // 0 → 0 (thinking disabled; Gemini 3 Pro will 400 here)
383
- // N > 0 → clamp to [0, maxOutputTokens - 1024] — the thinking
384
- // pool is carved out of the candidate-output allowance,
385
- // so reserving ≥1024 tokens for the answer prevents 400s
386
- // when a caller passes N ≥ maxOutputTokens.
387
- //
388
- // Schema `.refine()` guarantees `thinkingBudget` and `thinkingLevel`
389
- // are never both set, so we handle them in two mutually-exclusive
390
- // branches below.
391
- const rawThinkingBudget = input.thinkingBudget;
392
- const effectiveThinkingBudget = rawThinkingBudget === undefined
393
- ? THINKING_BUDGET_MODEL_DEFAULT
394
- : rawThinkingBudget === -1
395
- ? -1
396
- : rawThinkingBudget > 0
397
- ? Math.max(0, Math.min(rawThinkingBudget, effectiveOutputCap - 1024))
398
- : 0;
399
- // Cost-estimate thinking-token reserve — tier-aware when the caller
400
- // uses `thinkingLevel`, worst-case when they use `thinkingBudget: -1`
401
- // or omit both fields (Gemini 3 Pro always spends thinking tokens per
402
- // Google's docs, so we reserve dynamic-thinking headroom there too).
403
- // When the caller passes an explicit positive `thinkingBudget`, we
404
- // reserve only that much. Keeps `GEMINI_DAILY_BUDGET_USD` a TRUE
405
- // upper bound on completed spend without false-rejecting callers
406
- // who legitimately use a low tier (see `THINKING_LEVEL_RESERVE`
407
- // rationale above).
408
- //
409
- // We dereference `input.thinkingLevel` directly (rather than via
410
- // `usingThinkingLevel`) so TypeScript narrows the type inside the
411
- // ternary — `THINKING_LEVEL_RESERVE[undefined]` would be a compile
412
- // error. The `?? worst-case` fallback handles the HIGH tier (which
413
- // maps to `null` in the reserve table to mean "use the dynamic cap").
414
- const thinkingTokensForEstimate = input.thinkingLevel !== undefined
415
- ? (THINKING_LEVEL_RESERVE[input.thinkingLevel] ?? Math.max(0, effectiveOutputCap - 1024))
416
- : effectiveThinkingBudget === null || effectiveThinkingBudget === -1
417
- ? Math.max(0, effectiveOutputCap - 1024)
418
- : effectiveThinkingBudget;
359
+ // A positive `thinkingBudget` that leaves no answer reserve under the caller's own cap is refused
360
+ // (`thinkingBudgetRefusal`); below, the budget on the wire is clamped under the cap that runs.
361
+ const budgetRefusal = thinkingBudgetRefusal(input.thinkingBudget, input.maxOutputTokens, outputCap.effective);
362
+ if (budgetRefusal !== null)
363
+ return errorResult(`ask: ${budgetRefusal}`);
419
364
  // Estimated input-token count drives both the daily-budget reservation
420
365
  // (dollars) and the TPM throttle reservation (rate). Hoisted here so
421
366
  // both consumers share one fingerprint — no drift between what the $
@@ -454,6 +399,8 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
454
399
  // model has no advertised `inputTokenLimit`.
455
400
  let preflight;
456
401
  const contextWindow = resolved.inputTokenLimit;
402
+ // A decode that may take more than one request takes the exact count: its prompt is billed per request.
403
+ const preflightMode = preflightModeFor(input.preflightMode, outputCap);
457
404
  if (typeof contextWindow === 'number' && contextWindow > 0) {
458
405
  preflight = await countForPreflight(ctx.client, {
459
406
  files: scan.files,
@@ -463,7 +410,7 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
463
410
  // already encodes user-supplied includeGlobs/excludeGlobs — no
464
411
  // separate globsHash axis needed in the cache key.
465
412
  filesHash: scan.filesHash,
466
- ...(input.preflightMode !== undefined ? { preflightMode: input.preflightMode } : {}),
413
+ ...(preflightMode !== undefined ? { preflightMode } : {}),
467
414
  inputTokenLimit: contextWindow,
468
415
  extraTokens: attachmentTokens, // weighs on the heuristic-vs-exact decision, added to the count below
469
416
  // Thread the user's `timeoutMs` AbortSignal so a hung countTokens
@@ -484,9 +431,10 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
484
431
  // 'fallback-to-agentic'`. Default stays `'error'` so pre-v1.11.0
485
432
  // callers see identical behaviour. The fallback re-uses this
486
433
  // tool's resolved inputs (model / globs / thinking config /
487
- // workspace) and re-translates `timeoutMs` to `iterationTimeoutMs`
488
- // (per-iteration cap; total wall-clock can be N × that). Cost +
489
- // timing shape genuinely changes — opt-in is the right default.
434
+ // workspace) and forwards its resolved `timeoutMs` as `iterationTimeoutMs`
435
+ // (per-iteration cap; total wall-clock can be N × that) and `stallMs` as
436
+ // `stallMs` (`ask-agentic-fallback.ts`). Cost + timing shape genuinely
437
+ // changes — opt-in is the right default.
490
438
  if (input.onWorkspaceTooLarge === 'fallback-to-agentic' &&
491
439
  attachments.attached.length > 0) {
492
440
  return errorResult('ask: the workspace is too large for the eager path and `attachments` cannot follow the ask_agentic fallback (it sends no inline parts); drop the attachments or narrow the workspace', {
@@ -497,152 +445,15 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
497
445
  });
498
446
  }
499
447
  if (input.onWorkspaceTooLarge === 'fallback-to-agentic') {
500
- if (serviceTier === 'flex') {
501
- const note = 'ask: the ask_agentic fallback has no service tier — this call runs standard, not flex';
502
- emitter.emit(note);
503
- logger.warn(note);
504
- }
505
- logger.warn(`ask: WORKSPACE_TOO_LARGE (${preflight.effectiveTokens} > ${threshold}); falling back to ask_agentic per onWorkspaceTooLarge='fallback-to-agentic'`);
506
- // Translate `timeoutMs` AND `stallMs` to the agentic per-iteration
507
- // wall-clock cap (`iterationTimeoutMs`). Pre-v1.12.1 only `timeoutMs`
508
- // was forwarded — a user who set `stallMs: 60000` (the v1.12.0
509
- // recommended liveness watchdog) saw it silently dropped on the
510
- // fallback path. The agentic loop uses non-streaming
511
- // `generateContent`, so there's no chunk stream to reset on; we
512
- // collapse both knobs onto the per-iteration cap and use the
513
- // TIGHTER of the two as the bound. The CHANGELOG documents the
514
- // semantic divergence on `timeoutMs` (total → per-iteration).
515
- //
516
- // v1.12.2 — `iterTimeoutSource` is surfaced in metadata AND emitted
517
- // in the warn-log so users can see WHICH knob bound their call
518
- // (one of: 'timeoutMs', 'stallMs', or 'min(timeoutMs,stallMs)').
519
- // Pre-v1.12.2 the log only fired when stallMs was set alone; the
520
- // both-set case had no signal explaining the pick.
521
- const iterTimeoutPick = (() => {
522
- const a = input.timeoutMs;
523
- const b = input.stallMs;
524
- if (a !== undefined && b !== undefined) {
525
- const cap = Math.min(a, b);
526
- return { cap, source: 'min(timeoutMs,stallMs)' };
527
- }
528
- if (a !== undefined)
529
- return { cap: a, source: 'timeoutMs' };
530
- if (b !== undefined)
531
- return { cap: b, source: 'stallMs' };
532
- return { cap: undefined, source: 'none' };
533
- })();
534
- const tightestPerIterCap = iterTimeoutPick.cap;
535
- if (iterTimeoutPick.source !== 'none') {
536
- logger.warn(`ask: fallback to ask_agentic — \`iterationTimeoutMs\` set to ${tightestPerIterCap}ms (source: ${iterTimeoutPick.source}). The agentic loop uses non-streaming \`generateContent\` so there's no chunk stream to reset on; \`stallMs\` (when set) is treated as a per-iteration wall-clock cap, the same way \`timeoutMs\` is. Total wall-clock for the agentic call can reach \`maxIterations × iterationTimeoutMs\`.`);
537
- }
538
- const agenticInput = {
539
- prompt: input.prompt,
540
- ...(input.workspace !== undefined ? { workspace: input.workspace } : {}),
541
- ...(input.model !== undefined ? { model: input.model } : {}),
542
- ...(input.includeGlobs !== undefined ? { includeGlobs: input.includeGlobs } : {}),
543
- ...(input.excludeGlobs !== undefined ? { excludeGlobs: input.excludeGlobs } : {}),
544
- ...(input.thinkingBudget !== undefined ? { thinkingBudget: input.thinkingBudget } : {}),
545
- ...(input.thinkingLevel !== undefined ? { thinkingLevel: input.thinkingLevel } : {}),
546
- ...(input.maxOutputTokens !== undefined
547
- ? { maxOutputTokens: input.maxOutputTokens }
548
- : {}),
549
- // `timeoutMs` (ask wall-clock cap) → `iterationTimeoutMs`
550
- // (ask_agentic per-iteration cap). Total wall-clock for the
551
- // agentic call can be `maxIterations × iterationTimeoutMs`.
552
- // v1.12.1 fix: also collapses `stallMs` onto this cap (using
553
- // the TIGHTER of the two when both are set), so the fallback
554
- // path doesn't silently drop the user's stall watchdog.
555
- ...(tightestPerIterCap !== undefined ? { iterationTimeoutMs: tightestPerIterCap } : {}),
556
- };
557
- const agenticResult = await askAgenticTool.execute(agenticInput, ctx);
558
- // Wrap the agentic result so callers can audit the swap.
559
- // The text content (`content[0].text`) is the agentic prose
560
- // verbatim — same shape any direct `ask_agentic` caller would
561
- // see. The structuredContent is enriched with fallback-trail
562
- // metadata so orchestrators can distinguish a fallback-served
563
- // response from a direct one.
564
- //
565
- // Two load-bearing contracts the wrapper must preserve:
566
- // 1. `structuredContent.responseText` — the T23 wire-format
567
- // invariant (`textResult` / `errorResult` always write it
568
- // LAST; sub-agent orchestrators extract from this key
569
- // ONLY). We pull it from the agentic result's
570
- // `structuredContent.responseText` (or the prose text from
571
- // `content[0].text` as a fallback) and write it LAST so
572
- // it can't be clobbered.
573
- // 2. Top-level `errorCode` / `retryable` — orchestrator retry
574
- // policies switch on these. When the agentic call itself
575
- // fails (e.g. iteration budget exhausted), the failure
576
- // code lives in `agenticResult.structuredContent.errorCode`;
577
- // we lift it to the top level of `wrappedStructured` so
578
- // consumer policies see it without descending into nested
579
- // `agenticResult.errorCode` (which they wouldn't know
580
- // about pre-v1.11.0).
581
- const agenticStructured = agenticResult.structuredContent ?? {};
582
- const promptResponseText = typeof agenticStructured.responseText === 'string'
583
- ? agenticStructured.responseText
584
- : agenticResult.content?.[0]?.type === 'text'
585
- ? agenticResult.content[0].text
586
- : '';
587
- const wrappedStructured = {
588
- fallbackApplied: 'ask_agentic',
589
- serviceTier: 'standard', // ask_agentic has no tier: the fallback runs standard whatever was asked
590
- ...(serviceTier === 'flex' ? { serviceTierDowngraded: 'flex' } : {}),
591
- fallbackReason: 'WORKSPACE_TOO_LARGE',
592
- preflightEstimate: {
593
- tokens: preflight.effectiveTokens,
594
- threshold,
595
- source: preflight.method,
596
- cacheHit: preflight.cacheHit,
597
- rawTokens: preflight.rawTokens,
598
- },
599
- // v1.12.2 — surface the per-iteration cap selection so callers
600
- // can audit which knob bound the fallback call. Useful when the
601
- // user passed both `timeoutMs` and `stallMs` (we use the tighter
602
- // of the two — `iterTimeoutSource: 'min(timeoutMs,stallMs)'`).
603
- // Omitted when neither knob was set (`source: 'none'`).
604
- ...(iterTimeoutPick.source !== 'none'
605
- ? {
606
- iterTimeoutMs: tightestPerIterCap,
607
- iterTimeoutSource: iterTimeoutPick.source,
608
- }
609
- : {}),
610
- // Lift top-level error metadata from the agentic call's
611
- // structuredContent so orchestrator policies (e.g. retry on
612
- // BUDGET_EXHAUSTED, retry on stall but not total) keep working
613
- // without descending into `.agenticResult.errorCode`. v1.12.1
614
- // adds `timeoutKind` to the lift set — Phase 4 introduced
615
- // `timeoutKind` as a top-level field on direct-path TIMEOUT
616
- // errors; the fallback wrapper now mirrors that on the
617
- // fallback-served TIMEOUT path so the contract is uniform.
618
- ...(agenticResult.isError && typeof agenticStructured.errorCode === 'string'
619
- ? { errorCode: agenticStructured.errorCode }
620
- : {}),
621
- ...(agenticResult.isError && typeof agenticStructured.retryable === 'boolean'
622
- ? { retryable: agenticStructured.retryable }
623
- : {}),
624
- ...(agenticResult.isError &&
625
- (agenticStructured.timeoutKind === 'total' || agenticStructured.timeoutKind === 'stall')
626
- ? { timeoutKind: agenticStructured.timeoutKind }
627
- : {}),
628
- ...(Object.keys(agenticStructured).length > 0
629
- ? { agenticResult: agenticStructured }
630
- : {}),
631
- // T23 wire-format invariant — `responseText` is written LAST
632
- // so any caller-supplied / nested key with the same name is
633
- // intentionally clobbered by the canonical value. Sub-agent
634
- // orchestrators read the prose from this exact key.
635
- responseText: promptResponseText,
636
- };
637
- // `isError` lives at the ROOT of the MCP CallToolResult per spec
638
- // (sibling to `structuredContent`, not nested inside it). The
639
- // spread on the return statement carries the agentic root-level
640
- // `isError` automatically — no need to mirror it inside
641
- // `structuredContent`.
642
- return {
643
- ...agenticResult,
644
- structuredContent: wrappedStructured,
645
- };
448
+ return fallBackToAgentic({
449
+ input,
450
+ ctx,
451
+ emitter,
452
+ serviceTier,
453
+ preflight,
454
+ threshold,
455
+ timeout: timeoutController,
456
+ });
646
457
  }
647
458
  const attachmentsNote = attachments.attached.length > 0
648
459
  ? ' — it carries no attachments: drop them, or narrow the workspace and ask again'
@@ -667,6 +478,38 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
667
478
  else {
668
479
  logger.warn(`ask: resolved model ${safeForLog(resolved.resolved)} has no advertised inputTokenLimit — workspace size guard skipped. Request may fail downstream if the workspace exceeds the model's context window.`);
669
480
  }
481
+ const { cap: runCap, cut: inputWindowCut } = capThatRuns({
482
+ tool: 'ask',
483
+ cap: outputCap,
484
+ promptTokens: preflight?.effectiveTokens,
485
+ inputTokenLimit: contextWindow,
486
+ emit: (message) => emitter.emit(message),
487
+ });
488
+ // Normalise thinkingBudget:
489
+ // undefined → null (model-default path — `thinkingBudget` OMITTED on wire)
490
+ // -1 → -1 (explicit legacy dynamic; kept for Gemini 2.5 and
491
+ // power-users who want to force the legacy path)
492
+ // 0 → 0 (thinking disabled; Gemini 3 Pro will 400 here)
493
+ // N > 0 → clamp to [0, cap that runs - 1024] — the thinking
494
+ // pool is carved out of the candidate-output allowance,
495
+ // so reserving ≥1024 tokens for the answer prevents 400s
496
+ // when a caller passes N ≥ maxOutputTokens, and under a
497
+ // cap the input window cut to one request (its value on
498
+ // the wire) the budget stays below it.
499
+ //
500
+ // Schema `.refine()` guarantees `thinkingBudget` and `thinkingLevel`
501
+ // are never both set, so we handle them in two mutually-exclusive
502
+ // branches below.
503
+ const rawThinkingBudget = input.thinkingBudget;
504
+ const effectiveThinkingBudget = rawThinkingBudget === undefined
505
+ ? THINKING_BUDGET_MODEL_DEFAULT
506
+ : rawThinkingBudget === -1
507
+ ? -1
508
+ : rawThinkingBudget > 0
509
+ ? thinkingBudgetUnder(runCap, rawThinkingBudget)
510
+ : 0;
511
+ // The thinking the budget reservation counts, from the cap that runs.
512
+ const thinkingTokensForEstimate = thinkingReserve(input.thinkingLevel, effectiveThinkingBudget, runCap);
670
513
  // Atomic budget reservation. Happens AFTER scan (so the estimate is
671
514
  // accurate) but BEFORE any billable upload or generateContent call.
672
515
  // Concurrent tool calls cannot all pass a pre-check and then
@@ -683,7 +526,11 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
683
526
  // OR model's full limit) as `expectedOutputTokens` — a TRUE
684
527
  // upper bound, since Gemini's default stops at the model's
685
528
  // advertised `outputTokenLimit`.
686
- expectedOutputTokens: effectiveOutputCap,
529
+ expectedOutputTokens: runCap.effective,
530
+ // A cap above the per-request limit (long decode) adds the input every continuation request bills again.
531
+ perRequestOutputTokens: runCap.perRequest,
532
+ // The preflight's count, when it ran: the reservation takes the larger of it and the bytes/4 heuristic.
533
+ countedInputTokens: preflight?.effectiveTokens,
687
534
  thinkingTokens: thinkingTokensForEstimate,
688
535
  coverMs: timeoutController.timeoutMs,
689
536
  });
@@ -715,7 +562,8 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
715
562
  // AFTER `prepareContext` returns from `activePrep.inlineOnly` — see FN2 fix.
716
563
  const caching = cachingModeFor(input.cachingMode ?? ctx.config.cachingMode, resolved, emitter);
717
564
  const requestedCachingMode = caching.mode;
718
- const ctxPrep = await prepareContext({
565
+ // The context of the call: built here, and again by the stale-cache retry on a fresh cache.
566
+ const prepare = () => prepareContext({
719
567
  client: ctx.client,
720
568
  manifest: ctx.manifest,
721
569
  scan,
@@ -731,33 +579,7 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
731
579
  // v1.13.0 — opt-in implicit caching path (skip caches.create).
732
580
  cachingMode: requestedCachingMode,
733
581
  });
734
- // TPM throttle reservation — placed HERE (after `prepareContext`,
735
- // immediately before `generateContent`) so the reservation's `tsMs`
736
- // accurately reflects when tokens hit Gemini's quota counter. An
737
- // earlier reserve (before `prepareContext`) would stamp `tsMs`
738
- // minutes before actual dispatch on cold-cache calls; our window
739
- // would then expire before Gemini's, leaving a gap where we admit
740
- // concurrent calls that bust Gemini's per-minute quota. Trade-off:
741
- // two concurrent cold-cache callers will both complete
742
- // `prepareContext` before one backs off at `reserve` — mostly
743
- // idempotent via file-hash dedup, minor upload duplication.
744
- //
745
- // Extracted into a helper so the stale-cache retry branch can
746
- // cancel-and-re-reserve with an accurate tsMs rather than reusing a
747
- // stale reservation stamped before the first (failed) dispatch.
748
- const reserveForDispatch = async () => {
749
- if (ctx.config.tpmThrottleLimit <= 0)
750
- return;
751
- const reservation = ctx.throttle.reserve(resolved.resolved, estimatedInputTokens);
752
- throttleReservationId = reservation.releaseId;
753
- if (reservation.delayMs > 0) {
754
- emitter.emit(`throttle: waiting ${Math.ceil(reservation.delayMs / 1000)}s for TPM window…`);
755
- // Abortable sleep — without this a 60s TPM throttle wait would block
756
- // a 10s wall-clock timeout, defeating T19's whole-dispatch contract.
757
- await abortableSleep(reservation.delayMs, abortSignal);
758
- }
759
- };
760
- await reserveForDispatch();
582
+ const ctxPrep = await prepare();
761
583
  emitter.emit(usingThinkingLevel
762
584
  ? `generating response (thinking-level=${input.thinkingLevel})…`
763
585
  : effectiveThinkingBudget === null
@@ -790,9 +612,9 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
790
612
  // `input.maxOutputTokens` — otherwise we omit it and Gemini uses
791
613
  // its model-default cap (per docs, = the model's advertised
792
614
  // `outputTokenLimit`, currently 65,536 for pro-tier 3.x/2.5). Budget
793
- // reservation always uses `effectiveOutputCap` (= explicit cap OR
794
- // model limit) as worst-case upper bound, so the daily cap stays a
795
- // true ceiling regardless.
615
+ // reservation always uses the cap the decode runs under (explicit cap
616
+ // OR model limit, cut by the input window on a long decode) as the
617
+ // worst-case upper bound of what a completed request bills.
796
618
  // Cast to `ThinkingLevel` rather than indexing the runtime enum object
797
619
  // (`ThinkingLevel[input.thinkingLevel]`). For string enums the value
798
620
  // IS the member name, so passing `"HIGH"` as `ThinkingLevel` is
@@ -810,7 +632,7 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
810
632
  : { thinkingBudget: effectiveThinkingBudget, includeThoughts: true };
811
633
  const tierField = serviceTierConfig(serviceTier);
812
634
  const buildConfig = (cacheId) => {
813
- const maxOutputField = wireMaxOutputTokens !== undefined ? { maxOutputTokens: wireMaxOutputTokens } : {};
635
+ const maxOutputField = runCap.wire !== undefined ? { maxOutputTokens: runCap.wire } : {};
814
636
  return cacheId
815
637
  ? { cachedContent: cacheId, thinkingConfig, ...maxOutputField, ...tierField }
816
638
  : {
@@ -830,222 +652,70 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
830
652
  return attachments.parts.length === 0 ? input.prompt : [userTurn];
831
653
  return [...inline, userTurn];
832
654
  };
833
- // Track the prepared-context used for the FINAL successful call. Starts
834
- // pointing at the initial ctxPrep; retry branch below reassigns to the
835
- // rebuilt PreparedContext so post-call markHot + metadata reflect reality.
836
- let activePrep = ctxPrep;
837
- let retriedOnStaleCache = false;
838
- let response;
839
- try {
840
- // T20 (v1.7.0) — CRITICAL: withNetworkRetry wraps ONLY the stream
841
- // opening, NOT collectStream itself. If collectStream were inside the
842
- // retry closure, a mid-stream `TypeError: fetch failed` would discard
843
- // the partial response and re-issue a brand-new generateContentStream
844
- // — duplicating the model's billable work and emitting double
845
- // "thinking: …" progress lines. Mid-stream failure CANNOT be retried
846
- // (Gemini's generateContentStream doesn't support resume), so it must
847
- // propagate verbatim to the caller.
848
- const stream = await withNetworkRetry(() => ctx.client.models.generateContentStream({
849
- model: resolved.resolved,
850
- contents: buildContents(activePrep.cacheId, activePrep.inlineContents),
851
- config: { ...buildConfig(activePrep.cacheId), abortSignal },
852
- }), {
853
- attempts: networkAttempts(serviceTier), // a flex request is never re-sent
854
- signal: abortSignal,
855
- onRetry: (attempt, retryErr) => {
856
- logger.warn(`ask: retrying generateContent after transient network failure (attempt ${attempt}): ${retryErr instanceof Error ? retryErr.message : String(retryErr)}`);
857
- },
858
- });
859
- streamOpened = true;
860
- response = await collectStream(stream, {
861
- signal: abortSignal,
862
- onThoughtChunk: (text) => {
863
- // Surface the model's reasoning live as it arrives. Throttled
864
- // to ~1.5s by collectStream so the MCP host's progress channel
865
- // doesn't get flooded on long thinking bursts.
866
- const trimmed = text.trim().slice(0, 80);
867
- if (trimmed.length > 0)
868
- emitter.emit(`thinking: ${trimmed}…`);
869
- },
870
- // v1.12.0 — reset the stall watchdog on every chunk (text or
871
- // thought). Both prove the stream is alive. No-op when stall
872
- // is disabled (env + per-call both unset).
873
- onChunkReceived: () => timeoutController.recordChunk(),
874
- });
875
- }
876
- catch (err) {
877
- // Self-heal: if Gemini rejected our cached content (evicted, expired,
878
- // externally deleted), invalidate locally and retry ONCE with a fresh
879
- // cache. Prevents users from seeing hard failures they'd otherwise need
880
- // to clear/reindex manually. Budget: one retry; a second stale-cache
881
- // error propagates with both errors chained via Error.cause.
882
- if (activePrep.cacheId && isStaleCacheError(err)) {
883
- logger.warn(`Gemini rejected cached content ${activePrep.cacheId}; invalidating and retrying once.`);
884
- try {
885
- // Reset the cache pointer only — keep `files` rows so the rebuild
886
- // can reuse uploaded files via content-hash dedup instead of
887
- // re-uploading. The Gemini-side cache is already dead (that's
888
- // what triggered this branch), so no `caches.delete` needed.
889
- markCacheStale({ manifest: ctx.manifest, workspaceRoot });
890
- const rebuilt = await prepareContext({
891
- client: ctx.client,
892
- manifest: ctx.manifest,
893
- scan,
894
- model: resolved,
895
- systemPromptHash,
896
- systemInstruction: SYSTEM_INSTRUCTION_Q_AND_A,
897
- ttlSeconds: ctx.config.cacheTtlSeconds,
898
- cacheMinTokens: ctx.config.cacheMinTokens,
899
- emitter,
900
- allowCaching: !input.noCache && scan.files.length > 0,
901
- // v1.12.1 — propagate signal into stale-cache rebuild too.
902
- signal: abortSignal,
903
- // v1.13.0 — keep cachingMode consistent across happy + retry paths.
904
- cachingMode: requestedCachingMode,
905
- });
906
- // Cancel the original throttle reservation — its `tsMs` was
907
- // stamped at first-dispatch time, which is now seconds in the
908
- // past (the rebuild took real time). Re-reserve fresh so the
909
- // retry's tsMs reflects when the retry call actually hits
910
- // Gemini's quota counter. Without this, our window would expire
911
- // locally before Gemini's, leaving a gap where concurrent
912
- // callers bust the per-minute limit (PR #19 round-2 GPT review).
913
- if (throttleReservationId !== -1) {
914
- ctx.throttle.cancel(throttleReservationId);
915
- throttleReservationId = -1;
916
- }
917
- await reserveForDispatch();
918
- // The stale-cache retry itself can hit a transient network failure;
919
- // wrap it too so the rebuild isn't wasted on a one-off blip.
920
- // Stale-cache retry: discard partial response, open a brand-new
921
- // stream with the rebuilt cache. Gemini's stream API doesn't
922
- // support resume, so this is the only correct semantics.
923
- // Same retry-OPENING-only contract as the happy path — see above.
924
- const retryStream = await withNetworkRetry(() => ctx.client.models.generateContentStream({
925
- model: resolved.resolved,
926
- contents: buildContents(rebuilt.cacheId, rebuilt.inlineContents),
927
- config: { ...buildConfig(rebuilt.cacheId), abortSignal },
928
- }), {
929
- attempts: networkAttempts(serviceTier),
930
- signal: abortSignal,
931
- onRetry: (attempt, retryErr) => {
932
- logger.warn(`ask (stale-cache retry): retrying generateContent after transient network failure (attempt ${attempt}): ${retryErr instanceof Error ? retryErr.message : String(retryErr)}`);
933
- },
934
- });
935
- streamOpened = true;
936
- response = await collectStream(retryStream, {
937
- signal: abortSignal,
938
- onThoughtChunk: (text) => {
939
- const trimmed = text.trim().slice(0, 80);
940
- if (trimmed.length > 0)
941
- emitter.emit(`thinking: ${trimmed}…`);
942
- },
943
- // v1.12.0 — stall watchdog reset on retry stream too.
944
- onChunkReceived: () => timeoutController.recordChunk(),
945
- });
946
- activePrep = rebuilt;
947
- retriedOnStaleCache = true;
948
- }
949
- catch (retryErr) {
950
- // Re-throw the retry's TimeoutError directly so the outer catch's
951
- // `isTimeoutAbort` check sees it on the cause chain. Without this,
952
- // wrapping with `cause: err` (the original stale-cache error)
953
- // would mask the timeout — outer catch maps to UNKNOWN instead of
954
- // TIMEOUT, breaking the contract on stale-cache+timeout paths.
955
- if (isTimeoutAbort(retryErr))
956
- throw retryErr;
957
- // Otherwise preserve the ORIGINAL stale-cache error as `cause` so
958
- // ops can root-cause diagnostics across both the first 404 and any
959
- // rebuild-time failure.
960
- // an HTTP error on the retry (a 429/503) IS the error: thrown as the SDK made it, so its status and the
961
- // throttle's retry hint are read as usual; only a status-less failure is wrapped over the stale 404
962
- if (statusOf(retryErr) !== undefined)
963
- throw retryErr;
964
- throw new Error(`ask retry after stale cache failed: ${retryErr instanceof Error ? retryErr.message : String(retryErr)}`, { cause: err });
965
- }
966
- }
967
- else {
968
- throw err;
969
- }
970
- }
971
- if (activePrep.cacheId) {
972
- ctx.ttlWatcher.markHot(workspaceRoot, activePrep.cacheId, ctx.config.cacheTtlSeconds);
973
- }
655
+ // The decode — and, on a stale cache before any request completed, its one retry on a rebuilt cache. The TPM
656
+ // reservation is taken here, after `prepareContext` and immediately before the first request, so its timestamp
657
+ // is when the tokens reach Gemini's quota counter; only the OPENING of a request is retried (a mid-stream
658
+ // failure re-sent would bill the work again).
659
+ const decoded = await decodeOnContext({
660
+ client: ctx.client,
661
+ booking: { workspaceRoot, toolName: 'ask', model: resolved.resolved, serviceTier },
662
+ cachingMode: (prep) => cachingRan(prep, caching).mode,
663
+ request: (prep) => ({
664
+ contents: buildContents(prep.cacheId, prep.inlineContents),
665
+ config: buildConfig(prep.cacheId),
666
+ }),
667
+ cache: {
668
+ ttlWatcher: ctx.ttlWatcher,
669
+ manifest: ctx.manifest,
670
+ ttlSeconds: ctx.config.cacheTtlSeconds,
671
+ prepare,
672
+ },
673
+ openAttempts: networkAttempts(serviceTier), // a flex request is never re-sent
674
+ limits: { maxRequests: runCap.maxRequests, cap: runCap.effective },
675
+ timeout: timeoutController,
676
+ emit: (message) => emitter.emit(message),
677
+ completed,
678
+ tracking,
679
+ throttle: dispatch,
680
+ inputTokens: { estimated: estimatedInputTokens, counted: preflight?.effectiveTokens },
681
+ }, ctxPrep);
682
+ const { response, prep: activePrep, retriedOnStaleCache } = decoded;
974
683
  const text = response.text;
975
684
  // `collectStream` already aggregated thought-flagged parts and capped
976
685
  // the joined summary at 1200 chars (T20). Reuse directly — re-iterating
977
686
  // `response.candidates` would risk drift between in-flight thought
978
687
  // emit (live progress) and post-call summary.
979
688
  const thinkingSummary = response.thoughtsSummary;
980
- const usage = response.usageMetadata;
981
- const cached = typeof usage?.cachedContentTokenCount === 'number' ? usage.cachedContentTokenCount : 0;
982
- const inputTotal = typeof usage?.promptTokenCount === 'number' ? usage.promptTokenCount : 0;
983
- const uncached = Math.max(0, inputTotal - cached);
984
- const output = typeof usage?.candidatesTokenCount === 'number' ? usage.candidatesTokenCount : 0;
985
- const thinking = typeof usage?.thoughtsTokenCount === 'number' ? usage.thoughtsTokenCount : 0;
986
- // Finalise the throttle reservation with Gemini's actual
987
- // `promptTokenCount` — includes cached tokens, which DO count against
988
- // the per-minute quota (empirically confirmed 2026-04-20). If actual
989
- // < estimate, subsequent reserves see the freed headroom; if >, they
990
- // back off accordingly.
991
- if (throttleReservationId !== -1) {
992
- ctx.throttle.release(throttleReservationId, inputTotal);
993
- throttleReservationId = -1;
994
- }
995
- const cost = estimateCostUsd({
996
- model: resolved.resolved,
997
- serviceTier,
998
- uncachedInputTokens: uncached,
999
- cachedInputTokens: cached,
1000
- outputTokens: output,
1001
- thinkingTokens: thinking,
1002
- });
1003
- const costMicros = toMicrosUsd(cost);
1004
689
  // Telemetry and the response record the mode that ran, not the one asked for.
1005
690
  const ran = cachingRan(activePrep, caching);
1006
691
  const effectiveCachingMode = ran.mode;
1007
692
  const durationMs = Date.now() - started;
1008
- if (reservationId !== null) {
1009
- // The finalize UPDATE on a row we just inserted should be infallible
1010
- // in practice, but a disk-full / lock-contention edge could throw.
1011
- // If it does, we'd rather keep the reservation row (estimate stays
1012
- // billed, slight overcharge) than let the outer catch CANCEL it
1013
- // — cancelling would erase any record of this billable, completed
1014
- // call. Either way, drop `reservationId` so the outer catch doesn't
1015
- // touch it.
1016
- try {
1017
- ctx.manifest.finalizeBudgetReservation(reservationId, {
1018
- cachedTokens: cached,
1019
- uncachedTokens: uncached,
1020
- costUsdMicro: costMicros,
1021
- durationMs,
1022
- cachingMode: effectiveCachingMode,
1023
- cachedContentTokenCount: cached,
1024
- });
1025
- }
1026
- catch (finalizeErr) {
1027
- logger.error(`ask: finalizeBudgetReservation failed for id=${reservationId}; reservation row keeps the estimate (${costMicros} micros). Error: ${String(finalizeErr)}`);
1028
- }
1029
- reservationId = null;
1030
- }
1031
- else {
1032
- ctx.manifest.insertUsageMetric({
693
+ // Settled here, once: the outer catch then neither releases the reservation nor books the call again. What the
694
+ // API billed is every request that completed, over a long decode and a stale-cache retry alike.
695
+ const settledReservation = reservationId;
696
+ reservationId = null;
697
+ tracking.ledgerEntry = null;
698
+ const settled = settleAnswer({
699
+ manifest: ctx.manifest,
700
+ reservationId: settledReservation,
701
+ entry: {
1033
702
  workspaceRoot,
1034
703
  toolName: 'ask',
1035
704
  model: resolved.resolved,
1036
- cachedTokens: cached,
1037
- uncachedTokens: uncached,
1038
- costUsdMicro: costMicros,
1039
- durationMs,
1040
- occurredAt: Date.now(),
705
+ serviceTier,
1041
706
  cachingMode: effectiveCachingMode,
1042
- cachedContentTokenCount: cached,
1043
- });
1044
- }
707
+ },
708
+ completed,
709
+ response,
710
+ cap: runCap,
711
+ cut: inputWindowCut,
712
+ durationMs,
713
+ });
714
+ const finish = settled.finish;
1045
715
  // Settled and billed above; an empty answer at the cap is a failure the caller can act on, never a success.
1046
- const starved = emptyAtOutputCap(response, thinking);
1047
- if (starved !== null)
1048
- return errorResult(`ask: ${starved}`);
716
+ if (settled.outputCapStop !== null) {
717
+ return errorResult(`ask: ${settled.outputCapStop.message}`, settled.outputCapStop.fields);
718
+ }
1049
719
  if (scan.truncated) {
1050
720
  logger.warn(`workspace ${workspaceRoot} contains more files than GEMINI_CODE_CONTEXT_MAX_FILES (${ctx.config.maxFilesPerWorkspace}); the tail was dropped before indexing.`);
1051
721
  }
@@ -1060,11 +730,11 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
1060
730
  contextWindow: resolved.inputTokenLimit,
1061
731
  thinkingBudget: effectiveThinkingBudget,
1062
732
  thinkingLevel: input.thinkingLevel ?? null,
1063
- cachedTokens: cached,
1064
- uncachedTokens: uncached,
1065
- outputTokens: output,
1066
- thinkingTokens: thinking,
1067
- costEstimateUsd: Math.round(cost * 10000) / 10000,
733
+ ...settled.usage,
734
+ ...settled.cutFields,
735
+ // `truncated`: the answer stopped at the output cap (MAX_TOKENS, or a long decode at its limits) — it is cut
736
+ finishReason: finish.finishReason,
737
+ truncated: finish.truncated,
1068
738
  cacheHit: activePrep.reused,
1069
739
  cacheRebuilt: activePrep.rebuilt || retriedOnStaleCache,
1070
740
  retriedOnStaleCache,
@@ -1104,77 +774,47 @@ async function executeAskBody(input, ctx, workspaceRoot, model, started) {
1104
774
  }
1105
775
  catch (err) {
1106
776
  logger.error(`ask failed: ${safeForLog(err)}`);
1107
- // T22a — extract Gemini's `retryInfo.retryDelay` from 429 bodies and
1108
- // seed the throttle's per-model hint before we release the
1109
- // reservation. Google's hint is typically shorter (2-16s) than our
1110
- // pure-window math would compute (up to 60s+) so honouring it
1111
- // shortens the next caller's wait.
1112
- //
1113
- // Gated on `isGemini429` BEFORE parsing. The predicate requires BOTH
1114
- // `err instanceof ApiError` (SDK-provenance marker — user-controlled
1115
- // content can't forge an ApiError prototype) AND `err.status === 429`
1116
- // (typed field from the HTTP response). The earlier v1.3.2 draft had
1117
- // a `RESOURCE_EXHAUSTED` substring fallback; GPT + Grok round-2
1118
- // review (PR #21) showed that path was user-influenceable — echoed
1119
- // prompt content re-opened the hint-poisoning class the gate was
1120
- // meant to close. Removing the fallback drops hint extraction for
1121
- // errors that lose the ApiError shape in transit, but production
1122
- // 429s always arrive as real ApiError instances.
1123
- const retryDelayMs = isGemini429(err) ? parseRetryDelayMs(err.message) : null;
1124
- if (retryDelayMs !== null && resolvedModelKey !== null) {
1125
- ctx.throttle.recordRetryHint(resolvedModelKey, retryDelayMs);
1126
- }
1127
- // Release any unconsumed budget reservation so the failed call's
1128
- // estimate doesn't eat into future headroom for today. Wrap in its
1129
- // own try/catch — better-sqlite3 can throw SQLITE_BUSY or I/O errors
1130
- // and we must not let a rollback-time DB failure replace the real
1131
- // tool error the user is waiting on.
1132
- if (reservationId !== null) {
1133
- try {
1134
- ctx.manifest.cancelBudgetReservation(reservationId);
1135
- }
1136
- catch (cancelErr) {
1137
- logger.error(`ask: cancelBudgetReservation failed for id=${reservationId}; reservation row keeps the estimate. Error: ${String(cancelErr)}`);
1138
- }
1139
- reservationId = null;
1140
- }
1141
- // Drop the TPM reservation on any failure path. If Gemini actually
1142
- // consumed quota server-side before failing, the hint path above
1143
- // (T22a) compensates via `recordRetryHint` before we cancel. For
1144
- // non-429 errors (validation, transport), we under-count slightly
1145
- // but bounded — accepted trade-off documented in PR #19 round-2.
1146
- if (throttleReservationId !== -1) {
1147
- ctx.throttle.cancel(throttleReservationId);
1148
- throttleReservationId = -1;
1149
- }
777
+ // A 429's retry hint seeds the throttle; the ledger keeps what the completed requests billed and releases the rest
778
+ // of the reservation (all of it when none completed); their usage goes into the error result.
779
+ const failedUsage = settleFailure({
780
+ manifest: ctx.manifest,
781
+ failed: {
782
+ toolName: 'ask',
783
+ reservationId,
784
+ entry: tracking.ledgerEntry,
785
+ completed,
786
+ durationMs: Date.now() - started,
787
+ },
788
+ err,
789
+ model: resolvedModelKey,
790
+ throttle: ctx.throttle,
791
+ dispatch,
792
+ });
793
+ reservationId = null;
1150
794
  // Timeout-driven abort gets a dedicated errorCode so callers can
1151
795
  // distinguish "Gemini was slow" from "schema invalid" / "auth failed".
1152
796
  // v1.12.0 — `timeoutKind` ('total' | 'stall') distinguishes the
1153
797
  // wall-clock cap from the heartbeat-aware stall watchdog so callers
1154
798
  // can apply different policies (e.g. retry on stall but not on total).
1155
799
  if (isTimeoutAbort(err)) {
1156
- const ms = timeoutController.timeoutMs;
1157
- const stallMs = timeoutController.stallMs;
1158
- const kind = getTimeoutKind(err);
1159
- const limitMs = kind === 'stall' ? stallMs : ms;
1160
- emitter.emit(`ask: aborted after ${limitMs ?? '?'}ms ${kind ?? 'timeout'}`);
1161
- const message = kind === 'stall'
1162
- ? `ask timed out — no chunk received for ${stallMs ?? '?'}ms (stall watchdog). Increase \`stallMs\` per call or set \`GEMINI_CODE_CONTEXT_ASK_STALL_MS\` higher; default disabled. Stall watchdog is heartbeat-aware — it does NOT fire while the model is streaming chunks. A fired stall is usually a dead socket or a server-side hang. Note: AbortSignal is client-only — Gemini still finishes server-side and bills for completed work.`
1163
- : `ask timed out after ${ms ?? '?'}ms (total wall-clock). Increase \`timeoutMs\` per call or set \`GEMINI_CODE_CONTEXT_ASK_TIMEOUT_MS\` higher; default disabled. Consider \`stallMs\` instead — a heartbeat-aware watchdog that does NOT fire while the model is actively thinking. Note: AbortSignal is client-only — Gemini still finishes server-side and bills for completed work.`;
1164
- return errorResult(message, {
800
+ const timedOut = timedOutCall('ask', err, timeoutController);
801
+ emitter.emit(`ask: aborted after ${timedOut.limitMs ?? '?'}ms ${timedOut.kind}`);
802
+ return errorResult(timedOut.message, {
1165
803
  errorCode: 'TIMEOUT',
1166
- timeoutKind: kind ?? 'total',
1167
- timeoutMs: ms,
1168
- stallMs,
804
+ timeoutKind: timedOut.kind,
805
+ timeoutMs: timeoutController.timeoutMs,
806
+ stallMs: timeoutController.stallMs,
1169
807
  retryable: true,
1170
808
  serviceTier: reportedTier(),
809
+ ...failedUsage,
1171
810
  });
1172
811
  }
1173
812
  const httpStatus = statusOf(err);
1174
813
  return errorResult(`ask failed: ${err instanceof Error ? err.message : String(err)}`, {
1175
- ...tierErrorMeta(err, streamOpened),
814
+ ...tierErrorMeta(err, tracking.streamOpened),
1176
815
  serviceTier: reportedTier(),
1177
816
  ...(httpStatus !== undefined ? { httpStatus } : {}),
817
+ ...failedUsage,
1178
818
  });
1179
819
  }
1180
820
  finally {