crysk 0.7.0 → 0.8.0

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 (133) hide show
  1. package/README.md +27 -23
  2. package/dist/SKILL-1q3658n0.md +74 -0
  3. package/dist/SKILL-25j4q9b8.md +143 -0
  4. package/dist/SKILL-4747bhne.md +143 -0
  5. package/dist/SKILL-4jc3sv23.md +74 -0
  6. package/dist/SKILL-c0492rch.md +540 -0
  7. package/dist/SKILL-d1nzkhhr.md +74 -0
  8. package/dist/SKILL-f8p7kvx0.md +143 -0
  9. package/dist/SKILL-gxd45g38.md +553 -0
  10. package/dist/SKILL-rec41991.md +66 -0
  11. package/dist/SKILL-zze8zj5g.md +536 -0
  12. package/dist/cli.js +24658 -24654
  13. package/dist/index.js +492 -463
  14. package/dist/opencode-ci-qrcv7a99.md +106 -0
  15. package/dist/opencode-ci-rdpcc7xc.md +106 -0
  16. package/dist/opencode-ci-yt0k74t0.md +106 -0
  17. package/dist/workflow-andv2dvm.yml +158 -0
  18. package/package.json +1 -1
  19. package/prompts/default/PROVENANCE.md +1 -0
  20. package/skills/crysk/SKILL.md +6 -6
  21. package/skills/crysk-ci-setup/SKILL.md +32 -33
  22. package/skills/crysk-ci-setup/assets/workflow.yml +5 -10
  23. package/skills/crysk-ci-setup/references/ci-admission.md +5 -5
  24. package/skills/crysk-ci-setup/references/opencode-ci.md +6 -6
  25. package/skills/crysk-loop/SKILL.md +37 -54
  26. package/skills/crysk-pr/SKILL.md +1 -1
  27. package/src/asset-manifest.ts +0 -6
  28. package/src/assets.ts +1 -4
  29. package/src/brand.ts +20 -30
  30. package/src/ci/gates.ts +155 -201
  31. package/src/ci/outcome.ts +4 -31
  32. package/src/ci/platform-limits.ts +135 -0
  33. package/src/ci/platform-report.ts +212 -126
  34. package/src/ci/platform.ts +28 -26
  35. package/src/ci/reporter.ts +25 -26
  36. package/src/ci/review-admission.ts +1 -2
  37. package/src/ci/review-levels.ts +12 -53
  38. package/src/ci/setup.ts +96 -66
  39. package/src/ci/trace-bundle.ts +11 -27
  40. package/src/ci/trace-transcripts.ts +4 -6
  41. package/src/ci/trigger.ts +3 -5
  42. package/src/cli.ts +6 -5
  43. package/src/commands/ci-setup-questions.ts +2 -55
  44. package/src/commands/ci-setup.ts +2 -26
  45. package/src/commands/triage.ts +131 -38
  46. package/src/commands/usage.ts +0 -1
  47. package/src/compare/compare.ts +3 -3
  48. package/src/compare/floor-test.ts +30 -201
  49. package/src/config/config.ts +6 -30
  50. package/src/diversity/accounting.ts +0 -26
  51. package/src/diversity/benchmark.ts +1 -17
  52. package/src/diversity/identity.ts +0 -14
  53. package/src/diversity/pipeline-integration.ts +0 -29
  54. package/src/doctor.ts +3 -6
  55. package/src/execution/claude-session-trace.ts +7 -2
  56. package/src/execution/contracts.ts +15 -9
  57. package/src/execution/harness.ts +165 -187
  58. package/src/execution/spend-limiter.ts +39 -131
  59. package/src/execution/usage-facts.ts +397 -0
  60. package/src/execution/usage-normalized.ts +33 -86
  61. package/src/harness/materiality-judge.ts +0 -5
  62. package/src/harness/pipeline-plan.ts +0 -3
  63. package/src/harness/replay-rows.ts +0 -5
  64. package/src/harness/route-runner.ts +2 -2
  65. package/src/harness/single-step.ts +3 -9
  66. package/src/index.ts +1 -1
  67. package/src/mcp/preflight.ts +2 -2
  68. package/src/model/catalog.ts +1 -3
  69. package/src/model/free-discovery.ts +11 -78
  70. package/src/model/model-name.ts +14 -4
  71. package/src/model/profiles/claude.ts +1 -2
  72. package/src/model/profiles/openai.ts +1 -2
  73. package/src/model/profiles.ts +1 -2
  74. package/src/model/provider-capabilities.ts +16 -55
  75. package/src/model/routing.ts +39 -36
  76. package/src/pr/admission.ts +6 -6
  77. package/src/pr/answered-findings.ts +8 -4
  78. package/src/pr/ci-publish.ts +0 -2
  79. package/src/pr/discovery.ts +4 -25
  80. package/src/pr/plan.ts +40 -121
  81. package/src/pr/pr-status.ts +14 -2
  82. package/src/pr/preflight.ts +15 -12
  83. package/src/pr/publish-outcome.ts +1 -19
  84. package/src/pr/review-pr.ts +10 -47
  85. package/src/pr/target.ts +76 -82
  86. package/src/production-runtime.ts +52 -114
  87. package/src/progress.ts +0 -54
  88. package/src/rereview/prepare.ts +4 -1
  89. package/src/rereview/state.ts +1 -1
  90. package/src/review/boundary.ts +4 -8
  91. package/src/review/drafts.ts +3 -3
  92. package/src/review/findings-conformance.ts +1 -1
  93. package/src/review/findings.ts +0 -4
  94. package/src/review/pipeline.ts +57 -395
  95. package/src/review/preflight.ts +162 -201
  96. package/src/review/prompt-set.ts +0 -2
  97. package/src/review/report.ts +2 -162
  98. package/src/review/review.ts +16 -65
  99. package/src/review/route-preflight.ts +3 -35
  100. package/src/review/run.ts +37 -73
  101. package/src/review/size-gate.ts +3 -3
  102. package/src/review/spec.ts +6 -11
  103. package/src/review/step-runner.ts +4 -4
  104. package/src/runner-authority.ts +10 -9
  105. package/src/security/credential-broker.ts +5 -6
  106. package/src/server/preflight.ts +0 -1
  107. package/src/server/server.ts +17 -6
  108. package/src/store/activity.ts +1 -2
  109. package/src/store/backfill-preflight.ts +0 -3
  110. package/src/store/backfill.ts +0 -1
  111. package/src/store/metrics-preflight.ts +4 -14
  112. package/src/store/metrics.ts +5 -5
  113. package/src/store/preflight.ts +0 -13
  114. package/src/store/store.ts +11 -12
  115. package/src/transport-registry.ts +6 -30
  116. package/src/transports/claude-code-cli.ts +149 -215
  117. package/src/transports/opencode-client.ts +4 -24
  118. package/src/transports/opencode-sdk.ts +17 -112
  119. package/src/triage/plain-reply.ts +37 -0
  120. package/src/ui/activity.ts +1 -6
  121. package/src/ui/config-edit.ts +14 -67
  122. package/src/ui/config.ts +3 -8
  123. package/src/ui/plan.ts +11 -198
  124. package/src/ui/progress.ts +3 -49
  125. package/src/ui/result.ts +18 -77
  126. package/src/ui/review-menu.ts +3 -12
  127. package/src/ui/select.ts +9 -79
  128. package/src/updater.ts +1 -1
  129. package/src/usage.ts +1 -7
  130. package/prompts/default/deep-review-parity.md +0 -152
  131. package/prompts/scout.md +0 -140
  132. package/src/ci/platform-estimate.ts +0 -111
  133. package/src/review/scout.ts +0 -323
@@ -24,6 +24,10 @@ import {
24
24
  ACTIVE_CHILD_PROCS,
25
25
  type SpawnedProcess,
26
26
  } from "../execution/spawned-process";
27
+ import {
28
+ claudeUsageCounts,
29
+ type ModelUsageCounts,
30
+ } from "../execution/usage-facts";
27
31
  import type { NormalizedUsage } from "../execution/usage-normalized";
28
32
  import {
29
33
  envBillsMetered,
@@ -73,78 +77,45 @@ function hashPromptFile(promptPath: string): string {
73
77
  return createHash("sha256").update(readFileSync(promptPath)).digest("hex");
74
78
  }
75
79
 
76
- // #177, 2026-09-02. WHICH basis this attempt's money is filed on, decided from
77
- // the env the CHILD is spawned with — the one fact that separates a
78
- // subscription run (quota, nothing charged) from an API-key run (a real
79
- // per-token invoice).
80
+ // #177, 2026-09-02. WHICH credential class this attempt's usage is filed
81
+ // under, decided from the env the CHILD is spawned with — the one fact that
82
+ // separates a subscription run (quota) from an API-key run (a per-token
83
+ // credential).
80
84
  //
81
85
  // #173 applied the subscription rule STATICALLY here and was right about its
82
86
  // reasoning and wrong about its premise. Its premise was
83
87
  // `credentialKindForRoute` (runner-authority.ts), which returns
84
88
  // `claude_subscription_oauth` for this backend UNCONDITIONALLY without ever
85
89
  // reading the environment — so it cannot distinguish the two, and taking it as
86
- // proof made every API-key user's real spend render as "at list price, not
87
- // charged". Worse, §8 makes budget enforcement cash-only, so that user's spend
88
- // ceiling read $0 and effectively stopped existing. Under-reporting, shipped by
89
- // the change that set out to stop it.
90
+ // proof made every API-key user's usage file as subscription. Under-reporting,
91
+ // shipped by the change that set out to stop it.
90
92
  //
91
93
  // The fix is deliberately NARROW: it does not derive `credentialKindForRoute`,
92
- // and it does not touch route admission. See the guard on `envBillsMetered`.
93
- // #197 removed the REASON that guard used to give — a metered claude-code
94
- // route is no longer refused for lack of pricing, because `pricingReady` is
95
- // now `true` for this transport — but not the guard: which credential a route
96
- // runs on is `credentialKindForRoute`'s single decision (#161), and
97
- // deriving it from the env HERE would fork it. This only decides how an
98
- // attempt that already ran gets FILED.
99
- //
100
- // What each arm claims:
101
- // * metered — `costSource: "provider"` because the CLI's own figure is what
102
- // the API bills at (its per-model `costBasis` is `"list"`, and on a
103
- // per-token key list price IS the charge). This restores #173's cash
104
- // filing for this credential; `billingMode` is a truthful UPGRADE over the
105
- // pre-#173 shape, which said "subscription" on every route.
106
- // * subscription — #173's rule, unchanged: nothing is charged, so the list
107
- // figure is notional and `costSource` is `"subscription"` (not "provider":
108
- // it is not what the provider charged; not "versioned_rate_table": we
109
- // computed nothing).
94
+ // and it does not touch route admission. See the guard on `envBillsMetered`:
95
+ // which credential a route runs on is `credentialKindForRoute`'s single
96
+ // decision (#161), and deriving it from the env HERE would fork it. This only
97
+ // decides how an attempt that already ran gets FILED.
110
98
  //
111
- // Returned as ONE object rather than two reads so `noSpawnUsage` and the two
112
- // parse arms cannot drift: `sumNormalizedUsage` collapses `billingMode` AND
113
- // `costSource` to "unknown" the moment two attempts of one step disagree, and
114
- // this transport retries.
115
- function claudeCliCostBasis(
99
+ // Returned from ONE function so `noSpawnUsage` and the two parse arms cannot
100
+ // drift: `sumNormalizedUsage` collapses `billingMode` to "unknown" the moment
101
+ // two attempts of one step disagree, and this transport retries.
102
+ function claudeCliBillingMode(
116
103
  env: Readonly<Record<string, string | undefined>>,
117
- ): {
118
- readonly billingMode: NormalizedUsage["billingMode"];
119
- readonly costSource: NormalizedUsage["costSource"];
120
- } {
121
- return envBillsMetered(env)
122
- ? { billingMode: "metered", costSource: "provider" }
123
- : { billingMode: "subscription", costSource: "subscription" };
104
+ ): NormalizedUsage["billingMode"] {
105
+ return envBillsMetered(env) ? "metered" : "subscription";
124
106
  }
125
107
 
126
108
  // A denial before spawn, or a PGID proof failure that refuses to signal a
127
- // child we never trust, is genuine zero cost — no attempt reached the
128
- // provider, so nothing was spent. This is deliberately NOT
129
- // `normalizeUnavailableUsage`: "unavailable" means an attempt ran and its
130
- // cost is unknown, which would misfile a $0 refusal as an unresolved spend
131
- // once PR5's spend ledger reads completeness.
109
+ // child we never trust, used nothing — no attempt reached the provider. This
110
+ // is deliberately NOT `normalizeUnavailableUsage`: "unavailable" means an
111
+ // attempt ran and its usage is unknown, which would misfile a refusal as an
112
+ // unresolved spend once the spend ledger reads completeness.
132
113
  //
133
- // #173, 2026-09-02: `costSource` was `"provider"` here, which was never true
134
- // — no provider was contacted. It has to match the parse arms below for a
135
- // second, mechanical reason: an unclassified failure legacy-classifies as
136
- // "format" (step-runner.ts `classifyFailure`), which has a retry, so a denied
137
- // attempt 1 can be summed with a spawned attempt 2 — and `sumNormalizedUsage`
138
- // collapses `costSource` to "unknown" whenever two attempts disagree. Leaving
139
- // this one behind would have erased the run's cost basis on exactly the retry
140
- // path this transport already supports.
141
- //
142
- // #177, 2026-09-02: that obligation is now owed on TWO bases, so the basis is
143
- // no longer written out here — `claudeCliCostBasis` is the single source both
144
- // this function and the parse arms consume. Only `cashCostUsd: 0` stays local,
145
- // and it is a genuine zero on either basis: no attempt reached the provider.
146
- // With `tokens: {}` it also cannot trip the metered-zero rule
147
- // (`settlementFromUsage`), which needs output tokens to fire.
114
+ // It has to carry the same billing mode as the parse arms below, for a
115
+ // mechanical reason: an unclassified failure legacy-classifies as "format"
116
+ // (step-runner.ts `classifyFailure`), which has a retry, so a denied attempt 1
117
+ // can be summed with a spawned attempt 2 — and `sumNormalizedUsage` collapses
118
+ // `billingMode` to "unknown" whenever two attempts disagree.
148
119
  function noSpawnUsage(
149
120
  wallMs: number,
150
121
  env: Readonly<Record<string, string | undefined>>,
@@ -153,13 +124,11 @@ function noSpawnUsage(
153
124
  wallMs,
154
125
  tokens: {},
155
126
  completeness: "complete",
156
- ...claudeCliCostBasis(env),
157
- cashCostUsd: 0,
127
+ billingMode: claudeCliBillingMode(env),
158
128
  };
159
129
  }
160
130
 
161
131
  interface RawClaudeCliResult {
162
- readonly total_cost_usd?: number;
163
132
  // #299. Read through `toolInvocationsFromCliResult`, never directly: the
164
133
  // type is `unknown` because the only trustworthy shape is an integer >= 1.
165
134
  readonly num_turns?: unknown;
@@ -170,27 +139,28 @@ interface RawClaudeCliResult {
170
139
  readonly cache_read_input_tokens?: number;
171
140
  };
172
141
  // #175 half 2. The CLI reports, per model it actually ran, an entry keyed
173
- // on the exact snapshot. Only the two identity fields are declared:
174
- // `canonicalModel` (the family) and the key itself.
175
- //
176
- // The real block also carries `costUSD`, `costBasis` and token counts. They
177
- // are deliberately NOT read here. Cost already has ONE parse site
178
- // (`normalizeClaudeCliUsage`, from `total_cost_usd`) and one filing
179
- // decision; a second reader of a second cost field is exactly how two
180
- // numbers for one attempt start disagreeing.
181
- //
182
- // #173 closed the `costBasis: "list"` problem this comment used to defer,
183
- // and closed it WITHOUT reading the field. `costBasis` is in the response,
184
- // but the FILING decision follows the credential's billing mode rather than
185
- // the response's label: `normalizeClaudeCliUsage` files `total_cost_usd` as
186
- // notional on the strength of the route being a subscription, which is true
187
- // whatever any per-model entry says. Parsing `costBasis` per model would add
188
- // the second cost reader this paragraph exists to refuse. What is still unread and still
189
- // wanted is the per-model SPLIT (#173 records it): one `--model sonnet`
190
- // invocation ran two models, so a single route-level cost is an incomplete
191
- // provenance claim in the same way a single `modelSnapshot` is.
142
+ // on the exact snapshot. Read here: the two identity fields
143
+ // (`canonicalModel` and the key itself, `observedModelsFromCliResult`) and,
144
+ // since platform-owned-cost T6, the token counts (`modelUsageFromCliResult`).
145
+ // The CLI's per-model cost fields are never read: the engine reports usage
146
+ // facts and the platform prices them. One `--model sonnet` invocation can
147
+ // run two models, which is why the counts are per model.
192
148
  readonly modelUsage?: Readonly<
193
- Record<string, { readonly canonicalModel?: unknown } | null | undefined>
149
+ Record<
150
+ string,
151
+ | {
152
+ readonly canonicalModel?: unknown;
153
+ readonly inputTokens?: unknown;
154
+ readonly outputTokens?: unknown;
155
+ readonly cacheReadInputTokens?: unknown;
156
+ readonly cacheCreationInputTokens?: unknown;
157
+ readonly cacheCreationEphemeral1hInputTokens?: unknown;
158
+ readonly cacheCreationEphemeral5mInputTokens?: unknown;
159
+ readonly thinkingTokens?: unknown;
160
+ }
161
+ | null
162
+ | undefined
163
+ >
194
164
  >;
195
165
  }
196
166
 
@@ -201,7 +171,7 @@ interface RawClaudeCliResult {
201
171
  // key, and an EMPTY `modelUsage` all answer `undefined`. An empty array would
202
172
  // assert "we looked and nothing ran", which is a claim about the provider we
203
173
  // have no basis for — the same distinction `normalizeUnavailableUsage` draws
204
- // between an unknown cost and a zero one.
174
+ // between unknown usage and a zero one.
205
175
  //
206
176
  // Parsed separately from `normalizeClaudeCliUsage` rather than folded into
207
177
  // it: that function's contract is the §8 numeric shape, and its every branch
@@ -242,6 +212,66 @@ function observedModelsFromCliResult(
242
212
  return observed.length === 0 ? undefined : observed;
243
213
  }
244
214
 
215
+ // platform-owned-cost T6: the token counts of the same `modelUsage` block, per
216
+ // model, split into the platform's disjoint leaves (thinking out of output, the
217
+ // one-hour write out of the total write). The same absence rule as above: no
218
+ // parse, no block, or no usable entry answers `undefined`.
219
+ //
220
+ // The model is the provider's id, as the session transcript's `message.model`
221
+ // spells it, so the two can be joined: the CLI marks a context-window variant on
222
+ // its own model names with a bracketed suffix (`requestedModel:
223
+ // "claude-opus-5[1m]"` beside `message.model: "claude-opus-5"` in a real
224
+ // session), which names no provider model. A key carrying one is filed under the
225
+ // id without it, and entries that meet there are summed, so a summary key can
226
+ // never miss its transcript requests and count them twice as a remainder.
227
+ const CLI_MODEL_VARIANT = /\[[^\]]*\]$/;
228
+
229
+ function modelUsageFromCliResult(
230
+ rawStdout: string,
231
+ ): readonly ModelUsageCounts[] | undefined {
232
+ let parsed: RawClaudeCliResult;
233
+ try {
234
+ parsed = JSON.parse(rawStdout);
235
+ } catch {
236
+ return undefined;
237
+ }
238
+ const raw = parsed?.modelUsage;
239
+ if (raw === undefined || raw === null || typeof raw !== "object") {
240
+ return undefined;
241
+ }
242
+ const models = new Map<string, number[]>();
243
+ for (const [key, entry] of Object.entries(raw)) {
244
+ const model = key.replace(CLI_MODEL_VARIANT, "");
245
+ if (
246
+ model.trim().length === 0 ||
247
+ entry === null ||
248
+ typeof entry !== "object"
249
+ ) {
250
+ continue;
251
+ }
252
+ const counts = claudeUsageCounts({
253
+ uncached: entry.inputTokens,
254
+ cacheRead: entry.cacheReadInputTokens,
255
+ cacheWriteTotal: entry.cacheCreationInputTokens,
256
+ cacheWrite1h: entry.cacheCreationEphemeral1hInputTokens,
257
+ cacheWrite5m: entry.cacheCreationEphemeral5mInputTokens,
258
+ output: entry.outputTokens,
259
+ thinking: entry.thinkingTokens,
260
+ });
261
+ const sum = models.get(model) ?? [0, 0, 0, 0, 0, 0];
262
+ counts.forEach((n, i) => {
263
+ sum[i] = (sum[i] ?? 0) + n;
264
+ });
265
+ models.set(model, sum);
266
+ }
267
+ return models.size === 0
268
+ ? undefined
269
+ : [...models].map(([model, counts]) => ({
270
+ model,
271
+ counts: counts as unknown as ModelUsageCounts["counts"],
272
+ }));
273
+ }
274
+
245
275
  // #299 (2026-10-01): how many tool calls the model made, as far as the CLI's
246
276
  // result lets us say. `num_turns` counts ASSISTANT turns, and every turn after
247
277
  // the first begins from tool results, so `num_turns - 1` is a LOWER BOUND on
@@ -271,35 +301,6 @@ function toolInvocationsFromCliResult(rawStdout: string): number | undefined {
271
301
  : undefined;
272
302
  }
273
303
 
274
- // #173 (§8, docs/multi-runtime-model-diversity-design.md:462): "Subscription
275
- // OAuth may truthfully report `cashCostUsd: 0`; optional catalog cost is
276
- // `notionalCostUsd` and never mixed with cash." This is the one site that
277
- // applies that rule for this transport.
278
- //
279
- // #177, 2026-09-02: it is applied per-credential, NOT statically. #173 applied
280
- // it statically on the strength of `credentialKindForRoute`
281
- // (runner-authority.ts) returning `claude_subscription_oauth` for this backend
282
- // unconditionally — and "unconditionally" is the whole problem: that function
283
- // never reads the environment, so it says subscription for an
284
- // ANTHROPIC_API_KEY route that is spending real per-token money. The branch is
285
- // therefore neither unreachable nor untestable, and `claudeCliCostBasis` above
286
- // is what decides it, from the env this transport spawns the child with.
287
- //
288
- // The figure itself is list-basis, verified live 2026-09-02 rather than
289
- // inferred: the CLI's `modelUsage` block labels each model's `costUSD` with
290
- // `"costBasis": "list"`, and `total_cost_usd` is their sum. On a subscription
291
- // nothing is charged, so filing it as cash recorded spend that never happened.
292
- // On a per-token key, list price IS the charge, so cash is exactly where it
293
- // belongs — and `notionalCostUsd` stays absent rather than duplicating it,
294
- // because the same figure in both fields would report one attempt's tokens as
295
- // real spend AND as an untaken list price.
296
- //
297
- // Note the two fields answer to different things. `notionalCostUsd` follows
298
- // the provider: absent when the CLI reported no total, because inventing 0
299
- // there would claim the attempt consumed nothing. `cashCostUsd` follows the
300
- // CREDENTIAL: it is 0 on the subscription arm because the subscription charges
301
- // nothing, whatever the CLI said.
302
- //
303
304
  // §8: `--output-format json`'s usage block is already disjoint-additive —
304
305
  // input_tokens (uncached), cache_read_input_tokens, and
305
306
  // cache_creation_input_tokens sum to total input (verified against the real
@@ -307,50 +308,26 @@ function toolInvocationsFromCliResult(rawStdout: string): number | undefined {
307
308
  // leaves are read straight across rather than split from a total via
308
309
  // `normalizeInclusiveUsage`. Corrupted stdout, and valid JSON carrying no
309
310
  // `usage` block at all (a safety-blocked or otherwise usage-less response),
310
- // both yield `normalizeUnavailableUsage` — never a fabricated zero leaf, the
311
- // exact "$0 on parse failure" collapse this slice exists to kill.
311
+ // both yield `normalizeUnavailableUsage` — never a fabricated zero leaf.
312
+ //
313
+ // The credential class (`billingMode`) is filed per attempt from the env the
314
+ // child ran with (`claudeCliBillingMode`); the CLI's own cost figure is not
315
+ // read — the platform prices a run from the tokens and models reported here.
312
316
  function normalizeClaudeCliUsage(
313
317
  rawStdout: string,
314
318
  wallMs: number,
315
319
  env: Readonly<Record<string, string | undefined>>,
316
320
  ): NormalizedUsage {
317
- const basis = claudeCliCostBasis(env);
318
- // The two cost fields decided TOGETHER, so no future editor can move cash
319
- // without moving notional. `total_cost_usd` lands in exactly ONE of them —
320
- // as CASH when a per-token credential is paying for it, as NOTIONAL when a
321
- // subscription is — and never in both: the same figure in both fields would
322
- // report one attempt's tokens as spend AND as an untaken list price.
323
- //
324
- // Absence stays absence on either arm. A CLI that reported no
325
- // `total_cost_usd` gives a subscription no list figure to record, and gives
326
- // a metered route no answer to "how much real money was this?" — so the
327
- // metered arm leaves cash undefined and `settlementFromUsage` calls the
328
- // attempt unresolved, rather than settling a fabricated $0. That is the
329
- // same "$0 on parse failure" collapse §8 exists to kill, one door over.
330
- //
331
- // `cashCostUsd` on the subscription arm is the one unconditional value: it
332
- // is 0 because the subscription charges nothing, whatever the CLI said.
333
- const costFor = (
334
- total: number | undefined,
335
- ): { cashCostUsd?: number; notionalCostUsd?: number } =>
336
- basis.billingMode === "metered"
337
- ? { cashCostUsd: total }
338
- : { cashCostUsd: 0, notionalCostUsd: total };
321
+ const billingMode = claudeCliBillingMode(env);
339
322
  let parsed: RawClaudeCliResult;
340
323
  try {
341
324
  parsed = JSON.parse(rawStdout);
342
325
  } catch {
343
- return normalizeUnavailableUsage({
344
- wallMs,
345
- billingMode: basis.billingMode,
346
- });
326
+ return normalizeUnavailableUsage({ wallMs, billingMode });
347
327
  }
348
328
  const raw = parsed.usage;
349
329
  if (raw === undefined) {
350
- return normalizeUnavailableUsage({
351
- wallMs,
352
- billingMode: basis.billingMode,
353
- });
330
+ return normalizeUnavailableUsage({ wallMs, billingMode });
354
331
  }
355
332
  const leafValues = [
356
333
  raw.input_tokens,
@@ -367,8 +344,7 @@ function normalizeClaudeCliUsage(
367
344
  return normalizePartialUsage({
368
345
  wallMs,
369
346
  providerReportedTotal,
370
- ...basis,
371
- ...costFor(parsed.total_cost_usd),
347
+ billingMode,
372
348
  });
373
349
  }
374
350
  const inputUncached = leafValues[0] as number;
@@ -389,8 +365,7 @@ function normalizeClaudeCliUsage(
389
365
  providerReportedTotal: inputKnown + outputVisible,
390
366
  },
391
367
  completeness: "complete",
392
- ...basis,
393
- ...costFor(parsed.total_cost_usd),
368
+ billingMode,
394
369
  };
395
370
  }
396
371
 
@@ -482,7 +457,7 @@ export class ClaudeCodeCliTransport implements ProviderTransport {
482
457
  modelSnapshot: "sonnet",
483
458
  };
484
459
  get billingMode(): NormalizedUsage["billingMode"] {
485
- return claudeCliCostBasis(process.env).billingMode;
460
+ return claudeCliBillingMode(process.env);
486
461
  }
487
462
  private readonly spawnFn: typeof Bun.spawn;
488
463
  private readonly getPgid: (pid: number) => number | undefined;
@@ -535,9 +510,8 @@ export class ClaudeCodeCliTransport implements ProviderTransport {
535
510
  // contractual here, not probed: every execute() request must carry an
536
511
  // IsolationProjection (credentialProjectionId + syntheticHome) or admission
537
512
  // denies it. Unproven features are claimed false with a
538
- // non-blocking issue, never assumed green: no bounded event sink is wired
539
- // yet, no dedicated codegraph policy is enforced, and no pricing table
540
- // exists — hence status "degraded", not "ready".
513
+ // yet and no dedicated codegraph policy is enforced — hence status
514
+ // "degraded", not "ready".
541
515
  async capabilities(
542
516
  input?: ClaudeCliCapabilitiesInput,
543
517
  ): Promise<ProviderCapabilityReport> {
@@ -564,71 +538,29 @@ export class ClaudeCodeCliTransport implements ProviderTransport {
564
538
  conformance: CLAUDE_CAPABILITY_STATICS.cancellationConformance,
565
539
  },
566
540
  billing: {
541
+ // The backend-wide report is a STATIC claim made at admission, before
542
+ // any spawn and before any route (and so any credential) resolves.
543
+ // The per-route credential class lives on the exact binding
544
+ // (`FrozenRuntimeBinding.capabilities()`, production-runtime.ts),
545
+ // which upgrades this mode through `credentialKindBillsMetered`.
546
+ // That is #161's rule: `credentialKindForRoute` (runner-authority.ts)
547
+ // derives `provider_api_token` — a metered kind — for a claude-code
548
+ // route whose environment carries ANTHROPIC_API_KEY or
549
+ // ANTHROPIC_AUTH_TOKEN AND that has no credential broker projecting
550
+ // for it, because with no broker nothing strips the key before the
551
+ // child sees it (the common case on Linux/CI, which has no Keychain
552
+ // to project from). A route with a broker whose projection SUCCEEDS
553
+ // stays subscription whatever its env carries. A degraded projection
554
+ // is #279's known exception: admission cannot see that a projection
555
+ // failed and fell back to the unstripped env, so that route keeps
556
+ // saying subscription while the key reaches the child. That
557
+ // ADMISSION-time answer is deliberately left that way: the fix lives
558
+ // downstream, in `StepExecutionHarness.run` (harness.ts), which
559
+ // re-applies `envBillsMetered` to this SAME child env this module
560
+ // stamps `billingMode: "metered"` from, and opens a spend-ledger
561
+ // reservation for that one attempt so it is settled or fenced exactly
562
+ // like a metered route.
567
563
  mode: CLAUDE_CAPABILITY_STATICS.billingMode,
568
- // #197. WHAT THIS FLAG ASSERTS, which is narrower than it reads.
569
- // `pricingReady` lives inside `capabilities()`, built from
570
- // CLAUDE_CAPABILITY_STATICS — a STATIC claim made at admission,
571
- // before any spawn. No cost figure exists at the moment it is asked,
572
- // for any transport, so it cannot be answering "is this attempt's
573
- // cost a provider invoice?". It answers "will this transport tell you
574
- // what the attempt cost?", and this one will: the CLI reports
575
- // `total_cost_usd`, which #177 files as cash on a metered credential
576
- // and as notional on a subscription.
577
- //
578
- // That is why the list-vs-invoice objection this comment used to
579
- // carry does not survive. It rejected the CLI's figure for being a
580
- // LIST computation (`costBasis: "list"`) while the alternative it
581
- // preferred — the bundled `config/models/*-pricing.json` — was a
582
- // hand-transcribed snapshot of the same published list prices. The
583
- // choice was never invoice-over-list; it was OUR list price over the
584
- // CLI's, on a ground that convicted our own. The tables are gone
585
- // (#197); the CLI's figure tracks the tool the user already upgrades.
586
- //
587
- // THIS DELIBERATELY WIDENS ADMISSION, and that is the intent rather
588
- // than a side effect to discover later. `tokenPricingAvailable`
589
- // (production-runtime.ts) is now this flag and nothing else, so a
590
- // metered claude-code route that the old gate refused for a missing
591
- // or expired table is admitted and runs on the cost the CLI reports.
592
- // #161 landed on top of this: `credentialKindForRoute`
593
- // (runner-authority.ts) now derives `provider_api_token` — a metered
594
- // kind — for a claude-code route whose environment carries
595
- // ANTHROPIC_API_KEY or ANTHROPIC_AUTH_TOKEN AND that has no
596
- // credential broker projecting for it (no broker means nothing
597
- // strips the key before the child sees it — the common case being
598
- // Linux/CI, which has no Keychain to project from). A route that
599
- // DOES have a broker attached, and whose projection SUCCEEDS, stays
600
- // subscription regardless of what its env carries, because a
601
- // successful projection strips the key before spawn — so a
602
- // claude-code route CAN resolve metered today, not just in the
603
- // gate's hypothetical, but only when nothing is about to project a
604
- // credential for it. A degraded projection is #279's known
605
- // exception: `credentialKindForRoute` still cannot see that a
606
- // broker's projection failed and fell back to the unstripped env, so
607
- // that one route keeps saying subscription while the key actually
608
- // reaches and pays for the child — this ADMISSION-time answer is
609
- // deliberately left that way (#279, fixed 2026-09-24, option 2: the
610
- // fix lives downstream, in `StepExecutionHarness.run`, harness.ts,
611
- // which re-applies `envBillsMetered` to this SAME child env this
612
- // module stamps `billingMode: "metered"` from, and opens a
613
- // spend-ledger reservation for that one attempt so it is settled or
614
- // fenced exactly like a metered route). `pricingReady: true` here is what
615
- // makes a metered route ADMITTED rather than refused for lack of
616
- // pricing; #161 is what makes the route exist in the first place.
617
- // This BACKEND-WIDE report still claims `mode: "subscription"`
618
- // unconditionally — it is produced before any route resolves and has
619
- // no per-route credential to read — and it is the EXACT binding
620
- // (`FrozenRuntimeBinding.capabilities()`, production-runtime.ts) that
621
- // upgrades to the real per-route mode via
622
- // `credentialKindBillsMetered(credential.kind)`.
623
- //
624
- // The one gap this opens is settled, not hand-waved: `total_cost_usd`
625
- // is OPTIONAL on `RawClaudeCliResult`, so a metered attempt can
626
- // complete with no figure. It settles `unresolved` with no
627
- // `knownUsd` rather than a fabricated $0 (`costFor` above,
628
- // `settlementFromUsage` in spend-limiter.ts), pinned by "a metered
629
- // run reporting no cash figure at all" in
630
- // test/conformance/claude-cli-transport.test.ts.
631
- pricingReady: true,
632
564
  },
633
565
  ...(input !== undefined
634
566
  ? {
@@ -960,6 +892,7 @@ export class ClaudeCodeCliTransport implements ProviderTransport {
960
892
  request.isolation.env,
961
893
  );
962
894
  const observedModels = observedModelsFromCliResult(stdout);
895
+ const modelUsage = modelUsageFromCliResult(stdout);
963
896
  const toolInvocations = toolInvocationsFromCliResult(stdout);
964
897
 
965
898
  let stderrTail = stderr.slice(-4096);
@@ -1035,6 +968,7 @@ export class ClaudeCodeCliTransport implements ProviderTransport {
1035
968
  usage,
1036
969
  stderrTail,
1037
970
  ...(observedModels === undefined ? {} : { observedModels }),
971
+ ...(modelUsage === undefined ? {} : { modelUsage }),
1038
972
  ...(toolInvocations === undefined ? {} : { toolInvocations }),
1039
973
  ...(exitCode !== undefined ? { exitCode } : {}),
1040
974
  };
@@ -387,7 +387,6 @@ export interface OpenCodeTurnState {
387
387
  interface StepUsage {
388
388
  readonly inputTokens?: number;
389
389
  readonly outputTokens?: number;
390
- readonly costUsd?: number;
391
390
  // OpenCode's `info.tokens` leaves are disjoint-additive
392
391
  // (packages/schema/src/token-usage.ts): total = input + output + reasoning
393
392
  // + cache.read + cache.write. Carried so cache use is measurable on this
@@ -662,7 +661,6 @@ function addField(
662
661
  function addUsage(base: StepUsage, step: StepUsage): StepUsage {
663
662
  const inputTokens = addField(base.inputTokens, step.inputTokens);
664
663
  const outputTokens = addField(base.outputTokens, step.outputTokens);
665
- const costUsd = addField(base.costUsd, step.costUsd);
666
664
  const cacheReadTokens = addField(base.cacheReadTokens, step.cacheReadTokens);
667
665
  const cacheWriteTokens = addField(
668
666
  base.cacheWriteTokens,
@@ -672,7 +670,6 @@ function addUsage(base: StepUsage, step: StepUsage): StepUsage {
672
670
  return {
673
671
  ...(inputTokens !== undefined ? { inputTokens } : {}),
674
672
  ...(outputTokens !== undefined ? { outputTokens } : {}),
675
- ...(costUsd !== undefined ? { costUsd } : {}),
676
673
  ...(cacheReadTokens !== undefined ? { cacheReadTokens } : {}),
677
674
  ...(cacheWriteTokens !== undefined ? { cacheWriteTokens } : {}),
678
675
  ...(reasoningTokens !== undefined ? { reasoningTokens } : {}),
@@ -703,11 +700,6 @@ function rememberUsage(
703
700
  const wasCompleted = state.completedUsage.has(messageId);
704
701
  if (completed) state.completedUsage.add(messageId);
705
702
  if (existing !== undefined) {
706
- const costConflict =
707
- existing.costUsd !== undefined &&
708
- usage.costUsd !== undefined &&
709
- (usage.costUsd < existing.costUsd ||
710
- (wasCompleted && usage.costUsd !== existing.costUsd));
711
703
  const tokenConflict =
712
704
  (existing.inputTokens !== undefined &&
713
705
  usage.inputTokens !== undefined &&
@@ -717,7 +709,7 @@ function rememberUsage(
717
709
  usage.outputTokens !== undefined &&
718
710
  (usage.outputTokens < existing.outputTokens ||
719
711
  (wasCompleted && usage.outputTokens !== existing.outputTokens)));
720
- if (costConflict || tokenConflict) {
712
+ if (tokenConflict) {
721
713
  state.usageConflict = true;
722
714
  state.usageIncomplete = true;
723
715
  }
@@ -729,10 +721,6 @@ function rememberUsage(
729
721
  existing.outputTokens !== undefined || usage.outputTokens !== undefined
730
722
  ? Math.max(existing.outputTokens ?? 0, usage.outputTokens ?? 0)
731
723
  : undefined;
732
- const mergedCost =
733
- existing.costUsd !== undefined || usage.costUsd !== undefined
734
- ? Math.max(existing.costUsd ?? 0, usage.costUsd ?? 0)
735
- : undefined;
736
724
  // Cache and reasoning leaves merge like input/output (Math.max) but are
737
725
  // deliberately left out of the conflict check above: a snapshot that
738
726
  // momentarily reports fewer cache tokens must not flag the whole turn as
@@ -752,7 +740,6 @@ function rememberUsage(
752
740
  state.usage.set(messageId, {
753
741
  ...(mergedInput !== undefined ? { inputTokens: mergedInput } : {}),
754
742
  ...(mergedOutput !== undefined ? { outputTokens: mergedOutput } : {}),
755
- ...(mergedCost !== undefined ? { costUsd: mergedCost } : {}),
756
743
  ...(mergedCacheRead !== undefined
757
744
  ? { cacheReadTokens: mergedCacheRead }
758
745
  : {}),
@@ -1273,14 +1260,12 @@ export function reconcileMessages(
1273
1260
  const tokens = asRecord(info.tokens);
1274
1261
  const inputTokens = asNumber(tokens?.input);
1275
1262
  const outputTokens = asNumber(tokens?.output);
1276
- const costUsd = asNumber(info.cost);
1277
1263
  const cacheReadTokens = asNumber(asRecord(tokens?.cache)?.read);
1278
1264
  const cacheWriteTokens = asNumber(asRecord(tokens?.cache)?.write);
1279
1265
  const reasoningTokens = asNumber(tokens?.reasoning);
1280
1266
  if (
1281
1267
  inputTokens !== undefined ||
1282
1268
  outputTokens !== undefined ||
1283
- costUsd !== undefined ||
1284
1269
  cacheReadTokens !== undefined ||
1285
1270
  cacheWriteTokens !== undefined ||
1286
1271
  reasoningTokens !== undefined
@@ -1291,7 +1276,6 @@ export function reconcileMessages(
1291
1276
  {
1292
1277
  ...(inputTokens !== undefined ? { inputTokens } : {}),
1293
1278
  ...(outputTokens !== undefined ? { outputTokens } : {}),
1294
- ...(costUsd !== undefined ? { costUsd } : {}),
1295
1279
  ...(cacheReadTokens !== undefined ? { cacheReadTokens } : {}),
1296
1280
  ...(cacheWriteTokens !== undefined ? { cacheWriteTokens } : {}),
1297
1281
  ...(reasoningTokens !== undefined ? { reasoningTokens } : {}),
@@ -1415,7 +1399,6 @@ export function reconcileMessages(
1415
1399
  if (
1416
1400
  msg.time?.completed === undefined ||
1417
1401
  stepUsage === undefined ||
1418
- stepUsage.costUsd === undefined ||
1419
1402
  stepUsage.inputTokens === undefined ||
1420
1403
  stepUsage.outputTokens === undefined
1421
1404
  ) {
@@ -1659,14 +1642,12 @@ export function mapOpenCodeEvents(
1659
1642
  const tokens = asRecord(info.tokens);
1660
1643
  const inputTokens = asNumber(tokens?.input);
1661
1644
  const outputTokens = asNumber(tokens?.output);
1662
- const costUsd = asNumber(info.cost);
1663
1645
  const cacheReadTokens = asNumber(asRecord(tokens?.cache)?.read);
1664
1646
  const cacheWriteTokens = asNumber(asRecord(tokens?.cache)?.write);
1665
1647
  const reasoningTokens = asNumber(tokens?.reasoning);
1666
1648
  if (
1667
1649
  inputTokens !== undefined ||
1668
1650
  outputTokens !== undefined ||
1669
- costUsd !== undefined ||
1670
1651
  cacheReadTokens !== undefined ||
1671
1652
  cacheWriteTokens !== undefined ||
1672
1653
  reasoningTokens !== undefined
@@ -1677,7 +1658,6 @@ export function mapOpenCodeEvents(
1677
1658
  {
1678
1659
  ...(inputTokens !== undefined ? { inputTokens } : {}),
1679
1660
  ...(outputTokens !== undefined ? { outputTokens } : {}),
1680
- ...(costUsd !== undefined ? { costUsd } : {}),
1681
1661
  ...(cacheReadTokens !== undefined ? { cacheReadTokens } : {}),
1682
1662
  ...(cacheWriteTokens !== undefined ? { cacheWriteTokens } : {}),
1683
1663
  ...(reasoningTokens !== undefined ? { reasoningTokens } : {}),
@@ -2139,7 +2119,7 @@ const MCP_TOOL_ID_SET: ReadonlySet<string> = new Set(ALL_MCP_TOOL_IDS);
2139
2119
  // `tool.ids()`. "bash" has no entry in that table at all (no AgentSpec has
2140
2120
  // ever granted it), so it is unconditionally denied whenever this is called
2141
2121
  // with a non-empty gate; "read" is denied exactly when this call's canonical
2142
- // tools omit "Read" (a tool-less step: summarizer, scout).
2122
+ // tools omit "Read" (a tool-less step: the summarizer).
2143
2123
  function deniedFreeTierGatewayTools(
2144
2124
  canonicalTools: readonly string[],
2145
2125
  freeTierGatewayTools: readonly string[],
@@ -2382,11 +2362,11 @@ export function createOpenCodeClient(
2382
2362
  // FREE_TIER_GATEWAY_TOOLS). The registry caches ONE opencode client per
2383
2363
  // route (transport-registry.ts), so every step routed to the same
2384
2364
  // provider_free model shares this server — a hunter-shaped step (grants
2385
- // read) and a tool-less step (summarizer/scout) on that SAME route would
2365
+ // read) and a tool-less step (the summarizer) on that SAME route would
2386
2366
  // otherwise silently ride whichever one's ask set the server happened to be
2387
2367
  // launched with. This is fixed at spawn exactly like the MCP registry
2388
2368
  // above, and the pipeline's own phase boundaries (hunters all finish before
2389
- // dedupe releases the idle server, refuter/scout run in a later phase)
2369
+ // dedupe releases the idle server, the refuter runs in a later phase)
2390
2370
  // mean a real run only ever asks this server to relaunch, never to change
2391
2371
  // shape mid-flight — a genuine concurrent mismatch fails loud instead of
2392
2372
  // silently weakening.