wachiman 1.12.0 → 1.14.0

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 (46) hide show
  1. package/README.md +132 -101
  2. package/bin/wachiman.js +1 -1
  3. package/package.json +2 -3
  4. package/skills/wachiman-loop/SKILL.md +91 -0
  5. package/skills/wachiman-review/SKILL.md +2 -2
  6. package/src/call-message.ts +26 -3
  7. package/src/cli.ts +113 -18
  8. package/src/config.ts +99 -32
  9. package/src/connectors/local-git.ts +102 -14
  10. package/src/cost.ts +48 -1
  11. package/src/doctor/checks.ts +1 -1
  12. package/src/engine.ts +910 -89
  13. package/src/estimate.ts +64 -0
  14. package/src/fidelity/block.ts +263 -0
  15. package/src/fidelity/claims-call.ts +172 -0
  16. package/src/fidelity/claims.ts +69 -0
  17. package/src/fidelity/digest.ts +80 -0
  18. package/src/fidelity/grade.ts +242 -0
  19. package/src/fidelity/inert.ts +25 -0
  20. package/src/fidelity/rules.ts +166 -0
  21. package/src/fidelity/sink.ts +135 -0
  22. package/src/fidelity/summary.ts +16 -0
  23. package/src/fidelity/types.ts +93 -0
  24. package/src/formats.ts +16 -3
  25. package/src/history/fingerprint.ts +1 -1
  26. package/src/history/store.ts +44 -6
  27. package/src/init/options.ts +1 -1
  28. package/src/ledger.ts +306 -0
  29. package/src/local.ts +23 -5
  30. package/src/loop/command.ts +441 -0
  31. package/src/loop/state.ts +88 -0
  32. package/src/loop/stop-rule.ts +98 -0
  33. package/src/providers/claude-code.ts +48 -13
  34. package/src/providers/codex.ts +8 -7
  35. package/src/providers/index.ts +6 -3
  36. package/src/providers/types.ts +13 -3
  37. package/src/sinks/common.ts +3 -3
  38. package/src/sinks/json.ts +13 -2
  39. package/src/skills/install.ts +4 -4
  40. package/src/status/command.ts +1 -1
  41. package/src/store/command.ts +109 -0
  42. package/src/store/figures.ts +153 -0
  43. package/src/store/html.ts +138 -0
  44. package/src/store/index.ts +298 -0
  45. package/templates/caller-workflow.yml +84 -57
  46. package/templates/wachiman-review-caller.yml +13 -0
package/src/cli.ts CHANGED
@@ -10,10 +10,14 @@
10
10
  * when the head was already reviewed
11
11
  * wachiman local [paths…] review the committed branch of a clone on the signed-in
12
12
  * `claude`, and print it; no GitHub call (src/local.ts)
13
+ * wachiman loop [--base <ref>] one review round of a fix loop on the signed-in `claude`; exit 0
14
+ * converged, 10 findings open, 11 stopped (src/loop/command.ts)
13
15
  * wachiman ask [question…] answer a question about the local change, or explain it; one
14
16
  * call, no GitHub request, grounded in the stored review (src/local-ask.ts)
15
17
  * wachiman history list and judge what `local --history` kept on this laptop
16
18
  * wachiman status what the local history kept and its estimated cost; files only
19
+ * wachiman dashboard spend, findings and loops across every repository in the history, as one
20
+ * static HTML file or JSON, from a rebuildable index (src/store/)
17
21
  * wachiman setup [--profile <name>] bind this repository to an account profile (home config only);
18
22
  * on a terminal without --profile it asks
19
23
  * wachiman init write the GitHub caller workflow, or hand over to `setup`
@@ -23,8 +27,10 @@
23
27
  * wachiman skills install|uninstall copy the agent skills into place, or remove what was copied
24
28
  * wachiman --setup-profile [name] sign in an isolated Claude Code profile under the Wachiman home
25
29
  *
26
- * Exit codes: 0 posted or nothing to do; 1 engine error (nothing posted); 2 usage error;
27
- * 130 cancelled (`local`, `setup`, `init`).
30
+ * Exit codes: 0 posted or nothing to do; 1 engine error (nothing posted); 2 usage error; 3 the spend
31
+ * ledger stopped the run before any model call (a limit with `limit_mode: fail`, a repository it does
32
+ * not know, a job without an OIDC token, or an unreachable ledger with `ledger_unavailable: stop`);
33
+ * 130 cancelled (`local`, `loop`, `setup`, `init`); `loop` also exits 10 and 11 (see its help).
28
34
  * The settings are documented in docs/cli.md.
29
35
  */
30
36
  import { existsSync, readFileSync, statSync, writeSync } from "node:fs";
@@ -33,8 +39,10 @@ import { writeStepSummary, type CallCost } from "./cost";
33
39
  import { loadConfig, UsageError } from "./config";
34
40
  import { historyUsage, parseHistoryArgs, runHistory } from "./history/command";
35
41
  import { parseStatusArgs, runStatus, statusUsage } from "./status/command";
36
- import { comment, defaultGh, isThreadQuestion, review, threadReplies, type Deps, type ReviewLineComment } from "./engine";
42
+ import { dashboardUsage, parseDashboardArgs, runDashboard } from "./store/command";
43
+ import { LedgerStopped, SpendLimitReached, comment, defaultGh, isThreadQuestion, review, threadReplies, type Deps, type GhFn, type ReviewLineComment } from "./engine";
37
44
  import { localUsage, parseLocalArgs, runLocal } from "./local";
45
+ import { loopUsage, parseLoopArgs, runLoop } from "./loop/command";
38
46
  import { askUsage, parseAskArgs, runAsk } from "./local-ask";
39
47
  import { parseProfileArgs, parseProjectArgs, profileUsage, projectUsage, runProfile, runProject, setupUsage } from "./accounts/commands";
40
48
  import { wachimanHome } from "./accounts/config";
@@ -45,6 +53,7 @@ import { DEFAULT_WORKFLOW_REF } from "./init/options";
45
53
  import { runInitCommand } from "./init/run";
46
54
  import { projectCommands } from "./wizard/command";
47
55
  import { DEFAULT_ISOLATED_NAME, setupProfile } from "./profiles";
56
+ import { resetConfidence } from "./fidelity/sink";
48
57
  import { getProvider, PROVIDER_NAMES, UnknownProviderError } from "./providers";
49
58
  import { parseSetupFlags } from "./setup/args";
50
59
  import { runSetupCommand } from "./setup/run";
@@ -53,11 +62,20 @@ export const USAGE = `Usage:
53
62
  wachiman review <pr> [paths…] [--repo-dir <dir>] Review a pull request (or only the given paths)
54
63
  wachiman comment <pr> [--repo-dir <dir>] Handle the comment in $REQUEST (a review or an answer)
55
64
  wachiman threads <pr> [--repo-dir <dir>] Answer the pending Wachiman threads (an automatic run)
65
+ wachiman reset-confidence <pr> [--commit <sha>] Write "cannot say" into the PR Confidence block (no model call;
66
+ the first step of a review run, with WACHIMAN_CONFIDENCE=1)
56
67
  ${localUsage}
57
68
  Review the committed branch of a clone on your
58
69
  signed-in claude, and print it; nothing is posted.
59
70
  --commit: that one commit (<commit>^ → <commit>);
60
71
  --from: that commit forward (<commit>^ → HEAD)
72
+ ${loopUsage}
73
+ One round of a review loop on your signed-in claude:
74
+ --base starts it (round 1 reviews the whole range),
75
+ a call with a loop open continues it and reviews only
76
+ the commits since the last round. Fix, or judge a
77
+ finding wrong with a reason, between calls; the CLI
78
+ never edits code. Three rounds by default
61
79
  ${askUsage}
62
80
  Answer a question about the change the scope flags
63
81
  select, or explain it when no question is given. One
@@ -72,6 +90,17 @@ export const USAGE = `Usage:
72
90
  What this laptop kept of its local runs and their
73
91
  estimated cost, filterable; uses local history. GitHub
74
92
  Actions runs are not counted
93
+ ${dashboardUsage}
94
+ Write one static HTML page (or print JSON) with the
95
+ spend per day and repository, the findings by severity
96
+ and verdict and the loop rounds of every repository in
97
+ the local history. Built from an index under the
98
+ Wachiman home that is rebuilt whenever it is missing or
99
+ out of date. It writes only that index and, without
100
+ --json, the page, and, like \`wachiman history\`, moves
101
+ a damaged run file aside. No claude, no GitHub, no
102
+ network, no server, no browser. --repo-dir is accepted
103
+ and unused
75
104
  ${setupUsage}
76
105
  Bind this repository to an account profile; written
77
106
  in the Wachiman home, never in the repository. On a
@@ -92,7 +121,7 @@ export const USAGE = `Usage:
92
121
  Claude configuration directory, used in place
93
122
  ${projectUsage.split("\n").join("\n ")}
94
123
  ${skillsUsage.split("\n").join("\n ")}
95
- Copy the wachiman-review and wachiman-audit agent
124
+ Copy the wachiman-review, wachiman-loop and wachiman-audit agent
96
125
  skills into ~/.agents/skills and each Claude Code
97
126
  config directory; uninstall removes only the copies
98
127
  it marked. Never touches ~/.wachiman
@@ -107,17 +136,23 @@ Options:
107
136
 
108
137
  Environment:
109
138
  GH_TOKEN, GITHUB_REPOSITORY required (not read by \`local\`, \`ask\` or \`--setup-profile\`)
110
- ANTHROPIC_API_KEY required by the claude-code provider (not in a dry run);
111
- \`local\` and \`ask\` refuse to start when it is set
139
+ ANTHROPIC_API_KEY the claude-code provider needs this or CLAUDE_CODE_OAUTH_TOKEN
140
+ (not in a dry run); \`local\` and \`ask\` refuse to start when it is set
141
+ CLAUDE_CODE_OAUTH_TOKEN a Claude subscription token (\`claude setup-token\`); if set, it is
142
+ used and ANTHROPIC_API_KEY never reaches the agent
112
143
  REQUEST the comment text, for \`comment\`
113
- AGENT_REVIEW_COMMENT_AUTHOR the login of the comment's author, for \`comment\` and \`threads\`
114
- AGENT_REVIEW_* optional settings; see docs/cli.md
144
+ WACHIMAN_COMMENT_AUTHOR the login of the comment's author, for \`comment\` and \`threads\`
145
+ WACHIMAN_* optional settings; see docs/cli.md
115
146
 
116
- Exit codes: 0 posted or nothing to do, 1 engine error, 2 usage error.
147
+ Exit codes: 0 posted or nothing to do, 1 engine error, 2 usage error, 3 the spend ledger stopped the run
148
+ (a limit with \`limit_mode: fail\`, an unlinked repository, no OIDC token, or an unreachable ledger with
149
+ \`ledger_unavailable: stop\`); with \`limit_mode: notice\` a limit exits 0.
117
150
  \`local\`: 0 a report was printed, 1 no call returned a review, 2 usage error, 130 interrupted.
151
+ \`loop\`: 0 converged, 10 findings still open, 11 stopped without converging (round cap or review failed), 2 usage error, 130 interrupted. A stop is never an approval.
118
152
  \`ask\`: 0 an answer was printed (it may say what it could not answer), 1 no answer (the call failed or returned nothing usable), 2 usage error, 130 interrupted.
119
153
  \`history\`: 0 listed (or nothing to list), 2 usage error, including an unknown finding id.
120
154
  \`status\`: 0 printed (also with no run kept), 2 usage error.
155
+ \`dashboard\`: 0 written or printed (also with no run kept), 2 usage error, 1 the index or the page could not be written.
121
156
  \`doctor\`: 0 every required check passed, 1 one failed. \`setup\`: 1 when its closing doctor summary failed.
122
157
  \`setup\`, \`init\`: 130 cancelled (Escape, Ctrl-C, Ctrl-D); nothing pending is saved.
123
158
  \`skills\`: 0 done (copies left in place are warned about), 1 a copy could not be written, 2 usage error or a refused target.
@@ -129,7 +164,7 @@ export function readVersion(dir = import.meta.dir) {
129
164
  const path = join(dir, "..", "package.json");
130
165
  if (!existsSync(path)) return "unknown";
131
166
  const pkg = JSON.parse(readFileSync(path, "utf8"));
132
- return pkg.name === "wachiman" || pkg.name === "agent-review" ? String(pkg.version) : "unknown";
167
+ return pkg.name === "wachiman" ? String(pkg.version) : "unknown";
133
168
  } catch {
134
169
  return "unknown";
135
170
  }
@@ -194,7 +229,7 @@ export function routeMode(mode: string, commentLine: ReviewLineComment | undefin
194
229
  * comment's text is `REQUEST`.
195
230
  */
196
231
  export function readCommentLine(env: Record<string, string | undefined>): ReviewLineComment | undefined {
197
- const id = env.AGENT_REVIEW_COMMENT_ID ?? "";
232
+ const id = env.WACHIMAN_COMMENT_ID ?? "";
198
233
  if (id === "") return undefined;
199
234
  const number = (name: string) => {
200
235
  const raw = env[name] ?? "";
@@ -202,12 +237,12 @@ export function readCommentLine(env: Record<string, string | undefined>): Review
202
237
  if (!/^\d+$/.test(raw)) throw new UsageError(`${name} must be a comment id or a line number, got "${raw}".`);
203
238
  return Number(raw);
204
239
  };
205
- if (!/^\d+$/.test(id)) throw new UsageError(`AGENT_REVIEW_COMMENT_ID must be a comment id, got "${id}".`);
240
+ if (!/^\d+$/.test(id)) throw new UsageError(`WACHIMAN_COMMENT_ID must be a comment id, got "${id}".`);
206
241
  return {
207
242
  id: Number(id),
208
- replyTo: number("AGENT_REVIEW_COMMENT_IN_REPLY_TO"),
209
- path: env.AGENT_REVIEW_COMMENT_PATH || null,
210
- line: number("AGENT_REVIEW_COMMENT_LINE"),
243
+ replyTo: number("WACHIMAN_COMMENT_IN_REPLY_TO"),
244
+ path: env.WACHIMAN_COMMENT_PATH || null,
245
+ line: number("WACHIMAN_COMMENT_LINE"),
211
246
  };
212
247
  }
213
248
 
@@ -247,6 +282,37 @@ async function runWizard(mode: "setup" | "init", args: string[], repoDir: string
247
282
  });
248
283
  }
249
284
 
285
+ /**
286
+ * `wachiman reset-confidence <pr> [--commit <sha>]`: the first step of a review run (ADR 0033
287
+ * decision 14). It rewrites the PR Confidence block to `cannot say` for the commit under review
288
+ * before the push wait, so a run cancelled during it leaves no letter from an older commit. No
289
+ * model, no provider. A failure inside the reset itself (reading or writing the description) is a
290
+ * warning and the exit code stays 0. A usage or configuration error is not: a bad pull request
291
+ * number or flag, an invalid `WACHIMAN_*` value, or a missing `GH_TOKEN` or
292
+ * `GITHUB_REPOSITORY` exits non-zero. The workflow steps run it with `continue-on-error`.
293
+ */
294
+ export function runResetConfidence(args: string[], env: Record<string, string | undefined>, gh: GhFn = defaultGh): number {
295
+ const [prArg, ...flags] = args;
296
+ const pr = Number(prArg);
297
+ if (!/^\d+$/.test(prArg ?? "") || pr <= 0) throw new UsageError("reset-confidence needs a pull request number.");
298
+ let commit: string | undefined;
299
+ for (let i = 0; i < flags.length; i++) {
300
+ if (flags[i] !== "--commit" || !flags[i + 1]) throw new UsageError(`reset-confidence: unexpected "${flags[i]}".`);
301
+ commit = flags[++i];
302
+ }
303
+ const config = loadConfig(env);
304
+ if (!config.confidence) return 0;
305
+ required(env, "GH_TOKEN");
306
+ const repo = required(env, "GITHUB_REPOSITORY");
307
+ resetConfidence(
308
+ { gh, dryRun: config.dryRun, log: (line) => console.log(line), warn: (line) => console.error(line) },
309
+ repo,
310
+ pr,
311
+ commit,
312
+ );
313
+ return 0;
314
+ }
315
+
250
316
  export async function main(argv: string[], env: Record<string, string | undefined>): Promise<number> {
251
317
  const mode = argv[0];
252
318
  if (mode === "--help" || mode === "-h") {
@@ -281,6 +347,14 @@ export async function main(argv: string[], env: Record<string, string | undefine
281
347
  engineVersion: readVersion(),
282
348
  });
283
349
  }
350
+ if (mode === "loop") {
351
+ // Like `local`: no GitHub token, no `gh` function anywhere on this path.
352
+ return runLoop(parseLoopArgs(args.slice(1)), repoDir, env, {
353
+ stdout: (text) => writeAll(1, text),
354
+ stderr: (line) => console.error(line),
355
+ engineVersion: readVersion(),
356
+ });
357
+ }
284
358
  if (mode === "ask") {
285
359
  // Like `local`: no GitHub token, no `gh` function anywhere on this path.
286
360
  return runAsk(parseAskArgs(args.slice(1)), repoDir, env, {
@@ -298,9 +372,15 @@ export async function main(argv: string[], env: Record<string, string | undefine
298
372
  if ("help" in parsed) return (writeAll(1, `${statusUsage}\n`), 0);
299
373
  return runStatus(parsed, repoDir, env, { stdout: (text) => writeAll(1, text), stderr: (line) => console.error(line), version: readVersion() });
300
374
  }
375
+ if (mode === "dashboard") {
376
+ const parsed = parseDashboardArgs(args.slice(1));
377
+ if ("help" in parsed) return (writeAll(1, `${dashboardUsage}\n`), 0);
378
+ return runDashboard(parsed, env, { stdout: (text) => writeAll(1, text), stderr: (line) => console.error(line) });
379
+ }
301
380
  if (mode === "history") {
302
381
  return runHistory(parseHistoryArgs(args.slice(1)), repoDir, env, { stdout: (text) => writeAll(1, text), stderr: (line) => console.error(line) });
303
382
  }
383
+ if (mode === "reset-confidence") return runResetConfidence(args.slice(1), env);
304
384
  const [, prArg, ...rest] = args;
305
385
  if (mode !== "review" && mode !== "comment" && mode !== "threads") {
306
386
  throw new UsageError(mode ? `Unknown command "${mode}".` : "No command given.");
@@ -326,8 +406,10 @@ export async function main(argv: string[], env: Record<string, string | undefine
326
406
  const repo = required(env, "GITHUB_REPOSITORY");
327
407
  const request = effective === "comment" ? required(env, "REQUEST") : "";
328
408
  const costs: CallCost[] = [];
409
+ const ledgerNotes: string[] = [];
329
410
  const deps: Deps = {
330
411
  costs,
412
+ ledgerNotes,
331
413
  gh: defaultGh,
332
414
  provider,
333
415
  ...(verifier ? { verifier } : {}),
@@ -347,11 +429,11 @@ export async function main(argv: string[], env: Record<string, string | undefine
347
429
  }
348
430
  try {
349
431
  if (effective === "review") review(deps, repo, pr, rest);
350
- else if (effective === "threads") threadReplies(deps, repo, pr, env.AGENT_REVIEW_COMMENT_AUTHOR || undefined);
351
- else comment(deps, repo, pr, request, env.AGENT_REVIEW_COMMENT_AUTHOR || undefined, commentLine);
432
+ else if (effective === "threads") threadReplies(deps, repo, pr, env.WACHIMAN_COMMENT_AUTHOR || undefined);
433
+ else comment(deps, repo, pr, request, env.WACHIMAN_COMMENT_AUTHOR || undefined, commentLine);
352
434
  } finally {
353
435
  // Also after a failed run: the calls it made were paid for.
354
- writeStepSummary(costs, env);
436
+ writeStepSummary(costs, env, undefined, ledgerNotes);
355
437
  }
356
438
  return 0;
357
439
  } catch (error) {
@@ -359,6 +441,19 @@ export async function main(argv: string[], env: Record<string, string | undefine
359
441
  console.error(`wachiman: ${error.message}\n\n${USAGE}`);
360
442
  return 2;
361
443
  }
444
+ if (error instanceof SpendLimitReached) {
445
+ // The notice is already posted. `notice` mode keeps the job green.
446
+ if (error.mode === "notice") {
447
+ console.log(`wachiman: spend limit reached: ${error.message}`);
448
+ return 0;
449
+ }
450
+ console.error(`wachiman: spend limit reached: ${error.message}`);
451
+ return 3;
452
+ }
453
+ if (error instanceof LedgerStopped) {
454
+ console.error(`wachiman: ledger: ${error.message}`);
455
+ return 3;
456
+ }
362
457
  console.error(`wachiman: ${error instanceof Error ? error.message : error}`);
363
458
  return 1;
364
459
  }
package/src/config.ts CHANGED
@@ -1,4 +1,4 @@
1
- /** The `AGENT_REVIEW_*` settings, read once, with the old `CLAUDE_REVIEW_*` names as a fallback. */
1
+ /** The `WACHIMAN_*` settings, read once, with the old `CLAUDE_REVIEW_*` names as a fallback. */
2
2
  import { DEFAULT_MODELS } from "./providers";
3
3
  import { DEFAULT_CALL_TIMEOUT_S, DEFAULT_EFFORT, DEFAULT_MAX_TURNS, EFFORTS, type Effort } from "./providers/types";
4
4
 
@@ -14,19 +14,43 @@ export type Trigger = string;
14
14
  /** Whether a run started by `trigger` spends from the pull request's automatic ceiling. */
15
15
  export const isAutomatic = (trigger: Trigger | undefined) => trigger !== undefined && trigger !== "comment";
16
16
 
17
+ export type LimitMode = "fail" | "notice";
18
+ export type LedgerUnavailableMode = "run" | "stop";
19
+
17
20
  export interface EngineConfig {
18
21
  /** Total per review invocation, USD. */
19
22
  budgetUsd: number;
20
23
  /**
21
- * `AGENT_REVIEW_AUTO_BUDGET_USD`: the most all automatic runs on one pull request may spend
24
+ * `WACHIMAN_AUTO_BUDGET_USD`: the most all automatic runs on one pull request may spend
22
25
  * together (the first review, push reviews, thread runs), summed from the engine's own signed
23
26
  * footers. The operator's `comment` is not limited by it.
24
27
  */
25
28
  autoBudgetUsd: number;
26
- /** `AGENT_REVIEW_THREADS_BUDGET_USD`: the one provider call of a thread-only run. */
29
+ /**
30
+ * `WACHIMAN_SIZE_GATE_USD`: the most a review's estimate may be before a GitHub run skips it
31
+ * and posts one comment instead. The effective gate is never below what the run may spend
32
+ * (`sizeGate` in estimate.ts), so raising the budgets lifts it. Local runs only warn.
33
+ */
34
+ sizeGateUsd: number;
35
+ /**
36
+ * `WACHIMAN_LEDGER_URL`: the spend ledger service, trimmed, without a trailing slash. Unset
37
+ * means no ledger call of any kind. It is also the OIDC audience, as configured.
38
+ */
39
+ ledgerUrl?: string;
40
+ /** `WACHIMAN_DAILY_BUDGET_USD`: a stricter daily number for this repository; it never raises the account's limit. */
41
+ dailyBudgetUsd?: number;
42
+ /** `WACHIMAN_WEEKLY_BUDGET_USD`: the same, per week. */
43
+ weeklyBudgetUsd?: number;
44
+ /** `WACHIMAN_LIMIT_MODE`: when a limit refuses a run, `fail` exits 3 (red job); `notice` posts the comment and passes. */
45
+ limitMode: LimitMode;
46
+ /** `WACHIMAN_LIMIT_BYPASS`: `1` or `true` skips the ledger for this run. */
47
+ limitBypass: boolean;
48
+ /** `WACHIMAN_LEDGER_UNAVAILABLE`: with an unreachable ledger, `run` reviews anyway (and says so); `stop` exits 3 before any model call. */
49
+ ledgerUnavailable: LedgerUnavailableMode;
50
+ /** `WACHIMAN_THREADS_BUDGET_USD`: the one provider call of a thread-only run. */
27
51
  threadsBudgetUsd: number;
28
52
  /**
29
- * `AGENT_REVIEW_TRIGGER`: why this run started (see `Trigger`). Unset means a person ran the
53
+ * `WACHIMAN_TRIGGER`: why this run started (see `Trigger`). Unset means a person ran the
30
54
  * CLI by hand: not automatic, so the ceiling does not apply. `comment` mode ignores it.
31
55
  */
32
56
  trigger?: Trigger;
@@ -36,32 +60,32 @@ export interface EngineConfig {
36
60
  batchChars: number;
37
61
  dryRun: boolean;
38
62
  /**
39
- * `AGENT_REVIEW_FULL_REVIEW`: a dry run reviews the whole pull request, as if it had never been
63
+ * `WACHIMAN_FULL_REVIEW`: a dry run reviews the whole pull request, as if it had never been
40
64
  * reviewed, without reading the previous-review marker. Only together with `dryRun`.
41
65
  */
42
66
  fullReview: boolean;
43
67
  /** The provider's default (`DEFAULT_MODELS`) when unset. */
44
68
  model: string;
45
69
  provider: string;
46
- /** `AGENT_REVIEW_PRICE_PER_MTOK`: "input,cached_input,output" USD per million tokens, for a Codex model without a built-in price. */
70
+ /** `WACHIMAN_PRICE_PER_MTOK`: "input,cached_input,output" USD per million tokens, for a Codex model without a built-in price. */
47
71
  pricePerMTok?: string;
48
72
  /**
49
- * `AGENT_REVIEW_EFFORT`: the reasoning effort, always explicit so a CLI default change cannot
73
+ * `WACHIMAN_EFFORT`: the reasoning effort, always explicit so a CLI default change cannot
50
74
  * alter reviews. Passed as `--effort` (Claude Code) or `model_reasoning_effort` (Codex).
51
75
  */
52
76
  effort: Effort;
53
77
  /**
54
- * `AGENT_REVIEW_MAX_TURNS`: the most agent turns one provider call may take. Passed as
78
+ * `WACHIMAN_MAX_TURNS`: the most agent turns one provider call may take. Passed as
55
79
  * `--max-turns` (Claude Code); Codex has no equivalent and ignores it.
56
80
  */
57
81
  maxTurns: number;
58
82
  /**
59
- * `AGENT_REVIEW_CALL_TIMEOUT_S`: wall-clock limit on one provider call, in seconds. The call is
83
+ * `WACHIMAN_CALL_TIMEOUT_S`: wall-clock limit on one provider call, in seconds. The call is
60
84
  * killed when it passes and ends as `timeout`.
61
85
  */
62
86
  callTimeoutS: number;
63
87
  /**
64
- * `AGENT_REVIEW_ON_PROVIDER_UNAVAILABLE`: what a run does when the provider rejects every call
88
+ * `WACHIMAN_ON_PROVIDER_UNAVAILABLE`: what a run does when the provider rejects every call
65
89
  * because the reviewer's own account cannot pay or authenticate (or stays rate-limited). `warn`
66
90
  * posts a notice and finishes green; `fail` posts it and fails the job.
67
91
  */
@@ -70,64 +94,71 @@ export interface EngineConfig {
70
94
  /** Login allowed to command the tool. `comment` refuses to run when it is empty or a placeholder. */
71
95
  operator: string;
72
96
  /**
73
- * `AGENT_REVIEW_ASKERS`: who may ask a question (ADR 0032 decision 5). `operator` (default): only
97
+ * `WACHIMAN_ASKERS`: who may ask a question (ADR 0032 decision 5). `operator` (default): only
74
98
  * the operator. `write`: also a collaborator whose permission on the repository is admin,
75
99
  * maintain or write, checked by the engine on each question. Reviews stay operator-only.
76
100
  */
77
101
  askers: "operator" | "write";
78
102
  /**
79
- * `AGENT_REVIEW_ASK_CEILING_USD`: the most the questions of everyone but the operator may spend
103
+ * `WACHIMAN_ASK_CEILING_USD`: the most the questions of everyone but the operator may spend
80
104
  * on one pull request, summed from the engine's own footers. The operator's questions are not limited.
81
105
  */
82
106
  askCeilingUsd: number;
83
107
  /** Extra exclude globs on top of the built-in ones. */
84
108
  exclude: string[];
85
109
  /**
86
- * `AGENT_REVIEW_VERIFY`: after the batches, one more call tries to refute every finding above
110
+ * `WACHIMAN_VERIFY`: after the batches, one more call tries to refute every finding above
87
111
  * `info` before it is posted (ADR 0027). Off by default; enabled only by `1` or `true`.
88
112
  */
89
113
  verify: boolean;
90
- /** `AGENT_REVIEW_VERIFY_BUDGET_USD`: the verifier's reserve, taken out of `budgetUsd` (at most half of it). */
114
+ /** `WACHIMAN_VERIFY_BUDGET_USD`: the verifier's reserve, taken out of `budgetUsd` (at most half of it). */
91
115
  verifyBudgetUsd: number;
92
- /** `AGENT_REVIEW_VERIFIER_MODEL`: the verifier's model; it always runs at effort `high`. */
116
+ /** `WACHIMAN_VERIFIER_MODEL`: the verifier's model; it always runs at effort `high`. */
93
117
  verifierModel: string;
118
+ /**
119
+ * `WACHIMAN_CONFIDENCE`: PR Confidence (ADR 0033). A full review of a pull request lists the
120
+ * claims in its title and body with one more small call, grades the change against them and
121
+ * writes the grade into the description. Off by default; enabled only by `1` or `true`. Read by
122
+ * the pull request review only: `wachiman local` never grades.
123
+ */
124
+ confidence: boolean;
94
125
  }
95
126
 
96
- /** The verifier's model when `AGENT_REVIEW_VERIFIER_MODEL` is unset (Claude Code only). */
127
+ /** The verifier's model when `WACHIMAN_VERIFIER_MODEL` is unset (Claude Code only). */
97
128
  export const DEFAULT_VERIFIER_MODEL = "claude-opus-5-5";
98
129
 
99
130
  type Env = Record<string, string | undefined>;
100
131
 
101
132
  /**
102
- * `AGENT_REVIEW_<name>`, else the deprecated `CLAUDE_REVIEW_<name>` (with a one-line notice on
133
+ * `WACHIMAN_<name>`, else the deprecated `CLAUDE_REVIEW_<name>` (with a one-line notice on
103
134
  * stderr), else undefined. An empty value counts as unset.
104
135
  */
105
136
  export function readSetting(env: Env, name: string, warn: (line: string) => void) {
106
- const current = env[`AGENT_REVIEW_${name}`];
137
+ const current = env[`WACHIMAN_${name}`];
107
138
  if (current !== undefined && current !== "") return current;
108
139
  const old = env[`CLAUDE_REVIEW_${name}`];
109
140
  if (old !== undefined && old !== "") {
110
- warn(`CLAUDE_REVIEW_${name} is deprecated; use AGENT_REVIEW_${name}.`);
141
+ warn(`CLAUDE_REVIEW_${name} is deprecated; use WACHIMAN_${name}.`);
111
142
  return old;
112
143
  }
113
144
  return undefined;
114
145
  }
115
146
 
116
147
  /**
117
- * `AGENT_REVIEW_EFFORT`, else the deprecated `AGENT_REVIEW_REASONING_EFFORT` (Codex only before),
148
+ * `WACHIMAN_EFFORT`, else the deprecated `WACHIMAN_REASONING_EFFORT` (Codex only before),
118
149
  * else the default. An unknown value is a configuration error: a silent fallback would run the
119
150
  * review at an effort nobody chose.
120
151
  */
121
152
  function readEffort(env: Env, warn: (line: string) => void): Effort {
122
153
  const current = readSetting(env, "EFFORT", warn);
123
154
  const old = readSetting(env, "REASONING_EFFORT", () => {});
124
- let name = "AGENT_REVIEW_EFFORT";
155
+ let name = "WACHIMAN_EFFORT";
125
156
  let raw = current;
126
157
  if (old !== undefined) {
127
- warn("AGENT_REVIEW_REASONING_EFFORT is deprecated; use AGENT_REVIEW_EFFORT.");
158
+ warn("WACHIMAN_REASONING_EFFORT is deprecated; use WACHIMAN_EFFORT.");
128
159
  if (raw === undefined) {
129
160
  raw = old;
130
- name = "AGENT_REVIEW_REASONING_EFFORT";
161
+ name = "WACHIMAN_REASONING_EFFORT";
131
162
  }
132
163
  }
133
164
  if (raw === undefined) return DEFAULT_EFFORT;
@@ -144,44 +175,73 @@ export function loadConfig(env: Env, warn: (line: string) => void = console.erro
144
175
  if (raw === undefined) return fallback;
145
176
  const value = Number(raw);
146
177
  if (!Number.isFinite(value) || value <= 0 || (integer && !Number.isInteger(value))) {
147
- throw new UsageError(`AGENT_REVIEW_${name} must be a positive ${integer ? "integer" : "number"}, got "${raw}".`);
178
+ throw new UsageError(`WACHIMAN_${name} must be a positive ${integer ? "integer" : "number"}, got "${raw}".`);
148
179
  }
149
180
  return value;
150
181
  };
151
182
  const effort = readEffort(env, warn);
152
183
  const onProviderUnavailable = get("ON_PROVIDER_UNAVAILABLE") ?? "warn";
153
184
  if (onProviderUnavailable !== "warn" && onProviderUnavailable !== "fail") {
154
- throw new UsageError(`AGENT_REVIEW_ON_PROVIDER_UNAVAILABLE must be warn or fail, got "${onProviderUnavailable}".`);
185
+ throw new UsageError(`WACHIMAN_ON_PROVIDER_UNAVAILABLE must be warn or fail, got "${onProviderUnavailable}".`);
155
186
  }
156
187
  const provider = get("PROVIDER") ?? "claude-code";
157
188
  const dryRun = get("DRY_RUN") === "1";
158
189
  const fullReview = get("FULL_REVIEW") === "1";
159
190
  if (fullReview && !dryRun) {
160
191
  throw new UsageError(
161
- "AGENT_REVIEW_FULL_REVIEW=1 works only with AGENT_REVIEW_DRY_RUN=1: a full review ignores the previous-review marker, so it is printed, never posted.",
192
+ "WACHIMAN_FULL_REVIEW=1 works only with WACHIMAN_DRY_RUN=1: a full review ignores the previous-review marker, so it is printed, never posted.",
162
193
  );
163
194
  }
164
195
  const verifyRaw = get("VERIFY");
165
196
  if (verifyRaw !== undefined && verifyRaw !== "0" && verifyRaw !== "1" && verifyRaw !== "true" && verifyRaw !== "false") {
166
- throw new UsageError(`AGENT_REVIEW_VERIFY must be 1, true, 0 or false (unset means off), got "${verifyRaw}".`);
197
+ throw new UsageError(`WACHIMAN_VERIFY must be 1, true, 0 or false (unset means off), got "${verifyRaw}".`);
167
198
  }
168
199
  const verifierModel = get("VERIFIER_MODEL") ?? (provider === "claude-code" ? DEFAULT_VERIFIER_MODEL : undefined);
169
200
  // Verification is opt-in, including when a verifier model is configured.
170
201
  const verify = verifyRaw === "1" || verifyRaw === "true";
171
202
  if (verify && verifierModel === undefined) {
172
- throw new UsageError(`AGENT_REVIEW_VERIFY needs AGENT_REVIEW_VERIFIER_MODEL with the ${provider} provider: the default verifier is a Claude model.`);
203
+ throw new UsageError(`WACHIMAN_VERIFY needs WACHIMAN_VERIFIER_MODEL with the ${provider} provider: the default verifier is a Claude model.`);
204
+ }
205
+ const confidenceRaw = get("CONFIDENCE");
206
+ if (confidenceRaw !== undefined && !["0", "1", "true", "false"].includes(confidenceRaw)) {
207
+ throw new UsageError(`WACHIMAN_CONFIDENCE must be 1, true, 0 or false (unset means off), got "${confidenceRaw}".`);
173
208
  }
174
209
  const askers = get("ASKERS") ?? "operator";
175
210
  if (askers !== "operator" && askers !== "write") {
176
- throw new UsageError(`AGENT_REVIEW_ASKERS must be operator or write, got "${askers}".`);
211
+ throw new UsageError(`WACHIMAN_ASKERS must be operator or write, got "${askers}".`);
177
212
  }
178
213
  const trigger = get("TRIGGER")?.trim();
179
214
  if (trigger !== undefined && !/^[a-z_]+$/.test(trigger)) {
180
- throw new UsageError(`AGENT_REVIEW_TRIGGER must be a lowercase event action such as synchronize, got "${trigger}".`);
215
+ throw new UsageError(`WACHIMAN_TRIGGER must be a lowercase event action such as synchronize, got "${trigger}".`);
216
+ }
217
+ const limitMode = get("LIMIT_MODE") ?? "fail";
218
+ if (limitMode !== "fail" && limitMode !== "notice") {
219
+ throw new UsageError(`WACHIMAN_LIMIT_MODE must be fail or notice, got "${limitMode}".`);
220
+ }
221
+ const ledgerUnavailable = get("LEDGER_UNAVAILABLE") ?? "run";
222
+ if (ledgerUnavailable !== "run" && ledgerUnavailable !== "stop") {
223
+ throw new UsageError(`WACHIMAN_LEDGER_UNAVAILABLE must be run or stop, got "${ledgerUnavailable}".`);
224
+ }
225
+ const bypass = get("LIMIT_BYPASS")?.trim();
226
+ const ledgerUrl = get("LEDGER_URL")?.trim().replace(/\/+$/, "") || undefined;
227
+ // The job's OIDC token is sent to this address: never over cleartext, except to the local machine for tests.
228
+ if (ledgerUrl !== undefined && !/^https:\/\/\S/i.test(ledgerUrl) && !/^http:\/\/(localhost|127\.0\.0\.1)(?=[:/]|$)/i.test(ledgerUrl)) {
229
+ throw new UsageError(
230
+ `WACHIMAN_LEDGER_URL must start with https:// (http://localhost and http://127.0.0.1 are allowed for tests), got "${ledgerUrl}".`,
231
+ );
232
+ }
233
+ // With a ledger, `review-platform.yml` grants `id-token: write`, so the request token is in the
234
+ // engine's environment; a Codex sandbox could read it from /proc and mint tokens as this job.
235
+ // Without `ledger_url` nothing uses the token and the guard does not apply.
236
+ if (ledgerUrl !== undefined && provider !== "claude-code" && env.ACTIONS_ID_TOKEN_REQUEST_TOKEN !== undefined) {
237
+ throw new UsageError(
238
+ `WACHIMAN_PROVIDER ${provider} cannot run in a job that holds an OIDC request token: its sandbox can read it from the engine's environment. Use claude-code, or call review.yml instead of review-platform.yml.`,
239
+ );
181
240
  }
182
241
  return {
183
242
  budgetUsd: number("BUDGET_USD", 3),
184
243
  autoBudgetUsd: number("AUTO_BUDGET_USD", 6),
244
+ sizeGateUsd: number("SIZE_GATE_USD", 6),
185
245
  threadsBudgetUsd: number("THREADS_BUDGET_USD", 0.25),
186
246
  trigger: trigger || undefined,
187
247
  batchBudgetUsd: number("BATCH_BUDGET_USD", 0.75),
@@ -205,6 +265,13 @@ export function loadConfig(env: Env, warn: (line: string) => void = console.erro
205
265
  verify,
206
266
  verifyBudgetUsd: number("VERIFY_BUDGET_USD", 1),
207
267
  verifierModel: verifierModel ?? DEFAULT_VERIFIER_MODEL,
268
+ confidence: confidenceRaw === "1" || confidenceRaw === "true",
269
+ ledgerUrl,
270
+ dailyBudgetUsd: get("DAILY_BUDGET_USD") === undefined ? undefined : number("DAILY_BUDGET_USD", 0),
271
+ weeklyBudgetUsd: get("WEEKLY_BUDGET_USD") === undefined ? undefined : number("WEEKLY_BUDGET_USD", 0),
272
+ limitMode,
273
+ limitBypass: bypass === "1" || bypass === "true",
274
+ ledgerUnavailable,
208
275
  };
209
276
  }
210
277
 
@@ -218,10 +285,10 @@ const PLACEHOLDER_OPERATORS = ["your-github-login"];
218
285
  */
219
286
  export function operatorProblem(operator: string | undefined): string | undefined {
220
287
  const login = (operator ?? "").trim();
221
- if (login === "") return "The operator login is empty. Set `operator` (AGENT_REVIEW_OPERATOR) to your GitHub login.";
288
+ if (login === "") return "The operator login is empty. Set `operator` (WACHIMAN_OPERATOR) to your GitHub login.";
222
289
  const lower = login.toLowerCase();
223
290
  if (lower.startsWith("edit") || PLACEHOLDER_OPERATORS.includes(lower)) {
224
- return `The operator login is still the template placeholder "${login}". Set \`operator\` (AGENT_REVIEW_OPERATOR) to your GitHub login.`;
291
+ return `The operator login is still the template placeholder "${login}". Set \`operator\` (WACHIMAN_OPERATOR) to your GitHub login.`;
225
292
  }
226
293
  return undefined;
227
294
  }