@codedology/code-forge 0.9.0 → 0.11.0

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 (62) hide show
  1. package/CHANGELOG.md +93 -0
  2. package/README.md +49 -9
  3. package/docs/reference/blocks.json +587 -1
  4. package/docs/reference/config.md +36 -0
  5. package/package.json +2 -2
  6. package/schema/code-forge.schema.json +89 -0
  7. package/skill/SKILL.md +1 -1
  8. package/skill/references/adapters/claude-code.md +1 -1
  9. package/skill/references/adapters/solo.md +2 -2
  10. package/skill/references/adapters/subprocess.md +12 -2
  11. package/skill/references/code.md +14 -7
  12. package/skill/references/ledger.md +2 -2
  13. package/skill/references/proof.md +77 -2
  14. package/skill/references/review.md +39 -7
  15. package/skill/references/security.md +4 -4
  16. package/skill/scripts/forge +1 -1
  17. package/skill/templates/coder-brief.md +6 -6
  18. package/src/autopilot/binnacle.mjs +9 -3
  19. package/src/cli/block.mjs +262 -31
  20. package/src/cli/jev.mjs +22 -3
  21. package/src/cli/keys.mjs +156 -21
  22. package/src/cli/proof.mjs +35 -7
  23. package/src/cli/review-file.mjs +17 -11
  24. package/src/cli/spawn.mjs +44 -3
  25. package/src/config/validate.mjs +320 -5
  26. package/src/doctor/index.mjs +2 -1
  27. package/src/doctor/local.mjs +27 -6
  28. package/src/doctor/probes.mjs +38 -0
  29. package/src/engines/builders/codex.mjs +44 -4
  30. package/src/engines/sandbox-dirs.mjs +257 -0
  31. package/src/keys/store.mjs +128 -21
  32. package/src/proof/junit.mjs +530 -0
  33. package/src/proof/red-green.mjs +57 -9
  34. package/src/proof/red-parse.mjs +2 -0
  35. package/src/proof/runners.mjs +472 -0
  36. package/src/proof/xcresult.mjs +285 -0
  37. package/src/review/engine.mjs +141 -42
  38. package/src/review/fixloop.mjs +75 -23
  39. package/src/review/gate-check.mjs +3 -2
  40. package/src/review/lenses/A.md +2 -0
  41. package/src/review/lenses/B.md +2 -0
  42. package/src/review/lenses/full.md +2 -0
  43. package/src/review/lenses/quick.md +1 -0
  44. package/src/review/lenses/recheck.md +2 -0
  45. package/src/review/packet.mjs +48 -14
  46. package/src/review/paired.mjs +80 -0
  47. package/src/review/reopen.mjs +4 -6
  48. package/src/review/reviewed-lines.mjs +94 -0
  49. package/src/review/route.mjs +232 -0
  50. package/src/review/schema-fallback.mjs +77 -22
  51. package/src/review/session-retry.mjs +44 -9
  52. package/src/review/test-quality.mjs +201 -0
  53. package/src/review/verb-forms.mjs +39 -0
  54. package/src/session/spawn.mjs +39 -2
  55. package/src/state/block-end.mjs +10 -6
  56. package/src/state/block.mjs +13 -3
  57. package/src/state/config-snapshot.mjs +39 -0
  58. package/src/state/run.mjs +4 -0
  59. package/src/util/forbidden.mjs +2 -2
  60. package/src/worker/engine.mjs +43 -15
  61. package/src/worker/loop.mjs +98 -15
  62. package/src/worker/ticket.mjs +9 -1
@@ -0,0 +1,201 @@
1
+ /**
2
+ * The test-quality note (block B92, issue #10 item 3): a test file that reads a source file and
3
+ * asserts on its text proves text, not behaviour. Its review packets get a `## test-quality`
4
+ * section before the diff that names the lines (never their text), so the reviewer looks at them.
5
+ * The lenses tell the reviewer to raise such a test as a `warning` unless the read file is a
6
+ * fixture. The note is a hint, never a tool-made finding: no row, no gate change.
7
+ *
8
+ * The detector is pure (path and content in, line numbers out; no I/O):
9
+ * - it runs only on a test file: `isTestPath` (`test/ tests/ __tests__/ spec/` dirs, `*.test.*`,
10
+ * `*.spec.*`, `*Test.php`, `*_test.go`, `test_*.py`) plus a `Tests/` dir and a name ending in
11
+ * `Tests.<ext>` or `Test.<ext>` (`FooTests.swift`, `FooTest.kt`). A docs file (`.md`, `.txt`)
12
+ * and a file under a fixtures dir are never checked;
13
+ * - a READ line holds a read call ({@link READ_CALLS}) and, in the call's span, either a quoted
14
+ * path literal that ends in a source extension ({@link SOURCE_EXTENSIONS}), or a name bound on an
15
+ * earlier line to such a literal (`const SRC = 'src/a.mjs'` … `readFileSync(SRC)`). A name is
16
+ * bound only when its value IS the literal: the whole right-hand side, or the literal passed
17
+ * straight to `join`/`path.join`/`resolve`/`path.resolve`; the binding may carry modifiers and a
18
+ * type (`private val`, `const val`, `private static let`, `private static final String`,
19
+ * `private const string`, `export const`). The span is the call's line, plus up to
20
+ * {@link SPAN_LINES} following lines while the read call's own `(` is still open (a formatter
21
+ * split); the call's line is the one reported. A span with any string literal that has a
22
+ * fixtures dir segment (`fixtures`, `fixture`, `__fixtures__`) — `'test/fixtures/a.mjs'` or
23
+ * `join(dir, 'fixtures', 'a.mjs')` — or that names a name bound to such a literal
24
+ * (`const FIXTURES = join(dir, 'fixtures')`) is a fixture read, never a source read;
25
+ * - a TEXT-ASSERT line holds a contains or match call ({@link ASSERT_CALLS});
26
+ * - the file is flagged only when it has at least one of each. The flagged lines are every read
27
+ * line and every text-assert line, 1-based, sorted; the section lists at most
28
+ * {@link MAX_LINES} of them and counts the rest.
29
+ */
30
+
31
+ import { isTestPath } from '../gates/expand.mjs';
32
+
33
+ /** The section names at most this many line numbers. */
34
+ export const MAX_LINES = 10;
35
+
36
+ /** The file extensions of source code a test should not read for its text. */
37
+ export const SOURCE_EXTENSIONS = Object.freeze(['swift', 'ts', 'tsx', 'js', 'mjs', 'cjs', 'py', 'go', 'rs', 'kt', 'java', 'rb', 'php', 'cs']);
38
+
39
+ /** Common read calls, matched as plain text. */
40
+ export const READ_CALLS = Object.freeze(['readFileSync', 'readFile', 'open(', 'read_text', 'String(contentsOf', 'contentsOfFile', 'File.read', 'os.ReadFile', 'fs::read_to_string', 'readText(', 'readString(', 'ReadAllText(']);
41
+
42
+ /** A read call's path may sit on at most this many lines after the call (a formatter split). */
43
+ export const SPAN_LINES = 3;
44
+
45
+ /** Contains or match calls on text, matched as plain text (`XCTAssertTrue(x.contains(…))` holds `contains(`). */
46
+ export const ASSERT_CALLS = Object.freeze(['includes(', 'contains(', 'toContain(', 'toMatch(', 'assert.match(', 'assertIn(', 'strings.Contains(']);
47
+
48
+ const DOCS = /\.(md|txt)$/i;
49
+ const FIXTURE_DIRS = new Set(['fixtures', 'fixture', '__fixtures__']);
50
+ /** A quoted string literal (single, double or back quote) with no quote or newline inside. */
51
+ const LITERAL = /(['"`])([^'"`\n]*)\1/g;
52
+ const SOURCE_END = new RegExp(`\\.(${SOURCE_EXTENSIONS.join('|')})$`);
53
+ /**
54
+ * A binding at the line start: modifiers (`export`, `private`, `static`, `final`, `const`, …), then
55
+ * `const|let|var|val X` (JS, Swift, Kotlin) or an optional type and `X` (Java `String X`, Python
56
+ * `X`, Go `X :=`), an optional `: Type`, and `=` (not `==`). Group 1 or 2 is the name.
57
+ */
58
+ const BINDING = /^\s*(?:(?:export|public|private|protected|internal|fileprivate|open|static|final|readonly|lazy|const)\s+)*(?:(?:const|let|var|val)\s+([A-Za-z_$][\w$]*)|(?:[A-Za-z_][\w.<>[\],]*\s+)?([A-Za-z_$][\w$]*))\s*(?::[^=]*)?=(?!=)/;
59
+
60
+ /** @param {string} p @returns {string[]} the `/`-separated segments of a path. */
61
+ const segments = (p) => p.split('/');
62
+
63
+ /** @param {string} p @returns {boolean} whether `p` has a fixtures dir segment. */
64
+ const underFixtures = (p) => segments(p).some((s) => FIXTURE_DIRS.has(s));
65
+
66
+ /**
67
+ * Whether the detector checks `file`: a test file that is not a docs file and not a fixture.
68
+ * @param {unknown} file - a repo-relative POSIX path.
69
+ * @returns {boolean}
70
+ */
71
+ export function isCheckedTestFile(file) {
72
+ if (typeof file !== 'string' || file.length === 0) return false;
73
+ if (DOCS.test(file) || underFixtures(file)) return false;
74
+ if (isTestPath(file)) return true;
75
+ const parts = segments(file);
76
+ const name = parts[parts.length - 1];
77
+ return parts.slice(0, -1).includes('Tests') || /(Tests|Test)\.[^./]+$/.test(name);
78
+ }
79
+
80
+ /** @param {string} line @returns {boolean} whether the line holds a quoted source-path literal (not a fixture). */
81
+ function hasSourceLiteral(line) {
82
+ for (const m of line.matchAll(LITERAL)) {
83
+ const value = m[2];
84
+ if (SOURCE_END.test(value) && !underFixtures(value)) return true;
85
+ }
86
+ return false;
87
+ }
88
+
89
+ /** @param {string} line @returns {boolean} whether any string literal on the line has a fixtures dir segment. */
90
+ const hasFixtureLiteral = (line) => [...line.matchAll(LITERAL)].some((m) => underFixtures(m[2]));
91
+
92
+ /** @param {string} line @returns {number} `(` minus `)` outside string literals. */
93
+ const parenDelta = (line) => {
94
+ const bare = line.replace(LITERAL, '');
95
+ return (bare.match(/\(/g) ?? []).length - (bare.match(/\)/g) ?? []).length;
96
+ };
97
+
98
+ /**
99
+ * The lines of the read call on `lines[i]`: that line, plus up to {@link SPAN_LINES} more while
100
+ * the read call's own `(` stays open (the depth counts from that `(`, not the whole line).
101
+ * @param {string[]} lines @param {number} i @returns {string[]}
102
+ */
103
+ function callSpan(lines, i) {
104
+ const line = lines[i];
105
+ const span = [line];
106
+ const at = Math.min(...READ_CALLS.map((c) => line.indexOf(c)).filter((n) => n >= 0));
107
+ const open = line.indexOf('(', at);
108
+ let depth = open === -1 ? 0 : parenDelta(line.slice(open));
109
+ for (let j = i + 1; depth > 0 && j < lines.length && j <= i + SPAN_LINES; j += 1) {
110
+ span.push(lines[j]);
111
+ depth += parenDelta(lines[j]);
112
+ }
113
+ return span;
114
+ }
115
+
116
+ /** A call that passes a path straight on: `join(`, `path.join(`, `resolve(`, `path.resolve(`. */
117
+ const PATH_CALL = /^(?:path\.)?(?:join|resolve)\((.*)\)$/;
118
+
119
+ /**
120
+ * Whether a binding's right-hand side IS a source-path literal: the literal alone, or the literal
121
+ * passed straight to a path call. `spawnSync('node', ['bin/cli.mjs'])` is not.
122
+ * @param {string} rhs @returns {boolean}
123
+ */
124
+ function valueIsSourceLiteral(rhs) {
125
+ const value = rhs.trim().replace(/;$/, '').trim();
126
+ const whole = /^(['"`])([^'"`\n]*)\1$/.exec(value);
127
+ if (whole) return hasSourceLiteral(value);
128
+ const call = PATH_CALL.exec(value);
129
+ return call !== null && hasSourceLiteral(call[1]);
130
+ }
131
+
132
+ /** @param {string} line @param {ReadonlyArray<string>} calls @returns {boolean} */
133
+ const holdsAny = (line, calls) => calls.some((c) => line.includes(c));
134
+
135
+ /**
136
+ * The flagged line numbers of a test file that reads a source file and asserts on its text.
137
+ * @param {unknown} file - repo-relative path. @param {unknown} content - the file's current text.
138
+ * @returns {number[]} 1-based, sorted, no duplicates; [] when the file is not flagged.
139
+ */
140
+ export function testQualityLines(file, content) {
141
+ if (!isCheckedTestFile(file) || typeof content !== 'string') return [];
142
+ const lines = content.split(/\r?\n/);
143
+ /** @type {Set<string>} names bound to a source-path literal so far */
144
+ const bound = new Set();
145
+ /** @type {Set<string>} names bound to a literal with a fixtures segment so far */
146
+ const fixtureBound = new Set();
147
+ /** @type {number[]} */
148
+ const reads = [];
149
+ /** @type {number[]} */
150
+ const asserts = [];
151
+ /** @param {Set<string>} names @returns {(line: string) => boolean} whether a line names one of `names` */
152
+ const names = (names) => (line) => [...names].some((name) => new RegExp(`(^|[^\\w$])${name.replace(/\$/g, '\\$')}([^\\w$]|$)`).test(line));
153
+ const namesBound = names(bound);
154
+ const namesFixture = names(fixtureBound);
155
+ for (const [i, line] of lines.entries()) {
156
+ if (holdsAny(line, READ_CALLS)) {
157
+ const span = callSpan(lines, i);
158
+ const fixture = span.some((l) => hasFixtureLiteral(l) || namesFixture(l));
159
+ if (!fixture && span.some((l) => hasSourceLiteral(l) || namesBound(l))) reads.push(i + 1);
160
+ }
161
+ if (holdsAny(line, ASSERT_CALLS)) asserts.push(i + 1);
162
+ const m = BINDING.exec(line);
163
+ const name = m?.[1] ?? m?.[2];
164
+ if (name) {
165
+ if (hasFixtureLiteral(line)) fixtureBound.add(name);
166
+ else if (valueIsSourceLiteral(line.slice(m[0].length))) bound.add(name);
167
+ }
168
+ }
169
+ if (reads.length === 0 || asserts.length === 0) return [];
170
+ return [...new Set([...reads, ...asserts])].sort((a, b) => a - b);
171
+ }
172
+
173
+ /**
174
+ * The `## test-quality` section for `file`, or '' when no line is flagged. It names line numbers
175
+ * only — never the file's text.
176
+ * @param {string} file @param {ReadonlyArray<number>} lines - from {@link testQualityLines}.
177
+ * @returns {string}
178
+ */
179
+ export function renderTestQuality(file, lines) {
180
+ if (lines.length === 0) return '';
181
+ const shown = lines.slice(0, MAX_LINES);
182
+ const more = lines.length - shown.length;
183
+ return [
184
+ '## test-quality',
185
+ `file: ${file}`,
186
+ `lines: ${shown.join(', ')}${more > 0 ? ` (${more} more not listed)` : ''}`,
187
+ 'This test file reads a source file and asserts on its text (a hint from a text match, not a finding): check these lines and apply the lens rule on tests that read source text.',
188
+ ].join('\n');
189
+ }
190
+
191
+ /**
192
+ * The section for one file at its current content ('' for none, a deleted file, or any error).
193
+ * @param {string} file @param {string | null} content @returns {string}
194
+ */
195
+ export function testQualityNote(file, content) {
196
+ try {
197
+ return content === null ? '' : renderTestQuality(file, testQualityLines(file, content));
198
+ } catch {
199
+ return ''; // a hint only: no note, the review goes on
200
+ }
201
+ }
@@ -0,0 +1,39 @@
1
+ /**
2
+ * B80: the argument forms of the human verbs `block waive` and `block reopen`, in ONE place, so the
3
+ * check that runs before the terminal question (`src/cli/block.mjs`) and the verbs themselves
4
+ * (`waiveFinding`, `reopenFile`) can never drift apart. Every refusal is `usage` (exit 2).
5
+ *
6
+ * A control character (any char below 0x20, or 0x7f) is refused before anything else: such a
7
+ * value would reach the terminal question and the stderr messages that quote it.
8
+ */
9
+ import path from 'node:path';
10
+ import { StateError } from '../state/paths.mjs';
11
+
12
+ /** The finding id form a waive takes. */
13
+ const FINDING_ID = /^[A-Za-z0-9][A-Za-z0-9._:-]{0,63}$/;
14
+ // eslint-disable-next-line no-control-regex
15
+ const CONTROL = /[\x00-\x1f\x7f]/;
16
+
17
+ /**
18
+ * @param {unknown} finding
19
+ * @throws {StateError} `usage` for a value with a control character or not of the finding id form.
20
+ */
21
+ export function checkFindingId(finding) {
22
+ if (typeof finding === 'string' && CONTROL.test(finding)) throw new StateError('usage', 'a finding id must not contain control characters');
23
+ if (typeof finding !== 'string' || !FINDING_ID.test(finding)) throw new StateError('usage', 'a finding id is required (letters, digits, . _ : -)');
24
+ }
25
+
26
+ /**
27
+ * The `--file` form: no control character, then a repo-root-relative path (not absolute, no `..`
28
+ * segment, no leading `./`). `normalized` also refuses a path that `path.posix.normalize` changes
29
+ * — the rule `block reopen` has always had; `block waive` keeps its own (B80 does not change it).
30
+ * @param {unknown} file @param {{normalized: boolean}} rule
31
+ * @throws {StateError} `usage`.
32
+ */
33
+ export function checkFileForm(file, { normalized }) {
34
+ if (typeof file !== 'string' || file === '') throw new StateError('usage', '--file is required');
35
+ if (CONTROL.test(file)) throw new StateError('usage', '--file must not contain control characters');
36
+ if (path.isAbsolute(file) || file.split('/').includes('..') || file.startsWith('./') || (normalized && path.posix.normalize(file) !== file)) {
37
+ throw new StateError('usage', '--file must be repo-root-relative');
38
+ }
39
+ }
@@ -54,6 +54,7 @@ import { compileSchema } from '../config/schema-compile.mjs';
54
54
  import { buildArgv } from '../engines/builders/index.mjs';
55
55
  import { removeCodexHome } from '../engines/codex-home.mjs';
56
56
  import { cliNameForProvider } from '../engines/provider-cli.mjs';
57
+ import { DEFAULT_SANDBOX_PROFILE, SANDBOX_DIR_REFUSED, SandboxDirError } from '../engines/sandbox-dirs.mjs';
57
58
  import { parseSentinel } from '../engines/sentinel.mjs';
58
59
  import { parseUsage } from '../engines/usage-parse.mjs';
59
60
  import { budgetUsdOf, checkBudget, estimateSessionUsd, priceSession, reservedUsd, reserveBudget, SPEND_UNREADABLE_MESSAGE, withBudgetLock } from '../ledger/spend.mjs';
@@ -609,7 +610,7 @@ export async function spawnSession(opts, deps = {}) {
609
610
  if (hit) throw new SessionError('forbidden', `coder argv refused before spawn: forbidden entry ${hit.id}`);
610
611
  }
611
612
  writeSafe(stderr, `level=${level} provider=${step.provider} model=${step.model} effort=${step.effort ?? 'none'} fallback_step=${step.fallback_step}\n`);
612
- const base = { role, level, provider: step.provider, model: step.model, effort: step.effort ?? null, fallback_step: step.fallback_step, run: opts.run ?? null, block: opts.block ?? null, ...(role === 's2' ? { source: step.fallback_step > 0 ? 's2-fallback' : 's2' } : {}), ...(opts.rowExtra ?? {}) };
613
+ const base = { role, level, provider: step.provider, model: step.model, effort: step.effort ?? null, fallback_step: step.fallback_step, run: opts.run ?? null, block: opts.block ?? null, ...(role === 's2' ? { source: step.fallback_step > 0 ? 's2-fallback' : 's2' } : {}), ...sandboxRowField(opts, cli), ...(opts.rowExtra ?? {}) };
613
614
 
614
615
  if (opts.background) {
615
616
  keep = true;
@@ -829,6 +830,12 @@ function buildStep(opts, step, sessionDir, openBook) {
829
830
  if (cli === 'codex') params.outPath = path.join(sessionDir, 'out.json');
830
831
  // B32 opt-in: only a no-tools role needs it; coder and facts build exactly as without it
831
832
  if (cli === 'codex' && openBook && isNoToolRole(opts.role)) params.allowOpenBook = true;
833
+ // B93: the Codex coder's sandbox profile; no `coder.sandbox` key builds exactly today's argv
834
+ const sandbox = coderSandboxFor(opts, cli);
835
+ if (sandbox !== null) {
836
+ params.sandbox = sandbox;
837
+ params.runRoot = opts.runRoot ?? currentRunRoot();
838
+ }
832
839
  /** @type {Record<string, any> | null} */
833
840
  let compiled = null;
834
841
  const strict = variantFor(step.provider) === 'strict';
@@ -841,10 +848,40 @@ function buildStep(opts, step, sessionDir, openBook) {
841
848
  params.schema = compiled;
842
849
  }
843
850
  }
844
- const built = buildArgv(/** @type {any} */ (params));
851
+ let built;
852
+ try {
853
+ built = buildArgv(/** @type {any} */ (params));
854
+ } catch (err) {
855
+ // B93: a refused `add_dirs` entry; the message names the key path, never the value
856
+ if (err instanceof SandboxDirError) throw new SessionError(SANDBOX_DIR_REFUSED, err.message);
857
+ throw err;
858
+ }
845
859
  return { ...built, argv: [...built.argv], compiled, strict };
846
860
  }
847
861
 
862
+ /**
863
+ * B93: `{sandbox: <profile>}` for a Codex coder session row when `coder.sandbox` is set; `{}`
864
+ * otherwise, so a row without the key is byte-equal to the row before B93.
865
+ * @param {SessionOpts} opts @param {string} cli @returns {{sandbox?: string}}
866
+ */
867
+ function sandboxRowField(opts, cli) {
868
+ const sandbox = coderSandboxFor(opts, cli);
869
+ if (sandbox === null) return {};
870
+ return { sandbox: typeof sandbox.profile === 'string' ? sandbox.profile : DEFAULT_SANDBOX_PROFILE };
871
+ }
872
+
873
+ /**
874
+ * B93: the `coder.sandbox` value a Codex coder session builds with, or null (no key, another
875
+ * role, or another CLI — the Claude and Grok coders do not read it).
876
+ * @param {SessionOpts} opts @param {string} cli
877
+ * @returns {Record<string, any> | null}
878
+ */
879
+ function coderSandboxFor(opts, cli) {
880
+ if (cli !== 'codex' || opts.role !== 'coder') return null;
881
+ const sandbox = /** @type {Record<string, any> | undefined} */ (opts.cfg)?.coder?.sandbox;
882
+ return sandbox === undefined || sandbox === null ? null : sandbox;
883
+ }
884
+
848
885
  /**
849
886
  * @param {{cli: string, argv: string[], cwd: string, stdinFile?: string, outPath?: string, env?: Record<string, string>, compiled: Record<string, any> | null, strict: boolean}} built
850
887
  * @param {SessionOpts} opts
@@ -54,11 +54,13 @@ function safeId(id) {
54
54
  * @param {{
55
55
  * block: string, runId: string, status: 'closed' | 'stopped', event: string, reason?: string | null,
56
56
  * commit: string | null, spend: {usd: number, unknown: number} | null, findings: number | null, at: Date,
57
- * }} facts
57
+ * reviewed?: ReadonlyArray<{file: string, ticket: string, via: string}>,
58
+ * }} facts `reviewed` (B96): the close's reviewed lines, sorted by file; the key is added last, and only when given.
58
59
  * @returns {{block: string, run_id: string, status: string, event: string, reason: string | null, commit: string | null,
59
- * spent_usd: number | null, unknown_usd_sessions: number | null, findings: number | null, at: string}}
60
+ * spent_usd: number | null, unknown_usd_sessions: number | null, findings: number | null, at: string,
61
+ * reviewed?: Array<{file: string, ticket: string, via: string}>}}
60
62
  */
61
- export function doneFileContent({ block, runId, status, event, reason = null, commit, spend, findings, at }) {
63
+ export function doneFileContent({ block, runId, status, event, reason = null, commit, spend, findings, at, reviewed }) {
62
64
  return {
63
65
  block,
64
66
  run_id: runId,
@@ -71,6 +73,7 @@ export function doneFileContent({ block, runId, status, event, reason = null, co
71
73
  unknown_usd_sessions: spend && Number.isInteger(spend.unknown) ? spend.unknown : null,
72
74
  findings: Number.isInteger(findings) ? findings : null,
73
75
  at: at.toISOString(),
76
+ ...(Array.isArray(reviewed) ? { reviewed: reviewed.map((r) => ({ file: r.file, ticket: r.ticket, via: r.via })).sort((x, y) => (x.file < y.file ? -1 : x.file > y.file ? 1 : 0)) } : {}),
74
77
  };
75
78
  }
76
79
 
@@ -97,12 +100,13 @@ async function headCommit(cwd) {
97
100
  * @param {{
98
101
  * record: Record<string, any>, id: string, status: 'closed' | 'stopped', event: string, reason?: string | null,
99
102
  * openFindings: (rows: Record<string, any>[]) => Promise<number>, err: (s: string) => unknown, now?: Date,
103
+ * reviewed?: ReadonlyArray<{file: string, ticket: string, via: string}>,
100
104
  * }} opts - `record` is the run record read before the end (its `workspace` and the block's
101
105
  * `worktree` do not change at an end); `openFindings` counts the block's open findings exactly
102
- * as `code-forge status` does.
106
+ * as `code-forge status` does. `reviewed` (B96): the close's reviewed lines, kept as the last key.
103
107
  * @returns {Promise<string | null>} the done file's absolute path, or null when it was not written.
104
108
  */
105
- export async function writeDoneFile({ record, id, status, event, reason = null, openFindings, err, now = new Date() }) {
109
+ export async function writeDoneFile({ record, id, status, event, reason = null, openFindings, err, now = new Date(), reviewed }) {
106
110
  const runId = String(record?.run_id);
107
111
  try {
108
112
  const file = doneFilePath(runId, id);
@@ -127,7 +131,7 @@ export async function writeDoneFile({ record, id, status, event, reason = null,
127
131
  }
128
132
  const tree = record.blocks?.[id]?.worktree?.path;
129
133
  const commit = await headCommit(typeof tree === 'string' && tree.length > 0 ? tree : record.workspace);
130
- const content = doneFileContent({ block: id, runId, status, event, reason, commit, spend, findings, at: now });
134
+ const content = doneFileContent({ block: id, runId, status, event, reason, commit, spend, findings, at: now, ...(reviewed === undefined ? {} : { reviewed }) });
131
135
  const dir = path.dirname(file);
132
136
  await mkdir(dir, { recursive: true, mode: 0o700 });
133
137
  const tmp = path.join(dir, `.${path.basename(file)}.${process.pid}.${Date.now()}.tmp`);
@@ -18,7 +18,7 @@ import path from 'node:path';
18
18
  import { BLOCK_KINDS, blockKind, dispatchLevel } from '../decide/escalation.mjs';
19
19
  import { exec } from '../util/exec.mjs';
20
20
  import { DEFAULT_CONFIG_FILENAME } from '../config/load.mjs';
21
- import { blockEndHook, externalReviewer } from './config-snapshot.mjs';
21
+ import { blockEndHook, configInForce, externalReviewer } from './config-snapshot.mjs';
22
22
  import { ownedKinds, pathKind } from './owned-kinds.mjs';
23
23
  import { isAncestorArgv, revParseArgv } from '../util/git.mjs';
24
24
  import { StateError } from './paths.mjs';
@@ -263,12 +263,22 @@ function activeBlocks(record) {
263
263
  * subdirectory, relative to the repository top level (the close gate matches both, `ownsAt`).
264
264
  * Empty when the key is not set. B64b: `hooks.on_block_end` in the snapshot guards it the same way
265
265
  * — the file names a command the orchestrator runs at every block end — and `why` names the key
266
- * that guards it (`review.external.required` first, so its message is unchanged).
266
+ * that guards it (`review.external.required` first, so its message is unchanged). B87:
267
+ * `proof.runners` in the snapshot guards it too — the file names the commands that prove a block;
268
+ * any stored value guards it (fail closed: `validate` refuses an empty or malformed list anyway).
267
269
  * @param {Record<string, any>} record @returns {Promise<{paths: string[], why: string}>}
268
270
  * @throws {StateError} `config-snapshot`, `git-failed`
269
271
  */
270
272
  async function guardedConfigPaths(record) {
271
- const why = externalReviewer(record) !== null ? 'review.external.required is true' : blockEndHook(record) !== null ? 'hooks.on_block_end is set' : null;
273
+ const runners = configInForce(record)?.proof?.runners;
274
+ const why =
275
+ externalReviewer(record) !== null
276
+ ? 'review.external.required is true'
277
+ : blockEndHook(record) !== null
278
+ ? 'hooks.on_block_end is set'
279
+ : runners !== undefined && runners !== null
280
+ ? 'proof.runners is set'
281
+ : null;
272
282
  if (why === null) return { paths: [], why: '' };
273
283
  const top = await repoTop(record.workspace);
274
284
  // the repo-top spelling `<prefix>/.code-forge.yml` may also match a different file of the same
@@ -44,6 +44,11 @@ export const IMMUTABLE_KEYS = Object.freeze({
44
44
  // B64b: the block-end hook runs a command on the owner's machine at every block end, so only
45
45
  // the owner sets it, before the run starts
46
46
  'hooks.on_block_end': 'it runs a command on the owner\'s machine at every block end; fixed by the owner before the run starts',
47
+ // B87: the runner argv proves the coder's own work, so only the owner sets it, before the run starts
48
+ 'proof.runners': 'the commands that prove a block are fixed by the owner before the run starts',
49
+ // B93: the coder sandbox profile widens what the coder may write, so only the owner sets it,
50
+ // before the run starts (a coder can never own `.code-forge.yml`)
51
+ 'coder.sandbox': 'the coder sandbox widens what the coder may write; fixed by the owner before the run starts',
47
52
  });
48
53
 
49
54
  /** @param {unknown} v @returns {boolean} what `JSON.stringify` drops from an object. */
@@ -235,3 +240,37 @@ export function blockEndHook(record) {
235
240
  if (!Number.isInteger(timeoutS) || timeoutS < HOOK_TIMEOUT_MIN_S || timeoutS > HOOK_TIMEOUT_MAX_S) return null;
236
241
  return { argv: [...argv], timeoutS };
237
242
  }
243
+
244
+ /**
245
+ * B87: `proof.runners` of the run-start config snapshot (never the live config), as a deep copy
246
+ * (entries and argv arrays), or null when the key is absent or the record is from before
247
+ * snapshots. `proof/runners.mjs` checks each entry again before it runs one (fail closed).
248
+ * @param {Record<string, any>} record
249
+ * @returns {unknown[] | null}
250
+ * @throws {StateError} `config-snapshot` when the snapshot in force does not match its hash;
251
+ * `runner-invalid` when the stored value is not a list, or is an empty one.
252
+ */
253
+ export function proofRunners(record) {
254
+ const runners = configInForce(record)?.proof?.runners;
255
+ if (runners === undefined || runners === null) return null;
256
+ if (!Array.isArray(runners)) throw new StateError('runner-invalid', `run ${record.run_id}: proof.runners in the run snapshot is not a list`);
257
+ // `validate` refuses an empty list (minItems 1); a record that holds one anyway is refused too
258
+ if (runners.length === 0) throw new StateError('runner-invalid', `run ${record.run_id}: proof.runners in the run snapshot is an empty list`);
259
+ // a deep copy: a caller that changes an entry or an argv array never changes the snapshot
260
+ return structuredClone(runners);
261
+ }
262
+
263
+ /**
264
+ * B93: `coder.sandbox` of the run-start config snapshot (never the live config). A record from
265
+ * before snapshots has no run-start value: `{snapshot: false}`. Otherwise `{snapshot: true,
266
+ * sandbox}`, where `sandbox` is a deep copy of the value, or null when the snapshot has no key.
267
+ * @param {Record<string, any>} record
268
+ * @returns {{snapshot: false} | {snapshot: true, sandbox: Record<string, any> | null}}
269
+ * @throws {StateError} `config-snapshot` when the snapshot in force does not match its hash.
270
+ */
271
+ export function coderSandbox(record) {
272
+ const config = configInForce(record);
273
+ if (config === null) return { snapshot: false };
274
+ const sandbox = config.coder?.sandbox;
275
+ return { snapshot: true, sandbox: sandbox === undefined || sandbox === null ? null : structuredClone(sandbox) };
276
+ }
package/src/state/run.mjs CHANGED
@@ -269,6 +269,10 @@ export async function reloadRun({ runId, config, readPending = () => [], effecti
269
269
  if (config.context?.require_coder_tests === true) frozen.push('context.require_coder_tests');
270
270
  // B64b: and never had a block-end hook
271
271
  if (config.hooks?.on_block_end !== undefined) frozen.push('hooks.on_block_end');
272
+ // B87: and never had a proof runner
273
+ if (config.proof?.runners !== undefined) frozen.push('proof.runners');
274
+ // B93: and never had a coder sandbox profile
275
+ if (config.coder?.sandbox !== undefined) frozen.push('coder.sandbox');
272
276
  }
273
277
  if (frozen.length > 0) {
274
278
  const why = frozen.map((key) => `${key} cannot change mid-run (${IMMUTABLE_KEYS[/** @type {keyof typeof IMMUTABLE_KEYS} */ (key)]})`);
@@ -292,7 +292,7 @@ export const CODER_ONLY_FORBIDDEN = Object.freeze(
292
292
  ['code-forge', 'block', 'waive'],
293
293
  ['forge', 'block', 'waive'],
294
294
  ],
295
- description: 'Waive a review finding (forge block waive) — human-only, run by the orchestrator after the human said so (§4.9)',
295
+ description: 'Waive a review finding (forge block waive) — human-only, typed by the human in their own terminal (§4.9, B80)',
296
296
  },
297
297
  {
298
298
  id: 'code-forge-block-reopen-from-coder',
@@ -301,7 +301,7 @@ export const CODER_ONLY_FORBIDDEN = Object.freeze(
301
301
  ['code-forge', 'block', 'reopen'],
302
302
  ['forge', 'block', 'reopen'],
303
303
  ],
304
- description: 'Reopen the review of a stopped or lost file (forge block reopen) — human-only, run by the orchestrator after the human said so (B72)',
304
+ description: 'Reopen the review of a stopped or lost file (forge block reopen) — human-only, typed by the human in their own terminal (B72, B80)',
305
305
  },
306
306
  {
307
307
  id: 'code-forge-autopilot-from-coder',