Files
gstack/lib/autoplan-run.ts
T
Garry Tanandcapy-ai[bot] c71d764ba4 v1.91.115.0 feat: field list 2 wave A: autoplan round cap and public export, eval-plan diagnostics and model rules, --for review-voice (#3234)
* fix(capy): the install trailer says behind=<n> and its fix, so a stale install shows in every relayed trailer (field list 2 item 13, #2378)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(capy): a lease records its holder's start identity, so a pid reused after a wake is reclaimed; a skipped session-start upgrade ends fail behind, never ok (GBRA-105, GBRA-106)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(lease): stale needs proof (unreadable or malformed identities keep the lease), ps identity is timezone-independent, and reclaim never deletes a concurrent renewal

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* feat(autoplan): enforced round cap and spend budget, deferred close, public export without machine paths, opt-in durable checkpoints (field list 2 W5-W7)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(lease): one directory lock for publish, prune and reclaim; reclaim rereads the record judged stale, so a renewal is never deleted or hidden

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(lease): the directory lock names its holder; broken only when the holder is dead or its pid reused, unlock removes only its own lock, touch exits 3 under a live holder

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(lease): breakers serialize on .lock.break and reread the lock before removing it; gstack-skill-start reports a lease the held lock kept out

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* docs(lease): name the double-crash limit of breaking a dead lock

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* feat(capy): --for review-voice install profile, no-Drive bootstrap command, reviewer prompt trailer (field list 2 W11)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
(cherry picked from commit 89bf556c11)

* test: pin review-voice row, helpers row and the no-Drive fallback in existing install tests

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
(cherry picked from commit 7f6c70459a)

* test: the DRIVER template names the reviewer trailer relay (pregate lane)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
(cherry picked from commit 2545d9a6d2)

* test(context-budget): capy skill eager ceiling for the review-voice subagent paragraph (W11; ratchet protocol, capture value)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* feat(eval-plan): prereg diagnostics dry run (W8) and .gstack/eval-models.json lint (W9), WIP

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
(cherry picked from commit b509096218)

* feat(eval-plan): W8/W9 tests, docs/eval-models.md, troubleshooting anchors, /eval-plan steps

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
(cherry picked from commit 9335e45de2)

* test(eval-plan): eval-plan/SKILL.md names only defined codes and accepted flags

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
(cherry picked from commit a7a974d268)

* test(review-voice): the launch line points at the verdict's bun-only directory (#3233's launch contract)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* test(third-party-actions): the Aside drive-offer detector matches the product name, not the English idiom (paid run 38988927989 tpa-absent-linux false positive)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(autoplan): export refuses artifact paths outside the run before writing; a deferred close keeps earlier undecided findings; reopening CEO grants Eng its rerun (review of #3234)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* test(autoplan-budget): bind the reservation expectation to lib/autoplan-budget's default estimate (test-of-test ratchet)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(evals): the docs observer accepts the spawned-prefixed gstack-skill-start --help read

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(windows): curate capy-review-voice out of the Windows subset (POSIX launcher contract)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* docs: changelog entry for field list 2 wave A

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* chore: stamp v1.91.115.0

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

---------

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
2026-10-17 02:40:45 +00:00

486 lines
36 KiB
TypeScript
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
/**
* autoplan-run — the unattended /autoplan loop as an explicit, durable state
* machine (plan B2). The run directory the parent chose (`--out`) holds
* `run.json` (the B0 manifest, extended with the runner's state), an
* append-only `attempts.jsonl` journal (dispatch intent before anything
* else, so a crash is reconcilable), `run.lock` (one writer at a time),
* `spend.json` (lib/spend-ledger.ts) and every prompt, result and consensus
* file. `next` prepares the current phase (snapshot, prompts), journals one
* attempt per pending voice and stops with `awaiting_result`; `submit` binds
* a reviewer's result (receipt, canonical findings, model family) exactly
* once, settles its reservation and closes the phase through reconciliation
* (lib/autoplan-gate.ts) when both voices are terminal; `resume` takes
* ownership and marks attempts with intent but no terminal record
* EXECUTION_UNKNOWN, never redispatching them on its own. Thin bin:
* bin/gstack-autoplan.
*/
import * as crypto from 'node:crypto';
import * as fs from 'node:fs';
import * as os from 'node:os';
import * as path from 'node:path';
import { canonicalRows, parseReceipt, parseRecordFindings, parseReviewResult, type CanonicalFinding } from './autoplan-reconcile';
import { amendmentSteps, applyGuidance, closeDeferred, closePhase, finishGate } from './autoplan-gate';
import { GSTACK_VERSION } from './autoplan-migrate';
import { driftLine, normalizeFindingFile, reviewHash, scaffold } from './autoplan-plan-text';
import { PHASES, SNAPSHOT_TOOL, attemptFile, bindPromptResult, detectScope, driverFile, extractHeading, phaseOrder, reviewSkillFile, scopeLines, scopeSourceLabel, sectionLoadFor, snapshotTool, writePrompts, type Phase, type Scope, type ScopeFlags } from './autoplan-prompts';
import { recoverRevise } from './autoplan-revise';
import { FILES, RunError, journal, now, pendingAmendments, readDecisions, readJournal, readState, voiceTerminal, writeState, type AttemptState, type PhaseState, type RunState, type Voice, type VoiceState } from './autoplan-state';
import { checkPlanReality, checkRealityRows, PLAN_REALITY_CHECKER } from './plan-reality';
import type { RunStatus } from './headless-artifacts';
import { sha256File } from './headless-artifacts';
import { outsideConsent } from './outside-consent';
import { familyConflict, modelFamily } from './outside-voice-runner';
import { renderSpend, reserve, settle, summarize, readLedger } from './spend-ledger';
import { DEFAULT_MAX_ROUNDS, attemptEstimate, overCap, resumeStop, roundCap, roundOf, roundsLine, stopForBudget, stopLines } from './autoplan-budget';
import { annotateConfirm, confirmBlock } from './autoplan-confirm';
/** An unreadable run.lock younger than this is a lock some other gstack is still writing, never a stale one. */
const LOCK_WRITE_GRACE_MS = 30_000;
/** Exclusive ownership for one command: a lock whose pid is alive is never stolen; a dead pid's lock is reclaimed. */
export function withRunLock<T>(out: string, fn: () => T, opts: { steal?: boolean } = {}): T {
fs.mkdirSync(out, { recursive: true });
const lock = path.join(out, FILES.lock);
const mine = JSON.stringify({ pid: process.pid, host: os.hostname(), at: now() });
const staged = `${lock}.${process.pid}.${crypto.randomBytes(4).toString('hex')}`;
fs.writeFileSync(staged, mine);
for (let i = 0; ; i++) {
try {
fs.linkSync(staged, lock);
fs.unlinkSync(staged);
break;
} catch (e: any) {
if (e.code !== 'EEXIST') { fs.rmSync(staged, { force: true }); throw e; }
let owner: { pid?: number; host?: string; at?: string } = {};
let unreadable = false;
try { owner = JSON.parse(fs.readFileSync(lock, 'utf8')); } catch { unreadable = true; }
let alive = false;
if (unreadable) {
try { alive = Date.now() - fs.statSync(lock).mtimeMs < LOCK_WRITE_GRACE_MS; } catch { /* gone: retry the link */ }
} else if (owner.host === os.hostname() && typeof owner.pid === 'number') {
try { process.kill(owner.pid, 0); alive = true; } catch (k: any) { alive = k.code === 'EPERM'; }
} else if (owner.host && owner.host !== os.hostname()) alive = !opts.steal;
if (alive) { fs.rmSync(staged, { force: true }); throw new RunError('RUN_LOCKED', unreadable ? `${lock} is being written by another command` : `pid ${owner.pid} on ${owner.host} since ${owner.at}`, 3); }
try { fs.unlinkSync(lock); } catch { /* raced with its owner's exit */ }
if (i > 50) { fs.rmSync(staged, { force: true }); throw new RunError('RUN_LOCKED', `could not reclaim ${lock}`, 3); }
}
}
try { return fn(); } finally { try { if (fs.readFileSync(lock, 'utf8') === mine) fs.unlinkSync(lock); } catch { /* already gone */ } }
}
// ---------------------------------------------------------------------------
// init
// ---------------------------------------------------------------------------
export interface InitOptions { plan: string; out: string; cwd: string; scope?: ScopeFlags; planHeading?: string; outsideEnabled?: boolean; spendCap?: number | null; sessionKind?: string; deadlineMin?: number; light?: boolean; runId?: string; scaffold?: boolean; maxRounds?: number }
/** `lines` receives what init itself reports (SCAFFOLDED:). A plan with one section heading missing refuses, or with `scaffold` is reviewed from a scaffolded copy. */
export function initRun(o: InitOptions, lines: string[] = []): RunState {
const out = path.resolve(o.out);
const input = path.resolve(o.plan);
let source = input;
if (!fs.existsSync(source)) throw new Error(`plan not found: ${source}`);
let scaffolded = false;
if (o.planHeading) {
const section = extractHeading(fs.readFileSync(source, 'utf8'), o.planHeading);
if (section === undefined) throw new Error(`heading ${JSON.stringify(o.planHeading)} not found in ${source}`);
fs.mkdirSync(out, { recursive: true });
source = path.join(out, FILES.source);
fs.writeFileSync(source, section);
} else {
const diag = scaffold(fs.readFileSync(source, 'utf8'));
if (diag.problem && (!o.scaffold || !diag.text)) throw new RunError('PLAN_SECTIONS_INVALID', `${source}: ${diag.problem}`, 3);
if (diag.text) {
fs.mkdirSync(out, { recursive: true });
source = path.join(out, FILES.source);
fs.writeFileSync(source, diag.text);
scaffolded = true;
lines.push(`SCAFFOLDED: added "${diag.added}" (copy at ${source}; ${input} unchanged)`);
}
}
fs.mkdirSync(out, { recursive: true });
const active = path.join(out, FILES.active);
const restore = path.join(out, FILES.restore);
const init = snapshotTool(['init', source, active, restore], o.cwd);
const scope = detectScope(active, o.cwd, o.scope ?? {}, path.resolve(o.plan));
const order = phaseOrder(scope);
const run = o.runId ?? `${path.basename(out)}-${new Date().toISOString().replace(/[-:T.Z]/g, '').slice(0, 14)}`;
// codex_reviews gates the outside voice here, before any dispatch: a refusal is a terminal not_run outcome per phase.
const consent = (o.outsideEnabled ?? true) ? outsideConsent({ userInvoked: false }) : undefined;
const consentMissing = consent && !consent.allowed ? consent.lines[0] : undefined;
const sessionKind = o.sessionKind ?? process.env.GSTACK_SESSION_KIND ?? 'unattended';
const phases: Record<string, PhaseState> = {};
for (const p of PHASES) if (!order.includes(p)) phases[p] = { skipped: p === 'design' ? 'no UI scope detected' : 'no developer-facing scope detected', native: { status: 'skipped' }, outside: { status: 'skipped' } };
const state: RunState = {
schema_version: 1, run, status: 'running', required_phases: order, phases, artifacts: {}, revision: 0, out,
plan_source: init.sourcePlan, active_plan: init.activePlan, restore: init.restorePath, scope, phase_order: order,
outside_enabled: (o.outsideEnabled ?? true) && !consentMissing, attempt_counter: 0, spend_cap_usd: o.spendCap ?? null,
...(consentMissing ? { outside_not_run: { reason: 'consent_missing', detail: consentMissing } } : {}),
session_kind: sessionKind, host: 'gstack-autoplan', max_rounds: o.maxRounds ?? (sessionKind === 'unattended' ? DEFAULT_MAX_ROUNDS : null), history: [], created_at: now(), updated_at: now(),
deviations: ['autoplan guard: not enforced by this host; publication order is unverified (GUARD_NOT_INSTALLED)'],
...(o.deadlineMin ? { deadline_at: new Date(Date.now() + o.deadlineMin * 60_000).toISOString() } : {}),
...(o.light ? { light: true } : {}),
gstack_version: GSTACK_VERSION, plan_input: input, scope_flags: { ...(o.scope ?? {}) },
...(o.planHeading ? { plan_heading: o.planHeading } : {}), ...(scaffolded ? { scaffolded: true } : {}),
};
fs.writeFileSync(path.join(out, 'timing.json'), JSON.stringify({ schema_version: 1, run, phases: [], started_at: now(), session_kind: state.session_kind }, null, 2) + '\n');
journal(out, { event: 'init', run, plan_source: source, scope, phase_order: order, ...(consentMissing ? { outside: 'not_run', reason: 'consent_missing' } : {}) });
writeState(state);
return state;
}
// ---------------------------------------------------------------------------
// next
// ---------------------------------------------------------------------------
export function currentPhase(state: RunState): Phase | undefined {
return state.phase_order.find(p => !state.phases[p]?.closed_at);
}
export function voiceState(state: RunState, phase: Phase, voice: Voice): VoiceState {
const ps = (state.phases[phase] ??= {});
return (ps[voice] ??= { status: 'pending' });
}
const openAttempt = (v: VoiceState) => v.attempts?.find(a => a.state === 'dispatched' || a.state === 'execution_unknown');
export interface NextOptions { out: string; cwd: string; estimateUsd?: number; redispatch?: string[] }
export interface NextResult { state: RunState; phase?: Phase; attempts: Array<{ attempt: string; voice: Voice; prompt: string; result: string; model_family: string; new: boolean }>; lines: string[] }
function openPhase(state: RunState, phase: Phase, cwd: string): void {
const ps = (state.phases[phase] ??= {});
if (ps.snapshot) return;
const methodology = snapshotTool(['methodology', phase, reviewSkillFile(phase), state.restore], cwd);
const snap = snapshotTool(['create', phase, state.active_plan, state.restore, methodology.methodologyPath], cwd);
// Every later phase reviews with its closed predecessors' consensus as data
// (not only Eng): a DX or design reviewer who never saw the CEO consensus
// re-raises what the CEO phase already reconciled (field gate rev 2).
const prior: Array<{ phase: string; text: string }> = [];
for (const p of state.phase_order) {
if (p === phase) break;
if (!state.phases[p]?.closed_at) continue;
const file = path.join(state.out, `${p}-consensus.md`);
if (!fs.existsSync(file)) throw new RunError('CONSENSUS_MISSING', `${p} closed without ${p}-consensus.md`, 3);
prior.push({ phase: p, text: fs.readFileSync(file, 'utf8') });
}
const sections = sectionLoadFor(phase, state.scope, state.session_kind === 'unattended' || state.light ? 'light' : 'full');
const decided = (state.gate_rev ?? 0) >= 1 ? readDecisions(state.out).filter(r => r.chosen != null && r.status !== 'pending')
.map(r => `r${r.gate_rev} ${(r.label ?? r.id).toLowerCase()}${r.chosen} ${r.title}${r.chosen_text ? ` -> ${r.chosen_text}` : ''}`) : [];
const confirm = confirmBlock(state, phase, snap.snapshotPath, roundOf(state, phase));
const prompts = writePrompts({ phase, snapshot: { nativePrompt: snap.nativePrompt, snapshotPath: snap.snapshotPath, sha256: snap.sha256 }, outDir: state.out, priorConsensus: prior, sections, decided, confirm: confirm?.text });
if (confirm) ps.confirm = confirm.confirm;
ps.snapshot = snap.sha256; ps.snapshot_dir = path.dirname(snap.snapshotPath); ps.snapshot_path = snap.snapshotPath; ps.methodology = methodology.methodologyPath; ps.opened_at = now();
ps.sections = { loaded: sections.loaded, skipped: sections.skipped, checklist: !!sections.checklist };
ps.native = { status: 'pending', attempts: [] };
const notRun = state.outside_not_run as { reason: string; detail: string } | undefined;
ps.outside = state.outside_enabled ? { status: 'pending', attempts: [] }
: notRun ? { status: 'disabled', outcome: 'not_run', reason: notRun.reason, detail: notRun.detail } : { status: 'disabled' };
ps.prompts = prompts;
journal(state.out, { event: 'phase_opened', phase, snapshot: snap.sha256, prompts: Object.fromEntries(Object.entries(prompts).map(([k, p]) => [k, path.relative(state.out, p)])) });
}
export function nextStep(o: NextOptions): NextResult {
const state = readState(path.resolve(o.out));
const lines: string[] = [];
const result: NextResult = { state, attempts: [], lines };
if (state.status === 'incomplete' && state.stop) { lines.push(...stopLines(state)); return result; }
if (state.status === 'gate_pending' || state.status === 'complete' || state.status === 'incomplete' || state.status === 'refused') return result;
if (state.deadline_at && Date.parse(state.deadline_at) < Date.now()) { state.status = 'interrupted'; writeState(state); lines.push(`DEADLINE: passed at ${state.deadline_at}`); return result; }
const pending = pendingAmendments(state).map(a => a.decision);
if (pending.length) {
const each = pendingAmendments(state).map(a => `APPLY ${a.item ?? a.decision}: ${applyGuidance(a)}`).join('; ');
throw new RunError('PLAN_AMEND_PENDING', `accepted at the gate but not in the plan: ${pending.join(', ')} (${each}); ${amendmentSteps(state, state.out)} (or --withdraw <id>)`, 3);
}
let phase = currentPhase(state);
while (phase && !state.phases[phase]?.snapshot && overCap(state, phase)) {
lines.push(...closeDeferred(state, phase, { reason: `round cap: round ${roundOf(state, phase)} would pass max ${roundCap(state, phase)}`, by: 'runner', cause: 'cap' }).lines);
phase = currentPhase(state);
}
if (!phase) { finishGate(state); writeState(state); return result; }
const first = !state.phase_order.some(p => state.phases[p]?.snapshot);
openPhase(state, phase, o.cwd);
const ps = state.phases[phase]!;
result.phase = phase;
lines.push(`PHASE: ${phase} snapshot=${ps.snapshot} order=${state.phase_order.join(',')}`);
const drift = driftLine(state, 'PLAN_CHANGED');
if (drift) lines.push(drift);
if (first) lines.push(...scopeLines(state.scope), `DRIVER: ${driverFile()}`);
if (ps.outside?.reason === 'consent_missing') lines.push(`OUTSIDE: ${phase} not_run reason=consent_missing`, String(ps.outside.detail));
if (ps.sections) lines.push(`Skipped sections: ${ps.sections.skipped.length ? ps.sections.skipped.join(',') : 'none'} (scope: ui=${state.scope.ui ? 'yes' : 'no'},dx=${state.scope.dx ? 'yes' : 'no'},source=${scopeSourceLabel(state.scope)}; checklist=${ps.sections.checklist ? 'loaded' : 'none'})`);
const prompts = ps.prompts!;
const estimate = attemptEstimate(state, o.estimateUsd, process.env);
if (estimate.line) lines.push(estimate.line);
const fresh = (['native', 'outside'] as const).filter(voice => { const v = voiceState(state, phase!, voice); const open = openAttempt(v); return !voiceTerminal(v) && (!open || (open.state === 'execution_unknown' && o.redispatch?.includes(open.attempt))); });
const before = summarize(readLedger(path.join(state.out, FILES.spend), state.spend_cap_usd));
if (state.spend_cap_usd !== null && fresh.length && before.worst_case_usd + fresh.length * estimate.usd > state.spend_cap_usd + 1e-9) {
const stop = stopForBudget(state, phase, fresh.length * estimate.usd);
writeState(state);
throw new RunError('SPEND_CAP_EXCEEDED', `spent ${before.spent_usd} + reserved ${before.reserved_usd} + unknown ${before.unknown_usd} + ${fresh.length} x estimate ${estimate.usd} > cap ${state.spend_cap_usd}; the run stopped incomplete (budget_exhausted)`, 3, { status: 'incomplete', lines: [...lines, ...stop] });
}
for (const voice of ['native', 'outside'] as const) {
const v = voiceState(state, phase, voice);
if (voiceTerminal(v)) continue;
const open = openAttempt(v);
if (open?.state === 'execution_unknown') {
if (!o.redispatch?.includes(open.attempt)) throw new RunError('EXECUTION_UNKNOWN', `${open.attempt} (${phase} ${voice})`, 3);
open.state = 'superseded'; open.terminal_at = now();
settle(path.join(state.out, FILES.spend), open.attempt, 'unknown');
journal(state.out, { event: 'superseded', attempt: open.attempt, phase, voice });
} else if (open) {
result.attempts.push({ attempt: open.attempt, voice, prompt: open.prompt, result: open.result, model_family: familyHint(state, voice), new: false });
continue;
}
const attempt = `${state.run}-a${++state.attempt_counter}`;
try { reserve(path.join(state.out, FILES.spend), attempt, estimate.usd, { cap: state.spend_cap_usd, label: `${phase}/${voice}` }); }
catch (e: any) { if (e.code === 'SPEND_CAP_EXCEEDED') { state.attempt_counter--; writeState(state); throw new RunError('SPEND_CAP_EXCEEDED', e.message, 3); } throw e; }
const row: AttemptState = { attempt, state: 'dispatched', prompt: prompts[voice], result: path.join(state.out, attemptFile(phase, voice, attempt.slice(state.run.length + 1), 'md')), estimate_usd: estimate.usd, dispatched_at: now() };
bindPromptResult(row.prompt, row.result);
journal(state.out, { event: 'dispatched', attempt, phase, voice, snapshot: ps.snapshot, prompt: path.relative(state.out, row.prompt), prompt_sha256: sha256File(row.prompt), result: path.relative(state.out, row.result), estimate_usd: estimate.usd });
(v.attempts ??= []).push(row);
v.attempt = attempt; v.status = 'running';
result.attempts.push({ attempt, voice, prompt: row.prompt, result: row.result, model_family: familyHint(state, voice), new: true });
}
state.status = 'awaiting_result';
writeState(state);
for (const a of result.attempts) lines.push(`ATTEMPT: ${a.attempt} phase=${phase} voice=${a.voice} prompt=${a.prompt} result=${a.result} model_family=${a.model_family}${a.new ? '' : ' (open)'}`);
lines.push(renderSpend(summarize(readLedger(path.join(state.out, FILES.spend)))), roundsLine(state));
return result;
}
function familyHint(state: RunState, voice: Voice): string {
if (voice === 'native') return state.native_model ? `same as ${state.native_model}` : 'the parent’s own model (recorded at submit)';
return state.native_model ? `not ${modelFamily(state.native_model)}` : 'differs from the native reviewer';
}
// ---------------------------------------------------------------------------
// submit
// ---------------------------------------------------------------------------
export interface SubmitOptions {
out: string; phase: string; voice: Voice; model?: string; cwd: string; runner?: string; usageUsd?: number; reconcile?: string;
/** Exactly one of: a canonical result, a hand-run record, or a terminal outside outcome. */
result?: string; record?: string; unavailable?: string; failed?: string;
/** Optional when exactly one attempt is open for the phase and voice; required for an attested bind. */
attempt?: string; attest?: string;
}
export interface SubmitResult { state: RunState; closed: boolean; findings: number; lines: string[] }
/** Report-only PLAN_REALITY on the active plan, once per Implementation-plan hash and checker (revise clears it). */
function recordPlanReality(state: RunState): string[] {
let text: string;
try { text = fs.readFileSync(state.active_plan, 'utf8'); } catch { return []; }
const sha = reviewHash(text);
if (state.plan_reality?.plan_sha256 === sha && state.plan_reality.checker === PLAN_REALITY_CHECKER) return [];
const rows = checkRealityRows(text, { phase: 'plan', sessionKind: state.session_kind });
const plan = checkPlanReality(text, { sessionKind: state.session_kind });
const verdict = rows.verdict !== 'complete' ? rows.verdict : plan.verdict;
state.plan_reality = { plan_sha256: sha, checker: PLAN_REALITY_CHECKER, verdict, line: `${rows.line}; ${plan.line}` };
journal(state.out, { event: 'plan_reality', plan_sha256: sha, checker: PLAN_REALITY_CHECKER, verdict });
return [`${rows.line} (report only)`, `${plan.line} (report only)`, ...plan.numberLines];
}
function resolveAttempt(v: VoiceState, attempt: string | undefined, phase: string, voice: Voice): AttemptState {
const open = openAttempt(v);
if (!open || (attempt !== undefined && open.attempt !== attempt)) {
throw new RunError('ATTEMPT_MISMATCH', `${attempt ?? '(none given)'} is not the open attempt for ${phase} ${voice}${open ? ` (${open.attempt} is)` : ''}`, 3);
}
return open;
}
/** The record's receipt and findings: a matching INPUT line binds verified; none binds only on --attest with --attempt, never across a later revise. */
function readRecord(state: RunState, phase: Phase, open: AttemptState, o: SubmitOptions, text: string): { findings: CanonicalFinding[]; source: VoiceState['findings_source']; receipt: 'verified' | 'attested' } {
const receipt = parseReceipt(text, { phase, sha256: state.phases[phase]!.snapshot! });
if (receipt.present && !receipt.ok) throw new RunError('RESULT_RECEIPT_MISSING', receipt.error!);
if (!receipt.present) {
if (!o.attest) throw new RunError('RESULT_RECEIPT_MISSING', `${o.record} has no INPUT: ${phase} <sha256> line; bind it on your word with --attest "<reason>" --attempt ${open.attempt}, or have the reviewer start with the INPUT line its prompt names`);
if (!o.attempt) throw new RunError('ATTEMPT_MISMATCH', `an attested bind names its attempt: pass --attempt ${open.attempt}`, 3);
const revised = readJournal(state.out).find(r => r.event === 'plan_revised' && String(r.ts) > open.dispatched_at);
if (revised) throw new RunError('RESULT_RECEIPT_MISSING', `the plan was revised after ${open.attempt} was dispatched (${revised.ts}); an attestation cannot say which plan was read: redispatch with gstack-autoplan next --redispatch, or bind a record with the INPUT line`);
}
const parsed = parseRecordFindings(text);
if (parsed.errors.length) throw new RunError('RESULT_FINDINGS_MISSING', parsed.errors.join('; '));
return { findings: parsed.findings, source: parsed.source, receipt: receipt.present ? 'verified' : 'attested' };
}
export function submitResult(o: SubmitOptions): SubmitResult {
const state = readState(path.resolve(o.out));
if (!(PHASES as readonly string[]).includes(o.phase)) throw new RunError('PHASE_NOT_AWAITING', `${o.phase} is not a phase`, 3);
const phase = o.phase as Phase;
const ps = state.phases[phase];
const v = ps?.[o.voice];
if (ps?.skipped) throw new RunError('PHASE_NOT_AWAITING', `${phase} is skipped (${ps.skipped})`, 3);
if (!ps?.snapshot) throw new RunError('PHASE_NOT_AWAITING', `${phase} has not been opened; run gstack-autoplan next --out ${state.out} first`, 3);
if (!v || ps.closed_at || voiceTerminal(v)) throw new RunError('PHASE_NOT_AWAITING', `${phase} ${o.voice} is ${v?.status ?? 'not opened'}`, 3);
if (o.unavailable !== undefined || o.failed !== undefined) {
const lines = closeOutside(state, phase, o.unavailable !== undefined ? 'unavailable' : 'failed', (o.unavailable ?? o.failed)!);
writeState(state);
return { state, closed: !!ps.closed_at, findings: 0, lines };
}
const open = resolveAttempt(v, o.attempt, phase, o.voice);
const file = (o.record ?? o.result)!;
const owner = new RegExp(`^${phase}-${o.voice}\\.(a\\d+)\\.md$`).exec(path.basename(file));
if (owner && path.dirname(path.resolve(file)) === state.out && `${state.run}-${owner[1]}` !== open.attempt) {
throw new RunError('ATTEMPT_MISMATCH', `${path.basename(file)} is the result path of ${state.run}-${owner[1]}, not of the open attempt ${open.attempt} (result=${open.result})`, 3);
}
if (!fs.existsSync(file)) throw new RunError('RESULT_RECEIPT_MISSING', `result file not found: ${file}`);
const text = fs.readFileSync(file, 'utf8');
let bind: ReturnType<typeof readRecord>;
if (o.record) bind = readRecord(state, phase, open, o, text);
else {
const parsed = parseReviewResult(text, { phase, sha256: ps.snapshot });
if (!parsed.receipt.ok) throw new RunError('RESULT_RECEIPT_MISSING', parsed.errors[0]!);
if (!parsed.fenceFound || parsed.errors.length) throw new RunError('RESULT_FINDINGS_MISSING', parsed.errors.join('; '));
bind = { findings: parsed.findings, source: 'fence', receipt: 'verified' };
}
const model = o.model!;
let conflict: string | undefined;
if (o.voice === 'outside') conflict = familyConflict(model, state.native_model);
else if (state.outside_model) conflict = familyConflict(state.outside_model, model);
else if (modelFamily(model) === 'unknown') conflict = `native model ${JSON.stringify(model)} has no recognized family`;
if (conflict) throw new RunError('MODEL_FAMILY_CONFLICT', conflict, 3);
// Bind: a sealed copy of the reviewer's bytes under an attempt-owned name the reviewer never writes, per-voice
// canonical rows, journal, ledger, manifest. Nothing above this line wrote; a later edit of `file` changes nothing bound.
const bound = path.join(state.out, attemptFile(phase, o.voice, open.attempt.slice(state.run.length + 1), 'bound.md'));
const tmp = `${bound}.${process.pid}.tmp`;
fs.writeFileSync(tmp, text, process.platform === 'win32' ? {} : { mode: 0o444 });
fs.renameSync(tmp, bound);
if (process.env.GSTACK_AUTOPLAN_FAULT === 'submit-after-seal') throw new Error('fault injected: submit interrupted after sealing (GSTACK_AUTOPLAN_FAULT, tests only)');
const planFiles = [ps.snapshot_path, ps.snapshot_path && path.join(path.dirname(ps.snapshot_path), 'source-implementation.md'), state.active_plan].filter((p): p is string => !!p);
const pathCtx = { cwd: o.cwd, planFiles, planSource: state.plan_source, planName: path.basename(state.plan_input ?? state.plan_source) };
const round = roundOf(state, phase);
const rows = canonicalRows(bind.findings.map(f => { const { file, line, ...rest } = f; return { ...rest, ...normalizeFindingFile({ file, line }, pathCtx) }; }), { run: state.run, phase, voice: o.voice, model, round });
const confirm = round > 1 ? annotateConfirm(state, phase, rows, { planName: pathCtx.planName, planSource: state.plan_source, snapshotPath: ps.snapshot_path! }) : undefined;
fs.writeFileSync(`${bound}.findings.jsonl`, rows.map(r => JSON.stringify(r)).join('\n') + (rows.length ? '\n' : ''));
const sha = sha256File(bound);
const provider = o.runner ?? 'host-subagent';
const receipt = { receipt: bind.receipt, findings_source: bind.source, reviewed_sha256: ps.snapshot, ...(bind.receipt === 'attested' ? { attest_reason: o.attest } : {}) };
journal(state.out, { event: 'submitted', attempt: open.attempt, phase, voice: o.voice, result: path.basename(bound), result_sha256: sha, model, runner: provider, findings: rows.length, usage_usd: o.usageUsd ?? 'unknown', ...receipt });
settle(path.join(state.out, FILES.spend), open.attempt, o.usageUsd ?? 'unknown');
open.state = 'submitted'; open.terminal_at = now();
Object.assign(v, { status: 'completed', model, provider, output: path.basename(bound), result_sha256: sha, findings: rows.length, attempt: open.attempt, usage: { usd: o.usageUsd ?? 'unknown' }, ...receipt } satisfies Partial<VoiceState>);
if (bind.receipt === 'attested') (state.deviations ??= []).push(`receipt attested: ${phase} ${o.voice} ${open.attempt} bound from a hand-run record without a reviewer receipt (snapshot ${ps.snapshot.slice(0, 12)}; reason: ${o.attest})`);
state.artifacts[path.basename(bound)] = { path: path.basename(bound), sha256: sha };
if (o.voice === 'native') state.native_model = model; else state.outside_model = model;
const lines = [`BOUND: ${open.attempt} phase=${phase} voice=${o.voice} model=${model} findings=${rows.length} sha256=${sha.slice(0, 12)} receipt=${bind.receipt}${confirm ? ` round=${round} correlated=${confirm.correlated} settled=${confirm.settled}` : ''}`];
lines.push(...recordPlanReality(state));
const closed = ps.native?.status === 'completed' && voiceTerminal(ps.outside);
if (closed) lines.push(...closeAndAdvance(state, phase, o.reconcile));
else state.status = 'awaiting_result';
writeState(state);
return { state, closed, findings: rows.length, lines };
}
function closeAndAdvance(state: RunState, phase: Phase, reconcile?: string): string[] {
const close = closePhase(state, phase, { overridesFile: reconcile });
if (!currentPhase(state)) finishGate(state); else state.status = 'running';
return [`PHASE_CLOSED: ${phase} confirmed=${close.consensus.confirmed}${close.consensus.unverified ? ` unverified=${close.consensus.unverified}` : ''} disagree=${close.consensus.disagree} new=${close.consensus.new} native_only=${close.consensus.native_only} coverage=${close.consensus.coverage}`];
}
/**
* Record the outside voice of `phase` unavailable or failed: its open attempt
* is cancelled and the outcome journaled with that attempt id before the
* manifest is written (resume replays it); closes the phase when native is bound.
*/
export function closeOutside(state: RunState, phase: Phase, status: 'unavailable' | 'failed', detail: string): string[] {
const v = voiceState(state, phase, 'outside');
const open = openAttempt(v);
journal(state.out, { event: 'outside_' + status, phase, attempt: open?.attempt ?? null, detail });
if (open) { open.state = 'cancelled'; open.terminal_at = now(); settle(path.join(state.out, FILES.spend), open.attempt, 'unknown'); }
Object.assign(v, { status, detail });
const lines = [`OUTSIDE: ${phase} ${status} (${detail})`];
if (state.phases[phase]?.native?.status === 'completed') lines.push(...closeAndAdvance(state, phase));
else if (state.status !== 'gate_pending') state.status = 'awaiting_result';
return lines;
}
/** Mark the outside voice of the current phase unavailable/failed (the runner could not run); closes the phase if native is bound. */
export function markOutside(out: string, phase: Phase, status: 'unavailable' | 'failed', detail: string): RunState {
const state = readState(out);
closeOutside(state, phase, status, detail);
writeState(state);
return state;
}
// ---------------------------------------------------------------------------
// resume
// ---------------------------------------------------------------------------
export interface ResumeResult { state: RunState; unknown: string[]; rebound: string[]; lines: string[] }
export function resumeRun(out: string): ResumeResult {
const recovered = recoverRevise(out);
const state = readState(out);
const rows = readJournal(out);
const terminal = new Set(rows.filter(r => ['submitted', 'cancelled', 'superseded', 'outside_unavailable', 'outside_failed', 'attempt_deferred'].includes(r.event)).map(r => r.attempt));
const unknown: string[] = [];
const rebound: string[] = [];
// An outside outcome journaled before its manifest write: replay it onto the attempt it names, never onto a later round.
for (const r of rows) {
if (r.event !== 'outside_unavailable' && r.event !== 'outside_failed') continue;
const ps = state.phases[r.phase];
const v = ps?.outside;
const a = v?.attempts?.find(x => x.attempt === r.attempt);
if (!a || !v || voiceTerminal(v) || ps.closed_at) continue;
if (a.state === 'dispatched' || a.state === 'execution_unknown') { a.state = 'cancelled'; a.terminal_at = r.ts; }
Object.assign(v, { status: r.event.slice('outside_'.length), detail: r.detail });
}
for (const r of rows) {
if (r.event !== 'dispatched') continue;
const ps = state.phases[r.phase];
const v = ps?.[r.voice as Voice];
const a = v?.attempts?.find(x => x.attempt === r.attempt);
if (terminal.has(r.attempt)) {
// A submit that journaled but died before run.json: re-bind from the journal and the file on disk.
const sub = rows.find(x => x.event === 'submitted' && x.attempt === r.attempt);
const result = sub && path.resolve(out, sub.result);
if (sub && result && a && a.state !== 'submitted' && fs.existsSync(result) && sha256File(result) === sub.result_sha256) {
a.state = 'submitted'; a.terminal_at = sub.ts;
Object.assign(v!, { status: 'completed', model: sub.model, provider: sub.runner ?? 'unknown', output: path.basename(result), result_sha256: sub.result_sha256, findings: sub.findings, attempt: sub.attempt,
...(sub.receipt ? { receipt: sub.receipt, findings_source: sub.findings_source, reviewed_sha256: sub.reviewed_sha256, ...(sub.attest_reason ? { attest_reason: sub.attest_reason } : {}) } : {}) });
state.artifacts[path.basename(result)] = { path: path.basename(result), sha256: sub.result_sha256 };
if (r.voice === 'native') state.native_model = sub.model; else state.outside_model = sub.model;
rebound.push(r.attempt);
}
continue;
}
if (!a) { (v!.attempts ??= []).push({ attempt: r.attempt, state: 'execution_unknown', prompt: path.resolve(out, r.prompt), result: path.resolve(out, r.result), estimate_usd: r.estimate_usd ?? 0, dispatched_at: r.ts }); v!.status = 'running'; }
else if (a.state === 'dispatched') a.state = 'execution_unknown';
else continue;
journal(out, { event: 'execution_unknown', attempt: r.attempt, phase: r.phase, voice: r.voice, by: 'resume' });
unknown.push(r.attempt);
}
for (const phase of state.phase_order) {
const ps = state.phases[phase];
if (!ps?.snapshot || ps.closed_at) continue;
if (ps.native?.status === 'completed' && voiceTerminal(ps.outside)) closePhase(state, phase, {});
}
const stopped = resumeStop(state);
if (!currentPhase(state) && !['gate_pending', 'amend_pending', 'complete', 'incomplete', 'refused'].includes(state.status)) finishGate(state);
else if (state.status === 'interrupted' || state.status === 'running' || state.status === 'awaiting_result') {
const anyOpen = state.phase_order.some(p => (['native', 'outside'] as const).some(v => openAttempt(voiceState(state, p, v))));
state.status = anyOpen ? 'awaiting_result' : 'running';
}
journal(out, { event: 'resumed', unknown, rebound, status: state.status });
writeState(state);
const lines = [...recovered, ...stopped, `RESUMED: run=${state.run} status=${state.status} execution_unknown=${unknown.length} rebound=${rebound.length}`];
for (const a of unknown) lines.push(`EXECUTION_UNKNOWN: ${a}`);
return { state, unknown, rebound, lines };
}
export function statusLines(state: RunState): string[] {
const lines = [`RUN: ${state.run} status=${state.status} revision=${state.revision} out=${state.out}`, `SCOPE: ui=${state.scope.ui} dx=${state.scope.dx} order=${state.phase_order.join(',')} gate_rev=${state.gate_rev ?? 0}`];
for (const p of PHASES) {
const ps = state.phases[p];
if (!ps) { lines.push(`PHASE: ${p} pending`); continue; }
if (ps.skipped) { lines.push(`PHASE: ${p} skipped (${ps.skipped})`); continue; }
const v = (voice: Voice) => { const s = ps[voice]; return `${voice}=${s?.status ?? 'pending'}${s?.model ? `:${s.model}` : ''}${s?.attempt ? `@${s.attempt}` : ''}`; };
lines.push(`PHASE: ${p} ${ps.closed_at ? (ps.deferred ? `closed-deferred(${ps.deferred.reason})` : 'closed') : ps.snapshot ? 'open' : 'pending'} ${v('native')} ${v('outside')}${ps.consensus ? ` confirmed=${(ps.consensus as any).confirmed} disagree=${(ps.consensus as any).disagree} new=${(ps.consensus as any).new}` : ''}`);
}
for (const a of pendingAmendments(state)) lines.push(`AMEND_PENDING: ${a.decision} ${a.item ?? ''}${a.changes ? ` changes=${a.changes}` : ''} -> ${applyGuidance(a)}`);
for (const p of state.phase_order) for (const voice of ['native', 'outside'] as const) {
const open = state.phases[p]?.[voice] && openAttempt(state.phases[p]![voice]!);
if (open) lines.push(`PENDING: ${open.attempt} phase=${p} voice=${voice} result=${open.result}`);
}
if (fs.existsSync(path.join(state.out, FILES.spend))) lines.push(renderSpend(summarize(readLedger(path.join(state.out, FILES.spend)))));
lines.push(roundsLine(state), ...stopLines(state));
for (const d of state.capped_deltas ?? []) lines.push(`CAPPED_DELTA: v${d.k} ${d.phase} +${d.added} -${d.removed} ${d.status}`);
return lines;
}
export const TERMINAL: readonly RunStatus[] = ['complete', 'gate_pending', 'amend_pending', 'incomplete', 'refused', 'interrupted'];
export { SNAPSHOT_TOOL, RunError, FILES, readState, writeState, readJournal, journal };
export type { RunState, Voice, VoiceState, AttemptState };