Epic Autonomous "Marathon" mode (v2.10) — opt-in#2
Merged
Conversation
Marathon Loop + Heartbeat + Two-Tier Watcher + cross-model Arbiter Panel, all opt-in (default epic unchanged). Pre-implementation design; under cross-model + pre-flight review. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…rathon spec Fundamental external blockers (upstream API has no data) are world-facts, not quality failures: isolate ONLY the truly-impossible sub-tree, keep building everything else, escalate with multi-model proof at epic end. blocked_external requires >=2 external models to CONFIRM (opposite bias from halt) so it can't become a lazy escape from hard work. Archetype: astrology #150 horary sign-only cusps. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Folds in code-archaeology + doc-validator(live) + domain-expert(research) + Codex Sol xhigh. Major changes: execution lease w/ fencing token + atomic writes (kills split-brain double-exec), global circuit breakers in-script (total-attempts/resume-count/wall-clock dead-man), family-diverse arbiter voting (dedupe correlated backends), parse-failure=drop-not-halt, terminal running_with_failures vs blocked_needing_human, anti-reward-hacking PASS audit, Py3.9 ISO Z-normalization, corrected honesty boundary (durable/ 7-day-expiry/sleep/quota-auth/disable-cron). Includes expert audit + KB. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fixes 3 critical concurrency defects + 6 secondary: flock-guarded lease (os.replace != mutual exclusion), single atomic --claim-stale (kills the acquire-then-recheck-stale self-defeat), honestly-scoped worker fencing (reject stale completion at merge boundary, drop 'never double-executes'), family-unknown ineligible for 2-family bar (cursor auto), done_with_blockers terminal, arbiter classifies even at retry cap (only masks retry_fix->halt), caller-supplied durable attempt token, disposition->status transition table, core-contract serial unit owns all of epic-state.py + publishes before F3/F4. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ferred Codex Sol R3: every critical concurrency finding traces to the auto- resurrection watcher, whose correct lease needs generation-fencing threaded through the hardened dispatcher/worker/merge/commit layer. Decision: phase. v2.10 = single-process marathon loop + Codex+Claude arbiter panel (agy/cursor dropped from arbitration per Sol x3, kept as workers) + blocker ledger (suspected->blocked_needing_human, confirmed needs 2 families) + global breakers (--breaker-check read / --trip-breaker write) + driver + docs. Resume = existing /v:resume + human /v:epic re-entry (no auto-revive). All lease/flock/claim-stale/two-tier-watcher moved to a deferred v2.11 section. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…2.10 spec Final-review gate (done iff all-done AND final_review passed), remove caller-controlled blocker confirmation (v2.10 always suspected, hard-reject confirmed:true), arbiter ballot interface B emits prompt / accepts --claude-ballot from driver C, agents/spec-reviewer.md into C's scope for the anti-reward-hack gate. All 11 R4 findings bound as explicit plan constraints (truth table, breaker frequency + honest wall-clock, --record-progress-cycle, --init --stance, ledger lifecycle, feature-id audit validation, secret redaction). Sol confirmed the v2.11 deferral killed the whole concurrency race class — remaining findings are contract precision, resolved in the plan. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
4-unit plan: A core-contract (epic-state.py, serial), B arbiter panel, C driver+reviewer prose, D docs+CI+release. Every spec Component + all 11 Sol-R4 constraints mapped to a task with exact signatures + acceptance tests. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…act corrections) Partition confirmed disjoint; 13 runtime-contract findings bound as acceptance criteria: A2 blocker-precedence (independent work before terminal escalation), A2/A6 final-review roll-up (done only via --record-final-review passed), A3 transition-gating to marathon stance, total_attempts invariant + drop caller --attempt, B evidence-file containment + two-phase challenge-bound ballot API + bounded supervisor output, C stance-binding + running->failed + breaker cadence + crash-resume commit points. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds the opt-in marathon stance to the epic-state spine: --init --stance, next_feature_autonomous (DAG-transitive routing, terminal states), attempts + --can-retry + --last-error, --record-disposition, blocker-ledger lifecycle, --record-final-review gate, global breakers (--breaker-check read / --trip-breaker / --record-progress-cycle), Py3.9 ISO. Default checkpoint path byte-identical; all new fields marathon-gated. 182 selftests (was 33) green under LANG=C + Py3.9. Dogfooded via Compound V: Sonnet-5 implementer + 3 Codex gpt-5.6-sol xhigh review rounds (11->5->8->0 findings fixed: final-review bypass, runnable- excludes-transitive-dependents, fail-safe breakers, stance-binding, ledger reblock, idempotent-replay guards, malformed-state controlled-errors). Threat model pinned (single-process script-owned state) to converge. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New scripts/compound-v-epic-arbiter.py: two-phase prepare/classify API, HMAC- keyed challenge with durable issued->in_progress->consumed record (anti-replay), Codex read-only poll via the timeout supervisor (bounded --stdout/--stderr, final-message-only parse, garbled=drop-not-halt), fail-closed secret redaction (incl. escaped-quote/unclosed-block), O_NOFOLLOW dir-fd evidence containment bound to the attempt dir, family-diverse aggregation truth table (one ballot per family, unknown ineligible for the 2-family bar), Claude-only fallback capped to retry_fix|halt_feature, blocked_external always SUSPECTED on the Codex+Claude panel, byte-bounded audit (printed==persisted), agy/cursor excluded from arbitration. 147 selftests green under LANG=C + Py3.9. v2.11 multi-process concerns (CAS/fsync) documented out-of-scope. Dogfooded: Sonnet-5 implementer + 3 Codex gpt-5.6-sol xhigh security rounds (10->11->3->0), threat model pinned (single-process, user-owned tree). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Closes the sticky-breaker gap Codex flagged in the C driver review: a tripped breaker / halt_epic latched the epic at blocked_needing_human forever with no resume path but dropping to checkpoint. Adds --clear-breaker (clear latch + re-arm: --reset-wall-clock / --set-max-total-attempts, warns if it would re-trip), --clear-disposition (clear a sticky halt_epic so next_feature_ autonomous advances), and --init --start-sha (capture epic HEAD for the halt-page accumulated-diff command). Marathon-gated; checkpoint byte-identical. 214 selftests (was 182) green under LANG=C + Py3.9. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Wires the opt-in marathon path into /v:epic + epic-mode.md: stance-binding (persisted autonomy.stance authoritative), --next --autonomous routing, running->failed-before-arbitration, two-phase arbiter exchange (--prepare -> fresh adversarial Claude Task -> --classify), breaker gate before every model call (7 boundaries), deterministic PASS sample-audit (every 3rd done + always the first), final-review gate before terminal done, halt-page runbook (/v:resume <run-id>, autonomy.start_sha diff), human marathon-resume via --clear-breaker/--clear-disposition. /v:init writes epic.autonomy.stance. spec-reviewer.md gains PASS-2.5 anti-reward-hacking (ISSUE: REWARD_HACK). Default checkpoint path behaviorally unchanged; honest v2.10 boundary (in-session + human resume, no auto-resurrection — that's v2.11); no fabricated metrics. Dogfooded: Sonnet-5 implementer + Codex gpt-5.6-sol xhigh review (5 findings fixed incl. the sticky-breaker gap -> the A follow-up above). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
README Marathon Loop feature bullet with the honest boundary (opt-in; in-session + human /v:epic resume; NO auto-resurrection — deferred). CI (validate.yml) now runs epic-state.py + epic-arbiter.py --selftest under Python 3.9 (setup-python placed last so it doesn't repoint python3 away from the 3.12+PyYAML lint/manifest steps). CHANGELOG [2.10.0]; plugin.json + marketplace.json 2.9.0 -> 2.10.0 (lockstep). No fabricated metrics. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…le sample-audit (v2.10) Closes 2 cross-unit resume BLOCKERS from the Codex integration review: (1) a breaker trip / hard death between arbiter-approval and the retry re-run stranded the retry — next_feature_autonomous now routes a failed feature BY its stored disposition (retry_fix+can_retry -> runnable, so the persisted disposition IS the crash-safe retry intent; failed-without-disposition -> needs_arbitration, re-run the arbiter, never silent-abandon). (2) a done+commit before the PASS sample-audit let a crash skip the audit — new --mark/--clear-sample-audit-due persist the obligation; done/final_review are gated on no audit outstanding (record_final_review passed rejected while due; _recompute_top_status factors it so status can't sit stale-done). 244 selftests (was 214); checkpoint byte-identical. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Driver side of the 2 integration-review blockers: --record-disposition + commit BEFORE the pre-retry breaker gate (persisted disposition = durable retry intent); handle the needs_arbitration --next reason (re-run the idempotent arbiter on a crash-mid-arbitration feature); persist --mark-sample-audit-due together-with/before the done mark so a crash can't skip the audit, surface sample_audit_due before final_review on resume. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…esume) Integration review round 2 blockers (crash-window class): a disposition is now stamped with the attempt it was made for and next_feature_autonomous honors it ONLY on attempt match — a stale retry_fix from a prior attempt routes to needs_arbitration instead of launching a second unapproved retry. New --record-audit-failed does status=failed + last_error + clear sample_audit_due + invalidate final_review in ONE atomic write, closing the window where a crash mid audit-ISSUES left a bad done sticking. 264 selftests (was 244); checkpoint byte-identical. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
… (v2.10) Closes the resume deadlock: --prepare rejects an in_progress/consumed challenge, so a crash after --classify began had no recovery path. New --resume-challenge returns absent/in_progress(+prompt)/consumed(+persisted result) — a consumed challenge re-emits its verdict with ZERO model calls, so the driver records the disposition without a second egress. Same containment (validated epic dir, O_NOFOLLOW, size caps). 162 selftests (was 147). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ver (v2.10) Driver side of integration round-2 blockers: needs_arbitration first calls --resume-challenge and branches (consumed -> record-disposition from the recovered result, no re-egress; in_progress -> re-dispatch+classify; absent -> fresh prepare) so a crash mid-arbitration can't deadlock; sample-audit ISSUES uses the atomic --record-audit-failed; a crash-recovered marathon success routes through the sample-decide/mark-due-before-done handler instead of the checkpoint done path so the first-success audit is never skipped. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…xternal (v2.10) Integration round-3 tail: a crash after --record-disposition blocked_external but before --update --status blocked lost the ledger transition — resume treated failed+blocked_external as abandoned. next_feature_autonomous now surfaces needs_blocker_recording (symmetric to needs_arbitration) so the driver completes the idempotent blocked/ledger transition on resume. 275 selftests (was 264); checkpoint byte-identical. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…rds (v2.10) Integration round-3 tail: a crash after --classify persisted the frozen audit but before flipping the challenge to consumed left the record in_progress, so --resume-challenge re-dispatched Claude for an already-decided verdict. It now probes the canonical audit first and returns consumed (zero model calls) whenever the result exists. 168 selftests (was 162). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…r_recording (v2.10) Driver side of the round-3 tails: the halt-page recovery no longer bare-marks a recovered feature done (routes through the §4 sample-decide/mark-due handler so the first-success audit is never skipped); handles the new needs_blocker_recording --next reason by idempotently completing the interrupted --update --status blocked transition. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
copeus
added a commit
that referenced
this pull request
Jul 13, 2026
…; usage-object conformance; advisor_calls derived Codex findings #1(BLOCKER),#2,#3,#4: empty/string/negative usage no longer becomes measured:true 0/0; aggregate emits null (not real 0) when nothing measured; advisor_calls aggregates independent of token-measured; collect-results deep-validates the usage object + derives advisor_calls from the per-job advisor log (<run-dir>/logs/<job-id>.advisor.jsonl). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Epic Autonomous "Marathon" mode (v2.10) — opt-in
Makes
/v:epicoptionally chew through its whole runnable feature DAG in one invocation instead of stopping at every feature — the productized version of the manual "keep-drilling" overnight pattern. Opt-in; the default epic is byte-for-byte unchanged.What shipped (5 units, A→B→C→D)
compound-v-epic-state.py):--init --stance marathon,next_feature_autonomous(DAG-transitive routing, terminal states), attempts +--can-retry+--last-error,--record-disposition, blocker-ledger lifecycle,--record-final-reviewgate, global breakers (--breaker-check/--trip-breaker/--record-progress-cycle), and human-resume (--clear-breaker/--clear-disposition/--start-sha). 214 selftests, checkpoint byte-identical.compound-v-epic-arbiter.py, new): on a failed feature, Codex + a fresh adversarial Claude decide retry / abandon / halt / suspected-external-blocker. Two-phase challenge-bound API (anti-replay), fail-closed secret redaction,O_NOFOLLOWevidence containment, family-diverse aggregation (one ballot per model family), Claude-only fallback capped to retry/abandon. 147 selftests. (agy/cursor are workers, not arbiters — no kernel confinement.)v-epic.md,epic-mode.md,v-init.md,spec-reviewer.md): the autonomous loop (breaker gate before every model call,running→failedbefore arbitration, deterministic PASS sample-audit, final-review gate, halt-page runbook),/v:initstance config, and a new anti-reward-hacking reviewer pass (§2.5).The Blocker Ledger (the "do everything you can" credo)
A fundamental external blocker — an upstream API that has no data — is a fact about the world, not a quality failure. Marathon isolates only the truly-impossible sub-tree, keeps building everything else, and reports the blockers to a human at the end (with the arbiter's evidence). In v2.10 blockers are always suspected (→ human verify); auto-confirmation needs a 2nd independent model family and is deferred.
Honest boundary (shipped verbatim in the docs — no overclaim)
"Survives a fall" = the loop continues automatically within one live session, and a human re-invokes
/v:epicafter a hard death (closed terminal, crashed machine, quota) to resume from committed state. There is NO automatic resurrection while you're away. That auto-watcher (execution lease + two-tier scheduler + generation-fenced execution layer) is a deliberately deferred, separate feature — the concurrency it needs is a distributed-systems problem in its own right, not an overnight add-on.How it was built (dogfooded through Compound V itself)
Design converged over 4 Codex gpt-5.6-sol xhigh rounds (which drove the scope decision to defer the auto-watcher) + a plan-review round. Each code unit: a Sonnet-5 implementer + independent Codex security/correctness rounds (A: 11→5→8→0; B: 10→11→3→0), each fix re-verified, threat model pinned to converge. Cross-model review caught ~50 real defects a same-family pass missed (final-review bypass, runnable-vs-transitive-dependents, fail-open breakers, an escaped-quote secret-egress leak, evidence TOCTOU, canonical-audit clobber). All selftests green under
LANG=C+ Python 3.9;lint-frontmatter+ manifest-validator clean.Not merged — this is for your verdict.
🤖 Generated with Claude Code