Handle OpenCode session events that carry the session ID under properties.info.id or properties.info.sessionID so background tasks and continuation hooks do not miss idle/error/delete events.
Add regression coverage for nested session.idle events completing background tasks and waking continuation hooks.
Oracle review of PR #3943 surfaced that endTaskTimer never fires for real Prometheus plans because their canonical path is .sisyphus/plans/ and the snapshot capture was nested inside the !isSisyphusPath branch intended for direct-work warning suppression. Move the snapshot/path tracking out of the warning gate so all plan-file edits are snapshotted regardless of .sisyphus prefix. Keep the warning branch isSisyphus-gated so Atlas does not yell at legitimate plan edits.
Regression test now uses a real .sisyphus/plans/ path and fails against HEAD before the fix.
Three correctness fixes on top of the dispatch-before-commit invariant:
- ralph-loop-event-handler.ts: after idleSettleMs, also require state ownership and non-verification-pending to match the event source before dispatching. Applied to both the session.idle and session.error retry paths.
- verification-failure-handler.ts: if incrementIteration fails after a successful continuation injection, clear the loop state and emit a warning toast instead of returning success.
- session-reset-strategy.ts: catch thrown session.create errors so they route through the typed session_creation_rejected path instead of surfacing as an unhandled rejection.
Three additional invariant tests addressing the gaps surfaced by Cubic and the post-implementation review:
- idle path must not dispatch when state ownership changes during the idleSettleMs window
- verification-failure path must treat incrementIteration failure as a loud failure, not a success
- reset strategy must surface session.create rejections as session_creation_rejected even when the SDK throws instead of returning an error envelope
Split the verification-failure restart into clearVerificationState (clears\nthe verification flags so we cleanly transition back to the main loop)\nfollowed by injectContinuationPrompt, with incrementIteration only on\nsuccessful injection. On rejection: clear the loop state and emit a loud\nwarning toast. Mirrors the dispatch-before-commit contract enforced for\nthe idle and session.error paths.
Reorder the session.idle and session.error retry paths so the durable\niteration counter and the progress toast advance only when continueIteration\nreturns dispatched. On dispatch_rejected or session_creation_rejected,\nclear the loop state and emit a loud failure toast instead of silently\nlogging while the loop appears to make progress.\n\nAdds an explicit settle-window state check so a session.deleted firing\nduring the idleSettleMs sleep no longer feeds dispatch against a cleared\nloop. Keeps idleSettleMs intact for the original idle-settle race.
Replace silent returns in continueIteration with a discriminated union\n(dispatched | session_creation_rejected | dispatch_rejected). Wraps\ninjectContinuationPrompt in try/catch so reset-strategy createIterationSession\nreturning null and promptAsync rejections both surface as typed failures\nthe caller can react to.
Lock the contract that durable iteration state and visible UI must only\nadvance when the continuation dispatch is semantically accepted. Adds 4\npermanent invariant tests covering the idle, session.error retry, and\nverification-failure orchestration paths, plus the reset-strategy\nsilent-null path.
- Add /exceeded.*quota/i and /usage\s*quota/i to classifyErrorType quota block
- Align /usage.?quota/i -> /usage\s*quota/i in RETRYABLE_ERROR_PATTERNS for consistency
- Strengthen auto-retry-signal test assertion
- Add classifyErrorType assertion to Volcano Engine regression test
Ensures Volcano Engine errors are both retryable AND logged as
errorType: quota_exceeded.
Volcano Engine sends quota exceeded errors with the words in reverse
order: 'You have exceeded the 5-hour usage quota'. The existing
patterns required 'quota' to precede 'exceeded', so they never matched.
- Add /exceeded.*quota/i and /usage.?quota/i to RETRYABLE_ERROR_PATTERNS
- Add exceeded.*quota and usage\s*quota to AUTO_RETRY_PATTERNS
- Add regression tests for both detection paths
Fixes: runtime-fallback not triggering on Volcano Engine quota errors
Atlas was delegating tasks one-by-one because the workflow framed parallel
as a conditional ("if tasks can run in parallel..."), letting models default
to the safer sequential path. The new shared ATLAS_PARALLEL_BY_DEFAULT block
flips the default: parallel is mandatory; sequential requires a NAMED
blocking dependency (input dependency or file conflict).
Adds two new prompt variants — kimi (K2.6 thinking-mode calibration:
commitment framing + concrete budgets) and opus-4-7 (counters 4.7's lower
default subagent count and literal-following bias). Recalibrates default
(Claude 4.6 family), gpt (GPT-5.5 outcome-first / decision rules over
absolutes), and gemini (preserves TOOL_CALL_MANDATE; replaces stale
session_id with task_id). All five variants share the parallel mandate
positioned BEFORE the workflow so "mandate above" references resolve.
Strengthens the orchestrator-direct-edit reminder hooks
(ORCHESTRATOR_DELEGATION_REQUIRED + DIRECT_WORK_REMINDER) with the central
challenge "Do you ACTUALLY need to be the one doing this?" — replacing the
previous bullet-heavy framing.
Tests now parametrized over all 5 variants. Adds prompt-routing.test
covering GPT/Gemini/Kimi/Opus 4.7/default routing and edge cases, plus a
session_id rejection test (every variant must use task_id for retries).
Source code change:
- src/shared/model-requirements.ts: prepend claude-sonnet-4-6 to metis fallback
chain so Sonnet becomes the default. Opus 4.7 max remains as the immediate
fallback for callers who want extra reasoning.
- src/shared/model-requirements.test.ts: update assertion to expect Sonnet
primary + Opus secondary.
AGENTS.md accuracy fixes (verified against source):
- Agent modes: Sisyphus/Hephaestus are 'primary' (not 'all'); Sisyphus-Junior
is 'subagent' (not 'all'). Confirmed via 'const MODE: AgentMode = ...' in
each agent file. Also clarified Prometheus has no agentSources factory and
is built via buildPrometheusAgentConfig.
- Sisyphus fallback chain: corrected order to kimi-k2.6 → k2p5 → kimi-k2.5
→ gpt-5.5 medium → glm-5 → big-pickle (was missing kimi-k2.5).
- Librarian/Explore: added missing minimax-m2.7 step between -highspeed and
claude-haiku-4-5.
- Metis chain: removed fictitious gemini-3.1-pro entry.
- Sisyphus-Junior chain: spelled out the actual fallback (was 'user-configurable').
- Temperatures: Sisyphus/Hephaestus do not set explicit temperature (model
default); Sisyphus-Junior is 0.1 via SISYPHUS_JUNIOR_DEFAULTS.
- Quick category default: gpt-5.4-mini (not gpt-5.4-mini-fast).
Team-mode corrections:
- Eligibility registry has 3 verdicts: eligible (sisyphus, atlas, sisyphus-junior),
conditional (hephaestus — needs D-36 teammate permission), hard-reject
(oracle, librarian, explore, multimodal-looker, metis, momus, prometheus).
- Schema has 11 fields, not 4: added max_messages_per_run, max_wall_clock_minutes,
max_member_turns, base_dir, message_payload_max_bytes, recipient_unread_max_bytes,
mailbox_poll_interval_ms.
- Hooks: 'team-session-events' is 4 sub-handlers in src/plugin/event.ts
(team-idle-wake-hint, team-lead-orphan-handler, team-member-error-handler,
team-member-status-handler), not a single Continuation-tier hook.
- Tier counts now show base + team-mode: ToolGuard 14/15, Transform 5/7.
- Total: 52 base hooks, 59 with team-mode.
Doc cascade for the Metis change:
- docs/guide/orchestration.md, agent-model-matching.md, installation.md
- docs/reference/configuration.md, features.md
When user types /hyperplan something, two pipelines race for the message:
1. keyword-detector hook (chat.message): the regex \b(hyperplan|hpp)\b
matches /hyperplan because \b is satisfied by the / boundary, so the
hook prepends <hyperplan-mode> to the text part.
2. auto-slash-command hook (chat.message, runs immediately after):
detectSlashCommand() checks `text.trimStart().startsWith("/")`. After
keyword-detector's prepend, the part now starts with <hyperplan-mode>
and the slash check fails, so the builtin command template
(with $ARGUMENTS substituted) is never injected.
The visible symptom: /hyperplan refactor X never runs the actual
HYPERPLAN_TEMPLATE - the user sees only the keyword-detector wrapper, which
is similar but not identical, and the slash command's $ARGUMENTS payload
is silently lost.
Fix: at the top of the keyword-detector hook, after isSystemDirective() but
before any keyword scan, bail out if the prompt text starts with a slash
command (^\s*\/[a-zA-Z][\w-]*\b). Slash commands are explicit invocations
and own their own mode-injection path; the keyword detector must not race
them. Free-form mentions like "hyperplan: refactor X" still trigger
keyword detection - only the leading-slash form is suppressed.
Lock the contract with three regression tests:
- /hyperplan refactor ... must NOT inject <hyperplan-mode>
- /hpp investigate ... must NOT inject (shorthand slash command)
- "hyperplan: refactor src/auth/handler.ts" still injects (free-form)
Generated with assistance of [OhMyOpenCode](https://github.com/code-yeongyu/oh-my-opencode).
Adjacent "hpp ulw" or "ulw hpp" (8 form combinations: short/long, both
orders) triggers a fused mode that suppresses the standalone ultrawork
and hyperplan banners and toasts in favor of one combo banner. The
combo banner explicitly preserves hyperplan's mandatory adversarial
workflow contract (do NOT improvise, do NOT skip rounds) instead of
silently downgrading it.
Suppression runs as a named pipeline step (suppressComboStandalones)
immediately after detection and before all consumers (planner filter,
session filters, toasts, message injection), so standalone toast checks
naturally see the already-suppressed list. Combo is allowed in non-main
sessions like ultrawork, filtered for planner agents like both
standalones, and blocked in subagent sessions via the existing gate.
disabled_keywords uses the intersection rule: disabling either
"ultrawork" or "hyperplan" also disables the combo, so no flavored
content leaks via the combo embedding when either base keyword is
disabled.
Includes a same-PR refactor of KEYWORD_DETECTORS from {pattern, message}
to {type, pattern, message} tuple shape, dropping the parallel hardcoded
types array in detector.ts that previously coupled type assignment to
registry index. Future detector additions can no longer silently corrupt
DetectedKeyword.type via reorder or insertion.
10 behavioral contract tests in hyperplan-ultrawork.test.ts cover both
trigger orders, non-adjacent rejection, suppression of injection and
toast, intersection-rule disable behavior, session/agent policy, and
ultrawork variant routing through the combo. The pre-existing combined
"ultrawork hyperplan" assertion in hyperplan.test.ts is removed in
favor of the new file.
Plan distilled from a hyperplan adversarial review (5 members,
3 rounds: skeptic, validator, researcher, architect, creative).
Update both injection paths to reflect the 7-phase workflow with the
mandatory plan agent handoff in step 7. The keyword detector now lists
all 8 enumerated steps including the explicit "do not write the plan
yourself" guidance, and the /hyperplan command template references the
7-phase workflow count.
Generated with assistance of [OhMyOpenCode](https://github.com/code-yeongyu/oh-my-opencode).
YeonGyu-Kim