Add no-progress turn detector that inspects the most recent assistant
message: if finish is 'unknown', all token counts are zero, and there
is no meaningful content beyond step-start/step-finish markers, the
turn is classified as no-progress.
Integrate the check at all three idle/completion/error continuation
points in the event handler so the loop stops cleanly with a warning
toast instead of injecting another internal prompt.
/start-work was unresponsive under Atlas because ralph-loop and todo-continuation-enforcer normalized the inherited agent to a config key (e.g. 'atlas') while OpenCode only accepts the registered display name (e.g. 'Atlas (Plan Executor)'), producing 'Agent not found' on dispatch.
Prefer resolveRegisteredAgentName(agent) and fall back to normalizeAgentForPromptKey only when no registration exists, mirroring the start-work hook's resolution chain.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
The double-fire race fix (#4256) had handlePendingVerification return
early when verification_attempt_id was set but verification_session_id
was not, since the oracle dispatch is still in flight. That guard
introduced a permanent-stall failure mode: if tool-execute-after never
runs (oracle session hangs, crashes, OOM-killed, or tmux killed
externally), verification_session_id stays undefined forever and the
ralph-loop never escapes the pending-verification state.
Track verification_attempt_started_at as a state field, clear it on
restart/clear/setVerificationSessionID, and fall through to
handleFailedVerification when the attempt has been pending past
STUCK_VERIFICATION_TIMEOUT_MS (30 minutes). Legacy persisted states
without the timestamp continue to defer (no timeout to evaluate),
matching pre-fix behavior for that edge case.
Closes pre-publish blocker V25.
extractRetryableSignal returns the raw isRetryable hint from up to 5
nested AI SDK error paths. isRetryableError previously trusted any
true result blindly, which would burn every configured fallback model
in an infinite loop if a provider mis-tagged a 401, 403, or other
non-transient 4xx as retryable.
Honor the signal only when the status code is absent, in the configured
retry_on_errors list, in 5xx, or in {408, 425, 429}. Reject the signal
when the status code is a non-transient 4xx and log the rejection so
operators can debug provider mis-classifications.
Closes pre-publish blocker V8.
Closes five gaps in the ultrawork prompt versus codex-plugins' parallel
directive, applied to all three model variants (default/Claude, GPT, Gemini)
with prompt-engineering entropy gate (each addition encodes a distinct
binding boolean, not narrative reinforcement):
1. TDD-MANDATORY (was conditional "when test infrastructure exists"):
every production change follows RED -> GREEN -> SURFACE. Failing test
first, capture assertion msg, smallest change to flip green, exercise
real surface, capture artifact. Exemption whitelist: formatting /
comment-only / version bump / rename-only, each must be justified
in writing; unjustified exemption = rejection.
2. Scenario contract (was free-form Test Plan Template): require 3+
scenarios upfront covering happy path, edge (boundary / empty /
malformed / concurrent), adjacent-surface regression. Each scenario
binds a binary pass condition, a real-surface artifact source, and
a test file + test id written test-first.
3. RED->GREEN evidence capture (was "all tests pass"): every scenario
requires TWO captured artifacts -- RED assertion msg before the
change AND GREEN assertion msg after -- alongside the real-surface
artifact (tmux / curl / browser / Playwright / computer-use /
CLI stdout / parsed config / DB diff). Tests are the floor (always
required); surface artifact is the ceiling (also required).
4. Durable notepad: mktemp -t ulw-*.md with append-only sections
(Plan, Scenarios, Now, Todo, Findings, Learnings). Survives context
loss; resume by re-reading.
5. Reviewer gate: trigger when user said strictly / rigorously /
"deeply", or task touches 3+ files / 20+ turns / 30+ min, or it is
refactor / migration / perf / security work. Reviewer verdict is
binding ("looks good but..." = rejection). Loop until unconditional
approval.
Plus: TODO format upgraded from vague "track every step" to atomic
`path: <action> for <scenario-id> -- verify by <check>` with a GOOD
test-first / impl pair example and a BAD list including
"production code before its failing test".
Per-variant adaptation:
- default.ts (Claude): full structured sections.
- gpt.ts (GPT-5.x): outcome-first prose, shorter prose per gpt-5.5 guide.
- gemini.ts: explicit enforcement framing + anti-optimism checkpoint
upgraded with a TDD-violation question (#7).
Verified by:
- bun test src/hooks/keyword-detector/ (119 pass / 0 fail).
- lsp_diagnostics clean on all three files.
- Module-load smoke test confirms each exported message string parses
and contains the new section anchors (TDD MANDATORY, SCENARIO
CONTRACT, DURABLE NOTEPAD, REVIEWER GATE).
Char deltas (directive body only):
- default 13646 -> 17144 (+26%)
- gpt 6740 -> 9215 (+37%, was the leanest start)
- gemini 14196 -> 16136 (+14%)
Existing tests only assert presence of "ULTRAWORK MODE ENABLED!" which
is preserved verbatim in every variant.
The lead currently leaves teams alive after the task list drains because
none of the prompt surfaces tell it WHEN to close or HOW. omx-style
'self-closing' behavior was missing for four reasons (diagnosed via
prompt-engineering A/B/C: wrong / misframed / missing):
1. builtin team-mode skill 'Lifecycle' (B+C): 'phase ends / shape
outgrown' is qualitative, so the model maps it to 'wait for user'.
Step 6 jumped to team_delete without the request/approve pair the
tool contract requires. Replaced with a 'Closure Contract' (a
computable predicate over team_task_list + team_status) and an
explicit 'Closure Sequence' (request -> approve -> delete, with
force=true reserved for unrecoverable paths only).
2. TEAM_MESSAGE keyword injection (C): spent 100%% of its one-shot
budget on routing ('do not substitute delegate_task'), 0%% on
closure. Added the same closure rule in compressed form. Kept the
'NEVER substitute with delegate_task' literal that
keyword-detector/index.test.ts depends on.
3. team-mode-status-injector body (C): the only per-session injection
for team mode had no closure obligation. Replaced the optional
'load the team-mode skill ... otherwise use the team_* tools'
sentence with a 'Closure invariant' clause that ties the check to
every team_task_update.
4. member-guidance Wrap-up (A+B): step 3 said 'so the lead can decide
whether to request shutdown', but team_shutdown_request is
lead-only - members cannot initiate it. Step ordering also placed
the completion message before team_task_update, so the lead's
closable check would see stale data. Reordered to
task_update -> check task_list for new work -> if nothing left,
send a single 'closure-ready' message and idle. Test assertion
updated to match the new accurate contract.
Also: stripped Korean alternation from TEAM_PATTERN per directive
('절대로 코드 내에 한국어 적지 마라'). Pattern is now
/\\bteam[\\s_-]?mode\\b/i. Removed 4 Korean test cases
(2 positive triggers + 2 false-positive guards) that the pattern no
longer needs to defend, and updated the keyword-detector AGENTS.md
row.
Net: -71 lines across prompt surfaces. The Closure Contract is the
only addition; everything else tightened.
Tests: 428/428 pass across src/features/team-mode/,
src/features/builtin-skills/, src/hooks/keyword-detector/,
src/hooks/team-mode-status-injector/, src/hooks/team-mailbox-injector/,
src/hooks/team-tool-gating/, src/hooks/team-session-events/.
LSP: no errors introduced (one pre-existing error in
keyword-detector/index.test.ts confirmed pre-existing on dev).
When verification_pending is true and the agent has dispatched an Oracle
verification (verification_attempt_id is set), session.idle events that
arrive before tool-execute-after stores the Oracle session ID
(verification_session_id still undefined) caused handlePendingVerification
to fall through to handleFailedVerification. This injected a duplicate
'verification failed' continuation prompt, spawning a second Oracle.
The fix adds a guard in handlePendingVerification: when
verification_attempt_id is set but verification_session_id is not, Oracle
dispatch is in flight and the handler returns early instead of declaring
failure. The pending wake will retry on the next session.idle.
Regression test added in given/when/then style proving the race sequence:
1. ULW loop detects DONE, enters verification_pending
2. Oracle dispatch stamps verification_attempt_id (tool-execute-before)
3. Second session.idle fires before tool-execute-after stores session ID
4. Handler must NOT call handleFailedVerification
RED (before fix): 2 prompt injections (duplicate Oracle)
GREEN (after fix): 1 prompt injection (correct)
Fixes#4256Fixes#4019
OpenCode Desktop's Electron sidecar runtime can omit Bun's ctx.$ helper.
The sender previously called ctx.$ unconditionally, throwing
TypeError: ctx.$ is not a function as unhandledRejection and crashing
the sidecar with exit code 1.
Add a runtime guard at every call site, falling back to Node.js
child_process.execFile (with windowsHide: true) when ctx.$ is missing.
The Bun ctx.$ path remains preferred when available. Every notification
path is wrapped in try/catch so no failure escapes as unhandledRejection.
Fixes#4128Fixes#4061
Issue 1: hasNewCommentsOnly() now returns false when oldString and newString
both contain comment syntax and the new lines are a subset of old lines —
preventing the hook from firing on comment-only modifications.
Issue 2: Per-session deduplication via sessionLastWarning Map with a 30s
window (DEDUP_WINDOW_MS). At most one warning fires per session per
response turn, breaking the deadloop on consecutive edits.
notepad-write-guard:
- The hook was created by create-tool-guard-hooks but tool-execute-before
never invoked it, so the guard was inert.
- It also only matched .sisyphus/notepads, missing the current
.omo/notepads layout introduced by the workspace migration.
- Add the dispatch call alongside writeExistingFileGuard, and extend
NOTEPAD_ROOTS to cover both paths via normalize() + sep. New
integration test pins the wire and the .omo block; the existing unit
test now asserts both paths.
start-work session-plan-affinity:
- PLAN_PATH_PATTERN only matched .sisyphus/plans, so sessions referring
to plans under .omo/plans returned null and start-work missed the
current session's own plan.
- Extend the regex to .(sisyphus|omo)/plans and add findPrometheusPlans
in packages/boulder-state to scan both directories during the
transition. New regression test pins .omo/plans matching; legacy
.sisyphus/plans coverage preserved.
- prompt-async-gate.test.ts: refactor ced36bffc removed
promptAsyncAfterSessionIdle in favor of the unified
dispatchInternalPrompt({ mode: 'async', ... }). One call site at
line 1441 was left behind. Replace it with the current API and pass
the explicit dispatchTimeoutMs so the status-timeout semantics are
preserved. Also switch the surrounding tests to the third-argument
timeout form so Bun's typings stay happy.
- runtime-model-readers.test.ts: implementation moved to
packages/model-core during the layering refactor; the orphaned test
still pointed at './runtime-model-readers'. Switch to the package
export via getModelCapabilities and keep the modality-reader
coverage by deriving keys through the package API.
- prometheus-prompt.test.ts: close missing }) on the OpenSpec expanded
commands describe block (introduced by d66b6bcbf, parse error).
- agent-sort-shim/agent-config-integration/continuation-injection/
unstable-agent-babysitter/subagent-resolver/sync-executor/
resolve-caller-team-lead tests: expect 'Sisyphus - ultraworker'
(lowercase) to match production after cd39f8858, which lowercased the
display name to dodge a TUI ZWSP rendering glitch. Legacy uppercase
inputs that exercise the normalization path are preserved.
- sync-executor.ts + resolve-caller-team-lead.ts: route legacy display
name inputs through normalizeAgentForPrompt so prompt agent names and
caller team lead lookups produce the canonical lowercase form.
The ZAI (Zhipu) provider emits 'Weekly/Monthly Limit Exhausted. Your limit will reset at YYYY-MM-DD HH:MM:SS' when the coding-plan subscription quota is hit. None of the existing quota regex patterns (/quota.?exceeded/, /usage\s+limit/, /exhausted\s+your\s+capacity/, /credit\s+balance.*too\s+low/, etc.) match the 'Limit Exhausted' phrasing, so the runtime-fallback never fires and the user is stuck on the dead model.
Add /limit\s+exhausted/i to both pattern lists that gate fallback dispatch:
- RETRYABLE_ERROR_PATTERNS in constants.ts (text-pattern path used by extractStatusCode + retryable scan)
- classifyErrorType quota_exceeded branch in error-classifier.ts (typed classification path used by isRetryableError)
The pattern is intentionally narrow: it requires the literal token 'Limit' followed by whitespace then 'Exhausted'. It matches the ZAI weekly, monthly, and combined Weekly/Monthly variants but does not collide with unrelated phrases such as 'context limit' or 'rate limit' that already have their own dedicated patterns.
Regression coverage added to quota-error-classifier.regression.test.ts:
- 'Weekly/Monthly Limit Exhausted. Your limit will reset at 2026-05-20 15:43:27' -> quota_exceeded + retryable=true
- 'Weekly Limit Exhausted. Your limit will reset at 2026-05-28 10:30:00' -> quota_exceeded + retryable=true
Verification: 11/11 quota-error-classifier.regression.test.ts pass (was 9 pass + 2 fail pre-fix). Broader runtime-fallback suite goes from 135/196 pass to 137/198 pass (the 61 pre-existing failures are unrelated to this change and reproduce on a clean upstream/dev checkout). bun run typecheck clean.
src/hooks/comment-checker/apply-patch-edits.ts was already a pure re-export over @oh-my-opencode/comment-checker-core after the core extraction landed. Every importer now reaches into the package directly, so the shim has no remaining call sites and can be removed.
Verified: rg "comment-checker/apply-patch-edits" src/ packages/ returns no matches.
Promote the project-rule constants (PROJECT_MARKERS, PROJECT_RULE_SUBDIRS, PROJECT_RULE_FILES, OPENCODE_USER_RULE_DIRS, USER_RULE_DIR, GITHUB_INSTRUCTIONS_PATTERN, RULE_EXTENSIONS, AGENTS_FILENAME, etc.) and the findAgentsMdUp walk-up helper out of the agents-md-core and rules-injector adapters and into @oh-my-opencode/rules-engine, the single owner of rule discovery.
- packages/agents-md-core/ drops the findAgentsMdUp/AgentsMdDiscoveryInput wrappers (now sourced directly from rules-engine) and its constants module re-exports AGENTS_FILENAME from rules-engine instead of duplicating it.
- src/hooks/directory-agents-injector/finder.ts pulls findAgentsMdUp from rules-engine directly while still re-exporting resolveFilePath from agents-md-core.
- src/hooks/rules-injector/constants.ts becomes a pure re-export shim over the rules-engine constants.
Add packages/agents-md-core/src/injector.test.ts to lock the root-skipping AGENTS.md injection order so future changes to findAgentsMdUp cannot silently regress the [Directory Context: ...] block format the injector emits.
Tests: bun test packages/agents-md-core packages/rules-engine src/hooks/directory-agents-injector src/hooks/rules-injector
The hyperplan trigger \b(hyperplan|hpp)\b/i matched 'hpp' inside common C++ header references like 'check interface.hpp' or 'open buffer.hpp'. The leading '.' is a non-word character, so \b is already satisfied and the false positive fires the hyperplan-mode prompt on routine code questions.
Split the alternation so 'hpp' additionally requires that the preceding character is neither a word character nor a '.'. This preserves every existing trigger ('hpp do this', '/hpp ...', mid-sentence usage, mixed case) while rejecting filename uses of the .hpp extension. The longer 'hyperplan' keyword keeps the original \b boundary semantics.
Reproduction (added regression tests):
- 'please help to check interface.hpp' must NOT fire
- 'open src/include/audio/buffer.hpp and fix the leak' must NOT fire
All 14 cases in hyperplan.test.ts pass (12 existing + 2 new), broader keyword-detector suite stays green (92 pass), typecheck clean.
YeonGyu-Kim