A wait loop like `while pgrep -f "pytest evals"` matches its own shell
and runs until its timeout even after the job exits. Tell Claude and Codex
owners to save the job PID and wait with kill -0, or use a [p]attern that
cannot match itself, and to confirm the job finished afterwards.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When a paired task escalated to the user after a reviewer final, the
"⚠️ 자동 해결 불가" mention was emitted after the turn had already
published its final and released its attempt. MessageTurnController
dropped it ("Discarding late agent output after terminal final") and the
ownership guard would have suppressed it anyway, so rooms looked stalled
right after the reviewer's reply (javis_bot 2026-11-11 14:11, 15:28).
- Mark host post-completion notices (escalation, reviewer-unavailable
answer) and send them as a fresh message past both guards and the
reviewer direct-terminal skip.
- Reviewer prompt: NEEDS_CONTEXT is only for questions the user must
answer; ask the owner for evidence with DONE_WITH_CONCERNS.
- run_verification: load the helper and verifier from the EJClaw install
instead of the repo under review (was "Module not found" outside
EJClaw), and run the helper outside the repo so its .env is not read
by EJClaw's config loader.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Reviewer: DONE_WITH_CONCERNS now means changes required (unresolved
P0-P2); TASK_DONE means no unresolved P0-P2 (P3 may be appended).
Drop "same concerns 2+ turns -> BLOCKED" (system already sends
deadlocks to the arbiter).
- Reviewer: flag only change-introduced problems or unmet requested items
with a concrete failing path; P0-P3 severity; "지적 없음" when clean.
- Reviewer: inspect git diff/show itself, treat owner summary as claims,
per-requirement met / not met / cannot verify.
- Reviewer: first round reports everything; later rounds check fixes and
regressions only; verify rebuttals in code; never approve to end a loop.
- Reviewer: review only, no file edits or commits (overrides platform
"do file work directly"); trivial change still needs a diff read but a
verification profile only when the change can affect its result.
- Arbiter: PROCEED when only P3 remains, REVISE on evidenced P0-P2.
- Owner: answer each finding with fixed / rebutted + code evidence.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Owner: run needed checks once after a change, do not rerun passed checks
without a new change/failure/concern, no verification subagents (the
reviewer is the independent step), no unrequested tests/files, clean up
temp scripts. Reviewer: one run per profile per snapshot, do not block on
extra tests when existing checks cover the change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A sketchup-app turn spent on English login pages (Trimble/Edge) ended
with an English NEEDS_CONTEXT answer. The Korean rule was only a prompt
line, the answer went straight to the user (no reviewer gate), and it was
carried forward into the next task, so the next turn answered in English
too.
- Claude runner: when a final answer is mostly English (Hangul < 10% of
Latin letters, >= 150 letters, code/links/boilerplate ignored), ask the
same session once to rewrite it in Korean and emit only the rewrite.
Falls back to the original on error or early stream end; disabled when
AGENT_LANGUAGE is set to a non-Korean language.
- Let handleResultMessage settle the buffered final text instead of
pre-flushing it as progress, so the English draft never surfaces.
- Platform prompt: explicit "Reply language" section; owner rule no longer
says "unless the user wrote in another language".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Bots had to zip .pptx decks because the outbound allowlist only covered
images, audio/video, pdf, zip and plain-text files. Add docx/xlsx/pptx
(checked for the ZIP local-file-header signature) and legacy doc/xls/ppt
(checked for the OLE2 compound-file signature). The 8MB cap and the
allowed-directory rules are unchanged. Prompt format lists are updated so
agents stop zipping these files.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(a) When the reviewer approves the owner's first TASK_DONE answer on the
first review (one review round, outputs are exactly owner -> reviewer, no
revision or arbiter turns) and the owner workspace still matches the
approved tree with a clean working tree, no owner run happens: the task is
completed as done and the reviewed answer (status line stripped) is posted
as the final message. If the review was requested with a bare "검토해줘"
after a review-skipped answer, a one-line confirmation is posted instead of
repeating the answer.
(b) Every other finalize turn runs in a fresh, non-persisted session (no
resume of the task's long working session, no room memory briefing) with a
self-contained prompt: this task's user messages, the latest six turn
outputs and the finalize instructions/format. All three finalize entry
points (pending turn, inline bot-only follow-up, queued turn) use it.
The owner prompt now states that the first TASK_DONE answer must already
follow the final-message format because it may be posted as the final
message unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The owner may end a TASK_DONE answer with "리뷰 불필요: <reason>". The
host treats it as a request and honors it only when every guard holds:
- workspace content equals the task base (hasCodeChangesSinceRef), the
working tree is clean, and HEAD did not move during the owner run
- the run used no file-mutating tool (Edit/Write/MultiEdit/NotebookEdit),
no subagent (Agent/Task) and no state-changing or unknown MCP tool
- every Bash command is classified read-only by a conservative
quote-aware classifier (src/bash-command-classifier.ts); anything
unknown counts as mutating
- tool usage was reported and not truncated (agent-runner now reports
tool names and Bash commands as toolUsage on its terminal output)
- the user did not ask for a review in this task
When honored the task completes as review_skipped with no reviewer or
finalize turn; the answer is delivered with the marker line removed and a
small "-# 검토 생략 …" note. Otherwise the marker is stripped and the
normal review runs. A bare "검토해줘"-style message right after a
review-skipped answer reopens that task for the reviewer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When Claude hit a limit the run was handed to a Codex service
(codex-main/codex-review) through service_handoffs and a global owner
failover. That service does not exist in this deployment, so handoffs
stayed pending forever (2 rows since May) and the user got no answer.
Removed:
- handoff creation (message-agent-executor-handoff, paired-turn-fallback)
and the delegated paired-execution path
- the pending-handoff poll/claim worker (message-runtime-handoffs) and the
service_handoffs CRUD module (db/service-handoffs, db exports)
- global owner failover in service-routing, its recovery timer in index,
the dashboard failover line, the codex-review failover prompt pack and
prompts/codex-review-failover-platform.md
- failoverOwner in RoomRoleContext and the `failover=` ROOM_ROLE field
- OWNER_/REVIEWER_/ARBITER_FALLBACK_ENABLED and FAILOVER_MIN_DURATION_MS
- tests for the removed behaviour
Now, when a Claude run fails with a usage/rate limit (trigger reason
429 or usage-exhausted) the executor closes the open paired task as
completed/claude_usage_limit and sends the room one final notice:
"Claude 사용 한도에 도달했습니다. 21:40에 재개할 수 있습니다." The reset time
comes from the error text (CLI banner "resets 9:40pm (Asia/Seoul)", dated
banners, "try again at ...", ISO timestamps), else from the cached usage
API window at 100%, else "잠시 후 다시 시도해 주세요". Nothing retries
silently; the next user message starts a new task. Other Claude failures
(session failure, org access denied) now end as ordinary run errors.
The service_handoffs table, its columns on paired_turn_attempts and the
migrations stay untouched for the later migration squash; the runtime no
longer reads or writes them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MOA_* was never configured, so enrichArbiterPromptWithMoa always
returned the prompt unchanged and the Kimi usage fetch only produced an
empty dashboard row. Remove src/moa.ts, src/kimi-usage.ts, the arbiter
prompt enrichment, MoA config/schema, the Kimi row and MoA line in the
Discord status dashboard, and the MOA_* / unused FALLBACK_* provider
examples from the docs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The user should be able to choose from the last arbiter message alone, without
scrolling back through the owner↔reviewer conversation. Instruct the arbiter, on
ESCALATE, to restate the decision in one sentence and list the options as a
short numbered list with per-option consequences, including only the minimum
context needed. Prompt files are read per-turn, so this is live immediately.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Owner continuity:
- Seed a freshly created owner task with the previous task's latest owner
final on a cold start after the previous task already closed, so a user
reply to a finalized TASK_DONE no longer produces a "no context" answer.
Activated for this deployment via PAIRED_CARRY_FORWARD_LATEST_OWNER_FINAL
(.env); carried text is injected as clearly-marked background only.
- Skip intermediate STEP_DONE outputs when picking the carry-forward anchor.
Single-mode routing:
- enforceRoomModeOnLease strips a stale reviewer/arbiter lease from a room
switched back to single, preventing single-mode messages from stalling in
the paired path on a stuck execution lease.
Session auth / credentials:
- Pre-sync Claude credentials into each session dir before the agent spawns.
- Honor CLAUDE_CREDENTIALS_PATH in setup/login.ts (per-service isolation).
- Add a relogin-required gate so a permanently logged-out claude-code room
asks the user to re-login instead of spawning a doomed agent.
Other:
- Arbiter verdicts written in the user's language (verdict keyword stays EN).
- status-dashboard chatName field; runtime-inventory credential path resolver.
Tests (make suite fully green: 1595 pass / 3 skip):
- service-routing: default owner is now the claude service and reviewer is
codex-review; update the 7 failover/default expectations accordingly.
- migrate-room-registrations: owner inferred as claude-code (configured
OWNER_AGENT_TYPE) for a dual legacy room; reviewer becomes codex.
- register: mock paired-workspace provisioning + reload signal (registration
now provisions a workspace and hot-reloads); assert RELOADED status.
- paired-execution-context: force a claude-code reviewer to exercise the
Claude read-only branch regardless of the deployment default.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Agents ended turns with "I'll wait … I'll be notified" expecting a
background-completion callback that this environment never delivers, so
the turn hung and left the planning sentence stuck in the channel. Add a
claude-platform rule that no such notification exists — finish within the
turn (foreground/poll) or ask the user for a follow-up.
Also suppress the elapsed-time suffix on progress messages until at least
one 5s bucket has elapsed, so freshly-created progress no longer shows a
meaningless "0초". Update the affected message-runtime expectations.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Making a channel registration live no longer depends on an operator/agent
remembering to restart. setup/register.ts and scripts/deregister-room.ts now
call signalEjclawReload() after the DB write, which sends SIGHUP to the running
service's main PID (→ runtimeState.reloadRoomBindings). Best-effort and
injectable-for-tests; no-ops when the service isn't running.
Also document the canonical rule in the committed, agent-loaded platform prompt
(prompts/claude-platform.md): prefer these paths, never fake a restart, verify
via the "Room bindings reloaded" log; a plain in-turn systemctl restart is
forbidden (it kills the agent).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The user only sees the final owner message, not the owner↔reviewer loop. Add a
"Finalize message format" section mandating a direct answer, a consolidated
recap (no transcript, no narrating disagreement), an explicit 사용자 액션 아이템
section (with "없음" when nothing is needed), and no reviewer-loop meta phrases.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
Follow-up to e3c8de6: prettier --check failed on the new section
(missing blank line after the "Before accepting any proposal..."
sentence ahead of the numbered list, and the "Do not include:" header
needed a blank line before the bullets). Running prettier --write
fixes those without changing semantics.
bunx prettier --check prompts/owner-common-paired-room.md → clean.
The user only sees the final owner message, not the owner↔reviewer
back-and-forth. Recent turns leaked meta-narrative like "리뷰어 지적
반영해서 정정합니다" / "PROCEED 확정" into user-visible replies and
sometimes omitted explicit action items, leaving the user to guess
whether anything was needed from them.
Add a "Finalize message format (self-contained answer)" section to
owner-common-paired-room.md that mandates:
- direct answer first
- consolidated recap (no transcript, no narrating disagreement)
- explicit "사용자 액션 아이템" section, with "없음" when there is
nothing for the user to do
- no meta phrases referencing the reviewer loop
The runner reads this prompt fresh per agent spawn
(src/agent-runner-environment.ts), so the change takes effect on the
next owner turn without a service restart.
- Build both agent-runner and codex-runner into container image
- Select runner path based on group.agentType at docker exec time
- Update README to role-centric overview, generic MoA references
- Add MoA reference opinions guidance to arbiter prompt
Owner was "agreeing" with concerns instead of fixing code. Reviewer
was listing issues without directing the owner to fix them. Now:
- Reviewer: explicitly direct the owner to fix specific issues
- Owner: fix issues when flagged, don't just acknowledge
Both agents were repeating "agreed, 3 concerns remain" without
taking action. Now DONE_WITH_CONCERNS requires listing specific
actions, and repeating the same concerns for 2+ turns triggers
BLOCKED escalation.
All reviewers (claude, codex-review, codex-review-failover) now use
claude-paired-room.md. All owners use owner-common-paired-room.md.
Removed: codex-paired-room.md, codex-review-paired-room.md,
codex-review-failover-paired-room.md, codex-review-platform.md
10 prompt files → 6.
The anti-pattern guardrails ("It should work now" means run it, etc.)
and stagnation pattern names (Spinning/Oscillation/Diminishing/No progress)
are too useful to compress out — they prevent common LLM failure modes.