Eyejoker
cb79450572
fix: add arbiter verdict compliance to owner and reviewer prompts
2026-03-30 22:28:48 +09:00
Eyejoker
a043f2b79a
feat: implement MAGI 3-agent arbiter system for deadlock resolution
...
Introduces a third agent role (arbiter) that is summoned on-demand
when owner and reviewer reach a deadlock (same verdict 3+ rounds).
Architecture:
- 3 Discord bots: owner (codex), reviewer (claude), arbiter (claude/codex)
- Arbiter is NOT always-on — only invoked when deadlock detected
- Arbiter renders binding verdict: PROCEED/REVISE/RESET/ESCALATE
- Non-escalate verdicts reset round_trip_count and resume ping-pong
- Backward compatible: ARBITER_AGENT_TYPE unset = existing 2-agent mode
Changes across 13 source files + 7 test files:
- types.ts: PairedRoomRole += 'arbiter', new statuses, ArbiterVerdict type
- config.ts: ARBITER_AGENT_TYPE, ARBITER_SERVICE_ID, ARBITER_DEADLOCK_THRESHOLD
- db.ts: schema migration (arbiter columns in channel_owner + paired_tasks)
- service-routing.ts: arbiter_service_id in lease
- room-role-context.ts: arbiter role detection
- paired-execution-context.ts: deadlock->arbiter, verdict handling
- message-runtime.ts: arbiter turn routing, cursor, sender labeling
- message-agent-executor.ts: arbiter mode, failover
- agent-runner.ts + environment.ts: arbiter container mode
- platform-prompts.ts: arbiter prompt loading
New files:
- prompts/arbiter-paired-room.md: arbiter system prompt
- src/arbiter-context.ts: builds conversation context for arbiter judgment
2026-03-30 22:20:42 +09:00
Eyejoker
196c885c9e
fix: clarify owner/reviewer action expectations in paired room prompts
...
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
2026-03-30 20:49:02 +09:00
Eyejoker
fde29db227
fix: break DONE_WITH_CONCERNS echo loop in paired room prompts
...
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.
2026-03-30 20:41:32 +09:00
Eyejoker
760e222b54
fix: explicitly declare owner role in paired room prompt
2026-03-30 03:58:39 +09:00
Eyejoker
1031cd44dd
refactor: SSOT for paired room prompts, remove 4 duplicate files
...
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.
2026-03-30 02:39:04 +09:00
Eyejoker
17b60cb10f
fix: restore communication discipline details and stagnation pattern names
...
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.
2026-03-30 02:33:41 +09:00
Eyejoker
691be0dc86
refactor: streamline paired room prompts and fix role separation
...
Prompts:
- Remove model-specific names (클코, Codex) — use owner/reviewer only
- Compress duplicated sections (communication, stagnation, consensus)
while preserving 4 ontological questions and verdict protocol intact
- Remove verdict examples (redundant with status descriptions)
- claude-platform.md: strip SDK defaults, keep only send_message rules
Code:
- Remove reviewer paired room prompt from owner CLAUDE.md
(was causing role confusion — owner got "You are the reviewer")
- Remove "Provide feedback or approve" from review turn prompt
(conflicted with CLAUDE.md verdict protocol)
187 lines → 100 lines across 4 prompt files.
2026-03-30 02:31:04 +09:00
Eyejoker
88fa0858ea
feat: strengthen reviewer prompts with critical review, completion status, and stagnation awareness
...
- Expand claude/codex-review paired-room prompts with full review structure
- Add completion status protocol (DONE/DONE_WITH_CONCERNS/BLOCKED/NEEDS_CONTEXT)
- Add stagnation detection (spinning, oscillation, diminishing returns)
- Add communication discipline (fact-based, verification-based)
- Remove <internal> tag instructions from owner platform prompt
- SSOT: all 3 reviewer prompts share the same structure
2026-03-29 22:46:09 +09:00
Eyejoker
806e5e4ff1
fix: keep codex approval wording role-based
2026-03-28 17:02:50 +09:00
Eyejoker
772fea45df
refactor: share owner prompts and tighten silent output guidance
2026-03-28 07:46:13 +09:00
Eyejoker
ba9f6871b6
feat: add codex review failover and suppress output hardening
2026-03-28 05:40:38 +09:00
Eyejoker
bba19b0c0f
docs: clarify Claude internal visibility rules
2026-03-25 23:43:46 +09:00
Eyejoker
8b5624731c
docs: add GitHub CI watcher usage guide to platform prompts
...
Both Claude and Codex platform prompts now include a short
section explaining when to use structured GitHub fields
(ci_provider, ci_repo, ci_run_id) for the host-driven fast
path instead of the generic LLM-per-tick watcher.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com >
2026-03-25 23:19:26 +09:00
Eyejoker
094278b08e
feat: add ontology checklist, verification gate, completion status protocol, and structured stagnation escalation to paired-room prompts
2026-03-25 07:46:53 +09:00
Eyejoker
2865ba07ce
feat: add communication discipline to paired-room prompts (fact-based responses, verification-based completion, 3-strike architecture review)
2026-03-25 07:37:51 +09:00
Eyejoker
5e75222754
feat: strengthen paired-room prompts with critical review, consensus gate, and assumption surfacing
2026-03-25 07:31:34 +09:00
Eyejoker
767ab20d76
Restrict Codex task scheduling to CI watchers
2026-03-20 04:09:36 +09:00
Eyejoker
a293a701f4
fix: restore regressions from discord-only refactor
2026-03-20 03:41:53 +09:00
Eyejoker
bb0628e8f4
refactor: remove legacy container and non-discord remnants
2026-03-20 01:07:46 +09:00
