Insights — 2026-06

Theme of the month: 운영 안정화 + CHANGELOG 정합성 회복

Drewgent v0.7.2 (2026-05-13) 이후 약 19일간 진행된 6개 release가 모두 CHANGELOG에 누락된 상태였음. 2026-06-01에 v0.8.0~v0.8.5까지 한 번에 prepend.

v0.8.0 — Token Plan M3 (1M context) Default 전환

4개 auxiliary 스크립트 (run_kanban_worker, hindsight, auxiliary_client×2) default를 MiniMax-M2.7 → MiniMax-M3 (1M context, 1,048,576 tokens). 소모 방식이 per-call 정액제 → 입력/출력 단가 기반 + 1M context 활용으로 변경.

왜 M3인가: M2.7 superset (1M context 포함), drop-in 호환. 장기 운영 시 context 압축 부담 감소 + Token Plan credits 효율.

scripts/run_kanban_worker.py이 가장 큰 영향 — 워커가 1M context로 작업하면 작은 SEO 글 작성 같은 작업에서 compression 안 일어나도 됨.

v0.8.1 — M2.5 Legacy 보존

agent/model_metadata.py의 MiniMaxAI/MiniMax-M2.5=204800 entry는 legacy flag로 보존. 기존 사용자 호환. 문서만 4개 doc/providers/ 페이지 default M3로 업데이트.

hard-delete 안 함 이유: M2.5는 여전히 작동. breaking change 우려. legacy flag로 mark.

v0.8.2 — P0 ReBAC Neurons 2개 추가

OpenCrab의 ReBAC (Relationship-Based Access Control) 패턴을 Drewgent에 적용:

Neuron	대상	ReBAC 체크
`禁rebac_integration`	tool/skill integration	`INTEGRATION_PROTOCOL` 참조 이력
`禁rebac_kanban`	kanban 작업	`KANBAN_INDEX` 참조 이력

기존 禁rule들이 정적 (위반 후 감지)이라면, ReBAC는 예방 (작업 전 relationship check). signal_processor.py의 _on_integration_workflow_started + _on_kanban_operation handlers가 발화 시점에 awareness.integrity event emit.

hybrid 구조: 정적 禁rule (위반 감지) + 동적 ReBAC (작업 전 nudge) 두 트랙이 독립적으로 작동.

v0.8.3 — Memory & Session Edge Cases (8 bugs)

장기 운영에서 누적된 8개 memory 버그 fix:

drewgent_state.py 5 patches: session dedup, encoding, PII 해시
agent/auto_learn.py 2 patches: dedup race, family filter
acp_adapter/server.py 1 patch: race condition

모두 edge case 모음이라 사용자 1회 데이터 손실 + 중복 insert 가능성이었음. 일괄 처리.

v0.8.4 — M3 follow-up (orchestrator/bot.py)

source/_agent/orchestrator/bot.py의 call_minimax() 안에 audit_log + client.messages.create() 두 곳이 M2.7로 남아있었음 (사용자 4개 파일 scope 밖). M3 superset이라 안전하게 같이 patch. CHANGELOG v0.8.4 sub-bullet follow-up entry.

v0.8.5 — Kanban Dispatcher Board Scope Hardening (오늘 작업)

3개 board-scoped dispatcher (default / content / integrations)의 Phase 0 (watchdog) + Phase 1 (TTL reclaim) SQL에 board 필터 추가.

문제: 이전까지 모든 dispatcher가 같은 in_progress pool을 봐서, 한 dispatcher가 다른 board의 dead worker를 reclaim해버리는 cross-board race가 존재. 예: [WATCHDOG-TEST] (default) + 정상 content task + 정상 integrations task가 동시에 죽으면, default dispatcher가 content task를 자기 board로 오인하고 reclaim.

해결: 각 dispatcher의 Phase 0/1 SQL에 AND board = "self_board" 추가. content만 legacy 호환 (board = "content" OR board = "" OR board IS NULL).

검증: 3 board에 fake in_progress (dead PID, TTL expired) 동시 생성 → 3개 dispatcher 각각 실행 → 자기 board만 reclaim, 다른 board는 touch 안 함. 격리 확인.

P0 alignment: 禁kanban_worker_accountability의 board-scoped enforcement 강화.

문서 정합성 회복

File	Before	After
`CHANGELOG.md`	0.7.2에서 stop (5개 release 누락)	v0.8.0~v0.8.5 모두 기록
`P5-ego/SELF_MODEL.md`	dispatcher path 오류 + board scope 미기록	path 수정 + v0.8.5 board scope 라인
`P2-hippocampus/memories/MEMORY.md`	v0.8.0, v0.8.4만 1줄	v0.8.1~v0.8.5 4줄 추가
`kanban-dispatcher-hardening` SKILL.md	board 필터 노트 없음	board 필터 섹션 + legacy pitfall

Cross-reference

INTEGRATION_PROTOCOL — ReBAC integration check
KANBAN_INDEX — ReBAC kanban check
SELF_MODEL — P5-Ego dispatcher section (board scope updated)
~/.drewgent/skills/brain/kanban-dispatcher-hardening/SKILL.md — 운영 가이드
~/.drewgent/source/drewgent-agent/CHANGELOG.md — 전체 release history

Lessons learned

CHANGELOG는 commit-time이 아니라 work-time에: 작업을 하면 그 release 당일에 CHANGELOG도 같이 prepend. batch prepend는 drift가 누적되어서 위험. 지금 19일치 누락이 한 번에 발견됨.
Cross-cutting 변경은 영향 범위 전체를 audit: M3 전환 (v0.8.0) 시 4개 파일만 했지만 5번째 파일 (orchestrator/bot.py)이 v0.8.4에 발견됨. 한 곳에 모델 이름이 hardcoded되어 있다면 다른 곳에도 있을 가능성. “이 값을 쓰는 모든 곳” 검색이 안전.
Dispatcher의 SQL은 single source of truth: 3개 dispatcher가 같은 _reclaim_stale_tasks() SQL을 복붙하면서 board scope가 빠짐. 함수형으로 만들어서 한 곳에서 board를 파라미터로 받는 게 더 안전. 현재는 inlined SQL이지만 sub-tick에는 helper 함수로 분리 가능성 있음.
Legacy 보존 vs 하드 삭제: M2.5처럼 명시적으로 deprecated 되었지만 사용자 코드가 있을 수 있는 entry는 legacy flag + 문서만 업데이트. hard-delete는 surprise breakage.

2026-06-01

user, preference exist
user, preference trigger, only watchdog should reclaim)
user, communication from short input: Verifying Phase 0 watchdog in dispatch_once_default.py. PID=
environment, tool the
user, communication from short input: Review the conversation above and consider two things:

**Me

environment, tool Watchdog + board scope isolation regression test (v0.8.5 verification). Phase 0 watchdog (os.kill(pid, 0)) catches dead workers BEFORE TTL expires; each dispatcher only touches its own board’s tasks via AND board = "self_board" SQL filter. Test: inject 3 tasks (pid=9999999 DEAD, claim_expires=+1h far future) into default/content/integrations boards, run all 3 dispatchers, assert (1) each task appears in its owning dispatcher’s watchdog_targets, (2) NO cross-board leak, (3) TTL=0. Script at /tmp/test_watchdog_isolation.py. Expected behavior: watchdog reclaims to status=ready, then Phase 2 re-claims with new live worker in the same tick (pid changes from FAKE to real).
user, communication from short input: worker_pid=9999999 (non-existent). claim_expires=FUTURE so T
environment, project [TEST] B: dead worker + expired TTL verification — 2026-06-01. Inject task (status=in_progress, worker_pid=DEAD, claim_expires=PAST), run dispatcher, assert: (1) watchdog_reclaimed=1, (2) ttl_reclaimed=0 (Phase 0 already moved it to ready), (3) claimed>=1 (Phase 2 re-claims in same tick), (4) spawned>=1 (new live worker), (5) board scope isolation — only owning dispatcher touches the task, no cross-board leak. Script: /tmp/test_b_deadworker_expired_ttl.py. Extends test_watchdog_isolation.py (test A) with TTL=-1h injection.
user, preference exist)
environment, project [TEST] B result: dead worker + expired TTL PASS (2026-06-01). All 6 assertions across 3 boards (default/content/integrations): watchdog_reclaimed=1, ttl_reclaimed=0 (Phase 0 moves it to ready before Phase 1 sees it), claimed=1, spawned=1, no cross-board leak, new live pid 9999999→44100/44102/44105, spawned workers alive=True. Self-heal works in single tick even when both failure modes fire simultaneously. Script: /tmp/test_b_deadworker_expired_ttl.py. Pairs with test A (/tmp/test_watchdog_isolation.py) — together they cover the watchdog matrix (dead/future+TTL=0, dead/past+TTL=0 due to ordering).
user, preference exist), TTL=+1h
environment, tool what
user, communication from short input: Review the conversation above and consider saving or updatin- environment, tool open(‘/Users/drew/
user, communication from short input: [SYSTEM: Background process proc_2e8bfa8a478b completed (exi
user, communication from short input: SEO가 수집하는 내용이 진짜 Search Engine Optimization에 관련된 것들인지 점검해줘.
environment, project 2026-06-01: gateway decomposition plan vs reality 갭 발견. (1) run.py 9,876→8,752 lines (1,124 lines already extracted). (2) cron_runner는 이미 done — cron_runner.py root (9.9KB) + P3-sensors/gateway/cron_scheduler.py (162 lines, start/stop/execute). (3) sentinel key-prefix bug fix 미적용 — run.py:2080-2091은 여전히 sentinel check 없이 무조건 evict. (4) mock_gateway_runner fixture는 conftest.py에 미존재 — 실제 4-file test는 116/133 pass, 16 fail, 1 skip. plan이 진술한 131/131은 거짓. (5) 9/9 sentinel test가 아닌 6/9 (3 fail with 30s timeout). 실제 scope: dispatcher.py 신규 (1,777 LOC _handle_message+_handle_message_with_agent) + sentinel eviction fix + 선택적 conftest mock fixture 정식 추가 + 3개 신규 test 파일.
user, communication from short input: 목적 gateway_decomposition_plan.md에서 dispatcher (~2,000 LOC
user, communication from short input: 모델 호출 할 때 하나로 고정되어 있나? minimax가 새 모델이 추가되면서 차이가 생긴 것 같은데. 1M
user, communication from short input: 다음을 진행하기 전에 확인이 필요한데, 지금 콘텐트 파이프라인을 통해 쿼츠로 humanerd.kr 사이트에

[2026-06-01 16:16] Trend: vibe-kanban

Source: https://github.com/BloopAI/vibe-kanban
Score: 7.32 | Axes: {‘relevance’: 1.0, ‘direct_impact’: 0.8, ‘actionability’: 0.5, ‘novelty’: 0.5, ‘credibility’: 0.65}
Reason: keep: score=7.32, direct_impact=0.8, relevance=1.0

[2026-06-01 16:16] Trend: Anthropic, OpenAI를 제치고 가장 가치 있는 AI 스타트업이 됨

Source: unknown
Score: 6.61 | Axes: {‘relevance’: 1.0, ‘direct_impact’: 0.5, ‘actionability’: 0.5, ‘novelty’: 0.559, ‘credibility’: 0.65}
Reason: keep: score=6.61, direct_impact=0.5, relevance=1.0

[2026-06-01 16:16] Trend: Show GN: $300/월 AI 비용을 절반으로 줄인 OSS — claude-ns-hub

Source: unknown
Score: 6.68 | Axes: {‘relevance’: 1.0, ‘direct_impact’: 0.65, ‘actionability’: 0.65, ‘novelty’: 0.299, ‘credibility’: 0.5}
Reason: keep: score=6.68, direct_impact=0.65, relevance=1.0

[2026-06-01 16:16] Trend: Claude Code - 문서가 알려주지 않는 설정 가능한 모든 것

Source: unknown
Score: 6.75 | Axes: {‘relevance’: 0.95, ‘direct_impact’: 0.65, ‘actionability’: 0.5, ‘novelty’: 0.561, ‘credibility’: 0.5}
Reason: keep: score=6.75, direct_impact=0.65, relevance=0.95

[2026-06-01 16:16] Trend: MCP는 죽었나?

Source: unknown
Score: 7.15 | Axes: {‘relevance’: 1.0, ‘direct_impact’: 0.65, ‘actionability’: 0.65, ‘novelty’: 0.559, ‘credibility’: 0.5}
Reason: keep: score=7.15, direct_impact=0.65, relevance=1.0

[2026-06-01 16:16] Trend: docs

Source: https://github.com/github/docs
Score: 7.14 | Axes: {‘relevance’: 0.95, ‘direct_impact’: 0.5, ‘actionability’: 1.0, ‘novelty’: 0.5, ‘credibility’: 0.5}
Reason: keep: score=7.14, direct_impact=0.5, relevance=0.95

[2026-06-02 00:06] Trend: tiez-clipboard

Source: https://github.com/jimuzhe/tiez-clipboard
Score: 6.77 | Axes: {‘relevance’: 0.95, ‘direct_impact’: 0.5, ‘actionability’: 0.8, ‘novelty’: 0.5, ‘credibility’: 0.5}
Reason: keep: score=6.77, direct_impact=0.5, relevance=0.95

[2026-06-02 00:06] Trend: cli-printing-press

Source: https://github.com/mvanhorn/cli-printing-press
Score: 7.45 | Axes: {‘relevance’: 1.0, ‘direct_impact’: 0.65, ‘actionability’: 0.8, ‘novelty’: 0.5, ‘credibility’: 0.65}
Reason: keep: score=7.45, direct_impact=0.65, relevance=1.0

[2026-06-02 00:06] Trend: skills-for-copilot-studio

Source: https://github.com/microsoft/skills-for-copilot-studio
Score: 7.73 | Axes: {‘relevance’: 1.0, ‘direct_impact’: 0.65, ‘actionability’: 0.95, ‘novelty’: 0.5, ‘credibility’: 0.65}
Reason: keep: score=7.73, direct_impact=0.65, relevance=1.0

[2026-06-02 00:06] Trend: armbian

Source: https://github.com/sponsors/armbian
Score: 7.27 | Axes: {‘relevance’: 1.0, ‘direct_impact’: 0.5, ‘actionability’: 1.0, ‘novelty’: 0.5, ‘credibility’: 0.5}
Reason: keep: score=7.27, direct_impact=0.5, relevance=1.0

[2026-06-02 00:06] Trend: TradingAgents

Source: https://github.com/TauricResearch/TradingAgents
Score: 6.5 | Axes: {‘relevance’: 0.95, ‘direct_impact’: 0.5, ‘actionability’: 0.65, ‘novelty’: 0.5, ‘credibility’: 0.5}
Reason: keep: score=6.5, direct_impact=0.5, relevance=0.95

[2026-06-02 00:06] Trend: ruflo

Source: https://github.com/ruvnet/ruflo
Score: 7.18 | Axes: {‘relevance’: 1.0, ‘direct_impact’: 0.65, ‘actionability’: 0.65, ‘novelty’: 0.5, ‘credibility’: 0.65}
Reason: keep: score=7.18, direct_impact=0.65, relevance=1.0

2026-06-02

environment, tool what
user, communication from short input: Review the conversation above and consider saving or updatin
user, communication from short input: humanerd.kr 사이트에 올라가는 모든 콘텐트는 콘텐트 파이프라인을 거치고 특히 윤문 단계를 통해 잘 - user, communication from short input: 지금 내 토큰의 사용량과 리셋까지 남은시간, 주간 한도 등 정보를 터미널에 띄울 수 없을까? 매번 platf
user, communication from short input: Review the conversation above and consider two things: