869-bot-fleet-audit-2026-06

Doc 869 - ZAO Bot Fleet Audit (June 2026)

Date: 2026-06-17 Scope: bot/ - every Telegram/agent surface in the ZAOOS monorepo Trigger: "bots terminal" session - full audit before resuming ZOE social builds Tier: INFRA / AGENTS

Companion to doc 601 (agent-stack cleanup decision) and the bot/REGISTRY.md inventory. This is a point-in-time health check of the running fleet.

---

Fleet at a glance

Bot	Handle	Dir	Files	Status	Tests	Top risk
ZOE	@zaoclaw_bot	bot/src/zoe/	~87	ACTIVE	19 files	memory-block sync; 50-call/day cap not enforced
Hermes	@zoe_hermes_bot	bot/src/hermes/	11	ACTIVE	0	no tests on coder/critic loop + cost cap
ZAO Devz	@zaodevz_bot	bot/src/devz/	1	ACTIVE	0	dual-bot coupling; loopback HTTP :3007
ZAOstock	@ZAOstockTeamBot	bot/src/ (root)	~23	ACTIVE / graduating	0	snapshot deploy drift; weak 4-char auth
Team bots (Magnetiq/AttaBotty)	@zao_magnetiq_bot, @z_attabotty_bot	bot/src/teams/	5	DORMANT (decommissioned)	0	lingering decommissioned code
Fleet agent	(none)	bot/src/fleet-agent/	1	STAGED / disabled	0	enable deliberately only
Bonfire	@zabal_bonfire	external (bonfires.ai)	-	ACTIVE	-	recall is manual relay, no SDK
DeepMeeting	@zdeepmeeting_bot	external	-	DEGRADED	-	group routing broken

All on-VPS bots share the Hermes pattern: LLM calls go through bot/src/hermes/claude-cli.ts -> Claude Code CLI on Max-plan OAuth (no API key in process env). Three systemd --user units on zaal@31.97.148.88: zoe-bot, zao-devz-stack (Devz + Hermes in one process), zaostock-bot.

---

Top issues (ranked)

1. CRITICAL - Untracked snapshot deploy drift. Hermes + Devz + ZAOstock run from ~/zaostock-bot, which is not a git checkout - it's a hand-patched snapshot that can silently diverge from main. Not reproducible, no audit trail. Fix: convert to a real checkout or a tagged release + deploy script.
2. HIGH - ZOE VPS on a stale feature branch. ~/zao-os tracks an old claude/* branch; merged PRs on main aren't deployed. Fix: point at the canonical ZOE branch and fast-forward (use scripts/zoe-deploy.sh, which exists precisely to prevent the per-file drift that crashed the bot 2026-06-15).
3. HIGH - Decommissioned code lingering. bot/src/teams/ (Magnetiq/AttaBotty) is dormant + decommissioned per doc 601 but still in-tree. Same for FISHBOWLZ references outside bot/. Fix: delete or formally un-decommission with a doc.
4. MEDIUM - Zero tests on Hermes core. The coder->critic retry loop, cost cap (fleetDailyGuard is an in-memory counter that resets on restart), and the risk-pattern/secret detector are all untested. Fix: add runner + types tests.
5. MEDIUM - ZOE 50-call/day cap is documented but not enforced. No counter in the concierge path; the "alert if exceeded" never fires. Fix: per-day counter in the turn path with a warn + soft-block.
6. MEDIUM - Secret-hygiene guards are partial. .env is gitignored and Hermes detects some risk patterns, but there is no pre-commit hook (guard 2) and the detector misses PEM blocks and the sk-ant- Anthropic key shape. Fix: add a .husky/pre-commit running the doc-473 step-2 scan; widen the regex set.
7. MEDIUM - Bonfire recall is manual relay, not SDK. recall.ts hardcodes the relay path; a Bonfire API change blinds ZOE. Off-VPS bots have no heartbeat to the cowork board. Fix: migrate once the SDK lands; add a VPS relay heartbeat.
8. LOW - Unstructured logging. ~167 console.log/warn/error across the fleet - fine for the systemd journal, but no severity/context. Fix: thin structured logger.

---

Per-surface notes

ZOE (bot/src/zoe/)

Mature, the most-developed surface. Per-turn model is a single synchronous runConciergeTurn (one Claude CLI call) with a typing refresh + one 6s "Got it, working" ack. Multi-step work goes through dispatch.ts/decompose.ts with a onSubtaskStart progress hook and a y/n ambiguity gate. Memory is a Letta-style 4-block builder (persona.md, human.md, recent/<chat>.json, tasks.json). Good test coverage on the hard logic (decompose, dispatch, reflexion, extractors, pii); gaps on memory.ts, concierge.ts, scheduler.ts. The posts/ social drafter is its own subsystem (see Decision below).

Hermes (bot/src/hermes/)

Lean, integration-focused, no standalone entry - invoked via /fix in the Devz chat or a loopback HTTP dispatch. Coder (Opus) + Critic (Sonnet), retry to 3, PR on score >=70. Solid risk-pattern guard in types.ts. Zero tests; cost accounting is "notional" (public API pricing, but Max plan is flat-fee) and the daily guard is non-persistent.

ZAO Devz (bot/src/devz/)

Two grammY bots in one process narrating the Hermes loop as Coder/Critic. Clean, well-guarded loopback HTTP listener on 127.0.0.1:3007 (x-hermes-secret, strict path validation). Tight coupling: if either token dies, narration breaks. Phase 3 plan is to fold into Hermes.

ZAOstock (bot/src/ root)

Festival coordination; graduating to its own repo. Monolithic command dispatch, Supabase-backed. Weak 4-char login codes, fragile freeform-text action parser, still references Minimax/Ollama for classification (should move to the Claude CLI path the rest of the fleet uses).

Team bots / Fleet agent

teams/ is dormant + decommissioned - delete or re-doc. fleet-agent/ is a deliberately-disabled host supervisor with a hard-coded systemctl allowlist and no shell interpolation; safe, enable only on purpose.

---

What changed this session

This audit accompanied two ZOE social-drafter builds:

• Build 1 (shipped this session): ZOE now surfaces the single best post draft instead of a backlog of up to 30. New posts/select-best.ts judges the day's drafts (LLM judge + heuristic fallback, fresh-window filter) and both the daily notice and /drafts surface one winner, archiving the rest. See the PR.
• Build 2 (doc 872 - effectiveness): non-blocking clarify, live steering, progress narration. Scoped against the actual turn model above; tracked separately because doc 872 itself lives off-repo (local only) and live-steering is a concurrency change to the turn loop.

---

Source

• Sub-agent fleet sweep, 2026-06-17 (this session).
• Cross-ref: doc 601 (agent-stack cleanup), bot/REGISTRY.md, .claude/rules/secret-hygiene.md (doc 473).