- Removed actions/checkout@v4 (requires Node.js, unavailable on runner)
- Gitea act runner provides repo checkout automatically
- Added 'enforcement' to valid kind list for hermes-key-enforcement
- All steps are now pure shell — zero external dependencies
Previously the pipeline only triggered on pull_request events.
Direct pushes to master bypassed all validation (auth, lint, ai-review).
Now push to master also triggers the pipeline.
NEW: docs/AUTHORING-GUIDE.md — canonical style guide for writing contracts
Patterns borrowed from Anthropic's frontend-design skill:
- Ground it in the subject: verify live state before writing
- Two-pass process: draft → critique → revise → ship
- Restraint: one contract, one concern; cut one thing before shipping
- Self-critique: would an agent trust this at 3am with 30s of context?
Core rules:
- Specificity over generality (lists rot, vagueness is invisible)
- Live-state fields in tables, not prose (grep-able)
- Descriptions answer: what, to what, why care?
- Templates for function and responsibility contracts
- CI enforces structure; author enforces quality
Linked from AGENTS.md for agent discovery.
- New 'Operating Doctrine (read first)' section with the .117 rule, safe-mutate wrapper usage, field trust taxonomy, and contract-native verify-before-fix flow
- Added hermes-key-enforcement to templates table
- Added stirling-pdf-agent-access to functions table
- Noted VERIFY-BEFORE-USE labels on infrastructure-control
- Added FIELD TRUST convention to frontmatter (live-state vs policy fields)
- Marked credentials/IPs with VERIFY-BEFORE-USE labels
- Fixed 3 stale Zulip IP references (.117 → .19) that caused the 2026-07-03 outage
- Added verify-before-mutate skill reference
- See knowledge graph node #604 for post-mortem
- hermes-key-enforcement.prose.md: NEW contract enforcing api_key_env for all harness/LiteLLM providers, exempting external providers
- hermes-config-template.prose.md: Updated with full fix log (15 total fixes across Koby/Koonimo/Mumuni/Tanko), model:auto detection rule, key-to-agent mapping
- infrastructure-control.prose.md: Added Stirling-PDF service details (port 8989, credentials, API key, Swagger URL) replacing Bento-PDF
- REMOVED hardcoded Agent Keys table (keys became stale after LiteLLM downgrade).
- ADDED Key Management section: verify loop, rotate + deploy procedure, key
deployment rules (plaintext in /etc/environment only, api_key_env pattern),
current aliases (2026-07-02 rotation).
- LiteLLM DB is the single source of truth for keys; contracts never store plaintext.
- Mumuni sub-agent profiles documented (6, all inherit auth from main config).
- infrastructure-control: add Section 5.3 (Network Verification & Routing) with
dual-path checks, NetBird ingress health, DNS drift detection, and routing
regression alerts. Add Section 7 (Configuration Doctrine — IP-First) mandating
LAN IPs for all configs/agents, URLs reserved for human browser access only.
Also enrich access matrix (Mumuni, Koonimo, LiteLLM Admin, Grafana entries),
reachability matrix, and CT 116 nginx routing + Prometheus targets.
- build-zulip-plugin: fold Success Criteria into Maintains postconditions.
- pi-approval-architecture: new reference doc — pi approval model vs Hermes,
architectural limits, /approve and /deny stub commands.
- zulip-approval-fix: new responsibility — documents the Zulip HTML/prefix bug
that broke /approve and /deny, and the one-line fix applied.
Runaway request from .123 (Mumuni) decoded 39,868 tokens over 24 min,
pushing Tctl to 98°C (9°C past critical) and throttling 70->29 t/s.
Server-side -n 8192 hard cap now bounds generation. Documented connection
source (all from .116/LiteLLM) and the thermal constraint (fanless APU).
Revert the /grafana/ nginx sub-path I wrongly added — it broke the existing
http://192.168.68.116:3001/d/gpu-fleet URL. Restored direct 0.0.0.0:3001 LAN
access. Contract is now the source of truth: no GF_SERVER_SERVE_FROM_SUB_PATH,
no nginx /grafana/ route. gpu-fleet + 3 new dashboards all at :3001.
- gpu-fleet: documented parallel=2/batch=4096 optimizations for RTX 3090/5070
- gpu-fleet: added Strix Halo ROCm status (installed, blocked by HSA runtime on Debian 13)
- gpu-fleet: documented alert migration from DM → #agent-hub stream topics
- gpu-monitor: added Alert Delivery section with stream topic conventions
- All GPU alerts now go to #agent-hub for cross-agent visibility
- Rate limiting during reconnection (429): honor retry-after headers
- Placeholder edit fails silently: fallback to sending as new message
- Both fixes deployed to pi Zulip extension
- Polls GPU sidecars every 15s for temp, VRAM, power, utilization
- Creates responsive HTML dashboard at /root/dashboard/gpu-fleet.html
- Served via nginx at /gpu/ on CT 116
- PM2-managed gpu-monitor server on port 9100
- Tracks Strix Halo llama-server via SSH
- Exposes JSON API at /gpu-data for integrations
Documents lessons from building pi Zulip extension and Hermes Zulip plugin:
1. Queue registration content-type
2. Missing bot_user_id for @mentions
3. Zero stream subscriptions
4. Stale errors never cleared
5. Stuck detection too aggressive
6. Env var name mismatch
7. Poll interval too aggressive
8. A2A port conflict
Also: fixed Hermes adapter to clear stale errors on successful poll
Root cause: abiba-zulip extension had STUCK_THRESHOLD_MS=30min.
After 30min of inactivity (overnight), it reported stuck=True, which
triggered the health monitor to restart it — cycling restarts up to 18.
Permanent fixes:
1. STUCK_THRESHOLD_MS raised 30min → 4 hours in extension index.js
2. PM2 restart counter reset: delete+re-add abiba-zulip (was 18, now 0)
3. pm2-self-heal.sh alert threshold documented at 30 (implementation already)
4. Prose contract updated with historical note and threshold rationale
root
jerome