Plugin System Design
Design document for the Gas Town plugin system. Written 2026-01-11, crew/george session.
Problem Statement
Gas Town needs extensible, project-specific automation that runs during Deacon patrol cycles. The immediate use case is rebuilding stale binaries (gt, bd, wv), but the pattern generalizes to any periodic maintenance task.
Current state:
- Plugin infrastructure exists conceptually (patrol step mentions it)
~/gt/plugins/directory exists with README- No actual plugins in production use
- No formalized execution model
Design Principles Applied
Discover, Don’t Track
Reality is truth. State is derived.
Plugin state (last run, run count, results) lives on the ledger as wisps, not in shadow state files. Gate evaluation queries the ledger directly.
ZFC: Zero Framework Cognition
Agent decides. Go transports.
The Deacon (agent) evaluates gates and decides whether to dispatch. Go code provides transport (gt dog dispatch) but doesn’t make decisions.
MEOW Stack Integration
| Layer | Plugin Analog |
|---|---|
| Molecule | plugin.md - work template with TOML frontmatter |
| Ephemeral | Plugin-run wisps - high-volume, digestible |
| Observable | Plugin runs appear in bd activity feed |
| Workflow | Gate → Dispatch → Execute → Record → Digest |
Architecture
Plugin Locations
~/gt/├── plugins/ # Town-level plugins (universal)│ └── README.md├── gastown/│ └── plugins/ # Rig-level plugins│ └── rebuild-gt/│ └── plugin.md├── beads/│ └── plugins/│ └── rebuild-bd/│ └── plugin.md└── wyvern/ └── plugins/ └── rebuild-wv/ └── plugin.mdTown-level (~/gt/plugins/): Universal plugins that apply everywhere.
Rig-level (<rig>/plugins/): Project-specific plugins.
The Deacon scans both locations during patrol.
Execution Model: Dog Dispatch
Key insight: Plugin execution should not block Deacon patrol.
Dogs are reusable workers designed for infrastructure tasks. Plugin execution is dispatched to dogs:
Deacon Patrol Dog Worker───────────────── ─────────────────1. Scan plugins2. Evaluate gates3. For open gates: └─ gt dog dispatch plugin ──→ 4. Execute plugin (non-blocking) 5. Create result wisp 6. Send DOG_DONE4. Continue patrol ...5. Process DOG_DONE ←── (next cycle)Benefits:
- Deacon stays responsive
- Multiple plugins can run concurrently (different dogs)
- Plugin failures don’t stall patrol
- Consistent with Dogs’ purpose (infrastructure work)
State Tracking: Wisps on the Ledger
Each plugin run creates a wisp:
bd wisp create \ --label type:plugin-run \ --label plugin:rebuild-gt \ --label rig:gastown \ --label result:success \ --body "Rebuilt gt: abc123 → def456 (5 commits)"Gate evaluation queries wisps instead of state files:
# Cooldown check: any runs in last hour?bd list --type=wisp --label=plugin:rebuild-gt --since=1h --limit=1Derived state (no state.json needed):
| Query | Command |
|---|---|
| Last run time | bd list --label=plugin:X --limit=1 --json |
| Run count | bd list --label=plugin:X --json | jq length |
| Last result | Parse result: label from latest wisp |
| Failure rate | Count result:failure vs total |
Digest Pattern
Like cost digests, plugin wisps accumulate and get squashed daily:
gt plugin digest --yesterdayCreates: Plugin Digest 2026-01-10 bead with summary
Deletes: Individual plugin-run wisps from that day
This keeps the ledger clean while preserving audit history.
Plugin Format Specification
File Structure
rebuild-gt/└── plugin.md # Definition with TOML frontmatterplugin.md Format
+++name = "rebuild-gt"description = "Rebuild stale gt binary from source"version = 1
[gate]type = "cooldown"duration = "1h"
[tracking]labels = ["plugin:rebuild-gt", "rig:gastown", "category:maintenance"]digest = true
[execution]timeout = "5m"notify_on_failure = true+++
# Rebuild gt Binary
Instructions for the dog worker to execute...TOML Frontmatter Schema
# Requiredname = "string" # Unique plugin identifierdescription = "string" # Human-readable descriptionversion = 1 # Schema version (for future evolution)
[gate]type = "cooldown|cron|condition|event|manual"# Type-specific fields:duration = "1h" # For cooldownschedule = "0 9 * * *" # For croncheck = "gt stale -q" # For condition (exit 0 = run)on = "startup" # For event
[tracking]labels = ["label:value", ...] # Labels for execution wispsdigest = true|false # Include in daily digest
[execution]timeout = "5m" # Max execution timenotify_on_failure = true # Escalate on failureseverity = "low" # Escalation severity if failedGate Types
| Type | Config | Behavior |
|---|---|---|
cooldown | duration = "1h" | Query wisps, run if none in window |
cron | schedule = "0 9 * * *" | Run on cron schedule |
condition | check = "cmd" | Run check command, run if exit 0 |
event | on = "startup" | Run on Deacon startup |
manual | (no gate section) | Never auto-run, dispatch explicitly |
Instructions Section
The markdown body after the frontmatter contains agent-executable instructions. The dog worker reads and executes these steps.
Standard sections:
- Detection: Check if action is needed
- Action: The actual work
- Record Result: Create the execution wisp
- Notification: On success/failure
Escalation System
Problem
Current escalation is ad-hoc “mail Mayor”. Issues:
- Mayor gets backlogged easily
- No severity differentiation
- No alternative channels (email, SMS, etc.)
- No tracking of stale escalations
Solution: Unified Escalation API
New command:
gt escalate \ --severity=<low|medium|high|critical> \ --subject="Plugin FAILED: rebuild-gt" \ --body="Build failed: make returned exit code 2" \ --source="plugin:rebuild-gt"Escalation Routing
The command reads town config (~/gt/config.json or similar) for routing rules:
{ "escalation": { "routes": { "low": ["bead"], "medium": ["bead", "mail:mayor"], "high": ["bead", "mail:mayor", "email:human"], "critical": ["bead", "mail:mayor", "email:human", "sms:human"] }, "contacts": { "human_sms": "+1234567890" }, "stale_threshold": "4h" }}Escalation Actions
| Action | Behavior |
|---|---|
bead | Create escalation bead with severity label |
mail:mayor | Send mail to mayor/ |
email:human | Send email via configured service |
sms:human | Send SMS via configured service |
Escalation Beads
Every escalation creates a bead:
type: escalationstatus: openlabels: - severity:high - source:plugin:rebuild-gt - acknowledged:falseStale Escalation Patrol
A patrol step (or plugin!) checks for unacknowledged escalations:
bd list --type=escalation --label=acknowledged:false --older-than=4hStale escalations get re-escalated at higher severity.
Acknowledging Escalations
gt escalate ack <bead-id># Sets label acknowledged:trueNew Commands Required
gt stale
Expose binary staleness check:
gt stale # Human-readable outputgt stale --json # Machine-readablegt stale --quiet # Exit code only (0=stale, 1=fresh)gt dog dispatch
Formalized plugin dispatch to dogs:
gt dog dispatch --plugin <name> [--rig <rig>]This:
- Finds the plugin definition
- Slinga a standardized work unit to an idle dog
- Returns immediately (non-blocking)
gt escalate
Unified escalation API:
gt escalate \ --severity=<level> \ --subject="..." \ --body="..." \ [--source="..."]
gt escalate ack <bead-id>gt escalate list [--severity=...] [--stale]gt plugin
Plugin management:
gt plugin list # List all pluginsgt plugin show <name> # Show plugin detailsgt plugin run <name> [--force] # Manual triggergt plugin digest [--yesterday] # Squash wisps to digestgt plugin history <name> # Show execution historyImplementation Plan
Phase 1: Foundation
gt stalecommand - Expose CheckStaleBinary() via CLI- Plugin format spec - Finalize TOML schema
- Plugin scanning - Deacon scans town + rig plugin dirs
Phase 2: Execution
gt dog dispatch --plugin- Formalized dog dispatch- Plugin execution in dogs - Dog reads plugin.md, executes
- Wisp creation - Record results on ledger
Phase 3: Gates & State
- Gate evaluation - Cooldown via wisp query
- Other gate types - Cron, condition, event
- Plugin digest - Daily squash of plugin wisps
Phase 4: Escalation
gt escalatecommand - Unified escalation API- Escalation routing - Config-driven multi-channel
- Stale escalation patrol - Check unacknowledged
Phase 5: First Plugin
rebuild-gtplugin - The actual gastown plugin- Documentation - So Beads/Wyvern can create theirs
Example: rebuild-gt Plugin
+++name = "rebuild-gt"description = "Rebuild stale gt binary from gastown source"version = 1
[gate]type = "cooldown"duration = "1h"
[tracking]labels = ["plugin:rebuild-gt", "rig:gastown", "category:maintenance"]digest = true
[execution]timeout = "5m"notify_on_failure = trueseverity = "medium"+++
# Rebuild gt Binary
Checks if the gt binary is stale (built from older commit than HEAD) and rebuilds.
## Gate Check
The Deacon evaluates this before dispatch. If gate closed, skip.
## Detection
Check binary staleness:
```bashgt stale --jsonIf "stale": false, record success wisp and exit early.
Action
Rebuild from source:
cd ~/gt/gastown/crew/george && make build && make installRecord Result
On success:
bd wisp create \ --label type:plugin-run \ --label plugin:rebuild-gt \ --label rig:gastown \ --label result:success \ --body "Rebuilt gt: $OLD → $NEW ($N commits)"On failure:
bd wisp create \ --label type:plugin-run \ --label plugin:rebuild-gt \ --label rig:gastown \ --label result:failure \ --body "Build failed: $ERROR"
gt escalate --severity=medium \ --subject="Plugin FAILED: rebuild-gt" \ --body="$ERROR" \ --source="plugin:rebuild-gt"---
## Open Questions
1. **Plugin discovery in multiple clones**: If gastown has crew/george, crew/max, crew/joe - which clone's plugins/ dir is canonical? Probably: scan all, dedupe by name, prefer rig-root if exists.
2. **Dog assignment**: Should specific plugins prefer specific dogs? Or any idle dog?
3. **Plugin dependencies**: Can plugins depend on other plugins? Probably not in v1.
4. **Plugin disable/enable**: How to temporarily disable a plugin without deleting it? Label on a plugin bead? `enabled = false` in frontmatter?
---
## References
- PRIMING.md - Core design principles- mol-deacon-patrol.formula.toml - Patrol step plugin-run- ~/gt/plugins/README.md - Current plugin stub