agent

Initialize any agent by name -- looks up the dispatch table and runs the right primitive (skill, daemon, integration, external).

SKILL.md

GitHub

Agent Dispatch

Initialize, query, or interact with any named agent in the fleet. Resolves the agent name to its execution primitive and runs the appropriate action.

Usage

[agent] <name>                  # Default action (status for daemons, run for integrations)
[agent] <name> status           # Check agent status (process, bus, health)
[agent] <name> run [args]       # Execute the agent's primary function
[agent] <name> logs             # Tail recent logs or bus messages from this agent
[agent] <name> delegate <task>  # Prepare a delegation packet for external agents

Dispatch Table

The dispatch table lives in playbooks/AGENT_REGISTRY.md under the Dispatch Table section. It maps every agent name to:

  • Primitive: skill, agent-def, daemon, integration, ci-workflow, external, entry-script
  • Invocation: the actual command or action to take
  • Default action: what happens when no action is specified

Execution

0. Emit SKILL_INVOKE

Post SKILL_INVOKE to the bus before any stateful action.

Type: SKILL_INVOKE
To: all
Message: [skill=agent] [mode=side_effecting] [args_hash=<sha256>] [session=<session_id>]

(The skill invocation runtime injects the caller's canonical identity as from_id.)

1. Parse arguments

Extract $ARGUMENTS into:

  • AGENT_NAME: first word (required)
  • ACTION: second word (default: per dispatch table)
  • ARGS: remaining words

2. Resolve agent identity

# Check canonical agents, autonomous services, and aliases
python3 -c "
from hummbl_governance.services.agent_identity import canonicalize, get_status, get_trust_tier, AUTONOMOUS_SERVICES, CANONICAL_AGENTS
name = '$AGENT_NAME'
canonical = canonicalize(name)
print(f'canonical={canonical}')
print(f'status={get_status(name)}')
print(f'trust={get_trust_tier(name)}')
print(f'is_service={name in AUTONOMOUS_SERVICES or canonical in AUTONOMOUS_SERVICES}')
print(f'is_agent={canonical in CANONICAL_AGENTS}')
"

If not found in identity registry, check the dispatch table in AGENT_REGISTRY.md.

3. Look up dispatch table

Read playbooks/AGENT_REGISTRY.md and find the row matching AGENT_NAME in the dispatch table.

4. Execute by primitive type

Daemon (launchd)

Default action: status

# Check process
launchctl list | grep -i "$AGENT_NAME"
# Recent bus messages from this agent
grep "$AGENT_NAME" _state/coordination/messages.tsv | tail -10
# Logs (if available)
cat ~/Library/Logs/hummbl/$AGENT_NAME.log 2>/dev/null | tail -20

If action is run: warn that daemons are managed by launchd. Suggest launchctl kickstart if they want to force a run.

Integration (Python adapter)

Default action: run Execute the invocation command from the dispatch table and interpret the output.

External agent (codex, gemini)

Default action: delegate Chain to [delegate] skill with the agent name and any provided task description. Apply guardrails from ~/.agents/rules/ (gemini-guardrails.md, _archived/kimi-guardrails.md).

CI Workflow (GitHub Actions)

Default action: status

gh run list --workflow="$WORKFLOW_FILE" --limit 3

Entry Script (bin/)

Default action: run Execute the script from the dispatch table.

Agent Def (Claude Code session persona)

Default action: info Read the agent definition file and summarize capabilities. Note: Agent defs can only be initialized at session start, via your runtime's agent-init mechanism (see the runtime binding table in rules/skill-provider-neutrality.md).

Skill (existing skill)

Default action: run Chain to the skill: invoke [skill-name] directly.

5. Output format

## Agent: <display_name>
**Canonical ID**: <id> | **Primitive**: <type> | **Trust**: <tier> | **Status**: <status>

### Result
<action output, interpreted>

### Recent Bus Activity
<last 5 bus messages from this agent, if any>

Error Handling

  • Unknown agent: List similar names from the registry. Suggest [agent-roster] for full list.
  • Retired agent: Show retirement reason from registry. Do not execute.
  • Deprecated identity: Reject. Show canonical name if one exists.
  • Probation agent: Warn about guardrails before executing.

Examples

[agent] arbiter              → runs arbiter score, shows quality digest
[agent] lead-doctor          → shows daemon status + recent health transitions
[agent] lead-doctor logs     → tails lead-doctor bus messages and log files
[agent] codex delegate "run security audit on services/"
                            → prepares delegation packet with context
[agent] gemini status        → shows probation status, recent bus activity, guardrail summary
[agent] config-drift         → shows daemon status + last drift check result
[agent] quality-arbiter      → shows last CI run status for quality-arbiter.yml

Skill Chains

Mandatory

None — meta-skill; the dispatched skill carries its own mandatory chains. This skill only resolves the dispatch table and delegates.

Advisory

  • After [agent] <name> status → [agent-roster] for fleet-wide context
  • After [agent] <name> delegate → [delegate] to complete the delegation packet
  • After [agent] <name> logs → [debug-test] if logs reveal errors worth investigating
  • For Devin→opencode delegation → [cross-runtime-bridge] (python ~/bin/cross_runtime_bridge.py delegate "<task>" --workdir <dir> --auto)

Authority

  • T1 (TRUSTED): May run freely
  • T2 (Active/High): May run freely
  • T3 (Medium): Operator approval required for side_effecting targets; read-only actions (status, logs) permitted
  • T4 (Probationary): May run advisory targets only (status, info, logs); side_effecting targets BLOCKED
  • Operator: Override any restriction