Sub-agents panel

Sub-agents are child workers spawned by the main agent via agent_spawn / delegate_to_agent. CRAFT roles (Explore, Implementer, Review, …) appear here live; the bottom section summarizes CRAFT blackboard tasks.

Code mode only. Office has no sub-agents sidebar entry.


How to open

Entry Notes
Sidebar → Agents Full-height tracking
Audit grid → bottom-right With checklist, scratchpad, LHT (UI tour)

Title bar Show audit grid when harness data exists (Code).


Layout

Header stats (three columns)

Column Counts
Running spawned | running
Completed completed
Interrupted interrupted (failed/cancelled terminals)

Only rows with ids like agent_* (not parent call_* tool ids).

Agent cards (expandable)

Element Meaning
Status dot Running pulse / complete green / interrupted red
Title nickname or type + badge
agent_id Full monospace id
Objective First ~200 chars of spawn prompt
progressStatus Latest progress line while running (SSE)
Steps done/max vs max_steps (default cap 100)
Step cap Ns From step_timeout_ms
Stuck suspected No heartbeat past timeout + buffer
Work package task_id CRAFT blackboard / scratchpad run
Footer Tool count · tokens · duration
Expanded Full objective, tool calls, resultSummary

Updates via runtime SSE and polling.


Agent types and tools

Spawn type / agent_type (CRAFT):

Type Aliases Tool clip
explore exploration, explorer Read-only: list/read/grep/glob/file_search/web/note
review reviewer, code_review Read-only + exec_shell (verify commands)
implementer implement, builder Inherits parent
verifier verify, tester Inherits parent
auditor audit, fact_checker Audit prompts + tools
plan planning Planning
general default, worker Inherits parent
custom Requires explicit allowed_tools

Review uses exec_shell for cargo check, npm test, etc. Implementer may get a git stash snapshot before edits (non-fatal rollback).


CRAFT tasks section

When .zagens/blackboards/*.json exists, bottom CRAFT tasks list (sorted by task_id):

Field Source
task_id Blackboard filename
Explorer yes/— explorer partition present
Implementer rounds implementer.rounds length
Reviewer reviewer.verdict — PASS / BLOCKER / MAJOR / FAIL
Verifier verifier.summary if any

Quick fix-loop status without opening JSON.


Settings

Settings → System → Security:

Setting Range Default
Sub-agents enabled On/off On
Max sub-agents 1–20 From config
Step timeout 120–1800 s 600s

Spawn step_timeout_ms overrides one job. Large Explore: 10–30 min typical.

Strict: [features] subagents = false in config.

See execution policy.


Scratchpad / LHT linkage

Step journals (v0.8.6+)

Sub-agents persist a compact step journal under .zagens/state/subagent-journals/{agent_id}.json (tool name, path/pattern summary, result bytes — not full transcripts). Failed/interrupted cards can show step progress and export the journal.


Tips

  1. One task_id per fix-loop; order Explore → Implementer → Review → Verifier (CRAFT).
  2. On stuck suspected, ask parent to cancel or narrow scope.
  3. Finish review before new scope in the same macro cycle.
  4. Parallel Explore is read-only; concurrent Implementer writes follow runtime ownership rules.

Related: CRAFT · Audit scratchpad · LHT settings