subagent

name

subagent

description

The ONLY sanctioned way to delegate work to a lazar sub-agent. Uses locked bin/lazar-spawn (not agent-writable skills logic) — lints the prompt, bounds retries, wall-clock-kills wedged children, and gates success by re-checking declared claims against the filesystem. Raw `lazar -p` is vetoed by a pre-tool hook. Use whenever you need to fan out work, isolate context, or run something as a separate agent. Provider-agnostic.

subagent — how I extend myself

Delegation is a core skill. I never call lazar -p directly (a pre-tool hook vetoes it). I go through the locked spawn surface:

bash "$LAZAR_HOME/bin/lazar-spawn" <task-id> <prompt-file>
# equivalent thin re-exec (same binary):
bash skills/_meta/subagent/spawn.sh <task-id> <prompt-file>

bin/lazar-spawn is installed by setup.sh with the same immutability story as bin/lazar (not under sandbox-writable skills/). Budget state lives in logs/subagent/ (sandbox cannot write logs/). The copy under skills/ is only a re-exec stub — never the real control plane.

This is the same path for every model/provider. No provider-specific shortcuts.

The translation (run this on every delegation)

1. Gate. One task that needs my full context → I do it myself. Independent chunks or context-isolation helps → delegate.
2. Decompose. Units a stranger with zero context and only the filesystem can complete.
3. Translate each unit into a self-contained spawn prompt, written to a file:
   • Imperative objective.
   • Absolute paths, never names (/Users/.../x.md, not "the file").
   • A SUCCESS: line — the criterion in one sentence.
   • An EXPECT-VERIFY: block — claims that must become true, one per indented line: exists=/abs, contains=/abs:token, or size_at_least=/abs:bytes. I author these; the child cannot pass by declaring its own.
   • "Report honest failure; do not fabricate."
4. Dispatch: bash "$LAZAR_HOME/bin/lazar-spawn" <task-id> <prompt-file>. Never raw lazar -p.
5. Gate on the envelope. The wrapper re-checks EXPECT-VERIFY and returns {"status":"ok","matched":true} only if they hold. I trust matched, not the child's prose. On matched:false I fix the task and retry within bounds.
6. Aggregate + terminate on success-met or ceiling-hit. Never respawn the same failing prompt unbounded.

Prompt example

Write the current UTC date to /Users/jkneen/lazar/workspace/out/stamp.txt.
SUCCESS: stamp.txt exists and contains a 2026 date
EXPECT-VERIFY:
  exists=/Users/jkneen/lazar/workspace/out/stamp.txt
  contains=/Users/jkneen/lazar/workspace/out/stamp.txt:2026

Dispatch: bash "$LAZAR_HOME/bin/lazar-spawn" stamp /tmp/stamp-task.txt

Bounds (all models)

Three layers — do not remove one:

1. Kernel stall-nudges (in-child recovery, bounded)
2. Per-task ceiling LAZAR_SPAWN_CEILING (default 2)
3. Session budget LAZAR_SPAWN_BUDGET (default 10)

Wall-clock kill: LAZAR_SPAWN_TIMEOUT (default 300s).

Self-test

bash skills/_meta/subagent/selftest.sh — lint, ceiling, claim re-check, veto, and that production state is under logs/subagent/.