delegate

name	delegate
description	Route one task file to the cheapest capable model lane, validate the output, and return a single verdict line. Use when work is mechanical, scoped, or parallel and does not require a human judgement call. Keeps orchestrator context clean - worker output goes to a file, only PASS/BLOCKED comes back.

delegate

Send work out. Keep your context clean. Get back one line.

python scripts/delegate.py --task <file> --tier <tier> [--files a.rs b.rs] [--mode diff|full]

Exit 0 PASS, 3 no usable output after retries, 4 BLOCKED (no capable lane or task needs a shell), 5 task file missing.

Choosing a tier

The tier says what the work needs, not who does it. Routing is derived from that plus the live roster.

Tier	Use when	Notes
micro	Mechanical, one file, no design judgement	Sub-second lanes exist
scoped	A real diff against a written spec	The common case
reasoning	Diagnosis or design where being confidently wrong is the risk	Heavier, slower lanes
long-context	Many files must be in one prompt	512k and 1M ctx lanes
shell	Must actually RUN something (gh, cargo, adb, gradlew)	HTTP lanes cannot; goes to agy
verdict	Go/no-go with human consequences	Never delegate. That is your job.

delegate.py refuses to send a shell-needing task to an HTTP lane and tells you to re-dispatch with --tier shell. Three lanes each burned a call discovering this by hand before the guard existed.

Do not delegate

• One-liners. Dispatch overhead exceeds the task.
• Bounded 1-5 tool-call diagnostics. Cheaper inline.
• Anything whose failure mode is "confidently wrong and merged anyway."
• Verdicts. A worker cannot verify itself; that is the only reason your tier exists.

Writing a task file that works

The task file IS the prompt. Every failure traced so far came from an underspecified one.

1. State the output contract. delegate.py appends one, but say it again in your own words: a diff, a whole file, a table.
2. Give it an out. Tell it to reply BLOCKED: <reason> if it cannot comply. A model that says BLOCKED is worth more than one that invents a plausible answer, and the validator treats BLOCKED as a real result.
3. Pin the scope. Name the files it may touch. Say what NOT to reformat.
4. Split diagnosis from repair. One task extracts the error, a second fixes it and is forbidden from starting without that output. Fixing from a hypothesis is how repeated attempts fail.
5. Never let it self-certify. "Run the tests and confirm they pass" returns a claim. Require the command output, and have someone else run it.

Reading the result

PASS writes to tmp/delegate/<task>__<lane>.md with lane, model, and latency in the header. Read the diff before you trust it. A PASS means the output was well-formed, not that it was correct.

BLOCKED lists every lane tried and why. Escalate deliberately, in this order, stopping as soon as one works:

1. Re-dispatch with a better task file - most blocks are spec defects, not lane defects.
2. --tier long-context if it plausibly ran out of room.
3. agy-gemini - the only free lane with a shell, so it can verify itself. Always --add-dir <repo> and always pin --model to an exact name from agy models. On timeout use --continue, never re-dispatch fresh.
4. agy-claude - spends Anthropic quota from the same pool as your session. Only after free lanes failed twice on the same task with different contracts.
5. Native. You. Last.

Keeping the roster honest

python scripts/delegate.py --list-lanes      # current capacity + staleness date
python scripts/lane_probe.py                 # re-measure; run weekly or after any 401/404

Lanes die without warning. Between 2026-08-04 and 2026-08-15, Qwen and DashScope both went to 401 and OpenRouter retired four :free tiers. scripts/lanes.json carries an explicit expiry for that reason - if it is stale, re-probe rather than trusting it.

The failure mode that will bite you: free reasoning models spend their whole token budget on hidden reasoning and return empty content. That is not a refusal and not a broken lane. delegate.py sends reasoning:{effort:low} to OpenRouter automatically, and must never send it to Google, NVIDIA, Cerebras or Groq, which reject it. See traps in scripts/lanes.json.