fix(eval): drive CLI-harness agents (claude/codex) with the prompt at launch#126
Merged
Conversation
… launch The seg harness drove every agent via launch-empty-then-`session send` — the opencode *server* model (`/prompt`). But claude/codex are one-shot CLI harnesses (`agent -p PROMPT`): the prompt must be at LAUNCH, and a later `session send` can't drive a one-shot. So an `AGENT=claude` monolithic run booted a VM that never ran a turn → all cells scored 0.0 at $0 cost (real boot time, no work). Verified via a controlled run: `run --detach --agent claude -- "<prompt>"` → wait-idle → the work lands (model=claude-opus-4-8). Fix: - `pb_run_session` takes an optional launch prompt; for a CLI-harness agent it passes `--detach -- "$prompt"` (server agents unchanged — launch idle, drive via send). - `run_monolithic_cell` branches: opencode → launch idle + `drive_bounded`; CLI → launch-with-prompt + wait-idle (no send). Backwards-compatible (opencode path byte-identical). Scope: the monolithic arm. The chained/segmented arms still send per-segment (a one-shot CLI agent can't do multi-turn chaining anyway — a deeper question, left for later). NOTE (follow-up, not fixed here): `dispatch`'s WorkerDriver drives the same detach-then-send way, so `dispatch --agent claude` (the default) likely has the same can't-drive bug and needs the same launch-with-prompt treatment.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The seg harness drove every agent via launch-empty-then-
session send(opencode's server model). claude/codex are one-shotagent -p PROMPT— the prompt must be at launch, and a latersendcan't drive a one-shot. SoAGENT=claudebooted a VM that never ran a turn → all cells 0.0 at $0 cost.Verified live:
run --detach --agent claude -- "<prompt>"→ wait-idle → work lands (modelclaude-opus-4-8). Fix is scoped to the monolithic arm; opencode path byte-identical.This unblocked the Opus bar for the cost-quality comparison — Opus solo 0.880 beats the cheap heterogeneous best-of-3 fanout 0.819 (win concentrated on the hard
ap_dot_dsl: 0.639 vs 0.525), and Opus is ~free on subscription vs the fanout's real $0.264/run. Verdict: just use Opus.Follow-up (flagged, not fixed):
dispatch's WorkerDriver drives the same detach-then-send way, sodispatch --agent claude(the default) likely can't drive claude either — needs the same launch-with-prompt fix.