Step 6d.1 experimenter briefs must carry plan §9 parallelization shape — verbatim workload_cmd reuse silently runs shard 0 only on CVD-pin plans
Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.
No comments yet.
Goal
Close the experimenter-brief composition gap that reproduces shard-0-only launches on launcher-CVD-pin plans: when the /issue orchestrator (Step 6d.1) composes an experimenter brief by copying the dispatch handle's persisted workload_cmd verbatim, and the plan's §9 mandates launcher-env CUDA_VISIBLE_DEVICES pins with per-shard --shard-id invocations, the verbatim command is the plan fence's dry-parse form (--num-shards K, no --shard-id) and silently runs only shard 0 on 1 GPU of a K-GPU pod.
Incident
Task #2254, follow-up round first-k-answer-token-steering, 2026-08-23 (~10:50Z launch). The orchestrator's experimenter brief carried the handle's verbatim workload_cmd (uv run python scripts/issue2254_first_k_steering.py --phases stage_inputs,steer --num-shards 4). Executed as-is on pod-2254 (4× H100) this would have run 40/160 cells on one GPU while three H100s idled. The experimenter caught the drift against plan v10 §9 (lines 231/278 — launcher-env CVD pins, --shard-id {0..3} --num-shards 4 in parallel after a serial stage_inputs), composed /workspace/launch_issue_2254_firstk.sh, and stated the correction in its epm:progress marker (#2254 events v86). No compute was lost — but only because the subagent independently re-derived the plan's parallelization shape.
Fix shape (for the planner/implementer to refine)
The workflow surface at fault is the orchestrator-side brief-composition guidance (.claude/skills/issue/steps/ Step 6d.1 experimenter-brief text, plus any brief-template prose that says "launch the persisted workload_cmd"). The brief composer MUST carry plan §9's realized parallelization line (shard axis, width, CVD-pin form, serial-vs-parallel phase split) alongside the cmd=, and instruct the experimenter that the handle's workload_cmd is the FENCE DRY-PARSE form — the launcher composition is governed by §9, not by the persisted string. Consider: (a) a one-line duty in Step 6d.1 ("brief carries §9's parallelization block verbatim; verbatim workload_cmd reuse on a multi-shard plan is the #2254 shard-0 drift"); (b) optionally, dispatch_issue.py persisting a structured parallelization field in the handle sidecar so briefs can quote it mechanically.
Provenance
Surfaced as a prose workflow-surface observation by the experimenter subagent on task #2254 (round first-k-answer-token-steering, launch report 2026-08-23). Auto-filed by the #2254 followup orchestrator per .claude/rules/workflow-fix-on-bug.md (surfaced-prose follow-ups clause).