experiment-bridge skill

Workflow 1.5: Experiment Bridge

Implement and deploy experiments from plan: $ARGUMENTS

Overview

This skill bridges Workflow 1 (idea discovery + method refinement) and Workflow 2 (auto review loop). It takes the experiment plan and turns it into running experiments with initial results.

Workflow 1 output:                    This skill:                                    Workflow 2 input:
refine-logs/EXPERIMENT_PLAN.md   →   implement → GPT-6-Astra review → deploy → collect → initial results ready
refine-logs/EXPERIMENT_TRACKER.md     code        (cross-model)    /run-experiment     for /auto-review-loop
refine-logs/FINAL_PROPOSAL.md

Constants

• CODEREVIEW = true** — GPT-6-Astra xhigh reviews experiment code before deployment. Catches logic bugs before wasting GPU hours. Set false to skip.
• AUTODEPLOY = true** — Automatically deploy experiments after implementation + review. Set false to manually inspect code before deploying.
• SANITYFIRST = true** — Run the sanity-stage experiment first (smallest, fastest) before launching the rest. Catches setup bugs early.
• MAXPARALLELRUNS = 4 — Maximum number of experiments to deploy in parallel (limited by available GPUs).
• BASEREPO = false** — GitHub repo URL to use as base codebase. When set, clone the repo first and implement experiments on top of it. When false (default), write code from scratch or reuse existing project files.
• COMPACT = false — When true, (1) read idea-stage/IDEACANDIDATES.md instead of full idea-stage/IDEAREPORT.md if available, (2) append experiment results to EXPERIMENT_LOG.md after collection.

Override: /experiment-bridge "EXPERIMENT_PLAN.md" — compact: true, base repo: https://github.com/org/project

Inputs

This skill expects one or more of:

1. refine-logs/EXPERIMENTPLAN.md** (best) — claim-driven experiment roadmap from /experiment-plan
2. refine-logs/EXPERIMENTTRACKER.md** — run-by-run execution table
3. refine-logs/FINALPROPOSAL.md** — method description for implementation context
4. idea-stage/IDEACANDIDATES.md — compact idea summary (preferred when COMPACT: true) (fall back to ./IDEACANDIDATES.md if not found)
5. idea-stage/IDEAREPORT.md — full brainstorm output (fall back to ./IDEAREPORT.md if not found)

If none exist, ask the user what experiments to implement.

Workflow

Phase 1: Parse the Experiment Plan

Read EXPERIMENT_PLAN.md and extract:

1. Run order and milestones — which experiments run first (sanity → baseline → main → ablation → polish)
2. For each experiment block:

• Dataset / split / task
• Compared systems and variants
• Metrics to compute
• Setup details (backbone, hyperparameters, seeds)
• Success criterion
• Priority (MUST-RUN vs NICE-TO-HAVE)

1. Compute budget — total estimated GPU-hours
2. Method details from FINAL_PROPOSAL.md — what exactly to implement

Present a brief summary:

📋 Experiment plan loaded:
- Milestones: [N] (sanity → baseline → main → ablation)
- Must-run experiments: [N]
- Nice-to-have: [N]
- Estimated GPU-hours: [X]

Proceeding to implementation.

Research-contract fallback: if idea-stage/docs/researchcontract.md does not exist yet (idea selected outside /idea-discovery, or an older run), create it now from templates/RESEARCHCONTRACTTEMPLATE.md using the selected idea + claims from the experiment plan. Downstream /result-to-claim and /ablation-planner read this file as the claims source, and session recovery (docs/SESSIONRECOVERY_GUIDE.md) depends on it existing.

Phase 2: Implement Experiment Code

If BASEREPO is set** — clone the repo first:

git clone <BASE_REPO> base_repo/
# Read the repo's README, understand its structure, find entry points
# Implement experiments by modifying/extending this codebase

For each milestone (in order), write the experiment scripts:

1. Check existing code — scan the project (or cloned base_repo/) for existing experiment scripts, model code, data loaders. Reuse as much as possible.

1. Implement missing pieces:

• Training scripts with proper argparse (all hyperparameters configurable)
• Evaluation scripts computing the specified metrics
• Data loading / preprocessing if needed
• Baseline implementations if not already present
• Fixed random seeds for reproducibility
• Results saved to JSON/CSV for later analysis
• Proper logging (wandb if configured in CLAUDE.md)

1. Follow the plan's run order — implement sanity-stage experiments first, then baselines, then main method, then ablations.

1. Self-review before deploying:

• Are all hyperparameters from EXPERIMENT_PLAN.md reflected in argparse?
• Is the random seed fixed and controllable?
• Are results saved in a parseable format (JSON/CSV)?
• Does the code match FINAL_PROPOSAL.md's method description?

Phase 2.5: Cross-Model Code Review (when CODE_REVIEW = true)

Skip this step if CODEREVIEW is false.**

Before deploying, send the experiment code to GPT-6-Astra xhigh for review:

mcp__codex__codex:
  model: gpt-6-astra
  config: {"model_reasoning_effort": "xhigh"}
  prompt: |
    Review the following experiment implementation for correctness.

    ## Experiment Plan:
    [paste key sections from EXPERIMENT_PLAN.md]

    ## Method Description:
    [paste from FINAL_PROPOSAL.md]

    ## Implementation:
    [paste the experiment scripts]

    Check for:
    1. Does the code correctly implement the method described in the proposal?
    2. Are all hyperparameters from the plan reflected in the code?
    3. Are there any logic bugs (wrong loss function, incorrect data split, missing eval)?
    4. Is the evaluation metric computed correctly?
    5. **CRITICAL: Does evaluation use the dataset's actual ground truth labels — NOT another model's output as ground truth?** This is a common and severe bug.
    6. Any potential issues (OOM risk, numerical instability, missing seeds)?

    For each issue found, specify: CRITICAL / MAJOR / MINOR and the exact fix.

    === SCOPE LIMITS (these bound what you PROPOSE, never what you look for) ===
    Report anything that is actually wrong here — including a rare-looking case, if
    this repo actually produces it. Then keep t

On review results:

• No CRITICAL issues → proceed to Phase 3
• CRITICAL issues found → fix them, then re-submit for review (max 2 rounds)
• Codex MCP unavailable → skip silently, proceed to Phase 3 (graceful degradation)

Phase 3: Sanity Check (if SANITY_FIRST = true)

Before deploying the full experiment suite, run the sanity-stage experiment:

/run-experiment [sanity experiment command]

Wait for completion. Verify:

• Training loop runs without errors
• Metrics are computed and saved correctly
• GPU memory usage is within bounds
• Output format matches expectations

If sanity fails → auto-debug before giving up. Budget: up to 2 patch attempts on the same failure, then up to 2 clean reimplements (4 total):

read-the-primary-artifact discipline applies to surprising REVIEWER verdicts: see shared-references/review-tracing.md § Debugging With Traces.)

1. Read the error — parse traceback, stderr, and log files. (The same

1. Diagnose — classify the failure:

• OOM → reduce batch size or enable gradient checkpointing
• ImportError → install missing package
• FileNotFoundError → fix path or download data
• CUDA error → check GPU availability, reduce model size
• NaN/divergence → reduce learning rate, check data preprocessing

Before the next retry, invoke /codex:rescue to get a second opinion on the root cause. Codex independently reads the code and error logs — it may spot issues Claude missed (wrong tensor shapes, subtle import shadowing, config mismatches, etc.). Apply its suggested fix, then re-run.

1. Fix and re-run — apply the fix, re-run sanity
2. Attempt 2+ still failing? → Call in Codex rescue (if Codex plugin installed):

• If /codex:rescue is not available (plugin not installed), continue with Claude's own diagnosis

(up to 2 reimplements). Rewriting the failing script from EXPERIMENTPLAN.md / the research contract is a PEER move to another patch, not a last resort — a third patch on top of two wrong ones is usually worse than a clean rebuild. Delete ONLY the attempt's own code/scaffolding (scripts this phase generated); the plan, EXPERIMENTTRACKER.md, user-authored project source, collected data, and results are never deletable (see shared-references/external-cadence.md § Let a broken attempt restart, not just patch).

1. Both patch attempts failed on the same failure? → Discard and reimplement cleanly

way?** → stop, report the failure with all attempted fixes and error logs. Two clean reimplements failing identically usually means the plan or the environment is wrong — say so explicitly in the report, because that (not the broken build itself) is what needs the human. Do not proceed with broken code.

1. **Budget exhausted (2 patches + 2 reimplements), or two reimplements failed the SAME

Never give up on the first failure. Most experiment crashes are fixable without human intervention.

Phase 4: Deploy Full Experiments

Deploy experiments following the plan's milestone order. Route by job count:

Small batch (≤5 jobs per milestone) → use /run-experiment directly:

/run-experiment [experiment commands]

Large batch (≥10 jobs, multi-seed sweeps, or phase dependencies) → use /experiment-queue for proper orchestration: