messageboardbench/AGENTS.md

# Working in messageboardbench

This is the canonical workspace for the experiment. Read `EXPERIMENT.md` for the
current question, completed results, limitations and next work. The old
`../archive/messageboard-20260908` repository is retired; its plans and instructions
are historical. Do not create experiment files there or in a sibling scratchpad.

- Use Codex subagents for analysis and any substantial, independently delegable
  work: literature review, trajectory review, artifact validation, implementation
  and review. Assign clear ownership; the primary agent checks and integrates.
- These are internal research assistants, not experimental agents. Do not change
  the experimental population to include subagents without PJ requesting it.
- Keep root files limited to project entry points and configuration. Put reusable
  code in `src/` or `scripts/`, tests in `tests/`, maintained explanations in `docs/`,
  retained raw runs in `logs/`, and reviewed evidence in named `results/` bundles.
- Use ignored `work/` for disposable agent working files. Never use a user's
  personal notes or scratchpad for generated outputs. Promote useful work into an
  existing documented location; do not leave competing PLAN/STATUS documents.
- Treat migrated `results/` bundles as frozen evidence. Preserve transcript,
  artifact and provenance bytes, including historical paths. Use the migration
  manifest to resolve old paths; write recomputations to fresh output directories.
- Use the existing `.venv/bin/python` and the documented `just board*` workflow.
  `muse` means `openrouter/meta/muse-spark-1.3-contributor`; no silent model fallback.
- Keep source, Python execution, credentials, logs and results on this workstation.
  For every Docker-backed check or experiment, use only the remote x86-64 Docker
  daemon at `ssh://[email protected]`; do not copy or stage the repository on that host.
- Do not launch paid experiments merely to test code or reorganize files. Use
  offline tests and mock models. Existing user authorization governs actual runs.
- Exploratory experiments are one-off research code. Optimize for the first useful
  result, not production quality.
- Use simple, direct Python. Do not add frameworks, abstractions, schemas, manifests,
  provenance systems or reusable orchestration unless PJ explicitly requests them.
- Do not add tests, validation gates, preflight matrices, package builds or production
  hardening to an exploratory experiment unless PJ explicitly requests them. Let the
  run fail and debug it from the raw log.
- `just start` must start the experiment immediately and stream its logs. It must not
  validate populations, rebuild packages, screen tasks or run tests first.
- Never start, restart, resume or stop an experiment. PJ exclusively controls
  experiment processes and runs the launch commands. Prepare and verify commands,
  then hand them to PJ to execute.
- Treat infrastructure retries as attempts within the original experiment, not as
  new experiments with separate `experiments/` folders. Keep their working output
  under the original experiment's log namespace and promote successful scored
  attempts into its core results after the run completes.
- For diagnosed infrastructure failures, record the affected task IDs, attempt
  counts and error categories in the original experiment's existing recovery
  record. Once recovery is complete, remove failed retry artifacts and redundant
  retry folders rather than retaining full failed traces solely for provenance.
  Never delete artifacts from an active run. Do not put infrastructure-only retry
  diagnostics in `analysis.md`.
- Treat process preferences and future-work instructions as prospective, not as
  permission to change current work. If their timing is unclear, ask PJ explicitly
  before editing, staging, committing or running anything.
- Distinguish publication, delivered peer receipt, attributed adoption, rejection,
  and independent discovery. A passing contradictory test or similar code alone
  does not establish transmission, concealed collusion, or monitor evasion.
- After analyzing a completed experiment, write the maintained interpretation to
  lowercase `analysis.md` inside that experiment's `experiments/<id>/` folder. Keep
  automatic output distinct from reviewed behavioral conclusions and link the exact
  evidence used.
- Prefer a clean code/config commit before launching each new experiment. After the
  run and analysis, commit the associated code, frozen configuration, documentation,
  and `analysis.md` after checking for secrets and accidental raw logs. Record the
  commit hash in the handoff. The executed source snapshot remains authoritative if
  code changed during or after the run.