feat(campaign): experiment-campaign loop — controller skill, worker agents, kill-rule core

[sumtopmus]

What this adds

A Claude Code experiment-campaign loop for RL runs: a durable controller skill plus worker subagents that launch training, monitor W&B telemetry, early-stop bad runs, analyze results, design the next experiment, and continue — backed by one deterministic, unit-tested telemetry/kill-rule helper exposed as a read-only CLI command.

Implements docs/superpowers/plans/2026-06-08-experiment-campaign.md.

Components

Deterministic core (pure, unit-tested):

• core/campaign/killrules.py — evaluate(rows, rules, baselines) → KillVerdict; rule types nan / floor_at_step / collapse_vs_baseline / budget.
• core/campaign/telemetry.py — W&B history reader via an injectable api_factory (tests never touch the network). Fetches each metric key independently and merges by _step so an absent or step-disjoint metric can't zero out the fetch.

CLI:

• dsim campaign-status <run> [--rules f.json] — read-only telemetry + kill-rule verdict as JSON; the deterministic monitor tick. Invoked as python -m dsim … (repo sets package = false).

Agent layer (.claude/):

• run-campaign controller skill + experiment-analyst, experiment-monitor, experiment-coder worker agents.

Scaffolding & docs:

• docs/campaigns/TEMPLATE/{goal,ledger,frontier}.md, the smoke validation campaign (+ runbook), conf/experiment/smoke_canary.yaml, and a campaign-status entry in the README CLI surface.

Validation

Driven end-to-end on real W&B telemetry as the smoke campaign (interactive gate → launch → healthy monitor tick kill:false → forced abort kill:true/floor_at_step → analyst report card → ledger/frontier written). The smoke surfaced and fixed two real monitor-path defects, each with a regression test:

1. Objective metric never logged — the single-agent env emitted info['scored'] but SB3's EvalCallback keys on info['is_success'], so eval/success_rate (the campaign objective) was absent. Mirrored scored → is_success in envs/quidditch/simple_env.py.
2. Monitor was blind — telemetry.py requested a never-logged key in a single history(keys=…) call; wandb returns zero rows if any requested key is absent, silently neutering every kill-rule (kill:false). Fixed with the per-key fetch + merge described above.

Tests

• New: tests/core/campaign/*, tests/dsim/test_campaign_status_cli.py, tests/test_agent_assets.py, plus regression tests for both defects (test_simple_env_is_success.py, the absent-key telemetry test).
• Full fast suite + slow scoring-canary green on the author's machine. In a headless CI without an OpenGL/X11 display, the pre-existing test_render_smoke.py / test_video_callback_team.py abort on GL context creation — environmental, unrelated to this change.

Notes for reviewers

• The smoke trained but did not learn (success_rate ≡ 0.0 at lr 5e-5 / 200k steps) — expected; the smoke validated loop mechanics, not training quality. An lr-bump follow-up is queued in docs/campaigns/smoke/frontier.md.
• Clean fast-forward of the campaign work onto feature/experiment-campaign; the merge into develop was verified conflict-free via an in-memory trial merge.

https://claude.ai/code/session_018gE4JxQZ9dQSjuZJQGcm7R

---

Generated by Claude Code