Campaign runner: question-level autonomous orchestration layer (/campaign) above /issue

Built and merged 2026-06-10 (merge aaa510ce4, branch worktree-agent-a263b3c688c2d5b9d). Design approved in chat after a literature dive over Jun-2025–Jun-2026 autonomous-research systems (AIRA_2, Kosmos, ASI-Arch, Anthropic recursive-self-improvement study, Karpathy autoresearch, METR time-horizon data).

What it is

A campaign is a kind: campaign task pinned to ONE docs/open_questions.md anchor. A dedicated session runs /campaign <N> and loops: ingest landed child clean-results (critic-gated bodies only) into a writable world model (artifacts/world-model.md, claim → #task evidence) → check stop criteria → propose next experiments (info-gain per GPU-hour, shared-nothing arms for max parallelism) → file + spawn children as ordinary /issue <child> --auto sessions with the full adversarial-planner + critic stack → repeat.

Key properties

• Human gates: brief approval IN (executes only from status approved; gate campaign_brief_approval), daily digest + final report OUT. Per-child promotion stays user-only but never blocks the campaign.
• Harness-side enforcement: budget (GPU-hours, never dollars), max-experiments, wall-clock, dry-counter, and campaign-level working-belief confidence target live in campaign_state.py + artifacts/campaign-state.json; the watcher's campaign pass (pass 9) adds respawn, progress-based stall detection, a budget backstop, and terminal reap.
• Crash-safe filing: state persisted (filed + child id + committed hours) BEFORE spawning; children carry <!-- campaign_experiment_id --> for deterministic orphan adoption; committed-hours reconcile is idempotent.
• Defaults: 250 GPU-h total, 100 per child, 8 experiments, 4 concurrent, 5 days — overridable per-brief (frontmatter > spawn-registry caps > module defaults).

Surface

src/explore_persona_space/campaign_state.py; task.py list-children + kind: campaign; spawn_session.py spawn-campaign; watcher campaign pass; .claude/skills/campaign/SKILL.md + campaign-tick; workflow.yaml (6 epm:campaign-* markers, gate id=16); CLAUDE.md routing amendment (campaign exception to same-issue-only auto-execution, user-approved 2026-06-10). 25 new tests; reviewed in 2 rounds (2 blockers fixed + verified closed).

Usage

1. Create: task.py new --kind campaign --title "..." --goal "<question>", write ## Campaign Brief (question anchor, hypotheses, initial experiment table, budget overrides).
2. Review + task.py set-status <N> approved.
3. spawn_session.py spawn-campaign --issue <N> — runs unattended; watch digests on the dashboard.