/campaign — Autonomous Improvement Campaigns

A campaign is a durable, named, repeatable improvement run. It replaces the hand-authored standing prompt + ~/.claude plan files: a run becomes start(config). The campaign owns its scope/posture/decision-authority/stop-conditions, the Ralph loop(s) it drives, a decision log, an async parked-questions queue, and a progress ledger. /campaign run is the Ralph-pattern loop body — drive it repeatedly with /loop /campaign run <id>.

Dual surface — platform agents call the same actions through the MCP tools platform.campaign_start, platform.campaign_status, platform.campaign_answer_question, platform.campaign_stop and platform.campaign_resume.

Usage

/campaign start <name>     # create a campaign + its dedicated loop (interactive config)
/campaign run [<id>]       # drive ONE improvement iteration for the campaign (loop body)
/campaign status [<id>]    # progress + open questions + recent decisions (all, or one)
/campaign answer <id>      # answer a parked question to unblock the campaign
/campaign stop <id>        # stop the campaign + pause its loops + record a summary
/campaign resume <id>      # request a resume (human-confirmed; parks until a person approves)

start <name>

Surface the config, then create. Confirm these with the operator if not given:

• scope — which tree/area (e.g. a private extension, server/, worker/); stays in configuration.
• posture/ordering — bugs-first, dead-code, test-gaps, quality; free text in configuration.
• decision_authority — how much the loop decides vs parks (see table below). Default trusted.
• stop_conditions — e.g. { "max_failed": 3, "completion_pct": 100 }. Optional.

Call platform.campaign_start(name:, description:, decision_authority:, configuration:, stop_conditions:). It creates the Ai::Campaign and a dedicated campaign-scoped Ralph loop (branch campaign/<id>), starts the campaign, and snapshots initial progress. Report the campaign id + loop branch.

run [<id>]

One in-scope improvement, end to end — this is the loop body (compose, don't reinvent):

1. Read the campaign: platform.campaign_status(<id>) (or the most recent active campaign if id omitted). Honour decision_authority, configuration scope/posture, and stop_conditions.
2. Stop conditions / halt: if the campaign is terminal, should_stop? is met, or an emergency_halt / kill-switch is active → report and STOP (end /loop if active).
3. Refill if dry: if the campaign loop has no queued task, discover within scope — /improve discover <scope> (verify-before-offer gate), approve the next in-scope offer (individually, never batch — CLAUDE.md bulk rule), which promotes a task onto the campaign loop. If discovery is dry and nothing is pending → snapshot + report queue-empty + STOP.
4. Drive one task: run the /dev-loop <campaign-loop> iteration body on the campaign's loop (pull → re-verify claim → branch campaign/<id> → test-first red → minimal fix → verify → commit on the loop branch, STAGE-only, no push, no Claude attribution → report).
5. Record the decision: log what was decided (unblock/skip/build/remove/defer/policy) to the campaign so the operator sees the reasoning, not just the diff.
6. Park, don't stall: when blocked on something above your decision_authority (live-credential, irreversible external, business-policy/pricing value), park_question! on the campaign and move on or stop — never guess. Parked questions are answered async via /campaign answer.
7. Snapshot progress so the ledger + dashboard reflect the new state.

status [<id>]

platform.campaign_status(<id>) (or list active campaigns). Show completion %, task counts, open questions, and the recent decision log. Read-only.

answer <id>

List the campaign's open questions (status), then platform.campaign_answer_question(<id>, question_id:, answer:). The answer unblocks the next /campaign run for that area.

stop <id>

platform.campaign_stop(<id>, summary:) — pauses the campaign's loop schedules and marks the campaign completed with your summary. Use when scope is drained or the operator calls it.

resume <id>

Resumes a stopped or auto-completed campaign, optionally with updated stop_conditions (for example a higher max_failed). The action is human-confirmed: platform.campaign_resume(<id>, stop_conditions:) from any MCP caller parks as a pending approval, and a person confirms it from their own session (the approval queue, the campaign's Resume button, or POST /api/v1/ai/campaigns/:id/resume). The requester's MCP client cannot confirm its own request. Stop values must be valid for their type, and a resume never removes a stop condition.

Decision authority

Posture (inherits the loop body's rules)

• STAGE-only during a run — commit on the campaign branch, never push/land unless the operator says so.
• Test-first bug reproduction; minimal change tracing to the task; no scope creep.
• Core-purity: never make a core file depend on a private extension; extension findings stay extension-scoped.
• Crypto-material safety absolute; bulk-op confirmation for >5 items; 3-strikes → stop & ask.
• One improvement per run. Park genuine forks rather than guessing.

See autonomous-campaigns convention (tag guidance-autonomous-campaigns) for the model/driver/tool reference and the rationale.