dsl-campaign

name	dsl-campaign
description	Run ONE autonomous ability-DSL authoring campaign — inquisitor picks a ≤40-ability worklist from the sub-0.80 roundtrip corpus, the 16-agent suite (12 core + 4 kroot shape-scout) authors and adversarially verifies it, and the campaign closes as a draft PR. Use for "run a dsl campaign", "launch the authoring loop", "next dsl-campaign round". Must run from the jj workspace /Users/will.mitchell/40kdc-dsl (agents register at session start).
argument-hint	[--worklist-cap N] [--dry-run] [optional targeting bias, e.g. 'faction: orks' or 'shape: fight-eligibility-extension']

Skill: dsl-campaign

One launch = one campaign: prioritize → author in batches → apply+verify → close as a draft PR. The loop never self-terminates; relaunching is the human heartbeat, PR review is the second checkpoint. This file is the driver contract — a fresh session must be able to run a campaign from it alone.

Companion files: workflows/wf-prioritize.js, workflows/wf-author-batch.js, workflows/wf-verify-batch.js, and workflows/wf-shape-scout.js (invoke via Workflow({scriptPath, args}); each embeds the frozen agent Output contracts as JSON Schemas — never redesign those). Always pass repo_root: "/Users/will.mitchell/40kdc-dsl" in every workflow's args — subagents inherit the driver session's cwd, and the preamble it generates pins them to the workspace even when the driver was launched from another checkout. The worked example of one converged campaign is _private/loop-state/{roundtrip,inbox}-world-eaters.md.

Preconditions (fail loudly if unmet)

• cwd is /Users/will.mitchell/40kdc-dsl (jj workspace dsl). The 16 agents in .omp/agents/ (12 core + 4 kroot shape-scout) register only at session start — if data-enginseer isn't an available agent type, tell the user to restart the session here; do not simulate agents. Nested agent-spawning (the kroot suite, arch-magos) needs task.maxRecursionDepth >= 2 in .omp/config.yml; wf-shape-scout throws spawn-unavailable and fails loud if it is unset.
• Sibling repos: raw-text store ../40kdc-abilities; embeddings harness ../40kdc-embeddings (its own .venv).
• Toolchain in THIS workspace: tools/node_modules + tools/dist, python/.venv, target/release/wh40kdc-runner, go/wh40kdc-runner. Rebuild all three runners after ANY source edit before parity checks — stale runners give phantom verdicts.
• jj st clean apart from _private/ (reconcile, never clobber, if not); jj new before any work; never touch other workspaces' commits (<name>@); never move main.

Non-negotiables (pinned decisions)

• Prose is authoritative; placeholder lies are banned. An honest worse fit + a resisted_schema inbox block always beats a plausible wrong encoding.
• Canonical condition ids are cruncher levers and outrank cosine. Cogitator regressed fails the batch regardless of score gains.
• New shapes need tortured-fit + family evidence — the kroot shape-scout (wf-shape-scout.js) designs them: flesh-shaper must prove every nearest existing shape flattens the mechanic, and lone-spear faithful_family_size ≥ 4 (exact+near, flatten-excluded) gates the ship. Never flatten onto an over-similar neighbour.
• Cosine selects and measures; it never gates. The embeddings harness is advisory.
• warpsmith is the only agent that writes repo files. The driver writes loop-state, runs jj/gh. Workflow agents are read-only.
• IP boundary: GW prose may transit agent JSON, harness reports, and the session scratchpad; it must NEVER be written into any file inside this repo — including community_notes, [APPROX] notes, inbox entries, PR bodies, and commit messages (own-words paraphrase only). Breach ⇒ campaign abort + jj undo.

Stop condition

• Ability converged = eversor panel has no un-rebutted refuted:true ∧ skitarius overall_pass:true ∧ cogitator "clean" ∧ no psyker severity-3 ∧ inquisitor review accept ∧ re-scored cosine ≥ start (or a recorded correctness-first justification).
• Campaign done (draft PR opens) = every worklist entry has a terminal status — converged / improved (skeptic PASS) / needs-schema (inbox block filed) / abandoned (reason recorded) — ∧ full gates green at head ∧ whole-dataset prose diff touches only worklist ids ∧ faction mean not regressed ∧ draft PR opened.

Does NOT count as done (ten hard rejects — inquisitor enforces per batch AND at close)

1. Placeholder lies — valid DSL encoding a different mechanic.
2. Cosine-chasing lever drops — e.g. charged-this-turn → timing-is charge-move.
3. APPROX-stuffing — clauses evacuated into [APPROX] notes to dodge refutation.
4. needs-schema as escape hatch when an honest existing-shape fit exists.
5. Weakened verification — panel <2 voters, goldens loosened/deleted, unrun gates reported passed, "pending" counted as PASS.
6. Silent worklist shrinkage — any entry without a terminal status.
7. Collateral render drift — non-worklist describer output changed.
8. Invented metadata — applies_to on army-wide rules (pinned no-highlight), keyword gates or grant names not in prose.
9. IP breach — GW prose in any repo file ⇒ abort.
10. Unjustified cosine drop — ending below cos_start without a recorded correctness-first justification (the relentless-rage pattern).

Caps (bounded by construction)

Worklist ≤ 40 (default ~30; --dry-run caps at 5). ≤ 4 assembly attempts per ability, then forced terminal status. Eversor ≤ 3 voters (2 routine; 3 when new shape, arch-magos confidence < 0.7, or prior reject). ≤ 2 full gate re-runs per batch. Batch grain 5–6 (wf-author-batch hard-fails > 8). One campaign per launch. Abort (registry aborted, clean stop, never dressed as convergence): IP breach, gate failure surviving 2 re-runs, blocking escalation.

Shape-scout (wf-shape-scout.js): ≤ 3 cyclical review rounds per shape, then forced terminal; the family bar is lone-spear faithful_family_size ≥ 4 (exact+near, flatten-excluded); kroot leads spawn leaf helpers only (spawn tree depth ≤ 2).

Procedure

0 — Preflight

cd /Users/will.mitchell/40kdc-dsl
jj workspace update-stale 2>/dev/null; jj st   # reconcile surprises; never clobber
jj new -m "wip: dsl-campaign cNNN"             # NNN = next id from registry.json

Refresh the full-corpus scores (fast when embedding cache is warm):

cd /Users/will.mitchell/40kdc-embeddings && .venv/bin/python -m wh40kdc_embeddings \
  roundtrip --faction all --scope all --enrichment-dir /Users/will.mitchell/40kdc-dsl/data/enrichment

Snapshot _reports/roundtrip-all.json to the session scratchpad as prose-baseline.json — this is the whole-dataset prose-diff baseline (anti-condition 7): at close, a fresh --faction all run's describer outputs are diffed against it and any changed non-worklist id fails the campaign. (Coverage is store-paired abilities; the drift/conformance gates cover the goldens beyond it.)

1 — Prioritize

Compute a sub080_summary from the report (per faction: mean, count below 0.80, worst ~15 ids+scores). Then:

Workflow({ scriptPath: ".omp/skills/dsl-campaign/workflows/wf-prioritize.js", args: {
  scout_shapes: [ …recently shipped shapes + any user bias… ],
  artifacts: { roundtrip_report_path, sub080_summary, loop_state_paths, registry_excerpt },
  worklist_cap: 30 } })

From curation.priorities, materialize the worklist (ability_id, faction_id, cos_start, prior_reject from the ledger). Honor a user targeting bias from $ARGUMENTS as round-1 priority. Then: append the campaign entry to _private/loop-state/registry.json (status:"open", mean_before from the report), create roundtrip-<target>.md in the world-eaters table format, and file any escalate_to_user items into registry.escalations. If an escalation blocks target choice, stop and ask the user.

2 — Author (per batch of 5–6)

Workflow({ scriptPath: ".omp/skills/dsl-campaign/workflows/wf-author-batch.js",
  args: { batch_id: "cNNN-bK", new_shapes: […], abilities: […5–6 worklist entries…] } })

Statuses back: accepted → apply (step 3). needs-schema → step 2a (family check, then shape-scout or inbox). rejected → adjudicate: spawn inquisitor (mode:"review", the candidate + verdicts as agent_outputs); accept overrides a bad refutation (record why), otherwise terminal abandoned with reason — or needs-schema (→ step 2a) if the divergences show the schema can't express it. no-prose / agent-error → terminal abandoned (reason recorded). Each result's thread is the FULL per-round revision history (every attempt's divergences, not just the last) — record it in the ledger/roundtrip notes so cyclical revisions stay auditable.

2a — Shape-scout on a needs-schema result

A needs-schema is not automatically terminal. Spawn inquisitor to judge whether the resisted mechanic is a FAMILY (≥ ~4 abilities, exact+near) or a singleton:

• Singleton → append the resisted_schema block to inbox-<faction>.md (own words), ledger needs-schema; committed entry untouched (never a placeholder). Terminal.
• Family → fork to the shape-scout, seeding the resisted_schema block:

  Workflow({ scriptPath: ".omp/skills/dsl-campaign/workflows/wf-shape-scout.js", args: {
    repo_root: "/Users/will.mitchell/40kdc-dsl",
    seed: { ability_id, faction_id, raw_text, resisted_schema }, family_threshold: 4 } })
  

  The kroot suite (flesh-shaper → lone-spear → trail-shaper → war-shaper, each spawning its OWN helpers) returns status:
  • shipped-ready (war-shaper accept ∧ faithful_family_size ≥ threshold) → auto-apply: hand shape_package to warpsmith (implement) to land the schema oneOf branch + all four describer ports + conformance cases + SPEC bump + version lockstep, then re-author the seed AND every faithful_family member onto the new shape as accepted candidates (step 3). The campaign is now shape-led — expect a heavier PR, and the full close gate (step 4) must cover the port/SPEC work.
  • existing-fits → the scout proved an existing shape fits after all; re-enter step 2 authoring with that shape named (the resist was a false alarm).
  • rejected-sprawl / rejected-singleton / not-converged → file the inbox block, ledger needs-schema (terminal); record the scout's reason. A thrown spawn-unavailable means nested spawning is off — set task.maxRecursionDepth >= 2 in .omp/config.yml and restart; never hand-simulate the kroot agents (that is anti-condition 5, weakened verification).

3 — Apply + verify (per batch)

1. warpsmith applies the accepted candidates (Agent tool, sole writer): implement decisions referencing each candidate's dsl JSON; it edits data/enrichment/<faction>/abilities.json only.
2. Regenerate + rebuild what the edit touches (TS bundle at minimum: cd tools && npm run build); rebuild runners if any source changed.

3. Workflow({ scriptPath: ".omp/skills/dsl-campaign/workflows/wf-verify-batch.js",
     args: { batch_id, ability_ids, faction_ids, touched_files } })
   

   Fail the batch on !overall_pass (≤2 re-runs after fixes), verdict:"regressed" (drop/fix the offending ability — levers outrank cosine), or any severity-3 psyker finding. Psyker missing-clause-signal routes to inquisitor as fidelity, not to warpsmith as wording.
4. Re-score just the batch:

   cd /Users/will.mitchell/40kdc-embeddings && .venv/bin/python -m wh40kdc_embeddings \
     roundtrip --faction <f> --ids <id,id,…> --scope batch-cNNN-bK \
     --enrichment-dir /Users/will.mitchell/40kdc-dsl/data/enrichment
   

   Any ability below its cos_start needs a recorded correctness-first justification or another attempt (within its 4-attempt budget).
5. Inquisitor batch review (mode:"review", arch-magos outputs + verify results + scores). revise re-enters step 2 for that ability; reject → terminal per step 2.
6. On accept: jj commit -m "feat: dsl-campaign cNNN batch K — <target>", update roundtrip-<target>.md + registry.json (statuses, cos_best, attempts).

4 — Close

1. All worklist entries terminal (anti-condition 6) — else keep batching.
2. Full gate: PATH="$PWD/python/.venv/bin:$PATH" just regen fmt test-all version-lockstep in the workspace, plus parity differ with freshly rebuilt runners. Drift-check by committing regen output and confirming jj st clean.
3. Prose diff: fresh --faction all run vs prose-baseline.json; any non-worklist id whose describer output changed ⇒ anti-condition 7, fix before closing.
4. Inquisitor close-out (mode:"review" over the campaign summary): re-checks all ten anti-conditions campaign-wide; faction mean ≥ mean_before.
5. Ship: jj bookmark create wnmitch/dsl-cNNN-<slug> -r @- (the batch stack), jj git push --bookmark wnmitch/dsl-cNNN-<slug> --allow-new, gh pr create --draft — PR body: own-words summary, worklist table (ids, statuses, cos start→best), justifications for any cosine drops, inbox blocks filed, escalations. No GW prose, no personal info.
6. Persist: registry entry → converged (+pr, mean_after, finished); memory_store (tags: 40kdc-data, dsl-loop) one-paragraph campaign summary.
7. Report to the user: PR link, mean before→after, statuses, escalations, and anything blocked_shapes gained. If aborted instead: exact reason, what converged first, registry state — never dress an abort as convergence.

Registry (_private/loop-state/registry.json)

Machine source of truth; driver-written only; gitignored (never rides the PR). campaigns[] {id, kind: faction-led|shape-led|inbox-led|describer-led|shape-scout, target, bookmark, status: open|converged|aborted, pr, worklist_size, mean_before, mean_after, started, finished, notes} · blocked_shapes[] {proposal, why, reopen_when} — dedup source for swarmlord/warpsmith proposals (full seen-set; reopen only when reopen_when is met with cited new evidence) · ability_ledger {"faction/id": {status, campaign, cos_start, cos_best, attempts, justification}} · escalations[] {question, raised_by, campaign, resolved}.

Field notes

• Cull/stop driver-spawned agents as soon as their output is verified.
• execSync-based repo tools resolve path args against the repo root, not shell cwd.
• The dump/store/report debugging rule applies to prose lookups: before concluding an ability "has no prose", make data-enginseer show the failing grep, not the theory.
• A dry run (--dry-run) runs the identical procedure with worklist_cap 5 and should include one known-hard multi-clause aura ability to probe the eversor floor.
• The author loop records the FULL revision thread (every attempt's panel divergences, not just the last round); it feeds forward on revision and into loop-state, so a cyclical fix never silently re-breaks an earlier round's divergence.
• On a needs-schema FAMILY, the kroot shape-scout designs a distinct new shape — proving the nearest existing shapes flatten it — rather than reaching for an over-similar neighbour (the necron obelisk-vs-tau collision is the defect it prevents).