swarm: DAG swarm workflows (declarative task graphs: harness kind, executor, eval)

[sethkarten]

DAG Swarm Workflows — one consolidated PR

From the spec: Swarm DAGs: declarative orchestration in Continual Harness (https://app.notion.com/p/3da72940136f-81a8-8554-e6ec7119270e). This consolidates #2397 (harness kind + validator), #2401 (executor), and #2402 (capability eval) into a single review and merge unit, per the review plan.

What you get

1. The swarm harness entry kind — a versioned DAG of subagent nodes: stable ids, subagent references (harness entries or inline specs), task/resident lifecycles, typed input/output ports with from bindings, per-node and run budgets, retries, fail_fast/continue/escalate failure policies, and bounded foreach fan-out. Write-time dry run: validate_swarm_spec enforces 12 rule groups (ids, cycles over effective dependencies, port type matching, resident constraints, foreach bounds, budget sanity, 1024-node cap) — an invalid DAG never reaches the store.

2. The executor — rlm.swarm.run/status/stop/resume: nonblocking admission (re-validate, resolve references, start only ready nodes, end the model turn), then a kernel control loop resuming via bounded collect polls. Binds typed outputs into dependent prompts (json ports via fenced blocks), foreach fan-out with max clamp, retries, rate-limit backoff (respects the 32-concurrent glm-5.3-fast pool), injectable-clock budgets, failure policies (fail_fast cancels in-flight siblings; escalate pauses with one quiet notice), cancellation cascade, an event ledger with arrived/shown/read stages, and one swarm.progress notice per milestone — same injection path as bash.completed. The executor owns run state in kernel memory; the supervisor owns the children. asyncio imports are deferred to preserve the repl boot contract.

3. The capability evaluation harness — three reference swarms (review sweep with typed fan-in + escalation, N-wide builder with per-node budgets, resident watcher with teardown) vs hand-written baselines under identical inputs/model/budgets; deterministic replay + ledger checks, spec verdict rules, metrics report. Spends real tokens only when run deliberately; never in CI.

Review history baked in

• Executor round: two must-fix bugs (foreach mixed-instance policy bypass; stop() race admitting/finalizing after stop) — both fixed with regression tests and verified by repro re-runs.
• Eval round: replay seq-gap detection, retry settle accounting, baseline tautology removal, single-listing baseline fairness, fail-fast --swarms validation — all fixed with tests.
• Gates on this exact tree: 122 + 90 Python, 134 + 52 TS tests pass, tsgo and biome clean, import rlm boot contract verified. Only the two pre-existing test_bash failures remain anywhere (identical on clean main).

Merge-time note (reviewer decision)

Main's test-line budget (check:test-policy) fails this PR by design: the feature is contract-test-heavy (~1.8k net test lines over source) — the suites that caught two executor bugs and eight eval findings. The repo's reviewed-override mechanism lives on the unmerged security/release-hardening branch; if it lands first, register the overage there, or grant the exception however you prefer. No test coverage was deleted to fit the budget.

Out of scope

The communication series (#2351–#2356) is deliberately separate and deferred; this PR depends only on main. Supersedes and closes #2397, #2401, #2402.

Draft only — review withheld per workflow.

Note

Add DAG swarm workflows with executor, harness integration, and evaluation

• Adds swarm.py with validate_swarm_spec, canonicalization, topological ordering, and a lifecycle API (run, status, stop, resume) for executing declarative task graphs
• Registers swarm as a new HarnessKind in harness.py with CRUD methods that validate the DAG on create and update; exposes rlm.swarm.* operations via the rlm namespace
• Integrates swarm progress into AgentSession via a swarm.progress host handler and createSwarmProgressMessage in messages.ts, injecting progress notices with admission-pause tolerance
• Extends refinement in refinement.ts to accept swarm entries, requiring a non-null object DAG on create and update edits
• Adds swarm-dag-eval.ts, a CLI evaluation harness with reference swarms, deterministic ledger replay, verdict computation, and markdown reporting; includes scripts/ in tsconfig compilation
• Behavioral Change: startsAgentRun now returns true for swarm progress custom messages; emptyHarnessState and corrupt-state recovery now include a swarm entry collection

^{Macroscope summarized ea2a7e6.}

[github-actions]

Prime Agent performance — running

PR ea2a7e65 compared with main ff40ea24.

Benchmarking the latest PR commit. Results will appear here when this run finishes.

Run, logs, and downloadable raw results

[sethkarten]

Reverted: the user wants the DAG work as a STACKED PR series, not a single PR. Reopening #2397 -> #2401 -> #2402 as the review stack; this consolidated branch will be deleted.