feat(loops): recursive execution atom — budget-conserving Scope + Supervisor keystone

[drewstone]

What

The v1 recursive execution atom — the keystone for drivers-of-drivers: one self-similar Agent whose act spawns child agents through a Scope, run by a Supervisor that owns a conserved budget pool, an event-sourced journal, and an observability/conversation handle. The flat experiment harness is recovered as the simplest act (it does not compete with this — it is one program over it).

Design + decision record: docs/research/recursive-execution-atom.md (frozen contract, build order, the 4 resolved forks, the adversarial critique that shaped the surface).

The pieces

• src/loops/supervise/types.ts — the frozen contract. Agent, an open LeafExecutor interface (execute returns a promise or an async stream; router/inline + sandbox + cli are implementations, a user's own agent is first-class via the registry/BYO — no per-vendor adapters), Scope, Supervisor, Settled, Budget.
• supervise/budget.ts — a conserved reservation pool: atomic reserve-on-spawn, fail-closed admission, refund-on-settle. This is the load-bearing invariant: Σk(treatment) ≡ Σk(blind) holds by construction, so a steered arm can never silently out-compute blind (the confound that burned the earlier "+20pp steering" result).
• src/durable/spawn-journal.ts — event-sourced SpawnJournal + content-addressed ResultBlobStore + seq-ordered replay. Resumable, queryable, reproducible from one log.
• supervise/scope.ts — a ray.wait cursor over an in-memory nursery; spawn reserves budget and resolves the executor through the open registry; a Settled → Iteration adapter keeps defaultSelectWinner single-sourced.
• supervise/runtime.ts — the open executor registry; the sandbox executor composes runLoop (forwarding an optional lineage passthrough) rather than reinventing checkpoint/fork.
• supervise/supervisor.ts — nursery join barrier, abort cascade (incl. the acquire lifecycle), OTP intensity breaker, typed SupervisedResult, RootHandle (view/signal/abort) as the observability substrate.
• bench/src/drivers/ — flat-harness control (with the equal-k assertion), progressive-widening control, and the LLM-meta-driver treatment. WidenGate defaults to flat so the selector≠judge firewall stays dormant; widening reads trace findings, never the raw verdict, unless judgeExempt.
• program.ts — mapPool one-for-one failure semantics (a down child is excluded from the merge; an all-down batch re-throws the first original error so a maxDepth guard still propagates loud).

Verification

Independently re-run on this branch: typecheck ✓, lint ✓ (204 files), build ✓, full suite 642/642 (incl. 28 keystone property tests: conserved-budget fail-closed + refund, equal-k by construction, monotonic-seq cursor, abort/teardown, replay determinism), bench typecheck ✓.

Honest gaps (why draft)

• Live executor paths unproven. The router/sandbox/cli executors are exercised only through the offline mock LeafExecutor — they typecheck and the wiring is proven, but no real router-HTTP / sandbox-runLoop / cli-subprocess run has happened.
• The LLM-meta-driver is the gated treatment, not a default. It ships so the diverse-strategy-vs-blind gate can actually be run; it is not wired into any default path.
• Deferred (gated on a positive gate result): a tuned MCTS-PW algorithm, learned widening, a Temporal/DBOS durable backend, deleting runProgram's loop-layer parallel op.

Relationship to #150

#150 adds the leaf-level continued-session/fork lineage on runLoop. This is the driver layer on top — the sandbox executor forwards that lineage passthrough rather than duplicating it. #150's findings (esp. verify the client-minted sessionId) are load-bearing here too.

[tangletools]

🔍 Reviewing 78d80209

Pass	Status	ETA
opencode DeepSeek v4 Pro	Running (2 min)	~5-15 min
opencode GLM 5.1	Running (2 min)	~5-15 min

Agent review running. Reads the actual code. This comment updates in place.

_{tangletools · #151 · model: kimi-for-coding · started 2026-06-04T13:05:48Z}