Answer-state checkpoints, and finishing an interrupted run into an answer

[shanev]

A run that stops short used to hand the host a fatal error and throw away everything it had done. Three related changes make an interrupted run finish into an answer, plus one bug fix found while testing them.

Answer state no longer dies with the process

Hosts embedding the Pyodide substrate can be SIGKILLed mid-execution — Pyodide cannot be preempted, so a wedged run runs no finalizer and no signal handler. Until now every answer-critical value (the draft, and whatever host state backs it) existed only in process memory until project_run_response at the very end.

Trace ABI v7 adds a checkpoint event carrying the draft plus an opaque host payload, emitted after each executed step whose draft moved. A kill becomes a render rather than a recovery.

• Classified CONFIGURABLE, not DURABLE — it carries content, so it must be retention-gated.
• checkpoint_seq is strictly increasing per run_id, enforced in TraceRecorder (the only per-run authority); the stateless half stays in validate_event_body.
• payload is validated as Mapping | None and never inspected or schema-checked, matching how the relay already ferries adapter meta.
• A raising provider still emits a checkpoint with payload: null; an unrepresentable payload drops that one checkpoint. A checkpoint can never fail a run.

Wall-clock exhaustion routes to the extract fallback

Spending budget.wall_ms ended through early_result, which hands back a fatal error alongside whatever partial answer existed — and hosts discard the run. Every iteration of real work was thrown away over a time verdict, which says nothing about whether that work was good.

Deadline exhaustion now takes the same terminal handoff as an exhausted iteration budget. Token exhaustion deliberately stays fatal: there is no budget left to pay for the extract call it would trigger.

The extract call is reserved with deadline_exempt=True — call_root reserves through_deadline=True, which refuses once remaining.wall_ms hits 0, so without the exemption the fallback would be unreachable in exactly the case it exists to serve.

The host can now say what the engine cannot see

Two symmetric hooks, both opaque, both keeping the engine ignorant of what a host data layer is:

• extractable_work_probe — the engine decides whether extraction is worth attempting from a retained draft or a completed step. Neither exists when generated code calls an accessor and then raises before printing. Measured on a real 314k-message corpus: a run retrieved 84 GUIDs across four ledger records and returned a bare BudgetExhausted, never attempting extraction.
• extract_context_provider — unlocking the call was necessary but not sufficient. The extract prompt is built from draft + code + stdout, and a REPL only surfaces what the code chose to print, so extraction was handed failing code and answered InsufficientEvidence while the rows sat in the host's ledger.

extract_context_provider is deliberately narrow: observations are supplied only when the host is what made extraction possible. When the engine can already see the work, the prompt is byte-identical to before this PR — same answers, same spend, same benchmark scores. The behavior change is confined to runs that produce nothing today.

Bounded at 8k chars because the extract call is deliberately uncached (cache_anchors=None), so every character is paid at full prefill.

This is dspy.ReAct's shape rather than a new invention — it puts every tool's return value straight into the trajectory, so extraction always sees the data. dspy.RLM, being REPL-based, has exactly this gap: its variables_info is built once from the inputs and never reflects what execution retrieved.

Bug fix: emoji silently killed every checkpoint

events.ts compared Python's code-point draft_chars against JavaScript's UTF-16 String.length. Any non-BMP character made them disagree, so isRlmEvent rejected the frame and the relay dropped it before it reached the host — silently, and for every later checkpoint too, since the draft keeps the character.

Found by adversarial review, not by the suite: every fixture was ASCII. Regression tests now use emoji, astral CJK, math alphanumerics, a ZWJ sequence, and a regional-indicator flag, plus the negative case.

Testing

1261 passed, 3 skipped, ruff clean, deno test pyodide/ 60 passed.

Mutation-verified rather than assumed: removing the draft/payload gate, the provider guard, the emission guard, the repaired-code emission point, the draft_chars check, the strict-monotonic check, the host_context wiring, the 8k bound, or the narrow gate each fails a distinct test. The UTF-16 fix is verified against a real WASM-emitted frame.

Notes for review

• ABI 6 → 7 rewrites every content-addressed scaffold manifest id. Fixture ids were re-derived programmatically via ScaffoldManifest.from_dict, not hand-edited.
• The v7 golden corpus exercises the version bump but does not yet carry a checkpoint event — adding one renumbers seqs across a run. Say the word if the cross-language corpus should carry one.
• The terminal fail-closed finalization execute_step does not emit a checkpoint. That follows the spec's explicit enumeration, but that path can also mutate answer['content'] — a one-line addition if wanted.
• Cozy pins droste==0.21.1 and its host side is ready but blocked on a release carrying this.

🤖 Generated with Claude Code