Shared program execution

Decision and assessment

The 2026-09-07 Claude conversation proposed a thin program executor: shared files, prompt-level territories, a sole integrator after parallel work, and a result when the graph ends. The local Claude Stage B script corroborates that pattern: writers report cross-territory requests, then one integration agent applies them and runs repository gates. This is evidence of the authoring pattern, not a guarantee about every Claude run or provider.

The implementation at 5172e09 did not meet that direction:

Boundary	Before	New goal launches
Workspace	Isolation enabled unless setup explicitly disabled it	Shared target directory; --isolation is explicit
File scope	Every changed file had to be declared or the action failed	Advisory territory; no default manifest, copying, integration, or rollback
Shared concurrency	At most one writer	Disjoint territories run concurrently; overlapping writers serialize
Dependency execution	Awaited the entire selected batch	Starts a dependent as soon as its own inputs finish
Integrator	Required an exhaustive file list	Shared build/chore action with empty ownedFiles runs alone and may fix any file
Graph validation	Rejected ordering without an artifact or overlapping file	Accepts ordinary acyclic dependencies; validates declared artifact references
Completion	Required fresh passing evidence and opened gap rounds	Returns all action results when the graph ends
Verification	completed implied verified	verified separately reports passing mandatory requirement evidence
Unexpected action error	Could terminate the kernel	Records a failed action, retains shared edits, and continues independent branches
Quiet agent	Heartbeat depended on provider output	Independent 10-second kernel heartbeat

Execution contract

flowchart LR
  A1 --> B1[Integrator B1]
  A2 --> B1
  A3 --> B1
  B1 --> C1
  B1 --> C2

The caller authors the graph using workflow plan contract and launches with workflow goal --program. --orchestrator remains an explicit alternative that delegates planning. Routing, quota exclusion, recursion depth, content judgment, and mechanical retries still use the existing dispatcher.

Workers receive territory guidance, dependency output paths, and instructions to preserve user and sibling edits. The recommended integrator reads those outputs, applies requests for shared files, and runs the actual repository acceptance commands. An unrestricted integrator runs alone, including readers. The kernel never invents an integration task or acceptance command.

An action failure skips its success-dependent descendants; independent work continues. The terminal status is completed, partial, or cancelled. Negative requirement evidence is returned unchanged and does not trigger a planner round. Initial scouting and explicit user steering can still create a durable caller-planner pause. Steering waits for active workers before changing the program. Further repairs after a terminal result use a new program.

result.json includes every action’s status, failure, and output artifact, plus requirement evidence and verified. Workspace files remain in the target tree. The optional Git inventory lists current changed paths, the initial dirty-path list, and warnings. It includes pre-existing and concurrent changes; it is not per-worker attribution. Failure to obtain that inventory never fails the work.

Compatibility

New CLI launches persist config.settings.executionMode = "program" and workspaceMode = "shared". Saved V2 documents without executionMode retain their original requirement-gated completion, planner boundaries, and workspace policy. Resume does not reinterpret historical work under the new contract. Fixed authored graphs and their explicit verify steps retain their existing semantics.

--isolation preserves strict per-worker worktrees. Isolated writers must declare exact files; unrestricted writers are refused before dispatch. This opt-in path can still reject out-of-scope edits. The default shared path never discards work for that reason.

Evidence and limits

tests/workflow-program.test.js exercises shared-file retention, new files, overlap serialization, unrestricted integration, dependency-ready execution, failure propagation, optional/negative evidence, failed scouting, heartbeat, cancellation, resume, steering, literal filenames, unavailable Git inventory, and explicit isolation. Its real CLI fixture runs six routed local child processes in the graph above, proves a peak of three writers, checks recursion depth, and inspects the resulting files and all six durable outputs. The detached CLI test separately proves that negative evidence produces a stable result without another planner round. Legacy CLI recovery tests use saved requests without the new mode.

Validation commands:

npm test
node --test --test-timeout=30000 tests/workflow-program.test.js

The initial implementation passed 605/605 tests, including 17 program regressions; the untouched baseline passed 588/588. Release hardening adds the process-boundary recovery and dependency-display regressions described below.

These are deterministic local tests, including real subprocess dispatch; they do not call paid model providers or prove model-generated work quality. Territories are coordination guidance, not a security sandbox. A kernel or machine can still stop; recovery skips durable successes and replays unfinished actions, so external side effects are not exactly-once. Provider failures and bad plans remain possible. The simplification removes unnecessary engine failure paths rather than promising infallibility.

Live provider acceptance (2026-09-08)

Run ggtegs (wf-mts0h1pp-1228a7) used the installed CLI in a disposable calculator repository. Its exact six-action graph completed in 4m10s:

• Three Grok 4.6 builders overlapped in one shared tree and each added a focused test file outside its advisory source-file territory.
• One Grok integrator started after all builders finished, combined their modules, wrote usage documentation, and passed the 13-test suite.
• Two Claude Sonnet 5 reviewers overlapped after integration and returned passing evidence. There were six attempts, one submitted program, no planner dispatches, and no additional planning pause.
• Local inspection confirmed the files, preserved user note, and passing tests. Durable action timestamps confirmed the concurrency and dependency order.

The result reported completed and verified: true, but independent caller probes found an implementation bug missed by both reviewers: inherited operation names such as constructor, toString, and valueOf threw TypeError instead of RangeError. This is a model-output limitation, not evidence that the workflow kernel failed. Passing evidence verdicts are not a correctness guarantee. The explicit repair run nx5dyi (wf-mts0mwt1-4b4a74) completed with passing independent evidence. Its regression first reproduced the failure, then the repair passed all 18 tests. Caller probes also confirmed inherited names, coerced arrays/objects, symbols, null, and undefined now throw RangeError. Resuming the original completed run returned its existing result with exactly six attempts and no worker replay.

Fresh-controller acceptance

These controllers started in new sessions, read the installed skill, and received one short, human-style task prompt. No coordinator corrections were sent while they worked. Pool selection remained automatic; the controller’s model and the dispatched worker models are separate choices.

Controller	Task and graph	Durable result
Fresh Claude	Transaction CSV report CLI: three writers → sole integrator → two independent reviewers	ue5gbs, 6 attempts, 1 program, 18m17s, completed with passing evidence
Fresh Grok 4.6	Non-repository Temporal/Restate/Inngest research: three researchers → synthesis → independent source check	efy8c2, 5 attempts, 1 program, 23m06s, completed with passing evidence
Fresh Codex gpt-5.6-sol, medium	Existing repository, read-only cancellation/recovery audit: two parallel auditors → independent verifier	mkvn9a, 3 attempts, 1 program, 24m49s, completed with passing evidence

The Claude controller recovered from one rejected plan: a verifier prompt comparing documented product JSON/output was incorrectly treated as an attempt to replace the kernel’s evidence format. The release narrows that check; the exact rejected plan now validates, while explicit verdict-format overrides are still rejected. After its workflow, the controller used one bounded delegate to address reviewer concerns (test strength and README wording). It finished with 89/89 tests and reported 115 independent CLI probes with no failures. The coordinator separately ran the original 85 tests and nine adversarial CLI scenarios, then re-ran all 89 tests after the follow-up; all five protected seed files were byte-identical. The caller also verified that a deliberately reintroduced floating-point parsing mutation fails the strengthened tests. These checks are separate from the passing verdicts.

The Grok controller produced COMPARISON.md and three source reports in a non-Git research directory. The independent verifier inspected all three retry, recovery, and operational comparisons and checked official source links. Its report preserves pricing/documentation inconsistencies as uncertainties. The coordinator spot-checked the official Temporal, Restate, and Inngest retry documentation independently. This proves the non-repository research path, not universal correctness of every research claim or lasting accuracy of pricing.

The Sol controller kept its existing checkout unchanged at 89cda2c. The audit and independent verifier found concrete release blockers despite passing ordinary unit tests: cancellation could be overwritten, a successful attempt could replay before its action status was committed, isolated recovery could delete unfinished work, and V2 lacked reliable process ownership on interruption. The verifier corrected an overbroad submodule finding: default shared program mode did not enforce ownership, so that defect affected isolation/legacy modes. One audit-shell full-suite failure came from its inherited CLAUDE_CONFIG_DIR; the affected test passed with that variable removed.

Release hardening from the acceptance findings

• A separate cancellation-intent file prevents the operator from rolling back kernel state or having a request overwritten. Paused-run and watch readers overlay that intent immediately.
• An exclusive kernel lease fences state writes and caller submissions. Recorded process identities and process groups let resume drain surviving delegates before re-dispatch. SIGTERM/SIGINT now commit a resumable V2 interruption; watch names the resume command.
• A durable completion receipt precedes the successful attempt snapshot. Recovery resumes ownership/evidence/integration processing from either side of that write boundary. A partially applied isolated integration tolerates already-applied identical files but rejects different user edits.
• Each new isolated attempt gets a separate tree. Unfinished trees survive and are named in result warnings. Failed setup cleans its own registration, and manifests inspect submodule files without treating a directory as a file.
• Program presentation groups use topological dependency levels. They do not impose barriers: a ready dependent can run while an unrelated earlier-level action is still active. Saved runs are projected into the corrected view without rewriting their state or event history.

tests/workflow-v2-recovery.test.js crosses real process boundaries: it kills kernels after receipt/attempt persistence and during isolated integration, checks shared and isolated output is applied once, preserves conflicting user edits, forces the cancellation/persist interleaving, signals real CLI delegates with child processes, and races two resumes. Existing interrupted attempts may still have unknown external effects and can run again; this is not an exactly-once guarantee for arbitrary external services or power loss.

Final release validation on 2026-09-08: 621/621 repository tests passed. The coordinator re-ran the completed transaction CLI suite (89/89) and rendered both saved research/audit runs through the corrected dependency view.