Expose burst and resident CLaaS commands with existing-agent scaffolds

[kfallah]

Expose exp optimize claas burst CONFIG and exp optimize claas serve CONFIG, locally or on Modal. Both validate bounded configuration/imports and use shared spend consent before selecting compute. Local execution uses the process supervisor; Modal foreground ownership retains exact recovery identity and bounds interruption cleanup.

The learning workflow reuses the existing AgentRuntime/EnvironmentRuntime loop and collects response IDs for explicit delayed feedback. Failure mining, world-model scenarios, evaluation, and deployment stay outside CLaaS. An explicit zero estimate declares nonbillable existing local hardware; metered hosts require their conservative estimate and Modal always requires a positive estimate.

Validation: actual CLI help and empty-burst process execution, real SDK/API/scaffold integration, cancellation/recovery, release checks, and dependency isolation. Final full local suite: 4,945 passed, 8 explicit skips; whole Ruff/format/ty pass. Both real L4 backend integration tests pass; see #1021 for their scope.

Stack 7/7: #1018 → #1019 → #1020 → #1021 → #1022 → #1023 → this PR. No merge or production deployment requested.

The real local CLI also closes an empty burst successfully. The actual Modal launcher passes a live isolated empty-burst smoke with committed receipt/SQLite validation and verified cleanup.

[superagent-security]

Superagent didn't find any vulnerabilities or security issues in this PR.

[greptile-apps]

Confidence Score: 5/5

The current head appears safe to merge, with no outstanding correctness, security, or repository-rule findings.

Summary

This PR exposes bounded CLaaS burst and resident-serving commands through the existing optimize CLI and adds local and Modal launch orchestration.

• Validates configuration, imports, persistence mode, run identity, and spend consent before backend selection.
• Retains Modal recovery identity and bounds interruption cleanup and receipt transport.
• Adds an existing-agent learning scaffold that returns response IDs for application-owned delayed feedback.
• Hardens local process-session ownership and Modal Volume commits.
• Documents the supported CLaaS lifecycle, cost semantics, deployment requirements, and release boundaries.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    CLI[exp optimize claas] --> Validate[Validate run config, mode, imports, and run ID]
    Validate --> Consent[Shared spend consent]
    Consent --> Select{Execution target}
    Select -->|Local| Supervisor[Owned local process session]
    Select -->|Modal| Sandbox[Named Modal Sandbox]
    Supervisor --> Learner[Bounded CLaaS learner]
    Sandbox --> Volume[Modal Volume]
    Volume --> Learner
    Learner --> Report[Run report and checkpoints]
    Sandbox --> Recovery[Recovery receipt with exact Sandbox identity]
    Agent[Existing AgentRuntime] --> Scaffold[Learning episode scaffold]
    Environment[Existing EnvironmentRuntime] --> Scaffold
    Scaffold --> Learner
    Scaffold --> IDs[Response IDs for delayed feedback]

Loading

_{Reviews (4) · Last reviewed commit: "Expose finite CLaaS CLI runs and existin..."}

[kfallah]

@greptileai The stack has been updated with the shared contract and startup-ownership fixes. The remaining API/runtime/CLI findings are still being addressed; their completed fixes and regression evidence will be sent for another review. This update does not claim those findings are resolved.

[kfallah]

@greptileai Please review the current head after the stack update. The previous findings have been addressed with fixes and regression coverage, or an evidence-backed reply where local hardware initialization was mistaken for paid infrastructure allocation. Every review thread has a concrete response. Please check the current diff and report any remaining defect.

[kfallah]

@greptileai Please review the current head. This stack update only corrects and strengthens the CUDA checkpoint assertion, with six CPU corruption cases; public runtime code is unchanged. The previously addressed findings remain fixed.

[kfallah]

@greptileai Please review this current head after rebasing onto main's native0.3.81 integer Responses timestamps. The test conflict preserves all new main assertions and independent fixture isolation. CLaaS runtime source is unchanged by the rebase. Prior findings remain addressed.

[kfallah]

@greptileai Please review the current head. The security findings are fixed with regression coverage: retained unreaped session identity, no raw group-kill fallback, and fixed /usr/bin/sync for Modal commits. Previous review findings remain addressed. The focused final cleanup/Modal/CLI suite passed66 tests; full current-head CI is running.