feat(action-engine): add execution runtime and ActionAgent

[skywhite1024]

Description

Stack

• Stack: #543
• Layer: 6/7
• Base: ljd/gen-sim-refactor-05-runtime-grounding
• Depends on: feat(action-engine): add runtime grounding and action adapters #536
• Followed by: feat(task-engine): add orchestration and end-to-end integration #538

Add ProgramExecutor scheduling, per-environment state transitions, bounded retries and graph recovery, runtime recording/reporting, ActionAgent execution, isolated A/B evaluation, and the execution CLI.

Runtime plans consume main's endpoint-owned command sequences and retained joint trajectories while keeping SeedGraph execution reports tensor-free and auditable.

Refs #531

Type of change

• New feature (non-breaking change which adds functionality)

Validation

• Focused executor/agent/evaluation coverage - 176 passed
• Stack-tip GenSim regression - 706 passed, 9 warnings
• Stack-tip related AtomicAction coverage - 120 passed
• Black 26.3.1 - 855 Python files unchanged at the stack tip
• git diff --check - passed

Checklist

• Code passes Black 26.3.1.
• Tests cover the affected behavior.
• Stack-level architecture documentation is included by layer 7.
• No new third-party dependency is required.

[greptile-apps]

Greptile Summary

The PR adds an ActionAgent, an executable action runtime with recovery and recording, and offline/A-B CLI orchestration.

• Compiles grounded or unbound plans into executable seed graphs.
• Executes vectorized simulator actions with retries, recovery, and per-environment reporting.
• Adds process-isolated A/B evaluation and runtime artifact publication.
• Adds contract, recovery, agent, CLI, architecture, and evaluation tests.

Confidence Score: 4/5

The persistence failure path should be fixed before merging because a completed simulator execution can escape without the promised report; worker force-cleanup is also worth hardening.

Post-execution filesystem or serialization failures are outside ActionAgent.execute's handled exception set, so successful motion can end in a caller-visible exception rather than an execution report.

Files Needing Attention: embodichain/gen_sim/action_engine/agent.py; embodichain/gen_sim/action_engine/cli/run_agent.py

Important Files Changed

Filename	Overview
embodichain/gen_sim/action_engine/agent.py	Adds plan compilation and execution reporting, but successful runs can raise when post-execution trajectory persistence fails.
embodichain/gen_sim/action_engine/cli/run_agent.py	Adds offline and process-isolated A/B orchestration; worker cleanup lacks a final response for children that survive termination.
embodichain/gen_sim/action_engine/runtime/executor.py	Implements vectorized execution, retry, recovery, and result production with per-environment masking.
embodichain/gen_sim/action_engine/runtime/recovery.py	Adds per-environment retry and recovery decision tracking.
embodichain/gen_sim/action_engine/runtime/recording.py	Adds runtime graph, checkpoint, and revision artifact recording.
embodichain/gen_sim/action_engine/runtime/reporting.py	Adds strict execution-report validation and atomic report publication.
embodichain/gen_sim/action_engine/evaluation/ab.py	Adds strict offline-versus-online branch execution and comparison artifact generation.
embodichain/gen_sim/action_engine/environment/agent_env.py	Integrates the new execution runtime with the generated simulator environment.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Grounded task plan] --> B[ActionAgent plan and preflight]
    B --> C[ExecutionProgram]
    C --> D[ProgramExecutor]
    D --> E{Action outcome}
    E -->|Retryable| F[Retry and recovery]
    F --> D
    E -->|Complete| G[ExecutionResult]
    G --> H[Trajectory persistence]
    H --> I[ExecutionReport]
    I --> J[Runtime and Task Engine artifacts]
    C --> K[A/B branch workers]
    K --> L[Offline branch]
    K --> M[Online branch]
    L --> N[A/B comparison]
    M --> N

Loading

Prompt To Fix All With AI

### Issue 1
embodichain/gen_sim/action_engine/agent.py:414
**Persistence errors escape reporting**

When a successful execution cannot persist its trajectory because the record directory is unwritable, full, or otherwise unavailable, `_persist_executed_trajectory` raises an uncaught filesystem or serialization error, causing `ActionAgent.execute` to exit without the promised aborted `ExecutionReport`.

### Issue 2
embodichain/gen_sim/action_engine/cli/run_agent.py:531-534
**Stuck workers survive cleanup**

`_ABBranchWorker.close` stops after `terminate()` and a timed join without killing or reporting a child that remains alive, leaving a stuck simulator process able to retain GPU, shared-memory, and other runtime resources after the CLI exits.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (1): Last reviewed commit: "feat(action-engine): add execution runti..." | Re-trigger Greptile}

[greptile-apps]

Persistence errors escape reporting

When a successful execution cannot persist its trajectory because the record directory is unwritable, full, or otherwise unavailable, _persist_executed_trajectory raises an uncaught filesystem or serialization error, causing ActionAgent.execute to exit without the promised aborted ExecutionReport.

Prompt To Fix With AI

This is a comment left during a code review.
Path: embodichain/gen_sim/action_engine/agent.py
Line: 414

Comment:
**Persistence errors escape reporting**

When a successful execution cannot persist its trajectory because the record directory is unwritable, full, or otherwise unavailable, `_persist_executed_trajectory` raises an uncaught filesystem or serialization error, causing `ActionAgent.execute` to exit without the promised aborted `ExecutionReport`.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Stuck workers survive cleanup

_ABBranchWorker.close stops after terminate() and a timed join without killing or reporting a child that remains alive, leaving a stuck simulator process able to retain GPU, shared-memory, and other runtime resources after the CLI exits.

Prompt To Fix With AI

This is a comment left during a code review.
Path: embodichain/gen_sim/action_engine/cli/run_agent.py
Line: 531-534

Comment:
**Stuck workers survive cleanup**

`_ABBranchWorker.close` stops after `terminate()` and a timed join without killing or reporting a child that remains alive, leaving a stuck simulator process able to retain GPU, shared-memory, and other runtime resources after the CLI exits.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.