Skip to content

Latest commit

 

History

History
81 lines (63 loc) · 3.66 KB

manager.md

File metadata and controls

81 lines (63 loc) · 3.66 KB

version: 1.1 last_updated: 2025-11-09

Session Lifecycle

src/slop_code/execution/session.py replaces the earlier execution manager hierarchy with a lightweight orchestration class. A Session ties together:

  • A Workspace, which prepares an isolated directory, restores snapshots, and exposes helpers for reading tracked files.
  • Resolved static assets, including placeholder substitution for case payloads.
  • One or more SubmissionRuntime instances spawned on demand.

Sessions are context managers so evaluation adapters can prepare and tear down resources with a with block.

from pathlib import Path
from slop_code.execution import Session, LocalEnvironmentSpec

spec = LocalEnvironmentSpec()
session = Session.from_environment_spec(spec, base_dir=Path("/problem"), static_assets=None)

with session as active:
    runtime = active.spawn()
    result = runtime.execute("python main.py", env={}, stdin=None, timeout=30)

Core Workflow

  1. Construction – Callers provide an EnvironmentSpec, a Workspace, and optional static assets. Helper constructor Session.from_environment_spec builds a workspace snapshot function that respects spec filters and assets.
  2. Preparation – Entering the context calls workspace.prepare(), which materializes the initial snapshot and (for agent inference) copies static assets into place.
  3. Runtime Spawn – session.spawn() delegates to the runtime registry via spawn_runtime(). The runtime shares the workspace root and receives any additional mounts, ports, env vars, or setup commands requested by adapters.
  4. File Materialization – Request payloads are written with session.materialize_input_files(), which uses the unified file operations API to pick the correct handler based on file signatures.
  5. Execution & Streaming – The returned SubmissionRuntime handles process execution (see local.md and docker.md).
  6. Reading Results – session.get_file_contents() applies glob patterns to the workspace so adapters can retrieve tracked files after a run.
  7. Checkpoint Completion – Agent inference flows call finish_checkpoint() to update the workspace snapshot, export the archive, and compute a SnapshotDiff.
  8. Cleanup – Leaving the context calls runtime.cleanup() on every spawned runtime and disposes of the workspace directory.

Static Assets and Placeholders

Sessions keep a reference to the resolved static asset map. Two helpers make use of it:

  • session.resolve_static_placeholders(data) rewrites {{static:name}} placeholders so case fixtures can refer to assets using environment-appropriate paths (/static/... in Docker, absolute host paths locally).
  • The workspace materializes assets when running in agent inference mode, copying directories or files into the temporary working directory.

Agent vs Evaluation Sessions

The is_agent_infer flag changes a few behaviors:

  • Evaluation sessions leave static assets on the host (or mounted read-only into containers) and skip workspace asset copies.
  • Agent inference sessions allow finish_checkpoint() so runners can persist the resulting workspace and compute diffs for reporting.
  • Runtime spawn calls mark is_evaluation=not is_agent_infer, influencing setup command ordering and Docker user selection.

Logging Hooks

Every operation logs via structlog. Enable verbose logging in CLIs or tests to trace workspace preparation, runtime spawn parameters, and file access patterns.

Understanding the session lifecycle clarifies how adapters interact with the execution layer and where to introduce new instrumentation or backend features.