flywheel-bootstrap-staging 0.1.9.202602101920

Bootstrap runner for Flywheel provisioned GPU instances

Bootstrap

This package hosts the BYOC bootstrapper that:

• Ensures a pinned codex release is available (GitHub release assets).
• Fetches the bootstrap payload for a run from the Flywheel backend.
• Launches Codex app-server and streams notifications as logs.
• Collects artifacts (manifest-on-exit) and reports completion or error back to the backend.

Configuration

Bootstrap reads the user's Codex config.toml (Codex schema). See:

• Codex config basics: https://developers.openai.com/codex/config-basic
• Codex config reference: https://developers.openai.com/codex/config-reference

Flywheel adds a small extension under [flywheel] and requires one of:

[flywheel]
# inline instructions (host-specific tips, paths, sandbox notes)
workspace_instructions = """
Use /mnt/work as your workspace. Write artifacts under ./artifacts.
"""

# or: reference a file (relative paths are resolved against the config file directory)
workspace_instructions_file = "workspace_notes.md"

Rules:

• At least one of workspace_instructions or workspace_instructions_file is required; otherwise bootstrap exits before contacting the server.
• If both are set, the file wins and the inline value is ignored (warns once).
• File contents must be non-empty; the path is resolved relative to the config file if not absolute.

Bootstrap also reads a few optional Codex config fields because they affect workspace + sandbox behavior:

• cd / workspace_dir (run working directory; defaults to ~/.flywheel/runs/<run_id>)
• sandbox_mode and [sandbox_workspace_write].writable_roots

Prompt assembly (sent as the first Codex turn/start):

1. Flywheel engineer context (logging/artifact expectations).
2. Task Description (prompt fetched from the server).
3. Workspace Instructions (resolved from config as above).

Full example config (recommended starting point):

• https://github.com/paradigma-inc/paradigma/blob/main/project/bootstrap/examples/config.example.toml

Update the paths and instructions for your machine.

End-to-end flow (bootstrap.sh → Python bootstrapper)

1. User runs the bootstrapper on their BYOC machine:

   BOOTSTRAP_PACKAGE="${FLYWHEEL_BOOTSTRAP_PACKAGE:-flywheel-bootstrap}"
   uvx --no-cache --from "$BOOTSTRAP_PACKAGE" flywheel-bootstrap \
     --run-id <id> --token <token> --config /path/to/config.toml [--server <url>]
   

   Optional: if you have this repo checked out, you can run project/bootstrap/bootstrap.sh, which installs uvx if missing and then runs the same command. It defaults FLYWHEEL_BOOTSTRAP_PACKAGE to the local project/bootstrap path.

2. Python entrypoint (python -m bootstrap):

   • Parses args/env: requires run id + token, required --config, optional --server (default http://localhost:8000).
   • Loads Codex config.toml, enforces presence of workspace instructions (inline or file), extracts workspace/sandbox settings.

3. Workspace resolution:

   • Uses cd/workspace_dir from config if set; otherwise ~/.flywheel/runs/<run_id>.
   • Creates the workspace and validates the artifact manifest path is inside sandbox writable_roots when sandboxing is enabled; else exits with an error.

4. Engineer runtime availability:

   • If BOOTSTRAP_MOCK_ENGINEER is set, skips agent launch and runs a mock flow (used in tests).
   • Else, download a pinned Codex release asset and mark it executable.

5. Fetch bootstrap payload:

   • GET <server>/runs/<run_id>/bootstrap with X-Run-Token; payload contains the task prompt.

6. Build prompt text:

   • Combine base Flywheel engineer context, “Task Description” (server prompt), and “Workspace Instructions” (user config) into one prompt string.

7. Launch engineer via Codex app-server:

   • Create a per-run CODEX_HOME directory at <workspace>/.codex_home containing a copied config.toml (with ~/ expanded).
   • Best-effort copy auth credentials from ~/.codex/auth.json into <workspace>/.codex_home/auth.json so Codex can authenticate.
   • Launch codex app-server as a subprocess and speak JSON-RPC over stdio (newline-delimited JSON; "jsonrpc":"2.0" header omitted):
     • initialize + initialized
     • thread/start
     • turn/start (first prompt = full Flywheel prompt; subsequent prompts = autonomous resume)
     • turn/interrupt for takeover/resume
     • Stream turn/*, item/*, and delta notifications (item/agentMessage/delta, item/commandExecution/outputDelta, etc.)
   • Bootstrap auto-accepts command/file approvals (for-session) when Codex requests them.
   • Start a heartbeat thread posting /runs/{id}/heartbeat.
   • Start a run-control polling thread (/runs/{id}/control/poll) to handle:
     • take_over (interrupt autonomous prompt turn; enter interactive mode)
     • send_input (start a new Codex turn with user input in interactive mode)
     • resume_autonomous (return to autonomous mode)
   • Convert thread/tokenUsage/updated totals into a turn.completed usage log for billing.
   • Stream app-server notifications as logs to /runs/{id}/logs.

8. After the engineer prompt turn ends:

   • If code persistence is enabled, bootstrap commits/pushes non-artifact changes:
     • Excludes manifest-declared artifact paths from the git commit
     • Skips oversized files (default cap: 95 MiB, configurable via FLYWHEEL_GIT_MAX_COMMIT_FILE_SIZE_BYTES)
   • Artifact files referenced by payload.path/payload.file are embedded into artifact payloads when they are below the transfer threshold (default: 16 MiB, configurable via FLYWHEEL_ARTIFACT_BLOB_THRESHOLD_BYTES).
   • Read flywheel_artifacts.json; if missing/malformed/empty, retry up to MAX_ARTIFACT_RETRIES by prompting the engineer to fix/write the manifest.
   • POST artifacts to /runs/{id}/artifacts; POST /complete on exit 0, else /error with the exit code.
   • Stop/join the heartbeat thread.

9. Mock mode (BOOTSTRAP_MOCK_ENGINEER=1):

   • Sends a heartbeat, a few logs, writes a mock artifact manifest, returns 0 (used in e2e tests).

Next steps

• Iterate on prompts / general polish.