Module runner 

Task runner — executes tasks using AI agents.

Replaces run_task.sh with pure Rust. The runner:

1. Sets up a git worktree for isolation
2. Builds context (project instructions, skills, comments)
3. Renders prompts and spawns the agent in tmux
4. Waits for completion, collects and classifies the response
5. Handles errors (reroute on limits, fallback agents)
6. Auto-commits, pushes, and creates PRs
7. Updates status labels and posts result comments

Submodules handle the heavy lifting:

• task_init — guard checks, worktree setup, and invocation building
• session — tmux session lifecycle and output collection
• response_handler — success path: commit, push, PR, token storage
• fallback — error classification and recovery strategies

Re-exports

pub use response::WeightSignal;

Modules

agent

Agent command building + tmux invocation.

agents

Per-agent runner trait, error types, and agent registry.

context

Context building for task execution.

cow

Copy-on-write task worktrees cloned from a per-project template by the git-cow crate, see docs/content/cow-worktrees.md

diff

direct

One-shot agent invocation without tmux.

fallback

Error recovery and fallback strategies.

git_ops

Git operations — auto-commit, push, PR creation.

response

Response collection + error classification.

response_handler

Success-path response handler — commit, push, PR, token storage.

session

Agent session management — spawn, wait for completion, collect output.

task_init

Task initialization — guard checks, worktree setup, and invocation building.

worktree

Git worktree management for task execution.

Structs

RunAudit

Fully materialized task-run audit data.

RunExecution

Result returned by run() so callers can persist the audit trail.

TaskRunner

Task runner configuration.

Functions

parse_session_output

Tries parse_success_output first when stdout is non-empty. If that succeeds, returns success regardless of exit code. If it fails and exit_code != 0, checks for "terminal_reason":"completed" or cost-telemetry-only stdout ("costUSD":) before falling back to classify_error, so valid sessions that exit 1 produce InvalidResponse (soft error, re-route) instead of a garbled Unknown/Auth error (hard failure with spurious cooldown)