MoonBit AI Tasks: Implementation Inventory

This document captures the task-centric MoonBit implementation added in Flow, where .ai/tasks/*.mbt is the primary extension mechanism.

Scope

The implementation adds:

• discovery and execution of AI MoonBit tasks under .ai/tasks/
• CLI/docs updates to make tasks the primary interface
• legacy recipe command demoted to compatibility mode
• a concrete task pack for Flow self-development

Code Paths Added/Changed

Core runtime and wiring:

• src/ai_tasks.rs
• src/ai_taskd.rs
• src/tasks.rs
• src/bin/ai_taskd_client.rs
• src/palette.rs
• src/cli.rs
• src/lib.rs

Legacy compatibility updates:

• src/recipe.rs

Docs updates:

• docs/commands/tasks.md
• docs/commands/recipe.md
• docs/commands/readme.md
• docs/index.mdx
• docs/moonbit-ai-tasks-implementation.md

Workspace hygiene:

• .gitignore (ignore Moon generated dirs under .ai/tasks/**)

Runtime Behavior

Discovery

Flow scans .ai/tasks/ recursively for .mbt files and exposes selectors as ai:<path>.

Key behavior in src/ai_tasks.rs:

• ignores generated Moon artifacts during discovery (.mooncakes, _build)
• parses metadata from top comments:
  • // title: ...
  • // description: ...
  • // tags: [a, b]
• resolves stable task IDs and selectors from path layout

Selection

Task resolution accepts:

• full selector: ai:flow/dev-check
• scoped selector forms
• name-based matching with ambiguity detection

Execution

Flow executes AI tasks through a cache-first runtime.

Important execution details:

• auto-resolves nearest Moon workspace root (moon.mod.json / moon.mod)
• computes content hash (task source + Moon config + moon version)
• builds native artifact once (moon build --target native --release)
• reuses cached binary from ~/Library/Caches/flow/ai-tasks/<hash>/task-bin
• falls back to moon run for tasks without Moon workspace metadata
• optional mode control via FLOW_AI_TASK_MODE (dev, release, js, etc.)
• default frozen dependency behavior unless FLOW_AI_TASK_NO_FROZEN is set
• runtime override via FLOW_AI_TASK_RUNTIME=moon-run

Daemon

Flow now includes a lightweight Unix-socket daemon for repeated AI task runs:

• socket: ~/.flow/run/ai-taskd.sock
• lifecycle: f tasks daemon start|status|stop
• daemon execution: f tasks run-ai --daemon <selector>
• low-overhead client execution: ./target/release/ai-taskd-client <selector>

Recent runtime optimizations:

• daemon-side task discovery cache with TTL (FLOW_AI_TASKD_DISCOVERY_TTL_MS)
• daemon-side hot artifact reuse cache with TTL (FLOW_AI_TASKD_ARTIFACT_TTL_MS)
• fast exact selector resolution (skip full .ai/tasks scan for ai:scope/task)
• cache key generation optimized to use file metadata fingerprints + cached Moon version
• optional low-latency dispatch via fai with auto-preference from f for latency-tagged AI tasks in daemon mode

Task Pack Added

Flow-local task pack under .ai/tasks/flow/:

• .ai/tasks/flow/dev-check/main.mbt
• .ai/tasks/flow/pr-ready/main.mbt
• .ai/tasks/flow/regression-smoke/main.mbt
• .ai/tasks/flow/release-preflight/main.mbt
• .ai/tasks/flow/bench-cli/main.mbt
• .ai/tasks/flow/noop/main.mbt

Each task has its own Moon package/workspace files:

• .ai/tasks/flow/<task>/moon.mod.json
• .ai/tasks/flow/<task>/moon.pkg.json

Task Intents

• ai:flow/dev-check: fast quality gate (cargo check, targeted tests, CLI help smoke)
• ai:flow/pr-ready: pre-PR gate (dev-check + docs parity + gitignore hygiene)
• ai:flow/regression-smoke: temporary project smoke for task discovery/execution
• ai:flow/release-preflight: build release binary and run release-path smoke checks
• ai:flow/bench-cli: quick latency benchmark for high-frequency Flow CLI entry points

How To Run

From ~/code/flow:

f tasks list
f tasks build-ai ai:flow/dev-check
f tasks run-ai ai:flow/dev-check
f tasks run-ai --daemon ai:flow/dev-check
f tasks daemon start
f tasks daemon status
f tasks daemon stop
f ai:flow/dev-check
f ai:flow/pr-ready
f ai:flow/regression-smoke
f ai:flow/release-preflight
f ai:flow/bench-cli

Optional benchmark controls:

FLOW_BENCH_ITERATIONS=30 FLOW_BENCH_WARMUP=5 f ai:flow/bench-cli

Runtime-path benchmark harness:

f bench-ai-runtime --iterations 80 --warmup 10 --json-out /tmp/flow_ai_runtime_bench.json

Automated myflow Commit→Session Check

You can automate the exact flow:

1. Do real Claude/Codex work in a repo (for example ~/code/myflow).
2. Commit with sync (f commit --sync ...).
3. Verify that commit is visible in myflow and has attached sessions.

Flow task:

f myflow-commit-session-smoke --help

Common run for ~/code/myflow:

f myflow-commit-session-smoke --repo-path ~/code/myflow --require-sessions

What it checks:

• GET /api/commits?repo=<owner>/<repo> contains the target commit
• commit has sessionWindow metadata
• if --require-sessions is set, commit has sessions.length > 0
• first session id is fetchable via GET /api/sessions/:id (unless --skip-session-fetch)

Auth:

• uses MYFLOW_TOKEN if set
• otherwise falls back to ~/.config/flow/auth.toml token

Validation Commands

cargo check --all-targets
cargo build --release --bin f
f tasks list | rg '^ai:flow/'
f ai:flow/regression-smoke
f myflow-commit-session-smoke --repo-path ~/code/myflow --require-sessions

Generated Artifact Hygiene

To prevent accidental commit noise from Moon caches/build output, Flow ignores:

• .ai/tasks/**/.mooncakes/
• .ai/tasks/**/_build/

Notes for Commits

When committing this work, scope to the relevant code + docs only:

f commit --path src/ai_tasks.rs \
  --path src/tasks.rs \
  --path src/palette.rs \
  --path src/cli.rs \
  --path src/lib.rs \
  --path src/recipe.rs \
  --path docs/commands/tasks.md \
  --path docs/commands/recipe.md \
  --path docs/commands/readme.md \
  --path docs/moonbit-ai-tasks-implementation.md \
  "add task-centric moonbit ai task runtime and flow task pack"