dagger

Status: work in progress. APIs, flags, and file formats are unstable and will change without notice.

dagger is a CLI that builds software from a task graph described in markdown. Each node in the graph is handed to a pluggable coding harness, and the graph is executed maximally in parallel — a task fires the moment all of its parents succeed. A final harness call summarizes the run from an ephemeral state document, which is then deleted.

The project is written in Bun with Effect v4 (beta), and leans on structured concurrency, typed errors, and the platform FileSystem service rather than ad-hoc I/O.

Quick start

bun install
bun test
bunx @fuiste/dagger do path/to/plan.md --harness cursor --dry-run

A real run looks like:

dagger do plan.md \
  --harness codex \
  --model gpt-5.4 \
  --max-concurrency 3

--harness is the required run default and also the harness used for the final run summary. Individual tasks may still override harness, model, and thinking in the markdown graph; task metadata takes precedence over the CLI defaults.

See docs/task-graph-format.md for the markdown syntax and an example plan.

Install

dagger is published as @fuiste/dagger and expects Bun at runtime.

Run it ad hoc with:

bunx @fuiste/dagger do plan.md --harness codex --dry-run

Or install it globally:

npm install -g @fuiste/dagger
dagger do plan.md --harness codex --dry-run

Current fit

Live QA has been most encouraging for source-grounded tasks that benefit from explicit intermediate artifacts, for example:

• repo introspection or reporting tools built from multiple existing source files
• migration-readiness or compliance audits where partial analysis is useful on its own
• refactors where fact-extraction and implementation can be separated cleanly

Right now dagger looks more compelling as a control and observability tool than as a pure speed play. Its best runs make progress legible mid-flight through task-level artifacts and archived state, even when the overall run is slower than a single large prompt.

Architecture sketch

• src/cli — Effect CLI entry points and flag parsing.
• src/domain — RunConfig, harness/thinking schemas.
• src/parse — markdown → TaskGraph parser with schema-validated metadata.
• src/runtime — DAG scheduler, run finalization, state persistence.
• src/state — ephemeral run state service (single-writer queue, surfaced writer errors).
• src/harness — harness registry, Cursor/Codex adapters, shared prompts/process helpers, and the DAGGER_EVENT assistant-message protocol.
• test/ — unit and integration tests (scheduler, state service, parser, harness, end-to-end runDo).

Known limitations

• Worktree support is intentionally out of scope for this pass.
• The published package still expects Bun to be installed locally; this is a Bun-backed CLI, not a Node-native standalone binary.
• Cursor Agent CLI itself may need NODE_EXTRA_CA_CERTS set in environments with a corporate TLS proxy; streaming sessions can still be blocked by SSL inspection of long-lived HTTP/2 connections.
• Codex runs are ephemeral by default and currently rely on the local codex CLI being installed and authenticated.
• End-to-end latency can still be significantly worse than a single-shot Codex run when the graph is tightly coupled or the final integration task does too much work.
• Some Codex-backed runs have reached full task success but then hung in finalization instead of returning cleanly. The produced artifacts can still be correct, but runner completion reliability needs hardening.
• Task decomposition does not yet make much weaker models a safe drop-in replacement for stronger ones on environment-discovery-heavy tasks; early analysis mistakes can still invalidate the whole graph.

License

TBD.