liteflow

A ~1000-line C program that runs YAML-defined DAGs, where an LLM can edit the graph mid-run.

When a task fails, a planner LLM gets the stderr and emits one of four verbs: RETRY, PATCH, INSERT_BEFORE, ABORT. The runtime applies the mutation and continues. Every edit is recorded in an append-only event log, so you can replay any run and see exactly which graph changes the LLM made and why.

Not an agent. A runtime where the LLM is a peer of the scheduler.

Build

cc -std=c11 -O2 -o liteflow liteflow.c

Single file, no libraries. Shells out to curl for HTTP. Linux, macOS, modern Windows.

Run

export OPENAI_API_KEY=...
export OPENAI_BASE_URL=https://api.deepinfra.com/v1/openai   # any OpenAI-compatible endpoint

./liteflow run examples/demo.yaml
./liteflow replay logs/<run-dir>/events.jsonl

What a workflow looks like

name: release
tasks:
  - id: tests
    type: shell
    cmd: "pytest"

  - id: gate
    type: decision
    prompt: "Tests passed. Ship or hold?"
    branches: [ship, hold]
    depends_on: [tests]

  - id: ship
    type: shell
    cmd: "./deploy.sh"
    depends_on: [gate]
    when: gate                    # only runs if `gate` chose this id

  - id: archive
    type: shell
    cmd: "echo done > /var/log/release/last.log"
    on_failure:
      planner: gpt-4o-mini
      budget: 3                   # max LLM mutations applied to this task
    depends_on: [ship]

If archive fails because /var/log/release/ doesn't exist, the planner can INSERT_BEFORE a mkdir task. The graph grows by one node, the original retries, the run finishes green.

Task types

shell, llm, file_read, file_write, decision.

decision nodes constrain the LLM to picking one of the declared branches. Downstream tasks gate on the choice via when:. The unchosen branch ends in a gated_out state that doesn't poison its dependents.

Mutation grammar

Verb	Effect
RETRY	re-run unchanged
PATCH	modify one field (cmd/path/content/prompt/model), retry
INSERT_BEFORE	inject a shell remediation task, then retry
ABORT	give up

Anything else from the planner is logged and treated as ABORT. Per-task budgets cap the number of mutations.

Audit log

Every state change appends to events.jsonl:

run_started
task_started        task=archive origin=yaml
task_retry          task=archive rc=1
planner_invoked     task=archive budget=3
mutation_applied    verb=INSERT_BEFORE new_task=mkdir_dir cmd="mkdir -p ..."
task_started        task=mkdir_dir origin=planner
task_succeeded      task=mkdir_dir
task_started        task=archive origin=yaml
task_succeeded      task=archive
run_finished        succeeded=3 failed=0 skipped=0 gated_out=0

Tasks the planner created carry origin=planner along with the mutation id and parent task. You can always answer why is this task in the graph?

Limits

• YAML parser is a 2-space-indent subset; no block scalars (|, >), anchors, or multi-doc.
• Tasks share state via files on disk, not templating.
• Run-once CLI; no daemon, no scheduler.
• Open-ended graph synthesis is deliberately not in v1.