bug-fix-pipeline

bug-fix-pipeline

Example AgentFlow workflow demonstrating:

• Loop with session persistence — fix → eval retries up to 3× in shared conversation context
• HITL checkpoint — fixAgent pauses for human approval before applying a patch
• Typed ctx access — CtxFor<WorkflowTasks, "summarize"> for type-safe output access
• Test harness — full pipeline test with createTestHarness and mock agents

Pipeline

analyze ──→ fixLoop (fix → eval, up to 3×) ──→ summarize

1. analyze — scans a repo path for issues, returns a list
2. fix — generates a patch for the first issue (with HITL checkpoint); on retry, receives the previous patch as context
3. eval — scores the patch; if satisfied === false, loop retries
4. summarize — writes a final report using the original issue list and fix result

Running with mock CLIs

The __mocks__/ directory contains shell scripts that simulate claude and codex CLIs without network calls. Prepend the directory to PATH to use them:

PATH="$PWD/__mocks__:$PATH" agentwf run workflow.ts

Or use the smoke script:

bun run smoke

Running tests

bun run test

The test suite uses createTestHarness to mock all agents in-process — no subprocess or network calls needed.

Key patterns

Loop context

loop.input returns an object whose keys are merged into the inner task context. The fix and eval tasks access ctx["issue"].output to get the issue passed from the outer analyze result.

On retry (iteration ≥ 2), ctx["__loop_feedback__"].output holds the previous iteration's full output map, allowing fix to surface what didn't work.

Session persistence

fix and eval share fixSession = sessionToken("fix-context", "claude"). With context: "persistent", the runner reuses the same conversation handle across loop iterations so the model retains memory of its prior attempts.

HITL bypass in tests

fixAgent has hitl: { mode: "checkpoint" }. In tests, this is bypassed by wrapping the workflow with an auto-approving onCheckpoint hook:

const workflowForTest: WorkflowDef = {
  ...(workflow as WorkflowDef),
  hooks: {
    onCheckpoint: (_taskName, _message) => Promise.resolve(true),
  },
};

Type-safe ctx

CtxFor<WorkflowTasks, "summarize"> extracts the output types of all dependsOn keys. Because input function parameters are structurally typed (not generic), use as unknown as Ctx to safely narrow:

input: (ctx: unknown) => {
  type Ctx = CtxFor<WorkflowTasks, "summarize">;
  const typed = ctx as unknown as Ctx;
  return { originalIssues: typed.analyze.output.issues, ... };
},