BUG: strategy expansion discarded — raw body stored instead of expanded workflow (P0)

[addadi]

Bug: strategy expansion discarded — raw body stored instead of expanded workflow

Severity: P0 — all strategy features (sequential, parallel) are broken by this
Filed: 2026-07-28
Sibling: nullboiler-62s (parallel+reduce type error), nullboiler-9x1 (sequential chain order)

Summary

When submitting a workflow with "strategy": "sequential" via POST /runs, the expandStrategy function runs correctly — it adds depends_on between consecutive steps via applyChain. But the expanded steps are discarded. The API stores the raw HTTP request body as workflow_json, not the expanded version.

The engine then reads the raw body at runtime, sees no depends_on on any step, synthesizes all steps as graph roots (no edges between them), and dispatches them all as independent nodes.

Root cause

src/api.zig:585-586:

// Insert run — workflow_json = the original body (frozen snapshot)
ctx.store.insertRun(run_id, idempotency_key, "running", body, input_json, callbacks_json) catch {

body = raw HTTP request body. The expanded steps from expandStrategy (line 537, effective_steps) are used ONLY for step DB prepopulation (line 613). They are NOT serialized back into the stored workflow_json.

Evidence

Submitted workflow with "strategy": "sequential" and 3 steps. DB query:

SELECT workflow_json FROM runs WHERE id='05dcb8a3-...';

Result: original JSON with "strategy": "sequential" still present, all steps have no depends_on. The applyChain output was lost.

Fix

Build the workflow JSON from expanded steps (with depends_on), not from raw body. The expanded workflow should:

1. Drop the strategy field (consumed by expansion)
2. Use effective_steps (with depends_on from applyChain)
3. Preserve input, callbacks, and other top-level fields

Code location: src/api.zig around line 585. Replace body with a rebuilt JSON that uses effective_steps.

Acceptance

• Stored workflow_json has depends_on on steps when strategy: "sequential" is used
• Stored workflow_json does NOT have strategy field (consumed)
• Sequential chain enforced: step B doesn't dispatch until step A completes
• Manual depends_on (no strategy) still works (regression check)
• Parallel strategy + reduce works (regression check for nullboiler-62s)