feat(mcp): ground get_answer synthesis in the body it serves; slim low-confidence misses

[RaghavChamadiya]

What

Two get_answer quality fixes surfaced by dogfooding the MCP tools.

1. Synthesis now sees the same body it serves

For a question-matched symbol, the hydrator fed the synthesis LLM a 40-line source excerpt while the response inlined the symbol's full 120-line body in symbol_bodies. A docstring-heavy definition (its logic past line 40) was truncated before its answer-bearing code reached the model, so synthesis hedged with a low-confidence "not in the excerpts" answer on the very symbol whose full body the agent then received.

The top matched symbols are now read at the inline-body depth during hydration, so the model reasons over the same body the response serves. Bounded so a class-name flood (every sibling method matching through the parent's qualified name) cannot balloon the synthesis prompt.

2. Cheaper low-confidence misses

A low-confidence miss carried a full per-hit key_symbols dump that duplicated best_guesses and went unused. It is dropped (best_guesses + code_rationale carry the choosing signal), along with the page-excerpt DB round-trip that only fed it. The low-confidence answer branch keeps a grounding block but lean: hits with snippets, symbols pipeable, docstrings stripped.

Also extracts the symbol_bodies ranking into a tested _gather_body_candidates helper.

Testing

• New calibration test asserts the synthesis excerpt reaches the inline-body depth for a docstring-heavy matched symbol, and that a class-name flood does not upgrade every sibling.
• Unit tests for the body-candidate ranking and the leaner payload.
• Retrieval battery green; full mcp unit suite passing.

[repowise-bot]

✅ Health: 7.6 (unchanged)

📋 At a glance
1 file changed health · 4 hotspots touched · 4 new findings introduced · 5 co-change pairs left out · 2 files with recent fix history · 1 dead-code finding.

Files & modules (2)

• packages (3 files)
  • .../tool_answer/config.py
  • .../tool_answer/answer.py
  • .../tool_answer/symbols.py
• tests (1 file)
  • .../mcp/test_answer_payload.py

⚠️ Change risk: moderate _{(riskier than 55% of this repo's commits · raw 9.1/10)}
This change's risk is driven by:

• more lines added than baseline
• more scattered than baseline

🩹 Review priority _{(files here with the most recent bug-fix history — defects cluster, so review these first)}

• .../tool_answer/answer.py — fixed 2× in the last ~6 months
• .../mcp/test_answer_payload.py — fixed once in the last ~6 months

File	Score	Δ	Why
.../mcp/test_answer_payload.py	8.8 → 8.9	▲ +0.1	🔻 introduced primitive obsession

🔎 More signals (3)

🔥 Hotspots touched (4)

• .../tool_answer/config.py — 5 commits/90d, 5 dependents · primary owner: Swati Ahuja (91%)
• .../tool_answer/answer.py — 10 commits/90d, 2 dependents · primary owner: Swati Ahuja (61%)
• .../mcp/test_answer_payload.py — 2 commits/90d, 1 dependents · primary owner: Swati Ahuja (88%)

1 more

• .../tool_answer/symbols.py — 5 commits/90d, 3 dependents · primary owner: Raghav Chamadiya (53%)

🔗 Hidden coupling (3 files)

• .../tool_answer/config.py co-changes with .../mcp/test_answer_calibration.py (3× — 🟢 routine) — not in this PR.
• .../tool_answer/answer.py co-changes with these files (not in this PR):
  • .../mcp_server/tool_symbol.py (3× — 🟢 routine)
  • .../mcp/test_answer_calibration.py (3× — 🟢 routine)
  • .../mcp_server/tool_search.py (3× — 🟢 routine)
• .../tool_answer/symbols.py co-changes with .../mcp/test_answer_calibration.py (3× — 🟢 routine) — not in this PR.

💀 Dead code (1 finding)

• 💀 .../tool_answer/answer.py _json_default (confidence 0.65)

---

📊 Full report · ⭐ Star Repowise · 📥 Install bot · Last updated 2026-07-11 11:19 UTC
_{Silence on a single PR with [skip repowise] in the title · Per-repo toggle on repowise.dev/settings?tab=bot}