feat(mcp): graph-backed sections in get_context with structured handoff and VOR filter

[ManSio]

Summary

get_context v2 — the tool now assembles deterministic, graph-backed context sections (no LLM), following the Anthropic "code execution with MCP" / "writing effective tools for agents" guidance (tool consolidation, meaningful context, token efficiency).

New sections (all degrade gracefully to absent when data is unavailable)

• dataflow — ASSIGNED_FROM/TO chains with condition_path for local variables of the target file (candidate variables derived deterministically from local assignments in the source window; verified against the live graph: chain entries carry condition_path).
• writes — WRITES edges of the target symbol from PropertyGraph.
• receipts — recent ActionReceipts filtered by the target file.
• tests — affected tests derived from impact-analysis affected_files.

Structural improvements

• Structured handoff: _section_symbols now carries meta (file_path/line/symbol/affected_files) consumed by source/git and the new sections; regex text-parsing demoted to a fallback path (root cause of fragility identified in review).
• VOR filter in memory section: only VERIFIED/ACTIVE (or status-less) nodes are shown — REFUTED/INCONCLUSIVE/STALE are no longer surfaced (protocol §9.5 verify-on-read).
• Lazy SymbolIndexAdapter init; new sections wired into intent profiles (modify/debug/prepare_change/verify_change/test) with budgets/priorities; TOKEN_LIMIT kept at 2000.

Evidence

• 12 new tests (tests/test_context_sections.py): 14 passed together with existing context_tool tests.
• Full suite on this branch: 1659 passed / 9 skipped / 91 deselected, exit 0 (skips are env-conditional, zero failures).
• ruff clean; all 9 pre-commit checks passed (verify_diary, stale_detector, negative_controls, architecture_linter, ruff_gate, ...).
• Verified live-graph probes: variable flow chain entries expose condition_path; receipts ledger empty → section omitted.

Notes / follow-ups

• WRITES edges proved empty at method level on the current graph build — section degrades to omitted; revisit when more WRITES edges are indexed.
• Follow-up epic (separate PR): progressive disclosure tool-catalog search (detail=name|desc|full), narrow default toolset, tool-health consolidation report, E-11 "Bundle vs Relay" eval.
• Research basis: anthropic.com/engineering/code-execution-with-mcp, anthropic.com/engineering/writing-tools-for-agents, arxiv.org/abs/2605.24660 (tool-count accuracy impact).

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: 8ed94cfc-be55-4d4d-9877-208df596ffdf

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}