chore(skills): retire code-graph-context (CGC) - #45
Conversation
The code-graph-context MCP plugin is retired across all harnesses (Claude Code, Codex, OpenCode, Antigravity, desktop). Keeping the plugin running cost more maintenance effort than it returned, so structural code intelligence now runs on native harness tools. skills/ is a distribution source (README: `cp -r skills/* ~/.claude/skills/`), so leaving CGC here would reintroduce it on the next install. - audit: drop the CGC index gate and stats; dead code and complexity now come from Grep enumeration and Codacy's Lizard metric. - prime: drop the hasCGC docker probe and the Step 2.2 CGC index block. - discover, chat: drop CGC from allowed-tools; fold structural search into the Grep tier. - start: drop CGC from the "no indexing" rules. - .cgcignore: delete — CGC-only config, now dead. Search order is now claude-context (semantic) then Grep (structural/exact).
WalkthroughThe PR removes CGC references from skill tools, indexing workflows, code-health checks, search guidance, and report templates. It shifts structural discovery to grep/glob and keeps claude-context, Milvus, Codacy, Context7, and code-oracle guidance. ChangesIndexing and search workflow migration
Estimated code review effort: 2 (Simple) | ~10 minutes Suggested reviewers: 🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
Up to standards ✅🟢 Issues
|
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d236978cc6
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| > **Retired 2026-07-27 — `code-graph-context` (CGC).** This step used to index and health-check | ||
| > a second, structural graph backend. That backend was removed from every harness in favor of | ||
| > native harness tools. Do not re-add it or reference `mcp__code-graph-context__*`; those tools | ||
| > are unavailable. |
There was a problem hiding this comment.
When a Claude user installs the repository's raw assets, commands/audit.md:33-34,83 and commands/prime.md:135 still request CGC, while the dispatched agents/code-oracle.md:30-36,124-135 explicitly invokes its MCP tools. Since this block now declares those tools unavailable, /audit and /prime full will make failed calls and may omit dead-code and complexity results instead of using the new Grep/Lizard path; update these distributed assets as part of the retirement.
AGENTS.md reference: AGENTS.md:L28-L29
Useful? React with 👍 / 👎.
| Run the codebase search tier protocol (minimum tiers 1-2): | ||
|
|
||
| ### Tier 1: CGC — Structural | ||
|
|
||
| ``` | ||
| mcp__code-graph-context__find_code | ||
| query: "{key functions, components, or modules related to this issue}" | ||
| ``` | ||
|
|
||
| ``` | ||
| mcp__code-graph-context__analyze_code_relationships | ||
| query: "{affected area}" | ||
| ``` | ||
|
|
||
| ### Tier 2: Claude-Context — Semantic | ||
| ### Tier 1: Claude-Context — Semantic |
There was a problem hiding this comment.
When discovery is initiated through the npm MCP's spec-workflow-guide rather than this raw skill, src/tools/spec-workflow-guide.ts:249 still instructs the agent to run CGC, and src/markdown/templates/discovery-template.md:16 repeats that dependency. That independent path never reads this replacement tier, so clients from which CGC was removed still receive impossible tool instructions; update the MCP source and its template too.
AGENTS.md reference: AGENTS.md:L26-L31
Useful? React with 👍 / 👎.
![codacy-production[bot]](https://avatars.githubusercontent.com/in/56611?s=60&v=4){kind=small}
There was a problem hiding this comment.
Pull Request Overview
This PR retires the code-graph-context (CGC) plugin and transitions the project's skills to use Grep and Lizard-based analysis. While the removal of the Docker-based CGC dependency simplifies the environment, it introduces significant performance risks. Specifically, instructing an LLM to grep every symbol repo-wide in medium-to-large codebases will likely exceed tool-calling limits and context windows.
Although Codacy indicates the PR is up to standards, the effectiveness of the new Grep-based logic for dead code detection and structural search remains unverified by automated tests. These changes should be considered high-risk for larger repositories until the strategies are refined for scalability.
About this PR
- The PR lacks automated test cases to verify that the new Grep-based instructions provided to the agents are effective and produce the expected results for dead code and complexity analysis compared to the previous CGC implementation.
Test suggestions
- Verify that the audit skill correctly identifies dead code using the new Grep-based enumeration logic.
- Verify that the prime skill executes successfully without the CGC Docker container present.
- Verify that discover search results for structural queries (callers) are correctly handled by Grep after semantic search.
Prompt proposal for missing tests
Consider implementing these tests if applicable:
1. Verify that the audit skill correctly identifies dead code using the new Grep-based enumeration logic.
2. Verify that the prime skill executes successfully without the CGC Docker container present.
3. Verify that discover search results for structural queries (callers) are correctly handled by Grep after semantic search.
TIP Improve review quality by adding custom instructions
TIP How was this review? Give us feedback
| Focus on: | ||
| 1. Dead code detection (CGC if available) | ||
| 2. Top 5 most complex functions (CGC if available) | ||
| 1. Dead code detection (Grep-derived — grep each declared symbol repo-wide) |
There was a problem hiding this comment.
🟡 MEDIUM RISK
Performing a repo-wide grep for every declared symbol is a high-latency operation for an LLM agent. In larger codebases, this will likely lead to excessive tool usage and context window exhaustion. Suggesting a more targeted strategy (e.g., checking only exports) would be more robust.
| 1. Dead code detection — enumerate declarations, then grep each symbol repo-wide and flag | ||
| those with no reference outside their own definition. Check for dynamic references | ||
| (window.X, string lookups, HTML handler wiring) before calling anything dead. |
There was a problem hiding this comment.
🟡 MEDIUM RISK
Iterating over every declaration and performing a repo-wide grep is computationally prohibitive for an LLM agent and scales poorly. For medium-to-large codebases, this will likely hit tool-calling limits or context window constraints. Consider instructing the agent to focus on exported symbols or use specialized CLI tools (like unused or language-specific linters) if available.
| Take the identifiers Tier 1 surfaced and grep each one repo-wide to find every call site, | ||
| import, and reference. Also use Grep directly for specific identifiers, storage keys, and | ||
| constants you already know the name of. Semantic search ranks by similarity and will not | ||
| return all N references — Grep is what makes the impact list exhaustive. |
There was a problem hiding this comment.
⚪ LOW RISK
Nitpick: The claim that Grep makes the impact list 'exhaustive' is inaccurate. Grep fails to capture dynamic property access (e.g., obj[key]), reflection, or dependency injection, and it produces high noise for common names. It should be described as a fallback for literal reference checking rather than a complete structural solution.
There was a problem hiding this comment.
Pull request overview
Retires the code-graph-context (CGC) tool from the repo-distributed skills by removing CGC-specific checks, allowed-tools entries, and index/health-gate instructions, aligning the workflow around claude-context (semantic) plus Grep/Glob (structural/literal).
Changes:
- Removed CGC tool references from
start,prime,discover,chat, andauditskills (including probes, index/health steps, and report rows). - Reframed “structural search” guidance to use Grep/Glob instead of a graph index.
- Deleted
.cgcignoreas CGC-only configuration.
Reviewed changes
Copilot reviewed 6 out of 6 changed files in this pull request and generated 3 comments.
Show a summary per file
| File | Description |
|---|---|
| skills/start/SKILL.md | Removes CGC mentions from “indexing” guidance and rules. |
| skills/prime/SKILL.md | Removes CGC probe + index/health workflow; documents Grep/Glob structural search as replacement. |
| skills/discover/SKILL.md | Drops CGC from allowed-tools and search-tier protocol; promotes claude-context then Grep/Glob. |
| skills/chat/SKILL.md | Drops CGC from allowed-tools and exploration guidance; routes structural questions to Grep/Glob. |
| skills/audit/SKILL.md | Removes CGC index gate and CGC-based checks; replaces with Grep/Glob-based dead code/complexity guidance. |
| .cgcignore | Deletes CGC-only ignore config file. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| @@ -45,20 +45,13 @@ REMOTE=$(git rev-parse origin/$(git symbolic-ref --short HEAD) 2>/dev/null) | |||
|
|
|||
| ### Pre-flight: Ensure Codebase Indexes Exist | |||
|
|
||
| - **session-rag + mem0 MCP calls only.** No indexing MCPs (Milvus, CGC, Codacy). No agent dispatch. | ||
| - **No code indexing.** Do not trigger claude-context, CGC, or any indexing operation. | ||
| - **session-rag + mem0 MCP calls only.** No indexing MCPs (Milvus, Codacy). No agent dispatch. |
| > **Retired 2026-07-27 — `code-graph-context` (CGC).** This step used to index and health-check | ||
| > a second, structural graph backend. That backend was removed from every harness in favor of | ||
| > native harness tools. Do not re-add it or reference `mcp__code-graph-context__*`; those tools | ||
| > are unavailable. |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 3
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@skills/prime/SKILL.md`:
- Around line 359-360: Update skills/prime/SKILL.md lines 359-360 to require
dynamic-reference analysis before declaring symbols dead, and update lines
475-477 to use AST/code-oracle analysis for call chains, dead code, and
complexity while retaining Grep for literal references. In skills/chat/SKILL.md
lines 67-68, qualify “all structural questions” as literal-reference discovery.
In skills/discover/SKILL.md lines 115-118, qualify “exhaustive” as exhaustive
only for textual matches.
- Around line 359-360: Update the “Dead code detection” guidance in SKILL.md to
account for dynamic references before reporting symbols as dead, matching the
safeguards used by the audit skill for window.X access, string-based lookups,
and HTML event handlers. Keep the existing Grep-derived check and complexity
item unchanged.
- Around line 337-340: Update skills/prime/SKILL.md lines 337-340 so Phase 2
Step 2.1 waits for claude-context indexing to reach indexed before Phase
3/report assembly, and handles indexfailed as an error; retain Codacy execution
in parallel where possible. Apply the same polling and indexed/indexfailed
handling after degradation-triggered re-indexing at skills/prime/SKILL.md line
764.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: ASSERTIVE
Plan: Pro
Run ID: 7751673d-9f45-4d70-90ae-3b1a917e2344
📒 Files selected for processing (6)
.cgcignoreskills/audit/SKILL.mdskills/chat/SKILL.mdskills/discover/SKILL.mdskills/prime/SKILL.mdskills/start/SKILL.md
💤 Files with no reviewable changes (1)
- .cgcignore
| Kick off the claude-context indexing check immediately — it runs while everything | ||
| else gathers data. By the time we reach the report, index results are ready. | ||
|
|
||
| Run Phase 2 Steps 2.1, 2.2, and 2.3 (claude-context, CGC, Codacy) in parallel NOW. | ||
| Run Phase 2 Steps 2.1 and 2.3 (claude-context, Codacy) in parallel NOW. |
There was a problem hiding this comment.
🩺 Stability & Availability | 🟠 Major | ⚡ Quick win
Prime must await claude-context indexing completion. Starting or triggering indexing without polling for indexed or indexfailed allows report generation to race the index job.
skills/prime/SKILL.md#L337-L340: wait before Phase 3/report assembly.skills/prime/SKILL.md#L764-L764: apply the same wait/error handling after degradation-triggered re-indexing.
📍 Affects 1 file
skills/prime/SKILL.md#L337-L340(this comment)skills/prime/SKILL.md#L764-L764
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@skills/prime/SKILL.md` around lines 337 - 340, Update skills/prime/SKILL.md
lines 337-340 so Phase 2 Step 2.1 waits for claude-context indexing to reach
indexed before Phase 3/report assembly, and handles indexfailed as an error;
retain Codacy execution in parallel where possible. Apply the same polling and
indexed/indexfailed handling after degradation-triggered re-indexing at
skills/prime/SKILL.md line 764.
| 1. Dead code detection (Grep-derived — grep each declared symbol repo-wide) | ||
| 2. Top 5 most complex functions (estimated CCN, or Codacy's Lizard metric if available) |
There was a problem hiding this comment.
🎯 Functional Correctness | 🟠 Major | ⚡ Quick win
Grep cannot replace complete structural analysis. The affected instructions treat textual search as exhaustive for callers, call chains, dead code, imports, and complexity, which can miss aliases, dynamic references, generated code, and non-textual relationships.
skills/prime/SKILL.md#L359-L360: add dynamic-reference handling before declaring symbols dead.skills/prime/SKILL.md#L475-L477: use AST/code-oracle analysis for call chains, dead code, and complexity; retain Grep for literal references.skills/chat/SKILL.md#L67-L68: qualify “all structural questions” as literal-reference discovery.skills/discover/SKILL.md#L115-L118: qualify “exhaustive” as exhaustive only for textual matches.
📍 Affects 3 files
skills/prime/SKILL.md#L359-L360(this comment)skills/prime/SKILL.md#L475-L477skills/chat/SKILL.md#L67-L68skills/discover/SKILL.md#L115-L118
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@skills/prime/SKILL.md` around lines 359 - 360, Update skills/prime/SKILL.md
lines 359-360 to require dynamic-reference analysis before declaring symbols
dead, and update lines 475-477 to use AST/code-oracle analysis for call chains,
dead code, and complexity while retaining Grep for literal references. In
skills/chat/SKILL.md lines 67-68, qualify “all structural questions” as
literal-reference discovery. In skills/discover/SKILL.md lines 115-118, qualify
“exhaustive” as exhaustive only for textual matches.
🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win
Handle dynamic references before reporting symbols as dead.
The Grep-derived dead-code check omits the dynamic-reference safeguards present in skills/audit/SKILL.md Lines 87-89, so prime may falsely report symbols used through window.X, string lookups, or HTML handlers.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@skills/prime/SKILL.md` around lines 359 - 360, Update the “Dead code
detection” guidance in SKILL.md to account for dynamic references before
reporting symbols as dead, matching the safeguards used by the audit skill for
window.X access, string-based lookups, and HTML event handlers. Keep the
existing Grep-derived check and complexity item unchanged.
Summary
Retires the
code-graph-context(CGC) MCP plugin. It has been removed from every harness — Claude Code, Codex, OpenCode, Antigravity, and the desktop app — in favor of native harness tools. Keeping the plugin running cost more maintenance effort than it returned.This repo matters more than most for the retirement:
skills/is a distribution source (README:cp -r skills/* ~/.claude/skills/), so leaving CGC here would quietly reintroduce it on the next install.Changes
auditprimehasCGCdocker probe and the whole Step 2.2 CGC index/re-index blockdiscoverallowed-tools; fold structural search into the Grep tierchatallowed-toolsand the codebase-exploration liststart.cgcignoreSearch order is now
claude-context(semantic, to discover what to look for) then Grep (structural/exact, to find all of it).Notes
primeis 870 lines vs 370 globally, and still described the Docker/Neo4j-era CGC that hadn't been the deployment for months). I purged CGC in place and deliberately left the rest of that drift alone rather than re-syncing, to keep this diff to the retirement. The drift is worth a separate look..codex/hooks/block-generated-paths.pyhas an unrelated pre-existing modification in the working tree; it is intentionally not part of this PR.Summary by CodeRabbit
Improvements
Documentation