Skill Finder suggests Codex injected AGENTS.md context as a custom skill opportunity

[samuelkahessay]

Description

Skill Finder can surface Codex session-start injected context as a custom skill opportunity.

Example suggestion seen locally:

• load-agents-instructions
• Reason: "Repeatedly loads repo-specific AGENTS.md instructions before work, which is a consistent setup step across repositories."
• Examples begin with # AGENTS.md instructions for ... <INSTRUCTIONS>

These are not user prompts or repeatable developer workflows. They are harness-injected instruction/context payloads stored in Codex JSONL as user-role records.

Steps to Reproduce

1. Use the extension against Codex session logs that include session-start injected AGENTS.md context.
2. Open Skill Finder.
3. Run Analyze.
4. Observe a custom skill opportunity based on loading AGENTS.md instructions rather than on an explicit user task.

Expected Behavior

Skill Finder should exclude injected instruction/context payloads from repeated workflow clustering and AI triage.

Actual Behavior

Injected Codex context can be clustered as repeated prompts and triaged into a custom skill opportunity.

Evidence

Local Codex logs contain records like:

{
  "type": "response_item",
  "payload": {
    "type": "message",
    "role": "user",
    "content": [
      {
        "type": "input_text",
        "text": "# AGENTS.md instructions for /path/to/repo\n\n<INSTRUCTIONS>..."
      },
      {
        "type": "input_text",
        "text": "<environment_context>...</environment_context>"
      }
    ]
  }
}

Because these are recorded as role: "user", the current Codex parser can expose them as messageText, and WorkflowAnalyzer does not currently filter this pattern as noise.

Proposed Fix

Filter these payloads at parse time in parser-codex.ts, and defensively filter them in analyzer-workflows.ts in case similar injected context reaches workflow clustering from another parser.

Local patch tested:

• parser-codex.ts: ignore messages starting with # AGENTS.md instructions for ...\n\n<INSTRUCTIONS> and standalone <environment_context>...</environment_context> payloads.
• analyzer-workflows.ts: treat the same patterns as noise.
• Added regression tests in parser-codex.test.ts and analyzer-workflows.test.ts.

Targeted test run passed locally:

npm test -- src/core/parser-codex.test.ts src/core/analyzer-workflows.test.ts

Result: 2 test files passed, 21 tests passed.

Extension Version

0.1.0, locally built from source.

VS Code Version

1.123.2

Operating System

WSL/Linux remote extension host on Windows.

[notsalate]

Same root cause also affects Anti-Patterns / Repeated Prompts, because repeated-prompts uses request.messageText, and Codex injected AGENTS.md context can become messageText