route-gemini-context

Family: routing Status: Stable

Purpose

The execution arm for tasks that need a context window bigger than Claude's: whole- repo architecture maps, refactor-impact analysis, multi-document synthesis, and large structured-data passes. Gemini returns a synthesized result that Claude then integrates.

Triggers

• "scan the whole repo" / "long context" / "all the files" / "synthesize across these files"
• "gemini analysis" / "map the architecture"
• "trace dependencies" / "what changes if I refactor X"

Inputs

• Task description; the file/directory scope; the specific question

Steps

1. Check Gemini is available: gemini --version. If absent, tell the user and fall back to a scoped Claude pass over the most relevant files (narrower, but honest about the limit).
2. Scope the context — name the directories/files. For long media (video/PDF), preprocess first (cap duration/pages) so you don't blow the context.
3. Invoke Gemini with a focused ask and a required output format:

   gemini -p "Map the architecture. Return a file:line list of where X happens. Cap 800 words." @./src

   Prefer the adjacent cc-gemini-plugin (it packages directories) when available.
4. Demand citations — file:line, page numbers, or timestamps — not a flat summary, so the result is verifiable.
5. Integrate the synthesis into the conversation; spot-check key claims against the actual files before acting on them.

Outputs

• Gemini's synthesized analysis (with file:line / page / timestamp citations)
• A short note of what context was sent

Dependencies

• gemini CLI, or the adjacent cc-gemini-plugin (optional — adds a context window far larger than Claude's for whole-repo passes; free path: a scoped Claude pass over the most relevant files, narrower but honest about the limit).

Notes

Use when the task genuinely needs breadth. For a surgical lookup in a known file, use Claude directly — Gemini's strength is breadth, not precision. Verify before acting (recalled/long-context output can drift).