feat(mcp): reduce redundant diff resolution across multi-tool workflows

What problem does this solve?

A common agent workflow runs multiple operations on the same diff: recap to understand what changed, review to find issues, commit-draft to compose the message. Each call independently resolves the source (runs git diff, computes the digest, applies budget truncation). On a large repo this is 3x the git work and adds latency.

Proposed solution

Two complementary approaches (either or both):

Option A: Accept a pre-resolved digest to skip re-resolution

If a call returns meta.digest: "sha256:abc...", a subsequent call could pass source: { digest: "sha256:abc...", ... } as a hint. If the current diff still matches that digest, skip re-resolution and reuse the cached text. If it doesn't match (working tree changed between calls), resolve normally.

Option B: Composite operation

A new coco_batch tool (or operations array in the input) that runs multiple operations on a single resolved source:

{
  "source": "staged",
  "operations": ["review", "commit-draft"],
  "options": { "conventional": true }
}

Returns an array of envelopes, one per operation, all sharing the same meta.digest.

Trade-offs

• Option A is simpler to implement but still requires N tool calls
• Option B reduces round-trips but adds schema complexity
• Both maintain the read-only safety model

Acceptance criteria

• At least one approach implemented
• Shared diff resolution saves measurable git work (test with a large repo)
• Response envelopes include the shared digest for downstream chaining
• No regression in single-operation calls

Start by tracing how tool calls resolve the source and construct response envelopes, then compare the digest-hint and composite-operation options. Use the acceptance criteria to test repeated operations on a large repository, verify shared meta.digest values and measurable savings, and confirm single-operation calls remain unchanged.