Skip to main content

Roadmap

Consultation broker — done in v0.7​

The broker shipped. A peer no longer executes anything to consult another model: it requests a consultation, and the parent Model Peer invocation performs it.

                model-peer
                    |
          +---------+---------+
          v                   v
       Claude               Codex
          |
          | "I'd like a Gemini opinion"
          v
      model-peer broker
          |
          v
        Gemini

That removed the last capability grant. Claude no longer receives Bash at any depth, _delegate is gone, and the uneven provider matrix went with it — Gemini participates in chains again, because participation no longer requires shell access.

Peers remain read-only regardless of consultation depth.

See Peer-chain depth for how requests are validated and how the turn budget guarantees a chain terminates.

Consultation budgets​

Depth is not what actually bounds spend. A model may never appear twice on one path, so with three providers a chain rooted at one model can initiate at most two further consultations — the roster is the real limit, and --depth above 3 has never done anything.

That stops being true at a fourth provider, which is when budgets become a requirement rather than an option:

model-peer review --depth 2 --max-consultations 5 --models claude,codex,gemini
Maximum depth:          2
Maximum peer calls:     5
Maximum calls/model:    2
Cycle detection:        on (already enforced)
Write access:           never
Shell access to peers:  never

Cycle detection is done — the CHANGELOG records which release. A chain like:

Claude
 └─ Codex
     └─ Claude   <- rejected

is refused, because the guard tests membership across the whole chain rather than only against the immediately preceding model. What remains is the per-model and total call budgets above.

MCP​

All three vendor CLIs support stdio MCP servers, and MCP is a real structured tool interface — but it is deliberately not how the broker works. A JSON-RPC 2.0 server means either a JSON parser written in Bash 3.2 or a runtime dependency, and Model Peer is one dependency-free Bash script. It also inverts ownership: a stdio MCP server is spawned by the peer's CLI, so the broker would become a child of the peer rather than part of the parent.

If MCP proves useful, it belongs in a separate package wrapping the public Model Peer interface, not inside the core script.

Same-panel consultation​

review currently denies a reviewer's request for a fellow panel member. Allowing it is conceivable, but only alongside telling the synthesizer which findings are correlated — otherwise the report presents an echo as corroboration.

Not planned​

Applying fixes automatically. Model Peer analyzes; it does not write. Review before autonomy is a design principle, not a missing feature. The primary agent owns the decision, and peer output is evidence rather than a command.

Contributing​

Ideas and disagreement are welcome in Discussions. Security issues should go through private reporting instead.