Agent Switchboard v1.0.2

Straightforward CLI model + reasoning-effort selection, and a smallest-sufficient build rung.

A focused follow-up to v1.0.0. The model/effort fix below was tagged v1.0.1 in source but never shipped as a binary — v1.0.2 is the first release that includes it, so this is the build to download.

If this saves your agent workflow, please star the repo so others can find it.

---

Highlights

1. Pick the model and reasoning effort the obvious way

Choosing which model and how hard it thinks used to be tangled together, and asking for something like "5.5 extra high" could fail or stall. Now they're separate, first-class inputs:

• New effort field on consult_codex, consult_claude, and route_agent_task: minimal | low | medium | high | xhigh — plus natural phrases ("extra high" → xhigh, "ultra"/"max" → the family's top effort). It's passed to the CLI as its own flag (Codex -c model_reasoning_effort=, Claude --effort) and is never appended to the model name.
• Bare family = best available. Say just "codex" or "claude" and you get the flagship model at the highest available reasoning effort (Codex gpt-5.5 / xhigh, Claude opus / max) — no more model-selection stall.
• Specific model is honored verbatim. "send to sonnet 4.6 for implementation", "ask gpt-5.4-mini", etc. work as written, on the CLI (default route).
• Effort phrases are split out of the model text before matching. "5.5 extra high" now resolves cleanly to model gpt-5.5 + effort xhigh — instead of producing an invalid --model "gpt-5.5-codex xhigh" that Codex rejected with a 400.
• No surprise pinning. Auto-saving a topic's default model on every request is now opt-in (remember_model).
• Internals: new shared helper resolve_cli_model_and_effort(); Fable added to the Claude catalog.

Tip: because effort is its own input, you no longer need to overload target_model. consult_codex(target_model="gpt-5.5", effort="xhigh") is the clean form. Your ~/.codex/config.toml defaults (e.g. model, model_reasoning_effort) are respected when you don't override.

2. "Smallest sufficient implementation" rung in the build contracts

The implementation and implementation_plan task contracts now nudge the receiving agent to reach for what already exists before writing new code: prefer the standard library, a native platform feature, or an already-installed dependency over bespoke code or a new dependency.

It is deliberately scoped and guarded:

• Applies only to code-writing task kinds. consult, co_audit, debate, and review are untouched — second-opinion reasoning stays as detailed as before.
• It never licenses dropping required validation, error handling, security checks, or tests, and it must not be used to dispute an approved plan — if the plan looks unsafe or over-built, the agent stops and reports instead of silently trimming.

(This wording was sanity-checked via the broker itself against GPT-5.5 before shipping.)

3. Installer: more reliable Claude desktop detection

Setup now recognizes the Microsoft Store / MSIX "Cowork" Claude build (a registered AppX package) in addition to the %APPDATA%/Claude data dir and the legacy standalone installer — so Store users are no longer falsely detected as "not installed".

---

What's in this release

• agent-switchboard.exe — the self-contained installer and MCP server (recommended download).
• antigravity-agent-broker-bridge-1.0.0.vsix (optional) — the bridge extension (unchanged from v1.0.0; only needed for a manual/by-hand install).
• Source code (auto-attached by GitHub).

Upgrading

Run the new agent-switchboard.exe and choose Install (it backs up and re-registers as before), then restart your MCP server / reload your IDE so the updated broker is live. No data migration; existing ~/.agent-broker state is preserved.

---

Compatibility & notes

• Bridge extension is unchanged at 1.0.0; doctor will not report version drift.
• Everything from v1.0.0 still applies (CLI-default routing, debate, context snapshots, doctor, 36 MCP tools).

License

PolyForm Noncommercial 1.0.0 — noncommercial use allowed with the required notice; commercial use needs a separate written license from FutureisinPast / ChartTrades. See LICENSE.

⭐ If this helped you, please star the repo so others can find it!

Agent Switchboard v1.0.0

A local nervous system for AI coding agents.

Route tasks, run model debates, compact token-heavy context, and return answers across Claude Code, Codex, Gemini, Antigravity, and VS Code using the subscriptions you already pay for. No API keys. No extra billing. Local state.

This is the first public release. One self-contained .exe (or a Python install) wires the broker into every assistant you have, and a built-in uninstaller rolls it all back.

Recommended GitHub repo slug: mcp-agent-switchboard. Everything user-facing — binary, command, and MCP server key — is agent-switchboard; local state stays in ~/.agent-broker.

If this saves your agent workflow, please star the repo so others can find it.

---

What it does

• Cross-agent routing — from any assistant: "ask Claude Opus to audit this," "get Codex's second opinion," "have Gemini draft the plan." The broker delivers it to the right surface and brings the answer back.
• Subscription-only — it drives the assistants you're already logged into. No API keys, no metering.
• Local & private — all shared state lives in one local SQLite file under %USERPROFILE%\.agent-broker. It never scrapes private chat history.
• Token compaction is built in — compressed handoffs, compact context packs, word budgets, per-topic work memory, and retrievable originals instead of dumping entire transcripts.
• Honest model handling — when it can switch the answering model it does (Antigravity via UI automation, CLIs via --model); when it can't (the Codex/Claude extensions), it makes the agent stop and ask you to switch instead of letting a weaker model answer silently.
• Cross-model debate — agent-switchboard debate "<proposition>" runs two assistants (e.g. Codex vs Claude Opus) for N rounds headless on your subscriptions, each keeping real memory across rounds, then a synthesis judge writes a verdict. Pick model + reasoning effort per side; every line is labelled family/model (effort).
• See across chats (context snapshots) — pull a compact snapshot of what another open session knows (objective, plan, touched files, checks, next step), on request — opt-in and local, never silent scraping. Codex and Claude Code are read straight from disk (~/.codex / ~/.claude/projects, scoped to the matching project, redacted + truncated), so the broker returns their recent context immediately with no agent cooperation; it then folds into context packs and topic status so the next model picks it up automatically.
• doctor — agent-switchboard doctor tells you exactly what works on your machine per assistant (CLI + smoke test, extension, CDP, delivery route, reply path, debate readiness), plus a nerve-system view of which surfaces can feed context snapshots (on-disk fast-path vs live bridge vs push-only).
• Smart default routing — Codex/Claude default to the headless CLI; say "in app" for the IDE chat panel; Gemini and Antigravity-hosted models stay on in-app automation.
• 36 MCP tools plus CLI verbs (doctor, debate, status/result/cancel/reap, claude-responses); one dependency-free broker, one bridge extension (v1.0.0).

---

Two ways to install (pick one)

A — Self-contained agent-switchboard.exe (no Python). Download agent-switchboard.exe from this release and run it. One binary installs everything — the bridge extension is embedded — and the same binary is the MCP server (agent-switchboard.exe serve).

B — From source (Python 3.10+). Clone the repo and run install-agent-broker.ps1.

Both register the broker with every assistant detected (Codex, Claude Code, Antigravity, VS Code), install the bridge, back up every file they touch, and ship the same built-in uninstall.

---

Quick start (Windows)

1. Close Antigravity and VS Code. (The installer refuses to run while either is open, so nothing is left half-updated.)
2. Install:
   • Exe: run agent-switchboard.exe → choose Install (or agent-switchboard.exe install).
   • Source: powershell -NoProfile -ExecutionPolicy Bypass -File .\install-agent-broker.ps1
3. Reopen Antigravity / VS Code so the Agent Switchboard Bridge extension activates.
4. Try it in any assistant: "Use Agent Switchboard to ask Claude Opus to review this function."

---

Model selection — what actually happens

Target	Can the broker switch the model?
Antigravity (in-app Gemini/Claude)	✅ Yes — drives the model picker over CDP, sends into the live panel, and the structured reply returns to the broker. Verified end-to-end: "send to Antigravity Gemini 3.5 (High) and reply" auto-switched the model and returned the answer — the only fully programmatic in-app round-trip.
Claude / Codex CLI	✅ Yes — real switch via --model / -m
Claude / Codex extension	❌ No picker API — so the broker enforces it the safe way:

For the extensions, asking for a specific model now prepends a strict guard to the prompt:

"REQUIRED MODEL: Opus 4.8 — state your model; if you're not it, STOP and tell the user to switch."

…and the bridge pops a "select <model>" notification. ❌ Before: a request for Opus could be silently answered by whatever model was active (e.g. Sonnet). ✅ Now: the wrong model refuses and tells you to switch. Saying "get Opus's opinion" in the prompt is detected as a one-off and won't overwrite your topic default.

---

Uninstall / rollback

Run agent-switchboard.exe uninstall (or python setup.py uninstall, or pick Uninstall from the menu). It reverses MCP registration in all four assistants, removes the bridge extension, and removes the installed broker exe. Add --remove-data to also delete ~/.agent-broker. Every edited config was backed up under ~/.agent-broker\setup-backups\ first.

---

What's in this release

• agent-switchboard.exe — the self-contained installer and MCP server (recommended download).
• antigravity-agent-broker-bridge-1.0.0.vsix (optional) — the bridge extension, if you'd rather install it by hand or you installed from source.
• Source code (auto-attached by GitHub).

---

Requirements

• Windows 10/11 (the broker core is cross-platform; the installer/CDP layer is Windows-first today).
• Node.js on PATH — only needed for the CDP helpers (Antigravity model auto-select, Codex/Claude webview submit).
• Python is not required for the .exe; source install needs Python 3.10+.

---

Security & terms (please read)

• Subscription automation, not API. It drives logged-in assistant UIs, optionally via keystroke/CDP automation. This may conflict with a provider's terms and carries account risk — keep automation opt-in and review your providers' terms.
• Unsigned binary. agent-switchboard.exe isn't code-signed, so Windows SmartScreen may warn — choose More info → Run anyway if you trust it, or use the Python install.
• Local debug port is unauthenticated. CDP model auto-selection opens 127.0.0.1:9000; the installer asks before enabling it (default Yes when Antigravity is detected). The launcher patch is written to a durable path and is fully reversible — uninstall restores your original Antigravity shortcuts. Only turn it on if you're comfortable with a local debug port.
• No chat-history scraping. Only authenticated IDE surfaces and shared state you create are used.

---

Known limitations

• Only Antigravity has a fully structured in-app round-trip; the Codex/Claude extensions deliver + notify and return answers via respond_to_request (no native send/complete API).
• Antigravity model auto-select is best-effort UI automation and depends on the debug port + stable UI labels.
• A standalone Gemini CLI is optional and not bundled; Gemini is otherwise reached through Antigravity.
• Context snapshots can be read on disk for Codex and Claude Code only. The Claude desktop app can be registered to push context into the broker, but its chat is stored server-side/in Electron and can't be read on disk, so it can't be passively snapshotted — doctor flags this so the blind spot is visible.

---

License

PolyForm Noncommercial 1.0.0. Noncommercial use is allowed with the required copyright notice. Commercial use requires a separate written license from FutureisinPast. See LICENSE.

⭐ If this helped you, please star the repo so others can find it!