Objective
Make Claude and Antigravity agents usable from stock Codex through the MCP agent bridge that spike #949 proved, maintained as a sidecar on main like tools/account-router. The owner decided on 2026-09-26 that this bridge is the route for other providers' agents.
other providers' agents work through sidecars or small patches
The owner's bar from #964 is a richer experience "with callbacks and turns… to make it feel like a real codex agent", not just calling a helper through a CLI.
Finish Line
tools/agent-bridge is on main with tests and an install step. From an ordinary stock Codex CLI session the owner can hand a task to a Claude agent and an Antigravity agent and get their results back as progress and a final report, and can cancel them. The orchestrating model uses the bridge without being told to. The number of upstream files changed is recorded, and stays zero if possible.
Current Status
State: Active; this is the first item for other-provider agents in "Other needs on the fresh start".
Next action: bring tools/agent-bridge from spike branch work/spike-agent-bridge-20260919 (commit b9318ef) to main as a maintained Python sidecar with tests, then run the unprompted-delegation check.
Blocked by: none.
Last verified: 2026-09-26. The spike passed PONG from both providers through stock Codex 0.155.1 and through Claude Code, and cancel was verified (#949).
Scope
- In: the MCP tools
spawn, wait, cancel and list, plus progress reporting; the providers claude and agy; bounded output; cleanup of the whole process group; read-only and write modes; install and config (default_tools_approval_mode="approve" for non-interactive use); tests.
- Out: engine patches unless a measured gap requires one; TUI rendering; usage accounting; account handling.
Acceptance Criteria
Relationships
Validation
- The bridge's tests.
- A PONG check from stock Codex for
claude and agy, recorded with the exact commands.
- One real delegated task from an ordinary session.
Decisions
- 2026-09-26, owner: the MCP bridge sidecar is the route for other providers' agents.
Open Questions
- Can MCP progress notifications or elicitation give the "callbacks and turns" feel, or does that need a small patch?
Objective
Make Claude and Antigravity agents usable from stock Codex through the MCP agent bridge that spike #949 proved, maintained as a sidecar on main like
tools/account-router. The owner decided on 2026-09-26 that this bridge is the route for other providers' agents.The owner's bar from #964 is a richer experience "with callbacks and turns… to make it feel like a real codex agent", not just calling a helper through a CLI.
Finish Line
tools/agent-bridgeis on main with tests and an install step. From an ordinary stock Codex CLI session the owner can hand a task to a Claude agent and an Antigravity agent and get their results back as progress and a final report, and can cancel them. The orchestrating model uses the bridge without being told to. The number of upstream files changed is recorded, and stays zero if possible.Current Status
State: Active; this is the first item for other-provider agents in "Other needs on the fresh start".
Next action: bring
tools/agent-bridgefrom spike branchwork/spike-agent-bridge-20260919(commit b9318ef) to main as a maintained Python sidecar with tests, then run the unprompted-delegation check.Blocked by: none.
Last verified: 2026-09-26. The spike passed PONG from both providers through stock Codex 0.155.1 and through Claude Code, and cancel was verified (#949).
Scope
spawn,wait,cancelandlist, plus progress reporting; the providersclaudeandagy; bounded output; cleanup of the whole process group; read-only and write modes; install and config (default_tools_approval_mode="approve"for non-interactive use); tests.Acceptance Criteria
tools/agent-bridgeis on main with tests, runnable withuv.Relationships
Validation
claudeandagy, recorded with the exact commands.Decisions
Open Questions