agent-rally-point

Agent Rally Point

Run many coding agents in one repo without them clobbering each other — no server, no scheduler.

Rally is a repo-local coordination room. Agents — Claude Code, Codex, Gemini, Cursor — enter the room, claim the files they're about to touch, hand work off with delivery receipts, and read shared room state, all through one CLI (rally) backed by an append-only fact log in .rally/. There is no daemon to run and no service to reach: the coordination lives in your checkout.

Why it exists

When two agents work the same checkout in parallel, they overwrite each other's uncommitted edits, redo each other's work, and lose track of who decided what. The usual fixes are a human babysitting a queue or a heavyweight orchestration server. Rally is neither — it is a durable, git-friendly record of who is doing what, right now that any agent reads and writes in one command, and that advises, never blocks.

What you get

• Deconfliction before edits. rally check before-write warns when a file you're about to touch is claimed by another live agent — surfaced automatically through a PreToolUse hook.
• Handoffs that are actually received. A handoff waits for target-authored evidence (the receiver posts a receipt / artifact / resolve). Text landing in a pane is not proof; an unacknowledged handoff returns ack_state: "timeout" and a fallback plan.
• Room state on demand. rally room, rally risks, and rally next project current claims, decisions, risks, and the recommended next action from the fact log — no server round-trip.
• Host-neutral by design. One room coordinates Claude Code, Codex, Gemini, and Cursor; each agent passes its own --tool id.

Install

As a Claude Code plugin (recommended):

claude plugin marketplace add tyroneross/agent-rally-point
claude plugin install agent-rally-point@agent-rally-point

Hooks (SessionStart coordination + PreToolUse write-boundary checks) and skills (agent-rally-point, rally-workflows, mini-loop) activate on install. The rally CLI is auto-provisioned on first session — downloaded from GitHub Releases (SHA256-verified) or built from source as a fallback. Nothing else to run.

From source (manual / development):

git clone https://github.com/tyroneross/agent-rally-point.git
cd agent-rally-point
cargo install --path crates/rally-cli

Verify either path:

rally whoami --tool codex --json   # host runtime, room, lead, mission, ack status

The loop

What an agent does each turn — small on purpose:

whoami → enter → ack → next → (if actionable) claim → check before-write → edit
       → verify → say artifact|handoff|resolve|release → next

rally next returns actionable, requires_human, stop_reason, suggested_claims, suggested_commands, and completion — enough for a harness to act on its own without turning Rally into a scheduler or a coding agent. JSON contracts are designed for agents first; every command takes --json.

rally enter --tool codex --json
rally ack   --tool codex
rally next  --tool codex --json
rally check before-write --tool codex --path crates/rally-cli/src/main.rs --strict --json
rally say claim    --tool codex --subject "edit parser" --path crates/rally-cli/src/main.rs --json
rally say artifact --tool codex --subject "implementation complete" --uri crates/rally-cli/src/main.rs --evidence "cargo test" --json
rally say handoff  --tool codex --target claude_code --subject "review docs" --summary "Rally is now primary" --json
rally say resolve  --tool codex --ref <blocker-id> --subject "resolved" --json
rally room --json

Resolve handoff targets from live room state (rally whoami, rally lead show, rally room --json) — never from examples or old logs.

Room signal (v0.1.6)

rally room shows only human coordination risks. System-generated telemetry — unmanaged-agent, duplicate-active-squad-id, binary-drift, external-intake — projects into a separate, subject-deduped system_health bucket (count surfaced as system_health=N), so the risk view stays trustworthy. Read one kind at a time without hand-parsing JSON:

rally risks --json        # human coordination risks only
rally decisions --json
rally artifacts --json
rally claims --json

Managed sessions

Managed sessions make Rally part of normal agent behavior — live delivery into real panes, no setup step:

rally run --backend <auto|tmux|cmux|ptyd>   # start an addressable pane; auto = ptyd if live, else tmux
rally run claude                            # convenience: becomes claude-01, tool claude_code:01
rally inject <session|name|tool> --handoff <event-id> --json

Rally assigns readable per-agent ids from the room: rally run claude becomes claude-01 with tool claude_code:01.

Start here