Other agents

If your agent isn’t on the supported harness list, you can still wire PlanBridge in yourself. Any agent that can run a shell command works. Cursor, Aider, opencode, Gemini CLI, Aether, and most other tool-using agents fit this shape.

The setup is one instruction in your agent’s system prompt or rules file: pipe the proposed plan to contextbridge plan, then act on the markdown response on stdout.

Instructions to give your agent

Add something like this to your prompt, system instructions, or rules file (.cursorrules, AGENTS.md, project README, wherever your agent picks up persistent guidance):

Whenever you produce a plan, pipe it to `contextbridge plan` on stdin and read
the markdown returned on stdout. The response will tell you whether your plan
was approved or needs revision and what to change. Follow the instructions in
the response. If the command exits with a non-zero status, stop.

The agent runs something like:

agent shell

echo "$plan" | contextbridge plan

Your browser opens to the review UI. You annotate or approve. The markdown summary lands on stdout for the agent to read.

Open a file or arbitrary content

PlanBridge is not limited to plans. The same binary opens any markdown for review through contextbridge open, so you can give precise feedback on a file or on your agent’s last message, not just at plan time.

agent shell

contextbridge open path/to/spec.md

Or pipe content the agent already has in hand (a draft, a generated doc, its previous reply):

agent shell

printf "%s" "$content" | contextbridge open

The response shape and exit codes below are identical for plan and open.

Response shape

The agent receives markdown on stdout. It opens with one of two headers:

• # Plan review: approved. The human approved with no changes. Continue with implementation.
• # Plan review: changes requested. The human left feedback. Revise the plan to address every comment, then run contextbridge plan again with the revised plan.

The body of a “changes requested” response contains the human’s annotations as ## Comment (lines X–Y) sections, each with the relevant plan excerpt quoted and the human’s comment in a blockquote. A ## General feedback section appears for comments not anchored to a specific line.

The markdown carries its own agent-facing instructions (MUST revise, do not start implementing until approved, etc.). Most agents act on them directly. If yours doesn’t, paste the snippet above into its prompt.

Exit codes

Code	Meaning
0	Submission written to stdout
1	Input or runtime error (see stderr)
130	Interrupted (Ctrl-C) before the user submitted

Agents should treat 130 as “the user wants out” and stop the loop.

For the full flag and subcommand surface, point your agent at the CLI reference.