Skip to content

opencode.md

Latest commit

 

History

History
84 lines (64 loc) · 3.29 KB

opencode.md

File metadata and controls

84 lines (64 loc) · 3.29 KB
Raw

opencode

Probed against 1.18.18.

Seam

opencode documents that a custom tool whose filename matches a built-in takes precedence over it. That makes opencode the only harness where reach needs no workaround at all — no envelope parsing, no denied tools, no mirror. The model still sees read and bash, and they act on the target.

Confirmed by inspecting the shipped binary and the published plugin package:

  • built-in tool names present: bash, read, write, list, glob, patch
  • plugin hooks present: tool.execute.before, tool.execute.after, chat.message, chat.params, permission.ask
  • tool directories: .opencode/tool/ (project), ~/.config/opencode/tool/ (global)
  • the tool() helper signature, taken from @opencode-ai/plugin@1.18.18:
tool({
  description: string,
  args: ZodRawShape,          // note: a raw shape, not a ZodObject
  execute(args, ctx): Promise<ToolResult>,
})
// ToolResult = string | { title?, output, metadata?, attachments? }
// tool.schema is the zod namespace, so no separate zod import is needed

Install

reach up ssh://box/srv/app
reach opencode install          # writes ./.opencode/tool/*.ts
reach opencode install --global # or ~/.config/opencode/tool/
reach opencode uninstall        # restore the built-ins

reach generates the tool files rather than shipping them to be copied. Each is self-contained: opencode treats every file in the tool directory as a tool, so a shared helper module placed there would itself be loaded as one.

Verification status

Partially verified. Do not assume this path works end to end.

What was confirmed on this machine:

  • The generated files are accepted by opencode: it loads the project config and the tool directory and proceeds to model selection without a parse or registration error.
  • The reach fs verbs the tools invoke (read, write, ls, grep, glob) are covered by unit and integration tests against a real sshd, including binary-safe round trips.

What was not confirmed:

  • That opencode routes its built-in tool names to these files and executes them against the target in a real turn. One run using the mock model server (test/mockmodel/server.py) did produce a second conversation turn after the tool call, but the tool result could not be captured, and subsequent runs hung during opencode's initialisation before reaching the model. No valid model credential was available to complete the round trip another way.

Until an end-to-end run exists, treat the opencode adapter as untested code written against a documented API — which is exactly the standard of evidence the rest of this project avoids relying on.

Mock model server

test/mockmodel/server.py speaks enough of the OpenAI chat-completions streaming protocol to tell a harness "call this tool with these arguments" and report what came back. It exists so that agent-level tests can run in CI and for contributors without an API key.

python3 test/mockmodel/server.py --port 8911 --tool read \
    --args '{"filePath":"/srv/app/README.md"}'

Point any harness that accepts an OpenAI-compatible base URL at it. The server's protocol output is verified directly with curl in test/e2e/mockmodel_test.sh; what remains unproven is opencode's side of the exchange, not the server's.