nandi/ori · 2026-09-17-launch-for-testing.md

nandi/oripublic⑂ Fork 0
⑂ 34e69b510306161654f26903a269247aefa9c94d
✨ Workspace panel, selectors, previews, desktop app, sandbox template, resizable file tree, light/dark theme 76d62ac · on 34e69b510306161654f26903a269247aefa9c94d · k33g · yesterday
2026-09-17-launch-for-testing.md · 7 lines · 1.6 KBmarkdown Rendered  Raw
BlameHistoryOpen raw
Handoff — 2026-09-17 — ori launched for manual testing
ori runs in the claude-ori sandbox: env -u CLAUDECODE CLAUDE_CODE_EXECUTABLE=/home/agent/.local/share/claude/versions/2.1.275 ./bin/ori --addr 0.0.0.0:8888 (find it with pgrep -f bin/ori), log at /tmp/ori.log, real Claude Code adapter. Stop with pkill -f 'bin/ori ' (mind the pattern: it must not match the shell running pkill).
The user is testing the uncommitted work since 2434cc7 (@ file selector, / skill+command selector, image and draw.io previews, terminal, editor). Their feedback is the input for the next session.
Traps: bind 0.0.0.0 so the sbx ports forward reaches the server; env -u CLAUDECODE remains mandatory; a curl WS probe against /ws hangs on success.
Note for later: @zed-industries/claude-code-acp is deprecated in favour of @agentclientprotocol/claude-agent-acp — the default --agent-cmd, template/Dockerfile and the docs may want migrating (not done, user's call).
Trap (blocking, hit this session): the Zed adapter's bundled Claude Code CLI is 2.1.44 and rejects model: "fable[1m]" from ~/.claude/settings.json (-32603 … issue with the selected model). Workaround: CLAUDE_CODE_EXECUTABLE=<path to current claude binary> in ori's environment (adapter env passthrough → SDK pathToClaudeCodeExecutable). Candidate permanent fixes, user's call: set it in kits/ori/spec.yaml startup + template/Dockerfile (the image has the claude CLI at ~/.local/bin/claude), document it in the run how-to, and/or migrate to @agentclientprotocol/claude-agent-acp whose newer SDK may bundle a Fable-aware CLI (unverified).



1
2
3
4
5
6
7


# Handoff — 2026-09-17 — ori launched for manual testing

- ori runs in the `claude-ori` sandbox: `env -u CLAUDECODE CLAUDE_CODE_EXECUTABLE=/home/agent/.local/share/claude/versions/2.1.275 ./bin/ori --addr 0.0.0.0:8888` (find it with `pgrep -f bin/ori`), log at `/tmp/ori.log`, real Claude Code adapter. Stop with `pkill -f 'bin/ori '` (mind the pattern: it must not match the shell running pkill).
- The user is testing the uncommitted work since 2434cc7 (@ file selector, / skill+command selector, image and draw.io previews, terminal, editor). Their feedback is the input for the next session.
- Traps: bind `0.0.0.0` so the `sbx ports` forward reaches the server; `env -u CLAUDECODE` remains mandatory; a curl WS probe against `/ws` hangs on success.
- Note for later: `@zed-industries/claude-code-acp` is deprecated in favour of `@agentclientprotocol/claude-agent-acp` — the default `--agent-cmd`, `template/Dockerfile` and the docs may want migrating (not done, user's call).
- **Trap (blocking, hit this session)**: the Zed adapter's bundled Claude Code CLI is 2.1.44 and rejects `model: "fable[1m]"` from `~/.claude/settings.json` (`-32603 … issue with the selected model`). Workaround: `CLAUDE_CODE_EXECUTABLE=<path to current claude binary>` in ori's environment (adapter env passthrough → SDK `pathToClaudeCodeExecutable`). Candidate permanent fixes, user's call: set it in `kits/ori/spec.yaml` startup + `template/Dockerfile` (the image has the claude CLI at `~/.local/bin/claude`), document it in the run how-to, and/or migrate to `@agentclientprotocol/claude-agent-acp` whose newer SDK may bundle a Fable-aware CLI (unverified).