You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Research: full-text workspace search after OpenCode v2 (no v2 content-search route) — leads to a MADR/PLAN pair #34
The phone's workspace browser can search file contents today for OpenCode sessions. OpenCode v2 has no route for that, so moving the OpenCode provider to v2 drops it. This issue tracks the research needed to bring it back, and the MADR/PLAN pair that research should end in.
Until then, full-text search is a named regression for OpenCode:
OpenCode 1.x answers text search with GET /find and file search with GET /find/file (internal/provider/opencode/workspace.go:301-361). kilo, a 1.x-shaped fork, keeps both.
What OpenCode v2 offers
From the source at b44b7164ba ("release: v2.2.3"), packages/protocol/src/groups/fs.ts:
fs.list (/api/fs/list): direct children of a directory;
fs.read (/api/fs/read/*): a file's bytes;
fs.find (/api/fs/find): "Find recursively ranked filesystem entries", which is file names, not contents.
Upstream. Does any OpenCode v2 surface search contents? Check the experimental routes, the ACP tools (grep), and the changelog after 2.2.3. Is a content-search route planned? Would upstream take one?
The other engines. What does each give mcremote for content search: grok (ACP), codex (fuzzyFileSearch is file names; is there a content search?), kilo (1.x /find), gobble?
mcremote's own bounded search under the session root, if no engine route exists:
bounds: result count, bytes read, time, file size; binary files skipped;
scope: the session root only; no symlink or .. escape; .gitignore honoured or not;
engine: a Go walker with regexp, against an external tool such as ripgrep where installed;
Windows: paths, line endings, case;
where it runs: on the daemon host, which is the machine the agent edits.
The phone side. Does today's search UI assume anything that only /find gave, such as match offsets, context lines, or result caps?
Security. A daemon-side search reads files with the daemon's permissions, not the agent's. Is that acceptable inside the session root (0069's permission model)?
Deliverable
A MADR/PLAN pair choosing between an upstream route, an engine-provided search, and mcremote's own bounded search, for OpenCode first and for every provider that lacks one. The pair records the research above as its evidence.
Not this issue
Search by file name, which v2's fs.find covers.
Any other dropped surface in 0214 D11 (share, the todo panel, revert and undo, ! shell, LSP and formatter status).
Activity
docs(decisions): accept 0214 and add its PLAN, complete 0226
feat(opencode): serve workspace, diagnostics and diff over ACP