Project: claude-opencode-subagents (source private)  ·  Stack: TypeScript, OpenCode plugin API, Claude Agent SDK, Vitest

Timeline: 22 March to 26 April 2026  ·  244 commits  ·  96 tagged releases (v0.1.2 to v0.1.100)  ·  373 tests in 17 files, CI on lint, knip, typecheck, test, build

The deep dive covers the code and tradeoffs.

Problem

I wanted one agent to own a task while the hands-on work ran in Claude Code. One session that investigates, edits, and checks itself tends to edit early. A fresh session per step loses the context the last one built.

Constraints

• Investigation must not edit files, enforced in code, not in the prompt.
• Engineers keep their Claude session across assignments and process restarts.
• Two OpenCode sessions in one repo must not share engineers.
• Undo in the manager session must also undo the engineers’ work.

Design

cto (OpenCode, reads and delegates, never edits)
 ├─ dispatch_engineer ─> Tom | John | Maya | Sara | Alex   (persistent Claude Code sessions)
 ├─ plan_with_team ────> lead + challenger in parallel ─> write-restricted synthesis
 ├─ run_browser_qa ────> BrowserQA (Playwright, never writes)
 ├─ run_code_review ───> fresh read-only reviewer, focus inferred from changed paths
 └─ git, approval policy, team status, reset

• Modes: explore sessions get write tools and write-like Bash denied in the SDK canUseTool callback.
• State: one team file per CTO session, per-team write queue, atomic rename. Busy is a 15-minute lease, not a lock.
• Context: token, cost, then turn-count estimates drive warnings. Sessions reset only on a real context-exhausted error.
• Signals: implement raises review and verify flags. Verify clears only on VERIFY_STATUS: pass. Commits warn but do not block.
• Undo: a CTO revert sends /undo to each engineer’s Claude session once per later assignment.

What changed and why

Date	Change	Why
23 Mar	Removed per-run git worktrees	One shared worktree, one implementing engineer at a time
26 Mar	Worktree-wide “active team” replaced by one team per CTO session, 17 hours after shipping	Two sessions could share engineers
16 to 17 Apr	Removed the engineer wrappers, the architect/planner wrapper, and the claude bridge tool	Two model hops per assignment; assignments now reach Claude verbatim
23 to 26 Apr	Reviewer moved from Claude to an OpenCode agent on the CTO’s model, fresh session per review	A different model family reviews, without the CTO’s reasoning

Results

• 373 passing tests, including restart recovery, concurrent writes, and 19 undo cases.
• /reflect asks each engineer used in a session what in its instructions misled it.
• Known limits: single-writer is a prompt rule; Bash write detection is a regex list.