[teams] Live-verify OpenCode + internal Chimera teammates end-to-end

[0bserver07]

Status

Real-Codex integration is verified for the chimera-team-mcp server: I ran codex exec with the server wired in via -c overrides and confirmed it claims + completes a task end-to-end via real MCP tool calls. See examples/agent_teams/verify_integration.py for the protocol-faithful smoke test that runs on every PR.

Two runtime paths are wired correctly but not yet executed end-to-end:

1. OpenCode — examples/agent_teams/opencode.json follows OpenCode's documented MCP format. Couldn't run in our env because opencode auth list shows 0 credentials.

2. Internal Chimera (chimera code -p) — chimera code already loads MCP servers from ~/.chimera/mcp.json or <workdir>/.mcp.json (see chimera/cli/code.py:712). examples/agent_teams/chimera_mcp.json follows that format. Not run because the test env doesn't have a Chimera provider configured.

Scope

For each runtime, drop the example config in place, run the verify script with the appropriate --cmd template, capture results, fix anything that breaks:

# OpenCode
opencode auth login   # one-time
cp examples/agent_teams/opencode.json <project>/opencode.json
# adapt verify_integration.py or a sibling to use:
#   --cmd 'opencode run "{prompt}"'

# Internal Chimera
cp examples/agent_teams/chimera_mcp.json ~/.chimera/mcp.json
# verify with:
#   --cmd '.venv/bin/chimera code -p "{prompt}"'

Acceptance

• A verify_opencode.py script in examples/agent_teams/ that mirrors verify_integration.py but uses real OpenCode. Documented as opt-in (requires auth).
• Same for Chimera: verify_chimera_native.py.
• Any divergences from Codex's behavior are noted (system-prompt wording, tool-call ordering, sandbox quirks).
• Multi-model robustness check: pick 2-3 small models, run the same prompt, see if they all reliably call recv → claim → complete.

References

• The Codex-only verification is examples/agent_teams/verify_integration.py
• Sandbox-write gotcha hit on Codex; OpenCode and Chimera have different permission semantics — surface what they need

[0bserver07]

Status update (2026-07-25): the internal arm is DONE; only the OpenCode arm remains.

Shipped in 767ab03 + 7098f1f + f29e3f8 (T19) — #149 and #150 are closed. Internal Chimera teammates are live-verified end to end against the chimera-team-mcp server.

The live run earned its keep by finding two real bugs that every hermetic test had missed:

• chimera code -p never loaded MCP tools at all, so a teammate driven headlessly had no team surface.
• The team-policy allowance was keyed on a bare team_ prefix, but MCP tools arrive namespaced as mcp__chimera-team__team_*. Read-only teammates were stranded with every team call denied. A whole hermetic suite passed while production was broken — this is now a standing rule in CLAUDE.md: name-shaped guards must be tested against the name the LOOP sees, not the name you wrote them for, with the corollary that a feature which runs external agents is not verified until a real model has run it.

Remaining scope: the OpenCode teammate arm, blocked on credentials (no OpenCode account available here). Everything else in the original issue is satisfied.

Suggest either narrowing this issue's title to the OpenCode arm, or closing it and refiling that arm as its own issue — happy to do whichever you prefer. Leaving it open and unmodified for now, since which one is your call, not mine.