feat(team-mode): team_read_messages pull tool + top-level team_create args

[harshav167]

Summary

• Adds team_read_messages so a team lead can pull its own inbox bodies on demand — the biggest team-mode blocker was that team_status exposes only unread COUNTS and the injector only surfaces messages passively on the lead's next turn (and consumes them).
• Makes team_create accept the natural top-level { name, members } shape (folded into inline_spec), so agents stop hitting "exactly one of teamName or inline_spec" on the obvious call.

Changes

• team_read_messages({ teamRunId, sinceId?, mark_read? }) (new team tool, gated on team_mode.enabled): resolves the caller's member identity via resolveTeamRuntimeDetails (same path as team_send_message), reads its inbox via the existing listUnreadMessages oldest-first, optional sinceId returns only messages after that id, and by default acks read messages via ackMessages so they are not re-injected (mark_read:false to peek). Registered in tool-registry-team-tools.ts, tool-registry-factories.ts, and the tools barrel. The injector path is untouched.
• team_create top-level coercion: fold top-level spec fields (name/members/…) into inline_spec during arg preprocess when neither teamName nor inline_spec is supplied (leadSessionId preserved). Tool description + usage string updated to advertise all three shapes.

Testing

bun run typecheck   # 0 errors (tsgo, all packages)
bun test            # touched suites green; the 22 failing are pre-existing env failures
                    # (codex-installer / dist-bundle node-load) unrelated to this change

• read-messages.test.ts (4): bodies oldest-first + ack, sinceId filter, mark_read:false peek, empty inbox.
• lifecycle-top-level-args.test.ts (5): top-level coercion, leadSessionId preserved, inline_spec untouched, teamName path, empty still errors.
• Team-mode tool regression: 80/80. Registration verified by direct record build — createTeamModeToolsRecord returns 13 tools including team_read_messages.
• Isolated-XDG opencode boot; real ~/.local/share/opencode/opencode.db session count unchanged.

Related Issues

Team-mode tooling friction observed in a real hyperplan run:

• fix: correct OpenCode docs link (404 -> valid URL) #3 (lead can't read message bodies) / docs: improve README structure for better readability #4 (session_read hides tool args) / Recommended global Agents file #9 (lead-relay fidelity loss) → team_read_messages.
• docs: add 'For LLM Agents' setup guide section #1 (team_create schema/description mismatch) → top-level coercion.

Scoped to follow-up PRs (not here): #8/#2 (broadcast wakes idle members) and #5 (idle-wake determinism) touch the prompt-async-gate / live-delivery path and need their own PR with the mandatory duplicate-injection regression tests; #7 (list_models cold-cache) belongs on the per-call-model-override branch where list_models lives; #6 (filename ambiguity) is a hyperplan-skill prompt tweak.

---

Summary by cubic

Adds a team_read_messages tool for on-demand inbox reads, and lets team_create accept top-level name and members. This gives leads access to message bodies and removes a common arg mismatch.

• New Features
  • team_read_messages({ teamRunId, sinceId?, mark_read? }): returns the caller’s unread messages oldest-first, optionally after sinceId. Marks them read by default; set mark_read:false to peek. Gated by team mode; injector path unchanged.
  • team_create: accepts top-level name and members and folds them into inline_spec automatically. Preserves leadSessionId. Tool description and usage updated.

^{Written for commit 7c71765. Summary will update on new commits.}