Skip to content

DECISION_LOG.md

Latest commit

 

History

History
38 lines (27 loc) · 4.32 KB

DECISION_LOG.md

File metadata and controls

38 lines (27 loc) · 4.32 KB
Raw

TagTeam → opencode — Decision Log

Append-only. Every entry: ## <ISO date+time> — <Stage>: <one-line summary> then Decision / Evidence / Open question. Never rewrite history; if reversed, add a new entry referencing the old one.

2026-08-12 — Framing: opencode-only rebuild, zero-admin, zero-budget

  • Decision: Rebuild TagTeam on opencode (model-agnostic) on branch opencode. Drop Claude/Codex backends for now. No external admin or budget until a working POC exists — so no Entra/SSO, no Slack webhook, no K8s. Local-account auth (bcrypt) replaces SSO; SQLite (local file) replaces any managed DB. Boss agent orchestrates the build.
  • Evidence: Shash: "focus on opencode… I don't have the leverage to go buy new tools or get admin access until I have a working POC." Repo analysis: wire protocol, session model, invite lifecycle, turn engine, and web client are all agent-agnostic — only server/agent/* + a few UI strings are Claude-specific.
  • Open question: provider/model default — opencode Zen vs pinned Anthropic via opencode. (Deferred to M0 spike; Zen is the working assumption.)

2026-08-12 — Architecture: in-process opencode SDK, keep the shell

  • Decision: Use @opencode-ai/sdk createOpencode() in-process (Approach B). One TagTeam session = one opencode session. Map opencode SSE events onto the existing assistant_delta / tool_activity / assistant_complete frames. Wire protocol, HTTP surface, sessions/turns/protocol modules, and web client stay; server/agent/* is replaced.
  • Evidence: Current design is single-process, no-build, in-memory — in-process SDK preserves that. opencode SDK exposes session.prompt + event.subscribe SSE with text/tool parts that map cleanly to TagTeam frames. A separate opencode serve (Approach A) would add a second process and ops overhead for no benefit at this scale.
  • Open question: Does opencode's SSE actually emit per-token text deltas and paired tool start/end? M0 task 0.8 spikes this first.

2026-08-12 — Auth: local accounts now, OIDC-shaped interface, Entra later

  • Decision: M1 ships local-account auth (bcrypt + httpOnly cookie, users table). The auth interface is provider-agnostic so Microsoft Entra / Google / Okta is a later drop-in PR, not a rewrite. Invite tokens bind to a user identity on redemption.
  • Evidence: Shash has no admin access yet. Local accounts give real identity + accountability with zero infra. The POC's display-name-only model is not company-safe (no accountability, trivial impersonation).
  • Open question: Password policy + session-cookie expiry — security agent owns this in docs/design/security.md.

2026-08-12 — Spike: opencode SSE event shapes confirmed (plan A, in-process)

  • Decision: Use in-process createOpencode({ port: 0 }) — confirmed working end-to-end with provider bigmodel/glm-5.2. The adapter will create one opencode session per TagTeam session and pump the global event.subscribe() stream.
  • Evidence: docs/implementation/spike-opencode-sse.md (confirmed by running, SDK @opencode-ai/sdk@1.18.18). scripts/spike-output.txt holds the raw event stream.
  • Corrections to M0.md §3 (architect must apply):
    1. SDK wraps responses as { data, request, response } — use s.data.id, not s.id. Same for session.messages() / session.abort().
    2. createOpencode({ port: 0 }) returns { client, server } — use client for calls, server.close() to tear down.
    3. No discrete tool_start/tool_end event types. Tools are message.part.updated with part.type:"tool" and part.state.status transitions pending → running → completed|error. Adapter synthesizes tool_start/tool_end.
    4. message.part.delta (with field:"text") is a runtime event type NOT in the SDK's Event TS union — switch on ev.type at runtime; do not rely on the types.
    5. finish:"tool-calls" is NOT a stop — only finish:"stop" ends the turn.
    6. ~80 catalog/plugin events fire at session create — allowlist the ~6 event types we care about; ignore the rest.
    7. Provider retries (session.status type:"retry") are transient — do not treat as fatal.
  • Open question: Should reasoning parts be surfaced as text_delta or dropped? TagTeam's UI doesn't surface them separately. Recommend: route to text_delta (simplest, no UI change). Architect decides.