Persistent multi-turn opencode backend: one `opencode serve` HTTP server per BooCoder process, one opencode session per BooCode session (resumed on switch-back), single SSE read loop demuxed by session id. - backends/opencode-server.ts: AgentBackend implementation — spawn with waitForReady, session.next.* SSE event translation (text/reasoning/tool deltas), Paseo-ported reasoning dedup (streamedPartKeys), promptAsync fire-and-forget settled by session.idle, per-turn inactivity watchdog (180s) + reconnect reconciliation via session.messages, stale-session guard (crashed-not-resumed + config_hash fingerprint on model). - dispatcher.ts: opencode routes to pool backend (ensureSession→prompt); per-session concurrency Map replaces global running boolean (1.9); model coalesce (empty→DEFAULT_MODEL) + llama-swap/ prefix for opencode; diff-supersede (DELETE+INSERT pending_changes by session, stamp agent). - worktrees.ts: ensureSessionWorktree (session-keyed, captures base_commit, persists to session_worktrees); diffWorktree gains optional baseRef. - agent-probe.ts: mergeLlamaSwap branch fetches /v1/models, prefixes with llama-swap/, populates opencode's available_agents.models (was 0). - provider-snapshot.ts: export fetchLlamaSwapModels for probe reuse. - schema.sql: session_worktrees + agent_sessions tables (Phase 0) + config_hash column on agent_sessions, pending_changes.agent column. - package.json: @opencode-ai/sdk ~1.15.0 (resolved 1.15.12). Known Phase 1 limitation: single SSE stream scoped to most-recent session's directory; concurrent opencode sessions in different worktrees collide (warning logged, watchdog prevents hang). Phase 2 moves to per-session SSE. Smoke 1 verified: two turns in one session, both produce real tokens, same agent_session_id reused, same server port, turn 2 is 9x faster (no spawn). goose/qwen/claude paths untouched (runExternalAgent md5 identical). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
5.8 KiB
5.8 KiB