Unreleased source addition after 0.5.0. OpenCode 1.18.30 is the verified host; DeepSeek Flash is the verified native model. The adapter is model-independent: selecting Pro does not require another RAG installation.
From a retained checkout with its existing Python 3.13+ environment:
uv run rag install --opencode --check
uv run rag install --opencodeInstall on the Mac when T3 on Windows connects to that Mac environment. The loader lives at ~/.config/opencode/plugins/agentic-rag.js. XDG_CONFIG_HOME and OPENCODE_CONFIG_DIR are respected; an explicit --opencode-config-dir PATH wins. Check mode writes nothing. Keep the checkout and virtualenv at the same paths; the loader imports packaged JavaScript and invokes that virtualenv's Python. The bridge explicitly imports Python from the same package root as its JS.
Finish active work, refresh T3's provider status and start a new thread. A cached OpenCode helper may need to become idle or restart before loading the plugin. Ask the model whether it received agentic-rag startup context; skill loading and RAG hook loading are distinct checks. Existing Engineering Method can coexist.
The installer changes only its owned loader and creates missing parent directories. Foreign/modified loader bytes and symlinked destination directories are refused. A config hook adds agentic-rag-ro only if absent, using the reader role. Existing MCP entries, including disabled ones, providers, permissions, model choices and other plugins remain untouched. Automatic hooks use audited checkpoint/queue APIs; they do not grant model or subagent memory-write tools. The existing read-write MCP remains a separate, explicitly authorized surface.
| Native hook | RAG action |
|---|---|
| chat.message | Capture the real prompt and invalidate cached context for that turn. |
| experimental.chat.system.transform | Restore startup/resume context and selective recall as transient system context. Reconcile a completed summary before the next model request. |
| experimental.session.compacting | Persist deterministic checkpoint, enqueue existing asynchronous enrichment, append handoff instructions. |
| session.compacted | Attach matching bounded summary and invalidate cached context. |
| session.idle / idle session.status | Export completed prose/tool names and enqueue debounced mining. |
Child sessions (parentID present) are excluded from automatic context/projection and write paths; their read-only MCP access remains available. Session identity is namespaced as opencode: in RAG. Compaction is matched to the actual compaction request's message ID and manual/auto flag, never merely to the most recent checkpoint. Durable empty boundary records make subsequent enrichment cursors resolvable even across compactions with no new prose.
OpenCode has no reliable SessionEnd hook. Idle is the tail-capture boundary; a process killed before idle can miss its final tail until that session resumes. The same callbacks cover auto compaction, while the native smoke exercised manual compaction; real long-running automatic compaction remains an operational check. Experimental hook names must be revalidated when upgrading OpenCode.
The SDK reads at most 200 recent messages from the active main session. Projection merges them by stable identity into private mode-0600 JSONL files under ~/.agentic-rag/opencode/transcripts. It excludes reasoning, tool arguments/results, synthetic or ignored context, compaction summaries and incomplete assistant messages. Prose is secret-stripped before persistence. Each message contributes at most 16,000 characters and 100 tool names; the file is capped at 8 MiB, omitting oldest records with a logged warning. A single gap exceeding 200 host messages can omit older uncaptured material. This is bounded capture, not a full transcript archive. Canonical knowledge and checkpoints remain in the same PostgreSQL store.
Model-bound RAG context is secret-stripped again. Existing context budgets, project scope, exact-pin priority and selective recall gates remain authoritative. No provider call occurs inside a RAG hook. The existing worker still uses its configured Codex/Claude mining/enrichment provider: selecting DeepSeek for coding does not change that provider or replay historical sessions. Recalled knowledge is sent as context to the coding model selected in OpenCode.
Each SDK request has a five-second deadline; each Python hook process has a 15-second deadline and 256 KiB output limit, with a 4 MiB input bound. A model callback may perform several bounded operations. Errors do not abort the coding session; context failures are visible, and asynchronous failures are shown in subsequent context. Python diagnostics are secret-scrubbed in the existing hook log. SDK/spawn failures use a generic visible warning. No cloud RAG service, model routing, schema migration or extra scheduler is installed.
uv run rag install --opencode --uninstall --check
uv run rag install --opencode --uninstallOnly the exact generated loader is removed. Restart the helper afterward. Modified loaders require inspection; they are never blindly overwritten. To move the installation, uninstall from the old source/interpreter first. --restore is for the other host installers; this target uses owned-file uninstall. The existing AGENTIC_RAG_HOOKS_DISABLE environment switch stops its lifecycle actions.
A later production check used T3 0.0.40 with OpenCode 1.18.30 on the Mac and DeepSeek Flash. After provider refresh and a new T3 thread, the model recognized injected canonical context and completed an actual agentic-rag-ro_memory_search call with a one-time read approval. Native T3 tool events confirmed completion and a result payload. The user authorized cloud processing for this check. This verifies the T3 context/read path; compaction evidence below comes from the separate native OpenCode test, not a T3 compaction run.
See OpenCode evidence. Run Python tests plus node --test tests/test_opencode_plugin.mjs; Node is required for the JS adapter checks. Native testing used the real local OpenCode server and DeepSeek API, with synthetic data in agentic_rag_test. A test-only harness suppressed worker spawning; queue persistence, checkpoint persistence and handoff restoration were real. No synthetic data was written to the production knowledge store. A wheel must contain integrations/opencode/plugin.mjs and support check mode after an isolated installation.
Official references checked 2026-09-11: