refactor(backend): extract remote-server shared module from opencode + kilo

[claudiusthebot]

Summary

The OpenCode and Kilo backend server.ts files were 90%+ byte-for-byte duplicates — same MCP registration logic, same chat-visibility rotation, same session-permission ruleset, same provider catalog walk. Every fix had to land in both copies; exactly the recipe for the drift this PR closes.

Extracts common machinery into src/backend/remote-server/:

• client.ts — narrow RemoteAgentClient interface that both OpencodeClient and KiloClient structurally satisfy.
• state.ts — per-backend mutable RemoteServerState<TClient> container (config, gateway port, frontend label, MCP cache, provider cache).
• lifecycle.ts — ensureRemoteServer / stopRemoteServer (lazy spawn + reuse-existing-server via /global/health).
• mcp.ts — chat & plugin MCP registration, tool overrides, disconnect. Owns the chat-visibility rotation (disconnect rival chat servers before registering the current one — upstream's permission rules only block execution, not visibility, so this dance is the only way to hide cross-chat tools from the model's catalog).
• sessions.ts — ensureRemoteSession with Talon's standard permission ruleset (allow this chat, deny others, auto-allow built-ins).
• providers.ts — resolveProviderID with injected backend hooks for bucket priority + provider guessing.

Both backend server modules become thin wrappers:

File	Before	After	Delta
kilo/server.ts	745	357	−388
opencode/server.ts	475	224	−251
backend/remote-server/* (new)	—	962	+962
remote-server-mcp.test.ts (new)	—	372	+372

Net: 856 deleted, 1554 inserted; new code is mostly docstrings + the unit tests. Actual logic deduplicated: ~600 lines.

What's preserved

• All existing exports on kilo/server.ts and opencode/server.ts — the handler / heartbeat / test code that already imports KILO_BASE_URL, parseStoredKiloModelSelection, getRegisteredMcpServerNames, errMsg, the OPENCODE_-prefixed deprecated aliases, etc. all continue to work.
• All behaviour: chat-visibility rotation, heartbeat sentinel exemption, plugin server caching, idempotent mcp.add short-circuit, session expiry → recreate flow, provider catalog scoring heuristic.
• Existing kilo-server.test.ts and the integration/{kilo,opencode}-real-bootstrap.test.ts integration tests pass unchanged.

New tests

src/__tests__/remote-server-mcp.test.ts — 20 unit tests with a recording mock client, covering:

• Chat MCP server name derivation (Telegram numeric, supergroup negatives, Discord snowflakes, empty string).
• isTalonToolID matching for both _ and - variants.
• Visibility rotation: switching chatA → chatB disconnects chatA first.
• Heartbeat sentinel (talon-tools-heartbeat) is exempt from rotation.
• Plugin servers (mempalace, brave-search) stay connected across chat switches.
• Local cache short-circuits redundant mcp.add calls.
• ensureChatMcpServer returns the server name even when the upstream add throws.
• buildToolOverrides produces the correct enable/disable map for multi-chat catalogs.
• Returns undefined when no chat tools matched, only rival chats present, or tool.ids throws.
• disconnectChatMcpServer removes the cache entry on success, doesn't throw on upstream failure.

Full suite: 2228 passing (was 2208), 12 skipped (live tier, workflow_dispatch-only), 0 failing.

Upstream dissection notes

For the design check, shallow-cloned sst/opencode and Kilo-Org/kilocode to ~/.talon/workspace/builds/upstream-dissect/:

• Kilocode is a literal fork of opencode at packages/opencode/src level (diff -q confirms identical subdirectories). That's why their MCP wiring is identical — they're the same code with a few feature deltas.
• Upstream's packages/opencode/src/session/llm.ts:439 (resolveTools()) confirms the model's visible tool catalog comes from Object.keys(input.tools) (global registered MCP tools) filtered by input.user.tools?.[k] !== false (per-prompt overrides). Permission rules are evaluated separately at execution time. This is why Talon's chat-visibility hack (disconnect rival chat MCP servers) is necessary — it's the only way to keep tools out of the model's catalog, not just out of its allowed-to-execute set. The new mcp.ts documents this explicitly.

Why now

The June 1 Claude Max cancellation gives ~17 days runway. The OpenCode and Kilo backends are Talon's main fallback path when Claude goes offline; they need to be solid and easy to evolve. Reducing the duplicated surface from two parallel ~500-line copies to one tested shared module makes the next round of work tractable without a fork tax.

Deferred to follow-up PRs

• OpenCode → SSE/promptAsync — currently sync prompt hangs when upstream stalls. Kilo's SSE pattern in events.ts is the model to follow.
• Unified delivery routing — Kilo's 4-route delivery decision (tool / text-part / synthetic-error / empty) vs OpenCode's text-part-only vs Claude SDK's delivery-tool-delegation is its own design conversation.
• Backend conformance test matrix — same scenarios run against all three backends through the stub-client integration tier.
• Cross-backend system prompt — currently each backend has its own suffix.

Test plan

• npm run lint — 0 errors, 15 pre-existing warnings (unchanged)
• npx tsc --noEmit — clean
• npm test — 2228 passing, 12 skipped (live-tier), 0 failing (was 2208 — +20 new unit tests)
• npm run format:check — clean
• Manual: run prod against this branch and verify chat switching still disconnects the rival chat MCP server (Pandario → Dylan DM → Pandario should show the disconnect / re-register log lines)
• CI matrix on push

🤖 Generated with Claude Code