You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Extend server-side T3 MCP tool deferral beyond Copilot ACP to all other eager providers
#40640
T3 Code now ships around 75-90 built-in t3-code MCP tools (browser preview, device control, scheduling, PR guides, project admin, thread orchestration, etc.). Those tools are great to have when you need them, but in normal coding threads you rarely touch most of them.
Right now, if you use almost any provider other than Claude, Pi, or Copilot ACP, every single one of those tool schemas gets injected eagerly on every turn.
Why this matters in practice
Loading the full t3-code catalog upfront burns roughly 15k-25k tokens of context window before the conversation even starts (as measured in #34233 for Copilot before #34245, and in #40276 / #38976 for Antigravity). That hits users in three ways:
You pay a heavy prompt token and quota tax on every fresh session, especially on 5-hour/weekly plan limits.
Dumping 75+ unrelated tools into context crowds out actual repo context and hurts tool selection quality.
As T3 Code keeps adding toolkits, this upfront cost is only going to grow for every provider that still loads eagerly.
Where things stand on main (63fa4c892e)
T3 Code already has a working server-side deferral gate in apps/server/src/mcp/McpHttpServer.ts:1417-1429 (t3-tool-exposure: deferred + tool_search, added in #34245), so the core mechanism is already built. The gap is wiring it up consistently:
Deferred today: Claude (client-side ToolSearch), Pi (extension deferral), Copilot ACP (packages/provider-acp-registry/src/server/adapter.ts:335), plus Codex splitting t3-code and t3-code-direct.
Bridge gap: packages/provider-acp/src/server/adapter.ts:1060-1079 only sends t3-tool-exposure: deferred on HTTP MCP entries, and apps/server/src/mcp/AcpMcpStdioBridge.ts:235-241 does not forward t3-tool-exposure at all, so any ACP agent on the stdio fallback cannot opt into deferral yet.
Proposed direction
Since #34245 already proved the server-side model works cleanly for Copilot (and #39362 shows how to handle harnesses that freeze tools per turn), it would make a big difference across the whole app to:
Forward t3-tool-exposure through AcpMcpStdioBridge and ACP MCP bridges, not only direct HTTP entries.
Enable server-side deferral for the remaining ACP and HTTP providers that do not have their own client-side tool search (starting with the same small core set + tool_search).
That would let users pick Antigravity, OpenCode, Cursor, Muse Code, or any ACP Registry agent without paying a 15k-25k token penalty compared to Claude or Copilot.
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
Moving this here from #40630 as suggested.
T3 Code now ships around 75-90 built-in
t3-codeMCP tools (browser preview, device control, scheduling, PR guides, project admin, thread orchestration, etc.). Those tools are great to have when you need them, but in normal coding threads you rarely touch most of them.Right now, if you use almost any provider other than Claude, Pi, or Copilot ACP, every single one of those tool schemas gets injected eagerly on every turn.
Why this matters in practice
Loading the full
t3-codecatalog upfront burns roughly 15k-25k tokens of context window before the conversation even starts (as measured in #34233 for Copilot before #34245, and in #40276 / #38976 for Antigravity). That hits users in three ways:delegate_taskandt3_thread_launchto get dropped completely.As T3 Code keeps adding toolkits, this upfront cost is only going to grow for every provider that still loads eagerly.
Where things stand on
main(63fa4c892e)T3 Code already has a working server-side deferral gate in
apps/server/src/mcp/McpHttpServer.ts:1417-1429(t3-tool-exposure: deferred+tool_search, added in #34245), so the core mechanism is already built. The gap is wiring it up consistently:ToolSearch), Pi (extension deferral), Copilot ACP (packages/provider-acp-registry/src/server/adapter.ts:335), plus Codex splittingt3-codeandt3-code-direct.apps/server/src/orchestration-v2/Adapters/AntigravityAdapterV2.ts:218-225, tracked in [Bug]: Antigravity ACP drops the 8 T3 MCP tools >1,280B (incl. delegate_task & t3_thread_launch) without call_mcp_tool or T3_ACP_MCP_NODE fallback #38123 / fix(antigravity): defer optional T3 MCP tools via tool_call proxy #39362), Grok, Kiro,fx, every non-Copilot ACP Registry agent (Devin, Kimi, Droid, GLM, Mistral Vibe, Hermes,claude-acp, etc.), OpenCode v1/v2 (packages/provider-opencode/src/server/v2/adapter.ts:4405-4415), Cursor (packages/provider-cursor/src/server/adapter.ts:223), and Muse Code (packages/provider-muse/src/server/adapter.ts:1943).packages/provider-acp/src/server/adapter.ts:1060-1079only sendst3-tool-exposure: deferredon HTTP MCP entries, andapps/server/src/mcp/AcpMcpStdioBridge.ts:235-241does not forwardt3-tool-exposureat all, so any ACP agent on the stdio fallback cannot opt into deferral yet.Proposed direction
Since #34245 already proved the server-side model works cleanly for Copilot (and #39362 shows how to handle harnesses that freeze tools per turn), it would make a big difference across the whole app to:
t3-tool-exposurethroughAcpMcpStdioBridgeand ACP MCP bridges, not only direct HTTP entries.tool_search).deferred-proxyin fix(antigravity): defer optional T3 MCP tools via tool_call proxy #39362) for harnesses that freeze their tool list per turn and cannot use mid-turnnotifications/tools/list_changedupdates.That would let users pick Antigravity, OpenCode, Cursor, Muse Code, or any ACP Registry agent without paying a 15k-25k token penalty compared to Claude or Copilot.
All reactions