feat(server): delegation tools are always loaded and name the shell CLIs they replace

[lukemaj]

What changed: Claude Code agents in T3 see the full descriptions of spawn_thread, read_thread, message_thread and prism_submit from the first turn, and those descriptions say to use them instead of agent CLIs in the shell.
Why: The tools were deferred, so agents saw only names; a planner asked for a Luna review ran codex exec in the shell, invisible to the user.
So what: Independent review approved; you approve the merge and install the build, then a fresh thread confirms agents pick spawn_thread.

Closes #66.

• apps/server/src/mcp/alwaysLoad.ts: _meta["anthropic/alwaysLoad"] on the four delegation tools only; the other 29 t3-code tools stay deferred.
• spawn_thread and prism_submit descriptions rewritten trigger-first (toolboxmd/building-agentskills docs/05-authoring/triggers.md, obra/superpowers writing-skills): when to use, the shell alternative they replace, then the minimum mechanics. 528 and 470 characters.
• <t3_code_tool_use> launch paragraph names codex exec, claude -p, opencode run and grok in the shell. This is the only channel for Codex, OpenCode and Grok.

Elon record

• Requirements and who asked: the user (2026-09-29) asked to load the delegation tool descriptions instead of deferring them, and to write them with the skill-description guidance.
• Deleted: a PreToolUse hook blocking agent CLIs (user: last resort), server-wide alwaysLoad (all 33 tools every turn, startup waits up to 5 s), edits to other tools.
• Bottleneck: the planner never sees delegation guidance when it chooses how to run a review.
• Checked myself: a local stdio MCP server with one _meta-marked and one unmarked tool: Claude Code 2.1.284 debug log reported 0/19 deferred tools with the mark and 0/20 without. T3 connects over HTTP; the same behavior there is inferred until the post-install check.

Proof: vp test run src/mcp/alwaysLoad.test.ts src/mcp/toolInstructions.test.ts (10 passed); tsc --noEmit in apps/server passes; vp fmt --check and vp lint on apps/server/src/mcp clean apart from one existing warning in device/handlers.ts.

Review: Luna max (Codex gpt-6-luna). Round 1 on cb7ba42 requested one change (the spawn_thread description dropped that a role applies its kit); fixed in b8a109a. Round 2 on b8a109a: APPROVE.

🤖 Generated with Claude Code

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.5 KiB	13.5 KiB	−7 B (−0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.1 KiB	7.1 KiB	−9 B (−0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+2 B (+0.0%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	56.2 KiB	56.2 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	9	9	0 (0.0%)	21	✅
Claude	Total thread wire	13.5 KiB	13.5 KiB	−23 B (−0.2%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.1 KiB	7.1 KiB	0 B (0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.5 KiB	6.4 KiB	−23 B (−0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	57.1 KiB	57.0 KiB	−44 B (−0.1%)	66.4 KiB	✅
Claude	Live turn messages	10	9	−1 (−10.0%)	21	✅

Baseline: 26b5c3f · PR result: b8a109a · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 113.9 KiB
• Claude decoded thread snapshot: 114.6 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.