feat: open useful USER_TYPE-gated features to all users

[Flo5k5]

Summary

Remove 13 process.env.USER_TYPE === 'ant' gates that restricted useful features to Anthropic employees. These features work without Anthropic infrastructure and are now available to all open-build users.

Highlight: agent nesting unlocked

The most impactful change is a single line removed in src/constants/tools.ts:41:

...(process.env.USER_TYPE === 'ant' ? [] : [AGENT_TOOL_NAME]),

This line added AGENT_TOOL_NAME to the disallowed tools list for sub-agents, preventing non-ant users from using recursive agent workflows. Removing it unlocks recursive agent spawning: a sub-agent can now spawn its own sub-agents. This enables complex coordination patterns (agent teams, hierarchical delegation) that were previously reserved for Anthropic employees.

All changes

Features opened:

• Agent nesting — sub-agents can spawn sub-agents (src/constants/tools.ts)
• Effort 'max' persistence in settings (src/utils/effort.ts)
• Plan mode interview phase — now controlled by feature flags / env var (src/utils/planModeV2.ts)
• Sandbox disabled commands — configurable via ~/.claude/feature-flags.json (src/tools/BashTool/shouldUseSandbox.ts)
• All tips visible — plan mode, feedback, shift-tab tips shown to all users (src/services/tips/tipRegistry.ts)

Simplified:

• Fullscreen defaults to off for all (use /config to enable) (src/utils/fullscreen.ts)
• Explore agent always uses haiku model (src/tools/AgentTool/built-in/exploreAgent.ts)
• Plan mode tool uses conservative prompt for all (src/tools/EnterPlanModeTool/prompt.ts)

Continues the USER_TYPE cleanup from #637 (dead code removal) and builds on #639 (local feature flags).

Scope note: ~300 USER_TYPE checks remain — ~260 are intentionally kept (Anthropic infrastructure, security gates, API-specific fields, telemetry). Only the 13 gates above protect features that work independently.

Test plan

• bun run build — compiles successfully
• bun run smoke — smoke tests pass
• Verify agent nesting works without USER_TYPE=ant
• Verify effort 'max' persists in settings after restart
• Verify sandbox functions normally (default empty disabled commands)
• Verify tips appear for non-ant users

[Copilot]

Pull request overview

This PR removes several process.env.USER_TYPE === 'ant' gates to make previously internal-only (but infra-independent) features available to all users, including agent nesting, effort persistence, plan mode behavior, sandbox configuration, and tips.

Changes:

• Removes USER_TYPE==='ant' gating for plan-mode interview phase, effort 'max' persistence, sandbox disabled commands, and various tips.
• Enables nested agents by allowing the Agent tool inside sub-agents; standardizes explore agent model to haiku.
• Simplifies defaults by turning off fullscreen-by-default for everyone and using the external plan-mode prompt for all users.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/utils/planModeV2.ts	Removes ant-only always-on behavior for interview phase; relies on env var + feature flag.
src/utils/fullscreen.ts	Changes fullscreen default fallback to always off.
src/utils/effort.ts	Allows 'max' to be persisted for all users via toPersistableEffort.
src/tools/EnterPlanModeTool/prompt.ts	Removes ant-specific prompt path; uses the external prompt for everyone.
src/tools/BashTool/shouldUseSandbox.ts	Makes sandbox disabled-commands feature flag apply to all users.
src/tools/AgentTool/built-in/exploreAgent.ts	Forces explore agent to always use haiku.
src/services/tips/tipRegistry.ts	Removes ant-only tip filtering and drops internal-only tips.
src/constants/tools.ts	Allows Agent tool for agents (enables nested sub-agents).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 17 out of 17 changed files in this pull request and generated 4 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[kevincodex1]

looks good to me!

[Vasanthdev2004]

Review: Open useful USER_TYPE-gated features to all users

Reviewed on head 9b7c0e9. CI green ✅. 17 files, +75/-539.

✅ Scope and philosophy

The PR correctly targets only the 13 USER_TYPE === 'ant' gates that protect features working independently of Anthropic infrastructure. ~260 remaining gates (API-specific fields, telemetry, internal auth routing, classifier settings, model resolution) are intentionally left alone. This is the right boundary.

✅ Key changes verified

Agent nesting (src/constants/tools.ts):
Removing AGENT_TOOL_NAME from ALL_AGENT_DISALLOWED_TOOLS allows sub-agents to spawn their own sub-agents. CUSTOM_AGENT_DISALLOWED_TOOLS spreads from the same set, so it's automatically consistent. AGENT_TOOL_NAME was already in ASYNC_AGENT_ALLOWED_TOOLS (line 108), so async agent paths work too. This is the most impactful change and it's clean.

Effort 'max' persistence (src/utils/effort.ts + src/utils/settings/types.ts):
toPersistableEffort now returns 'max' for all users (not just ants). The Zod schema effortLevel enum is updated to ['low', 'medium', 'high', 'max'] (was conditional). Schema and persistence logic now match. ✅

Sandbox disabled commands (src/tools/BashTool/shouldUseSandbox.ts):
Good runtime validation added for the tengu_sandbox_disabled_commands feature flag value: typeof raw === 'object' && raw !== null guard + Array.isArray() checks on substrings and commands before iteration. Prevents crashes on malformed config. ✅

Plan mode interview (src/utils/planModeV2.ts):
Removed the if (USER_TYPE === 'ant') return true fast-path. Now falls through to the env var / feature flag check, which is the correct behavior — users can enable via CLAUDE_CODE_PLAN_MODE_INTERVIEW_PHASE=1 or the tengu_plan_mode_interview_phase feature flag (available via PR #639's ~/.claude/feature-flags.json).

Fullscreen default (src/utils/fullscreen.ts):
Defaults to false for all users instead of true for ants. JSDoc updated. Users can still enable via /config or CLAUDE_CODE_NO_FLICKER=1. ✅

Tips (src/services/tips/tipRegistry.ts):
Removed USER_TYPE === 'ant' → return false guards from plan-mode, shift-tab, and feedback tips. Removed internal-only tips (important-claudemd, skillify). External tip content now shown uniformly. ✅

OSC tab status (src/ink/termio/osc.ts):
supportsTabStatus() now returns false unconditionally instead of true for ants. JSDoc updated from "Ant-only" to "Currently disabled". This is correct — the spec is unstable and was never intended for external use. ✅

Dead code removal (src/tools.ts):
REPLTool and SuggestBackgroundPRTool set to null instead of being conditionally required. ConfigTool and TungstenTool imports removed entirely (they were ant-only). ✅

Insights remote host collection (src/commands/insights.ts):
Removed ~174 lines of Coder/homespace/SCP remote data collection. This was deeply Anthropic-internal infrastructure (SSH into Coder workspaces, SCP session files). The --homespaces flag and all related plumbing are gone. Clean removal. ✅

Slow operations tracking (src/bootstrap/state.ts):
addSlowOperation is now a no-op, getSlowOperations returns EMPTY_SLOW_OPERATIONS. Callers still work — they just never track anything. This was an internal dev-bar metric, not user-facing. ✅

Bridge config (src/bridge/bridgeConfig.ts, bridgeMain.ts, initReplBridge.ts):
CLAUDE_BRIDGE_OAUTH_TOKEN and CLAUDE_BRIDGE_BASE_URL overrides no longer require USER_TYPE === 'ant. These are dev overrides useful for anyone running local bridge endpoints. ✅

Explore agent (src/tools/AgentTool/built-in/exploreAgent.ts):
Model always 'haiku' — explore is a fast read-only search agent, haiku is the right choice. The ant-specific 'inherit' path used getAgentModel() + a GrowthBook flag, which is internal infrastructure. ✅

BashTool readOnlyValidation (src/tools/BashTool/readOnlyValidation.ts):
Removed ANT_ONLY_COMMAND_ALLOWLIST (66 lines) — gh read-only commands and aki (internal knowledge-base CLI). These are Anthropic-internal tools. ✅

🟡 Nit: dead code in tools.ts

Line 221 still has ...(process.env.USER_TYPE === 'ant' && REPLTool ? [REPLTool] : []) — but REPLTool is now always null, so this is dead code that never executes. The condition short-circuits to false, spreading []. Harmless but could be cleaned up to just ...(REPLTool ? [REPLTool] : []) or removed entirely. Non-blocking.

✅ All 6 Copilot comments resolved

1. Zod schema effortLevel now includes 'max' for all users ✅
2. Array.isArray() guards on substrings/commands ✅
3. typeof raw === 'object' guard for malformed config ✅
4. isFullscreenEnvEnabled JSDoc updated ✅
5. supportsTabStatus JSDoc updated ✅
6. PR description inaccuracy about feature-flags.json — noted as doc issue, not code ✅

---

Verdict: Approve-ready ✅

Well-scoped USER_TYPE cleanup that only opens features working independently of Anthropic infrastructure. All Copilot issues addressed with proper runtime validation. Agent nesting is the headline change — clean and correct. One minor dead code nit in tools.ts. Kevin already approved. No blockers.