[bassilkhilo-ag2]

Summary

• Adds PRESENTER_ACTIVE sentinel check at module load in src/inline-tools.ts (closes voice-agent: copres_next_anchor fires on greetings — inline tools needing mode-gating should gate at exposure, not just effect #1171)
• slideControlTool is spread into inlineTools and ownerOnlyTools only when state/presenter-mode.sentinel is present
• When sentinel is absent, Gemini never sees the slide_control description → can't fire it on greetings or filler utterances
• Sentinel check uses existsSync(join(WORKSPACE_DIR, 'state', 'presenter-mode.sentinel')) — workspace-resolved, not repo-relative

Background

slide_control was unconditionally registered as a global inline tool. Gemini saw its description every turn and fired it spuriously on greetings/silence even when no presentation was active. The tool body had no intent guard, and the spurious tool_call entries landed in voice sqlite, wasting tokens and noising up observability.

Susan's MBP bot (Maddy) filed this after observing 3 unprovoked copres_next_anchor firings in <40s on a greeting — the OSS-visible analog is slide_control.

Trade-off

Toggling presenter-mode.sh start N mid-session won't hot-reload the tool list (voice-agent must restart). The existing workflow already runs presenter-mode.sh before the voice agent, so the contract holds in normal use.

Test plan

• python3 tests/inline-tools-presenter-gate.test.py — PASS
• Manual: start voice agent without sentinel → slide_control absent from tool list
• Manual: bash scripts/presenter-mode.sh start 60, restart voice agent → slide_control present

Related

• discord-voice: tighten system prompt against tool-narration and capability hallucinations #1102 — discord-voice system prompt tightening (adjacent observability angle)
• Closed presenter-mode sentinel: bare string compare in check-pending-questions.py:37 (same bug class as PR #432/#435) #436 — same sentinel, different bug class

🤖 Generated with Claude Code

[VasiliyRad]

The gate itself is correct and minimal: PRESENTER_ACTIVE is checked once at module load via existsSync(join(WORKSPACE_DIR, 'state', 'presenter-mode.sentinel')), workspace-resolved, and the conditional spread ...(PRESENTER_ACTIVE ? [slideControlTool] : []) is applied consistently in both inlineTools and ownerOnlyTools. Structural test passes. The fix is the right minimal change.

One concern: the PR branch has 4 commits over main, and 3 of them are unrelated (d9a384c health-check multi-pid fix + its test, and two chore(log) commits). These should be on their own PRs or squashed out before merging — the health-check change is substantive enough to need its own review context. build_log.md is gitignored but still shows in the diff, likely force-added by a prior chore commit.

If you can rebase/squash this to just d94dac6 (the actual fix commit), it's good to approve. As-is I'd hold off on merging until the branch is cleaned up.

[bassilkhilo-ag2]

Rebased to just the fix commit d94dac6 (cherry-picked onto current main). The unrelated health-check fix and chore(log) commits have been dropped — they were already merged or not appropriate here. PR is now a single-commit diff: gate slideControlTool on the presenter-mode sentinel.

[bassilkhilo-ag2]

Branch has been rebased — the 3 unrelated commits VasiliyRad flagged are now stripped. The PR is now a single commit:

fix(inline-tools): gate slideControlTool exposure on presenter-mode sentinel

Same state as the end of that review comment's suggestion. Ready for re-review / approval.

[bassilkhilo-ag2]

Closing — PR #1197 is the canonical fix for #1171 with a TypeScript test and VasiliyRad's approval. Both apply the same sentinel gate; #1197 has better variable naming (presenterActive camelCase) and a more complete test file.