Add agentic mode with MessageOrchestrator

[RichardAtCT]

Summary

• Add AGENTIC_MODE env var (default: true) that switches the bot to a minimal conversational interface — 3 commands (/start, /new, /status), no inline keyboards, no follow-up suggestions. All other input goes straight to Claude.
• New MessageOrchestrator replaces inline handler registration in core.py, routing to agentic or classic handlers based on mode. Classic mode behavior is fully unchanged.
• Feature registry gates conversation enhancer, quick actions, and session export behind agentic_mode=false.

Test plan

• 252 tests passing (11 new orchestrator tests)
• Lint clean (flake8)
• Manual test with AGENTIC_MODE=true: send messages, verify no buttons, Claude responds
• Manual test with AGENTIC_MODE=false: verify full classic experience unchanged
• Test /start, /new, /status in both modes
• Test file upload in agentic mode

Follow-ups addressed

Closes #23
Closes #24

[RichardAtCT]

PR Review
Reviewed head: bf98f601e2b0fac2347eeb7d193041665d3587a0

Summary

• Nice refactor: moving mode-specific routing into MessageOrchestrator is a clean separation and makes core.py much slimmer.
• Agentic mode behavior is implemented consistently (minimal commands, no keyboards, direct passthrough for text/file/photo).
• Test coverage for registration and core agentic paths is solid for a first pass.

What looks good

• src/bot/core.py now delegates command setup and handler wiring cleanly to one component.
• src/bot/features/registry.py gates classic-only features in a straightforward way.
• tests/unit/test_orchestrator.py covers the high-risk branch split (agentic vs classic) and key UX expectations.

Issues / questions

1. [Blocker] src/bot/orchestrator.py (agentic_start) — Markdown parse risk from unescaped user names.
   user.first_name is interpolated into a parse_mode="Markdown" message. Names containing Markdown metacharacters (_, *, [, etc.) can break rendering or fail message send. Please escape user-provided text (or switch to plain text / MarkdownV2 with proper escaping).

2. [Important] src/bot/orchestrator.py (_register_agentic_handlers, _agentic_callback) — callback handler is unscoped and catches all callback queries in agentic mode.
   This is likely survivable but can cause noisy UX and accidental interception of unrelated callback payloads. I opened a non-blocking issue to track hardening: Agentic mode: restrict callback handler to cd:* pattern #23

3. [Important] src/bot/orchestrator.py (agentic_text) — audit logging marks failures as success.
   Current code logs success=True unconditionally after handling, including error paths. I opened a non-blocking issue: Agentic text handler logs failed runs as success #24

Suggested tests (if needed)

• Add a test for /start with a first name containing Markdown metacharacters (e.g. A_B) to confirm safe rendering.
• Add a callback registration/dispatch test proving only cd: payloads are handled in agentic mode.
• Add a test that failed claude_integration.run_command yields audit_logger.log_command(..., success=False).

Verdict

• ⚠️ Merge after fixes (blocker: Markdown escaping in agentic_start)