feat(agents): split chat agent into task-specific agents with prompt profiles

Before: monolithic ChatAgent with 13K-token system prompt caused 95s TTFT
for a simple "Hi!" on Gemma-4-E4B. Eval scenarios timed out at 610s.

After: 5 focused agents (chat, doc, file, data, web) + lite variants,
each with a lean prompt profile. TTFT drops from 95s to 0.12s (chat)
and 3-10s (doc). Eval pass rate: 89% judged (34/38), avg score 9.4/10.

Agent architecture:
- chat: conversation only, ~2K tokens, no tools
- doc: RAG + file search, ~5K tokens, hallucination prevention
- file: filesystem ops + discovery, ~4K tokens
- data: CSV/Excel analysis with scratchpad, ~3K tokens
- web: browser tools, ~2K tokens
- Each has a -lite variant using ~4B model for low-memory hardware

Eval framework updates:
- Per-scenario agent_type field in YAML (overrides --agent-type CLI)
- Latency validation: warns when TTFT > 30s
- Preserve eval sessions for review (no delete_session)
- Increased startup timeout 120s → 240s for Windows
- Fixed shutil.which("claude") for Windows .cmd resolution

Author: Ovtcharov

eval/scenarios/adversarial/empty_file.yaml

@@ -1,6 +1,7 @@
 id: empty_file
 name: "Empty File Handling"
 category: adversarial
+agent_type: doc
 severity: medium
 description: |
   User asks the agent to index and read a completely empty file. Agent must

eval/scenarios/adversarial/large_document.yaml

@@ -1,6 +1,7 @@
 id: large_document
 name: "Buried Fact in Large Document"
 category: adversarial
+agent_type: doc
 severity: high
 description: |
   A specific fact is buried deep within a large document. Tests whether the

eval/scenarios/adversarial/topic_switch.yaml

@@ -1,6 +1,7 @@
 id: topic_switch
 name: "Rapid Topic Switch"
 category: adversarial
+agent_type: doc
 severity: medium
 description: |
   User rapidly switches topics between two different documents across four turns.

eval/scenarios/captured/captured_eval_cross_turn_file_recall.yaml

@@ -1,6 +1,7 @@
 id: captured_eval_cross_turn_file_recall
 name: "Captured: Cross-Turn File Recall"
 category: captured
+agent_type: doc
 description: 'Captured from session: Eval: cross_turn_file_recall'
 note: "Subset of cross_turn_file_recall (2 of 3 turns captured from a real session)"
 persona: casual_user

eval/scenarios/captured/captured_eval_smart_discovery.yaml

@@ -1,6 +1,7 @@
 id: captured_eval_smart_discovery
 name: "Captured: Smart Document Discovery"
 category: captured
+agent_type: file
 description: 'Captured from session: Eval: smart_discovery'
 persona: casual_user
 setup:

eval/scenarios/context_retention/conversation_summary.yaml

@@ -1,6 +1,7 @@
 id: conversation_summary
 name: "5-Turn Conversation Summary"
 category: context_retention
+agent_type: doc
 severity: medium
 description: |
   A 5-turn conversation that tests the agent's ability to accumulate facts across

eval/scenarios/context_retention/cross_turn_file_recall.yaml

@@ -1,6 +1,7 @@
 id: cross_turn_file_recall
 name: "Cross-Turn File Recall"
 category: context_retention
+agent_type: doc
 severity: critical
 description: |
   User indexes a document in Turn 1, then asks about its content in Turn 2

eval/scenarios/context_retention/multi_doc_context.yaml

@@ -1,6 +1,7 @@
 id: multi_doc_context
 name: "Multi-Document Context"
 category: context_retention
+agent_type: doc
 severity: high
 description: |
   Two documents are indexed simultaneously. Agent must answer questions from each

eval/scenarios/context_retention/pronoun_resolution.yaml

@@ -1,6 +1,7 @@
 id: pronoun_resolution
 name: "Pronoun Resolution"
 category: context_retention
+agent_type: doc
 severity: critical
 description: |
   User asks follow-up questions using pronouns ("it", "that policy").

eval/scenarios/error_recovery/file_not_found.yaml

@@ -1,6 +1,7 @@
 id: file_not_found
 name: "File Not Found -- Helpful Error"
 category: error_recovery
+agent_type: doc
 severity: medium
 description: |
   User asks to read a nonexistent file. Agent must report the error gracefully