@yeongjaeyou/claude-code-config 0.9.0 → 0.9.2

package/.claude/agents/web-researcher.md

@@ -40,7 +40,11 @@ Detect keywords in the prompt to determine execution mode:
 ## Sub-Agent Output Schema
 
 **MANDATORY**: Each platform search Task **MUST** return results in this exact structure.
-Sub-agents that fail to follow this schema will have their results rejected by the coordinator.
+
+**On schema violation:**
+- Missing required fields → retry (max 2 attempts)
+- Invalid YAML → skip with warning, note in report
+- Empty findings → accept with `confidence: 0.1`
 
 ### Required Fields (Quick & Deep)
 
@@ -156,6 +160,11 @@ Parallel execution (Task tool, run_in_background: true):
 └── Task: arXiv + general web (if needed)
 ```
 
+**Execution policy:**
+- Timeout: 300000ms (5min) per agent
+- Min success: 3/5 agents required to proceed
+- On partial failure: continue with warning in report
+
 **Each Task prompt MUST include:**
 ```
 Format your response according to the Sub-Agent Output Schema.
@@ -187,7 +196,7 @@ Return findings as a YAML block with ALL fields:
 - gaps, conflicts, suggested_followups (Deep Mode required)
 ```
 
-Each Task returns results following Sub-Agent Output Schema. **Reject non-compliant responses.**
+Each Task returns results following Sub-Agent Output Schema. **On non-compliance: retry once, then skip with warning.**
 
 ---
 
@@ -227,6 +236,11 @@ Selective execution (1-2 platforms only):
 └── Task: {Second most suitable platform} (if needed)
 ```
 
+**Round context to include:**
+- Unresolved `gaps` list from Round 1
+- Detected `conflicts` requiring verification
+- Top 3 key findings (title + URL only)
+
 Repeat Round 1.5 analysis → Convergence check
 
 ---
@@ -271,17 +285,16 @@ Validation execution:
 | Condition | Value |
 |-----------|-------|
 | Max rounds | 3 |
-| Max total time | 10 min |
-| Token budget per round | 50k |
+| Max total time | 15 min |
+| Min successful agents | 3/5 per round |
 
 ### Soft Limits (Convergence Conditions)
 
 Terminate when any of the following is met:
 
-- **New information < 10%**: Compared to previous round
-- **All gaps resolved**: All identified gaps addressed
-- **Question fully covered**: All aspects of user question answerable
-- **High confidence**: Core answer confidence > 0.9
+- **New information < 10%**: `(new_unique_urls / prev_total_urls) < 0.1`
+- **All gaps resolved**: `len(unresolved_gaps) == 0`
+- **High confidence**: `avg(agent_confidences) > 0.9`
 
 ### Forced Termination
 

package/.claude/commands/council.md

@@ -216,12 +216,19 @@ Task(subagent_type="general-purpose", run_in_background: true):
 
 **Important:**
 - Use `run_in_background: true` for true parallelism
-- Timeout: 120000ms per call
-- Continue with successful responses if some models fail
+- Timeout per model (use same value for TaskOutput):
+  | Model | Timeout | Reason |
+  |-------|---------|--------|
+  | Opus/Sonnet | 300000ms (5min) | Direct execution |
+  | Codex | 480000ms (8min) | MCP tool + deep reasoning |
+  | Gemini | 600000ms (10min) | CLI invocation + long thinking |
+- Continue with successful responses if some models fail (min 2/4 required)
+- **TaskOutput must use matching timeout**: `TaskOutput(task_id, block=true, timeout=600000)`
 
-### Round 1.5: Coordinator Analysis
+### Round 1.5: Coordinator Analysis (MANDATORY)
 
-After collecting responses, the main agent (coordinator) performs analysis:
+**⚠️ DO NOT SKIP**: After collecting responses, the coordinator **MUST** perform this analysis before synthesis.
+Skipping Round 1.5 defeats the purpose of multi-round deliberation.
 
 **1. Anonymize Responses:**
 ```
@@ -261,7 +268,7 @@ conflicts_detected:
     resolution_needed: true
 ```
 
-**4. Convergence Check:**
+**4. Convergence Check (REQUIRED before synthesis):**
 ```yaml
 convergence_status:
   confidence_average: 0.815
@@ -271,6 +278,11 @@ convergence_status:
   decision: "proceed_to_round_2" | "terminate_and_synthesize"
 ```
 
+**Decision logic:**
+- If `gaps_remaining > 0` AND `confidence_average < 0.9` → `proceed_to_round_2`
+- If `conflicts_remaining > 0` AND round < 3 → `proceed_to_round_2`
+- Otherwise → `terminate_and_synthesize`
+
 ### Round 2: Targeted Re-queries (Conditional)
 
 If convergence criteria not met, re-query only models with gaps/conflicts:
@@ -342,8 +354,10 @@ After convergence or max rounds:
 | Condition | Value |
 |-----------|-------|
 | Max rounds | 3 |
-| Max total time | 10 min |
-| All models failed | immediate |
+| Max total time | 20 min |
+| Max per-model timeout | 10 min (Gemini) |
+| Min successful models | 2/4 (proceed with partial results) |
+| All models failed | immediate termination |
 
 ### Soft Limits (Convergence - any triggers termination)
 | Condition | Threshold |

package/.claude/guidelines/work-guidelines.md

@@ -88,6 +88,14 @@ Serena provides LSP-based semantic code analysis. Use it for precise code naviga
 3. Prefer symbolic editing over file-based editing when modifying functions/classes
 4. Always check `find_referencing_symbols` before renaming/removing symbols
 
+### Large File Handling
+- Files exceeding 25000 tokens cannot be read at once (Claude Code internal limit)
+- When encountering "exceeds maximum allowed tokens" error:
+  1. Use Grep to locate relevant content first
+  2. Use Read tool with `offset` and `limit` parameters to read in chunks
+  3. Example: `offset=0, limit=1000` → `offset=1000, limit=1000` sequentially
+- Never attempt to read entire large files without chunking
+
 ### Python Development
 - **Virtual Environment (MANDATORY)**: Always use virtual environment (uv) when running Python
 - **NEVER** use system Python (`python`, `python3`) directly

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeongjaeyou/claude-code-config",
-  "version": "0.9.0",
+  "version": "0.9.2",
   "description": "Claude Code CLI custom commands, agents, and skills",
   "bin": {
     "claude-code-config": "./bin/cli.js"