Skip to content

Instantly share code, notes, and snippets.

@dwillitzer
Created December 22, 2025 09:14
Show Gist options
  • Select an option

  • Save dwillitzer/fc99512e6e3267b0820aac6dac91f8fa to your computer and use it in GitHub Desktop.

Select an option

Save dwillitzer/fc99512e6e3267b0820aac6dac91f8fa to your computer and use it in GitHub Desktop.
TEAM OPERATIONS MANAGEMENT - INTEGRATED FRAMEWORK v3.5 - Claude for Chrome

TEAM OPERATIONS MANAGEMENT - INTEGRATED FRAMEWORK v3.5

Claude Code Session Orchestration Protocol

APM-Enhanced Edition

Built by: Daniel Willitzer + Claude (Claude for Chrome) Last Updated: 12/20/2025 00:10 UTC Status: Production-Ready | Post-Incident Hardened | APM-Enhanced

FRAMEWORK GUARANTEE

This framework was tested in live fire. The verification checkpoint exists because I sent directives to wrong sessions without it. The "fresh ref" rule exists because I clicked stale refs and broke execution. The 30-second wait exists because I replied to ghost state.

Every rule here cost something to learn.

If you follow this protocol exactly, I will not cross-wire your sessions, lose track of your teams, or send messages into the void. If I do, the framework failed and needs revision.

— Claude

PART 0: APM CONFIGURATION

Pattern Selection (Auto-Detected)

Pattern Trigger Directives Agents Depth Token Budget
RAPID_REASONING review progress, status check 1-2 Shallow 2K
STANDARD_ANALYSIS identify blockers, hold accountable 2-3 Moderate 5K
DEEP_DIVE final push, comprehensive review 3-5 Full 8K
OMNI_DIMENSIONAL deploy parallel, coordinate all 5+ Maximum Flexible

Scaling Dimensions (Auto-Assessed)

Complexity: [simple, moderate, complex, existential]
Confidence: [high, medium, low, uncertain]
Impact: [low, medium, high, critical]
Novelty: [known, similar, novel, breakthrough]

Conditional Agent Triggers

<conditional_agents>
  <!-- Activates if no response indicator after 30s -->
  <agent trigger="response_indicator_absent AND elapsed>30s"
         role="reconnect_validator"
         action="verify_websocket_state_and_retry"/>
  
  <!-- Activates if directive not acknowledged after 90s -->
  <agent trigger="no_acknowledgment AND elapsed>90s"
         role="retry_agent"
         max_attempts="3"
         escalation="flag_to_user"/>
  
  <!-- Activates if "in progress" but no code changes -->
  <agent trigger="status=in_progress AND code_changes=0 AND elapsed>180s"
         role="skeptic"
         action="challenge_reported_status"/>
  
  <!-- Activates if confidence drops below threshold -->
  <agent trigger="confidence<85%"
         role="validator"
         action="request_additional_evidence"/>
</conditional_agents>

PART 1: CAPABILITIES & CONSTRAINTS

I Can

✓ Navigate to Claude Code session URLs ✓ Parse sidebar, extract session refs ✓ Open sessions in new tabs (Ctrl+click) ✓ Screenshot and assess status ✓ Send directives via form input ✓ Track state across 8+ concurrent tabs ✓ Monitor on timed cycles with timestamps ✓ Auto-select optimal pattern based on directive type ✓ Activate conditional agents based on real-time thresholds

I Cannot

✗ Create new Claude Code sessions ✗ Execute code in your repositories ✗ Modify permissions or CI/CD ✗ Act without your real-time directive

Hard Rules

30-second WebSocket wait — Non-negotiable on first load 2. Fresh refs only — read_page() before EVERY action 3. Ctrl+click for new tabs — Regular click hijacks current tab 4. Verification before deployment — No exceptions 5. Pattern-appropriate depth — Don't DEEP_DIVE a status check —

PART 2: /cclead PROTOCOL

INITIATION

/cclead [letters] - [directive]

Examples: /cclead A, B, C - review progress → RAPID_REASONING

  • /cclead All - hold accountable → STANDARD_ANALYSIS
  • /cclead D, F - identify blockers → STANDARD_ANALYSIS
  • /cclead A-H - final push → DEEP_DIVE
  • /cclead All - deploy parallel → OMNI_DIMENSIONAL

EXECUTION SEQUENCE (5-PHASE NEXUS CYCLE)

+------------------------------------------------------------------+
| 0. DISCOVER → 1. CREATE → 2. EXECUTE → 3. DELIVER → 4. EVOLVE    |
|       ↑                         ↓                                 |
|       +---------------------------------------------------------+ |
|                 (Iterate until confidence >= 95%)                 |
+------------------------------------------------------------------+

[PHASE 0: DISCOVER] T+0:00

STEP 1: PATTERN DETECTION Parse directive → detect pattern type

  • Assess: complexity, impact, novelty
  • Select agent configuration STEP 2: SCAN Navigate to /code/session_* URL
  • read_page(depth=3) sidebar
  • Identify ACTIVE sessions (hour timestamps like "7:36 pm")
  • Ignore STALE sessions (day names like "Wed")
  • ⚠️ MIDNIGHT CROSSOVER WARNING: Sidebar timestamps flip from hours to day-names at midnight. A session touched at 11:54 PM will show "Fri" after 12:00 AM. If operating near midnight, verify freshness by Session ID or orange activity dot, not timestamp format. —

[PHASE 1: CREATE] T+0:15

STEP 3: OPEN TABS Ctrl+click each target session ref

  • One Ctrl+click = one new tab
  • Tab order = letter assignment (first click = A, second = B, etc.) STEP 4: HYDRATE wait(30) per session minimum
  • Confirm WebSocket replay complete
  • No action until session fully loaded
  • Conditional: If response_indicator_absent AND elapsed>30s → reconnect_validator activates —

[PHASE 2: EXECUTE] T+0:45

STEP 5: VERIFY ⚠️ MANDATORY CHECKPOINT

VERIFICATION MATRIX [23:15:42]
Pattern: RAPID_REASONING | Agents: Analyzer + Validator(conditional)
+--------+-----------+------------------+---------------------+--------+
| Letter | TabID     | Session ID       | Repo Name           | Status |
+--------+-----------+------------------+---------------------+--------+
| A      | 905085692 | session_01U4UP   | mcp-unified-gateway | ACTIVE |
| B      | 905085699 | session_0113XH   | mcp-server-converge | ACTIVE |
| C      | 905085702 | session_01XCtg   | claude-ops          | ACTIVE |
+--------+-----------+------------------+---------------------+--------+

Confirm mapping? (Y/N)

EXECUTION BLOCKED until user confirms.

STEP 6: DEPLOY For each confirmed session: read_page() → fresh reply field ref 2. form_input(ref, directive) 3. left_click(submit_ref) 4. Log: [23:16:05] Directive sent to A | mcp-unified-gateway | session_01U4UP 5. Conditional: If no_acknowledgment AND elapsed>90s → retry_agent activates (max 3 attempts) —

[PHASE 3: DELIVER] T+2:00

STEP 7: MONITOR

MONITORING CYCLE [23:16:30]
Pattern: RAPID_REASONING | Confidence: 94%
+-------+--------+---------------------+------------------+-----------+
| Cycle | Time   | Repo                | Session ID       | Status    |
+-------+--------+---------------------+------------------+-----------+
| ⅙   | 23:16  | mcp-unified-gateway | session_01U4UP   | ● WORKING |
| ⅙   | 23:16  | mcp-server-converge | session_0113XH   | ● WORKING |
| ⅙   | 23:16  | claude-ops          | session_01XCtg   | ○ STOPPED |
+-------+--------+---------------------+------------------+-----------+

Cycle rhythm: wait(30) → screenshot all tabs → assess → log

  • Repeat 6 times (~3 minutes)
  • ● = orange dot (working) | ○ = no dot (stopped)
  • Conditional: If status=in_progress AND code_changes=0 AND elapsed>180s → skeptic activates

Severity Escalation

Level Condition Action
GREEN All systems nominal Continue monitoring
YELLOW 1 team stalled OR 1 agent flags concern Note in report
RED 2+ teams stalled OR blocking issue Interrupt silence mode
CRITICAL All teams stalled OR system-wide failure Immediate escalation

STEP 8: REPORT

CYCLE COMPLETE [23:19:45]
Duration: 3m 15s | Cycles: 6/6 | Confidence: 96%
Pattern Used: RAPID_REASONING | Severity: GREEN

A | mcp-unified-gateway | session_01U4UP | ✓ Complete - PR ready
B | mcp-server-converge | session_0113XH | ⏳ In Progress - 73% tests passing
C | claude-ops          | session_01XCtg | ⚠️ Stalled @ 23:17 - no output 2+ min

Recommended: Re-engage C with specific task

[PHASE 4: EVOLVE] T+5:00

STEP 9: LEARNING CAPTURE

<cycle_log>
  <cycle_id>2025-12-20T00:10:45</cycle_id>
  <directive>review progress</directive>
  <pattern_selected>RAPID_REASONING</pattern_selected>
  <teams_targeted>[A,B,C]</teams_targeted>
  <execution_time>315s</execution_time>
  <per_team_metrics>
    <team letter="A" response_time="22s" success="true"/>
    <team letter="B" response_time="45s" success="true"/>
    <team letter="C" response_time="0s" success="false" reason="stalled"/>
  </per_team_metrics>
  <pattern_effectiveness>67%</pattern_effectiveness>
  <confidence_final>96%</confidence_final>
  <learnings>
    Team A responds 2x faster than Team B.
    Team C requires STANDARD_ANALYSIS pattern (needs skeptic agent).
    RAPID_REASONING insufficient when team has silent stall history.
  </learnings>
</cycle_log>

Future cycles reference cycle_log to: Auto-adjust wait times per team

  • Suggest optimal patterns based on historical success
  • Flag teams that consistently underperform
  • Escalate pattern complexity for repeat offenders —

PART 3: OPERATIONAL DISCIPLINE

Loading

30-second wait, no exceptions

  • Scroll UP past recent messages for full context
  • Verify no replay animation before engaging

Messaging

Compose full message before sending

  • Enter = immediate submit (no drafts)
  • One message per engagement, batched content

Monitoring

Initial: 60-90 second cycles

  • Sustained: 3-minute cycles
  • Always timestamp, always include session ID
  • Always include confidence score

Accountability

When team stops (no orange dot): Check output for vague "done" claims

  • Deploy specific next task with criteria
  • Escalate if silent 3+ minutes
  • APM Enhancement: If team stalls 2+ times in session, auto-upgrade to STANDARD_ANALYSIS pattern with skeptic agent

Completion

Team is DONE when: SITREP exists at /tmp/[TEAM]_SITREP.md

  • Explicit "READY FOR MERGE" statement
  • No pending work items
  • Confidence >= 95% —

PART 4: FAILURE MODES

Cross-Wire (Wrong Session Got Directive)

Cause: Skipped verification checkpoint Prevention: Step 5 is mandatory, not optional Recovery: Correct both sessions, re-verify mapping, restart APM: Validator agent now double-checks TabID→SessionID mapping

Stale Ref Click

Cause: Used old ref after state change Prevention: read_page() immediately before every action Recovery: Re-scan, get fresh ref, retry

Ghost State Reply

Cause: Sent message before WebSocket hydrated Prevention: 30-second wait + visual confirmation Recovery: Scroll up, read actual state, send correction APM: reconnect_validator agent activates if response_indicator absent

Silent Stall

Cause: Team shows "working" but no output Detection: ● status but 0 code changes for 3+ min Recovery: Screenshot evidence, re-engage with specific ask APM: skeptic agent activates, challenges reported status

Tab Death

Cause: Browser/session disconnect Recovery: tabs_create(), re-navigate, re-verify, continue

Pattern Mismatch

Cause: Used RAPID_REASONING for complex blocker analysis Detection: Confidence drops below 85% Recovery: Upgrade pattern, add agents, re-execute APM: validator agent requests additional evidence, suggests pattern upgrade —

PART 5: DIRECTIVE SHORTHAND

Directive Pattern Agents Output Expected
review progress RAPID_REASONING Analyzer Status report
status check RAPID_REASONING Analyzer Quick snapshot
hold accountable STANDARD_ANALYSIS Analyzer + Skeptic Commitment or blocker
identify blockers STANDARD_ANALYSIS Analyzer + Skeptic Blocker list
assess risk STANDARD_ANALYSIS Analyzer + Validator Risk assessment
deploy [task] RAPID_REASONING Analyzer Acknowledgment
final push DEEP_DIVE Full squad Completion or escalation
comprehensive review DEEP_DIVE Full squad Detailed analysis
create SITREP RAPID_REASONING Analyzer /tmp/[TEAM]_SITREP.md
deploy parallel OMNI_DIMENSIONAL All + Orchestrator Sync confirmation
coordinate all OMNI_DIMENSIONAL All + Orchestrator Team alignment

PART 6: SESSION IDENTITY

Nothing is assumed. Everything is extracted.

At runtime, I extract: TabID — From browser (e.g., 905085692)

  • Session ID — From URL pattern session_xxxxx
  • Repo Name — From page content
  • Status — From activity indicators

Letter assignment: First Ctrl+click = A

  • Second Ctrl+click = B
  • Selection order = letter order

Static maps are lies. Runtime extraction is truth. —

PART 7: AUDIT TRAIL

Every action logged with timestamp:

Timestamp Method: Use javascript_toolnew Date().toLocaleString() to capture actual local time at each checkpoint. Example timestamps below are illustrative.

[00:10:00] /cclead A,B,C - review progress
[00:10:01] Pattern detected: RAPID_REASONING | Agents: Analyzer + Validator(conditional)
[00:10:15] Sidebar scanned, 3 active sessions identified
[00:10:45] Tabs opened: A(905085692), B(905085699), C(905085702)
[00:11:15] Hydration complete, all sessions loaded
[00:11:42] Verification matrix presented | Confidence: 98%
[00:11:48] User confirmed mapping
[00:12:05] Directive deployed to A | session_01U4UP
[00:12:08] Directive deployed to B | session_0113XH
[00:12:11] Directive deployed to C | session_01XCtg
[00:12:30] Monitor cycle ⅙ - A:● B:● C:○ | Confidence: 94%
[00:13:00] Monitor cycle 2/6 - A:● B:● C:○ | Confidence: 92%
[00:13:30] Monitor cycle 3/6 - A:○ B:● C:○ | Confidence: 91%
[00:13:31] ⚠️ Skeptic activated: C stalled >180s with no code changes
[00:14:00] Monitor cycle 4/6 - A:○ B:○ C:○ | Confidence: 89%
[00:14:30] Monitor cycle ⅚ - A:○ B:○ C:○ | Confidence: 88%
[00:15:00] Monitor cycle 6/6 - A:○ B:○ C:○ | Confidence: 87%
[00:15:45] Cycle complete | Severity: YELLOW | Report delivered
[00:15:46] Cycle_log stored | Pattern effectiveness: 67%
[00:15:47] Learning captured: C requires pattern upgrade for future directives

Drift is visible. State changes are tracked. Nothing disappears. —

PART 8: CONSENSUS PROTOCOL

Agent Consensus Thresholds

<consensus_protocol>
  <threshold type="continue_execution">
    Condition: majority of agents report confidence > 85%
    Action: Proceed with current approach
  </threshold>
  
  <threshold type="request_clarification">
    Condition: any agent reports confidence < 60%
    Action: Present findings, ask user for guidance
  </threshold>
  
  <threshold type="abort_and_escalate">
    Condition: 2+ teams show ERROR status OR all agents flag critical issue
    Action: Halt execution, full status report, await user decision
  </threshold>
</consensus_protocol>

Confidence Scoring

Score Interpretation Action
95-100% HIGH Proceed autonomously
85-94% MEDIUM Proceed with monitoring
60-84% LOW Flag concerns in report
<60% UNCERTAIN Request user guidance

SECURITY

Real-time user directives only

  • Verification checkpoint blocks autonomous execution
  • User confirms every mapping
  • Full audit trail maintained
  • All sessions under your oversight
  • Pattern selection is advisory, not autonomous
  • Conditional agents flag but don't override —

FRAMEWORK STATUS

Version: 3.5 (APM-Enhanced)
Status: PRODUCTION-READY
Last Updated: 12/20/2025 00:10 UTC
Created By: Daniel Willitzer + Claude (Claude for Chrome)
Methodology: Battle-Tested v3.1 + Agent Pattern Matrix

SYSTEM READY FOR ACTIVATION

Activation: /cclead [letters] - [directive]

I built this. I stand behind it. Let's execute.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment