operational-protocols

name: operational-protocols description: Mechanical guardrails for Claude Code — Slack gateway, infra gate, activity log, context recovery. Read when diagnosing performance failures or setting up a new environment. type: reference

Operational Protocols — Claude Code

Created: 2026-04-21
Reason: April 20 sessions had 4 systemic failures. These protocols replace memory-based rules with mechanical enforcement.

The core principle

Rules that live only in memory fail when context compacts. Rules enforced by hooks, scripts, and scheduled tasks survive compaction.

Every protocol below has a MECHANICAL layer (hook or scheduled task) plus a BEHAVIOURAL layer (what Claude must do when the gate fires).

Protocol 1 — Slack Send Gate

Failure it prevents: Composing and sending Slack messages without Samer's review, especially after context compaction when prior conversation is lost.

Mechanical layer:
PreToolUse hook on slack_send_message — fires on every call, injects mandatory checkpoint.

Behavioural rule:
Before calling slack_send_message, Claude must:

1. Show Samer the exact message text in conversation and receive explicit approval ("send it", "yes", "go ahead")
2. State which channel/Jira thread was searched to confirm consistency with prior comms
3. If either is missing → use slack_send_message_draft instead, never slack_send_message

Escalation: If violated, this is a repeat offense (3rd recorded instance as of 2026-04-20). Next violation → Samer reviews hook config to block sends entirely.

Protocol 2 — Infra Destructive Action Gate

Failure it prevents: Unilateral deletion of Azure resources, Kubernetes objects, or files without explicit per-action approval.

Mechanical layer:
PreToolUse hook on Bash — bash-destructive-gate.py scans command for destructive patterns and injects a STOP checkpoint if found.

Patterns monitored:

• az … delete, az … remove, az group delete
• kubectl delete
• terraform destroy, terraform apply
• rm -rf, rm -fr
• SQL DROP TABLE, DROP DATABASE
• gh repo delete
• git push --force, git reset --hard

Behavioural rule:
When the gate fires, Claude must answer all 3 before proceeding:

1. What exactly will be deleted — name the specific resource
2. Is this reversible? (Yes/No — be explicit)
3. Did Samer explicitly approve this specific action in this session? ("clean up the infra" ≠ approval to delete a named NIC

If any answer is uncertain → stop and ask Samer.

Protocol 3 — Activity Log Enforcement

Failure it prevents: Sessions running without any log entry, making EOD wrap impossible and breaking cross-session continuity.

Mechanical layer:
Scheduled task activity-log-completeness-check runs at 23:00 local time every day. If no entry for today → DMs Samer to reconstruct.

Vault path:
/Users/samertadross/Library/Mobile Documents/iCloud~md~obsidian/Documents/Scrips Vault/01-operating-state/daily-activity-log.md

Behavioural rule:
Claude must append a log entry:

• At every natural topic shift within a session
• Before ending any session (before the Stop hook fires)
• Format: ### HH:MM — claude-code — [short title] [TAG]

Minimum viable entry (if session was simple): one Done: line. Never skip entirely.

Protocol 4 — Context Compaction Recovery

Failure it prevents: After context compaction, Claude proceeds on stale or fabricated facts because it doesn't know what it has lost.

Mechanical layer:
UserPromptSubmit hook fires on every prompt, injecting a rules reminder into Claude's context. This survives compaction because it re-injects on every message.

Behavioural rule:
If Claude detects that context was compacted (conversation summary visible, prior tool results gone):

1. Read daily-activity-log.md to reconstruct what happened today
2. Read any handoff doc at ~/claude-os/docs/handoffs/ dated today
3. Search Slack/Jira for the active topic before sending any message or taking any infra action
4. Do NOT proceed on memory of what "seemed" to be decided — verify from live sources

Specific rule for Manish/infra communications:
After any compaction in sessions involving Manish, KV access, or PROD infra — always search #general and the relevant Jira ticket before composing any message.

Protocol 5 — Session Wrap (existing, strengthened)

Failure it prevents: Work done without traceability.

Mechanical layer:
Stop hook fires session-end.sh which invokes session-capture skill.

Behavioural rule:
If Claude ends a session without a log entry, the Stop hook will attempt capture — but it needs context. Claude should write the activity log entry BEFORE the session naturally ends (not rely on the hook to reconstruct).

Monitoring

Hook file locations