Deploy — Post-Merge Deployment Workflow

Structured guidance for everything between /pr-merge (code merged) and /ops (long-term observability). Covers feature flags, gradual rollout, post-deploy verification, and rollback conditions. Auto-detects context from project signals or accepts explicit subcommand.

Usage

/deploy              # Auto-detect from project signals
/deploy flags        # Explicit: feature flag setup (LaunchDarkly, Flipt, env-based)
/deploy canary       # Explicit: canary / blue-green gradual rollout
/deploy verify       # Explicit: post-deploy smoke tests and SLO gate
/deploy rollback     # Explicit: rollback trigger conditions and strategies

Workflow

Step 1: Determine Context

Parse $ARGUMENTS for explicit mode. If no arguments, run context detector:

DETECTED=$(bash "$HOME/.claude/skills/deploy/lib/context-detector.sh")

Step 2: Show Relevant Guidance

Based on detected or explicit mode, use progressive disclosure:

Level 1 — Summary (default, ~15 lines): Read and present the summary template. This is the entry point.

Level 2 — Patterns (on request, ~60 lines): When the user asks for more detail, "show me patterns", code examples, or "Don't/Do/Best" style guidance, read the patterns template.

Level 3 — Full SOP (future): Not yet available. Dive into patterns and answer specific questions instead.

Step 3: Handle Multiple Contexts

If context-detector returns multiple modes (e.g., "verify,rollback"):

• Show summary for each detected mode
• Ask: "Which area do you want to dive deeper into?"

Step 4: No Context Detected

If detection returns "none" and no explicit mode was given:

• Read and show templates/overview.md
• Let the user choose which mode to start with

Template Locations

Summary Templates (~15 lines each — L1 default)

Patterns Templates (~60 lines each — L2 on request)

Reference Templates

Progressive Disclosure Rules

1. Always start at L1 — never dump 300+ lines unprompted
2. Read templates on demand — use the Read tool to load templates, don't memorize content
3. User drives depth — "show me patterns", "give me code", "how do I set up X" triggers L2
4. Cross-reference, don't duplicate — for secrets in flags, cite /security operations; for post-deploy metrics, cite /ops metrics

Relationship to Other Skills

• /pr-merge: Merge is the trigger. /deploy is the next step after the PR lands on main.
• /ops: Covers long-term observability (logging, metrics, dashboards, alerting). /deploy verify is the immediate post-deploy gate; /ops sets up the sustained monitoring layer.
• /security operations: Covers secrets management and deployment security. Use before configuring flag SDK credentials.
• /implementation: After implementing a flagged feature, use /deploy flags to wire the flag evaluation into the request path.

Examples

Auto-detected flags context (launchdarkly in deps):

→ context-detector returns "flags"
→ Read and show templates/flags/summary.md
→ User: "show me the SDK patterns"
→ Read and show templates/flags/patterns.md

Explicit canary request:

User: /deploy canary
→ Read and show templates/canary/summary.md

No context, overview:

User: /deploy
→ context-detector returns "none"
→ Read and show templates/overview.md

Post-merge verification flow:

PR merged → /deploy verify → smoke tests pass → promote to 100% traffic
→ if any check fails → /deploy rollback

Gotchas

• Never roll back a DB migration without a tested down-migration — data loss risk. Stop traffic and escalate instead.
• Automatic rollback triggers on 5xx > 1% absolute, p99 > baseline × 1.5, or health endpoint non-200 — these fire without manual confirmation; know them before deploying.
• Flag-gated and canary rollbacks are instant (flip flag / drain traffic); only a full-deploy rollback needs git revert + redeploy (5–15 min) — pick the fastest path that applies.
• Every automatic rollback in production requires a post-mortem, and the rollback event must be logged (version numbers, trigger, timestamp).