Session Workflow

Session Workflow Agent

Enforces consistent startup and shutdown discipline across all development sessions regardless of project. Invoked at the start and end of every working session.

For AI session file formats (WORKLOG, CHANGES, COMMANDS, COMMIT_INFO), see ../ai-session-files/SKILL.md.

For branch naming and commit format, see ../git-conventions/SKILL.md.

Session Startup Protocol

Run silently at the start of every session. If a file is missing, note the gap and continue.

Step 1 — Orientation (read in order)

Open with one sentence summarising the current state from AI/nextSteps.md. Example: "Session 5 continues from merge of PR #12; next item is EF Core migration for the new NetworkTest entity."

Step 2 — Git State Check

git status
git log --oneline -5

Report uncommitted changes from a prior session. If stale changes exist, ask Ryan whether to stash, commit, or discard before proceeding.

Step 3 — Dependabot PR Check

gh pr list --state open --author app/dependabot --json number,title,createdAt

If new Dependabot PRs are not in AI/nextSteps.md Tier 4, add them. Review strategy: GitHub Actions bumps → NuGet minor/patch → npm minor/patch → major version bumps (require manual test).

Step 4 — Branch State

If on main and there is in-progress work noted in nextSteps.md, create or switch to the appropriate feature branch before making any changes. main is always protected — no direct commits.

Step 5 — Surface relevant past KNOWLEDGE

Before any implementation work begins, recall past causal entries relevant to the current session topic. This runs every session — it is the low-cost guard against re-diagnosing problems already solved.

# Derive a keyword from the branch name or feature described in nextSteps.md
TOPIC="<branch-name-or-feature-keyword>"

# Find SKILL.md files whose KNOWLEDGE sections mention the topic
grep -ril "$TOPIC" .github/agents/ --include="SKILL.md"

# Print matching KNOWLEDGE entries
grep -h "^- " .github/agents/*/SKILL.md | grep -i "$TOPIC"

# Also scan the evolution log for cross-session patterns
grep -i "$TOPIC" .github/agents/agent-evolution/references/evolution-log.md

Summarise any matches in one sentence before proceeding to implementation. If no matches are found, proceed directly. Step 5 adds 30 seconds; it saves hours the first time it catches a repeat.

Session Shutdown Protocol

Run at the end of every session before the conversation ends.

Step 1 — Update AI Session Files

Update AI/WORKLOG.md, AI/CHANGES.csv, AI/COMMANDS.sh, AI/COMMIT_INFO.txt per the format in ../ai-session-files/SKILL.md.

Step 2 — Commit and Push

git add -A
git commit -m "type: short description

- detail 1
- detail 2"

# If on a feature branch:
git push origin <branch>

# If no PR exists yet:
gh pr create --title "type: short description" --base main

Never commit directly to main. Use github-workflow agent for PR management.

Step 3 — Verify AI/nextSteps.md State

• Mark completed rows ✅ with strikethrough
• Add any newly discovered tasks
• Confirm the "next session start" state is accurate

Step 4 — Invoke agent-evolution (mandatory)

After COMMIT_INFO is written, always invoke the agent-evolution agent:

Invoke agent-evolution to capture session discoveries and update KNOWLEDGE sections.

This step is not optional. If no new patterns were discovered, agent-evolution will record that explicitly in the evolution log. The log entry cost is negligible; the cost of skipping accumulation is not.

Step 5 — Confirm remote is up to date (push verification)

After agent-evolution runs and any evolution-log commits are made, verify nothing remains local-only:

# Show commits that exist locally but have not been pushed
git log --oneline origin/$(git branch --show-current)..HEAD

# If output is non-empty, push is incomplete — push now:
git push origin $(git branch --show-current)

If the remote tracking branch does not exist yet (new branch), push with -u:

git push -u origin $(git branch --show-current)

This is the final guard. A session is only complete when git log origin/<branch>..HEAD returns empty output.

Process Rules

1. Never skip startup orientation — AI has no memory between sessions. nextSteps.md IS the memory.
2. Branch before code — main is protected; create a feature branch before any file change.
3. Log everything — WORKLOG, CHANGES, COMMANDS, COMMIT_INFO are the audit trail.
4. One session = one commit set — Don't accumulate uncommitted work across sessions.
5. Ask about stale changes — Never silently overwrite or discard uncommitted work.
6. End every session with a push — Nothing should exist only on local disk at session end.

AI File Maintenance (quick reference)

PROCESS_KNOWLEDGE

Append new session process discoveries here. Format: YYYY-MM-DD: <discovery>

• 2026-02-27: Main branch is protected via GitHub Ruleset (ID 13316979 on HelloNetworkWorld). All changes must go through PRs. Required checks: .NET Build & Test, Frontend Build & Test.
• 2026-02-27: Runtime config fetch in main.jsx uses top-level await — requires build.target: 'esnext' in vite.config.js. Default esbuild target (es2020) does not support top-level await.