Session Wrap-Up
Summarize what was done, record what was learned, and update every project document the session affected, so the next session (you, another agent or a human) can pick up where this one stopped by reading the Session Continuity section of docs/context/STATUS.md. The wrap is done when the docs are committed and the final report is shown.
Working folder. Blueprint working files live under .agent-blueprint/ in the project root. Before the first write there, make sure .agent-blueprint/.gitignore exists and lists run/, team/, review-runs/, cache/ and .gitignore, so run state and the ignore file itself stay out of commits while plans and notes stay tracked.
Provenance record. Before any other work, run this as written from the project root: mkdir -p .agent-blueprint/run/provenance && date -u +'{"skill": "ab-session-wrap", "version": "4.2.1", "started_at": "%Y-%m-%dT%H:%M:%SZ", "helper_steps": []}' > .agent-blueprint/run/provenance/ab-session-wrap.json. It replaces any older record, and the Working folder step above has already made sure of .agent-blueprint/.gitignore. Each Helper step adds {step, path: helper|inline} to helper_steps. The record tells a run, and the smoke test, which skill ran and how; it is not a security control.
Documentation only. Apart from the working files above, change no source code, tests, configs or infrastructure, since a wrap that edits code ships an unreviewed change; a needed code change goes into BACKLOG.md. Git history and the file system are the ground truth: record only work you can find there. The remaining constraints are in references/step-checklists.md § Constraints.
Step 1: Gather Context
Before writing anything, read everything in references/step-checklists.md § Step 1 files to read, in parallel where you can, and run the commands in references/templates.md § Step 1 git state.
Step 1.5: Classify Session Type
Run git log --diff-filter=AM --name-only --since="8 hours ago" --format="" | grep -v -E '^(docs/|\.agent-blueprint/|AGENTS\.md|CLAUDE\.md|BACKLOG\.md|blueprint\.local\.md)'. Empty output means a planning session (only plans, research, design docs, decision records or ideation changed); anything else is an implementation session, which runs every step.
A plan is not delivery, so a planning session marks no goal, milestone, task or plan as done anywhere, records the plan as planned work in Steps 4, 6, 8 and 10, and skips Steps 7 and 11. Details: references/step-checklists.md § Planning session rules.
Step 2: Analyze Session Work
Before writing, answer the questions in references/step-checklists.md § Step 2 analysis questions.
Step 3: Present Summary to User
Present the summary in references/templates.md § Step 3 summary, with file paths, commit hashes and numbers. Ask whether it is accurate and what to add or correct, and change no file before the answer: the user may know things git does not.
Asking the user. Ask with your question tool if you have one, offering at most three options; otherwise ask in plain text with a numbered list. In a headless or unattended run nobody will answer: take the default named below, say so in your output, and log it in the run state's decisions if there is a run state.
Options: 1. Accurate, update the docs. 2. Correct or add something first. Default when nobody answers: update the docs from the summary as shown, and say in the Step 17 report that nobody confirmed it.
Step 4: Session Continuity in docs/context/STATUS.md
Rewrite the Session Continuity section at the top of docs/context/STATUS.md (create it if missing) from references/templates.md § Step 4 Session Continuity template. Write plain text with no HTML comments, because Hermes refuses a context file that has one. Regenerate it from the project state (git log, tests, files), never from the previous section, because a summary of a summary drifts. Rules and reasons: references/step-checklists.md § Step 4 rules.
Step 5: Update docs/learnings/LEARNINGS.md
This step always runs: judge whether any learning clears the bar, and report the answer even when it is no. Append entries per references/step-checklists.md § Step 5 entry format and rules. If none clears it, append nothing and state "No durable learnings this session" in the Step 17 report only, never in LEARNINGS.md.
Step 6: Update docs/context/STATUS.md
Update every table per references/templates.md § Step 6 status tables, and the "Last updated" date at the top.
Step 7: Update docs/context/CONVENTIONS.md (if needed)
Only on a trigger in references/step-checklists.md § Step 7 triggers; otherwise skip the file.
Step 8: Update docs/context/GOALS.md (if needed)
Only on a trigger in references/step-checklists.md § Step 8 triggers; otherwise skip the file.
Step 9: Update BACKLOG.md
Update Inbox, Triaged and Parked per references/step-checklists.md § Step 9 backlog sections.
Step 10: Update Active Plans (if applicable)
Update a plan the session followed per references/step-checklists.md § Step 10 plan updates; skip if none.
Step 11: Update Active Specs (if applicable)
Update a spec being implemented per references/step-checklists.md § Step 11 spec updates; skip if none.
Step 12: Create ADRs (if applicable)
Create an ADR only for a decision that is hard to reverse, would surprise a reader of the code, and involved a real trade-off (all three; full criteria in docs/decisions/README.md, "## When to Create an ADR"). Format: references/step-checklists.md § Step 12 ADR format.
Step 13: Update Auto-Memory (if it exists)
If your host keeps a project memory file (such as Claude Code's auto-memory), update it per references/step-checklists.md § Step 13 memory updates; skip if none.
Step 14: Clean Up Temporary Artifacts
Run the checks and actions in references/templates.md § Step 14 cleanup.
Step 15: Stamp STATE.md (if it exists)
If docs/context/STATE.md exists, use the ab-session-continuity skill to record the HEAD sha (head:) and timestamp (last-updated:) in its frontmatter, the stamp ab-resume-session checks next time. Do not create STATE.md: a project with no execution state stays without one.
Step 16: Commit Documentation Updates
Stage BACKLOG.md and docs/ and commit with the message in references/step-checklists.md § Step 16 commit messages.
No-commit mode. If the user says not to commit, AGENT_BLUEPRINT_GIT_WRITABLE is 0, or .git is read-only, make no commits: leave changes in the working tree and add the intended message to .agent-blueprint/run/commit-msg.md. With AGENT_BLUEPRINT_GIT_WRITABLE=0, the ship runner commits after the session, so it is not for work the user wants uncommitted. Reviews use the working tree and untracked files against the merge base instead of a commit range.
Step 17: Final Verification
Run references/templates.md § Step 17 verification, then give the user the report in references/templates.md § Step 17 report.
Success Criteria
Before ending the session, tick every item in references/templates.md § Success criteria.