| name | retro |
| description | Run a daily deployment retrospective. Use when the user says /retro, finishes a day of a deployment, hits a checkpoint or milestone, or asks to write a deployment retro. Accepts either an interview with one engineer or a paste of meeting notes / a chat scroll / individual notes from multiple engineers. Writes the retro file and optionally submits it to the retro-ingest workflow. |
Deployment retro interviewer
Your job: run today's deployment retro, write the file, and offer to ship it into the feedback funnel.
Step 1 - locate the template and prior retros
- Use the repo's own retro template (
RETRO.md or protocols/RETRO.md). If
neither exists, tell the user to add one and stop.
- Look in
retros/ for the most recent retro (usually yesterday's for the
same customer) to check continuity: which issues were open, which debt was
still _pending_, what decisions were due today.
- Read
DEPLOY.md's Debt Log entries added since the last retro - each one
needs a verdict today.
Step 2 - pick an input mode
Ask which shape the raw input takes, and adapt:
- Interview (one engineer, live): default. Ask in batches of 2-3
questions, not one giant form. Keep answers verbatim where possible.
- Meeting notes / chat scroll / shared doc paste: user pastes the raw
text. Extract structure from it; ask short follow-ups only for missing
fields (severity, count, sanitization gaps).
- Multiple engineers, async: collect what each contributed, then
synthesize into one retro file. Record every contributor by name in the
Contributors field.
Step 3 - walk the template
Walk the fields of RETRO.md (or protocols/RETRO.md) in the order they
appear. The template is the source of truth for what to capture, when each
field applies (e.g. success criteria only on the end-of-deployment retro),
and how to sanitize.
Interview mechanics on top of the template:
- Push for count estimates on issues - "a few times" becomes "~3".
- When walking the Debt Log review, re-check the
Sanitized context field
of every entry before recording its verdict. If a customer name, host, ID,
or literal record slipped in, ask the engineer to rephrase - describe the
shape of the data, not the data - and update the entry before continuing.
- If a contributor is clearly drained, accept short answers and mark gaps
with
<!-- TODO --> rather than dragging the retro out.
Step 4 - write the file
Write to retros/YYYY-MM-DD-<customer>.md (today's date, customer slug),
following the template's structure exactly - especially the ### Issue:
heading format, which downstream tooling parses. Update the Debt Log verdicts
in DEPLOY.md to match what was decided.
Step 5 - submit to the feedback funnel
If the environment variable DEPLOYKIT_RETRO_WEBHOOK_URL is set (or the user
provides a webhook URL), offer to submit the retro to the retro-ingest
workflow:
curl -X POST "$DEPLOYKIT_RETRO_WEBHOOK_URL" \
-H "Content-Type: application/json" \
--data @- <<'EOF'
{"customer": "<customer-slug>", "retro_markdown": <the full retro file contents, JSON-encoded>}
EOF
Confirm with the user before sending - the retro leaves the repo at this
point. If no webhook is configured, move on; the file itself is the artifact.