Retro
Last updated 2026-08-21
Once entries pile up they become material. Pick a period on the Retro screen (⌘6) and it pulls deterministically computed signals out of that period's records — shown before any AI is involved, with narrative layered on top only if you want it.
Periods are Last 7 days · 14 days · 30 days.
Four numbers first
| Metric | Meaning |
|---|---|
| Entries | Records left in the period |
| Shipped | Completed feature-side work |
| Resistance | Bugs and error cycles — how much wall you hit |
| Agents | How many agents contributed |
The ratio of Shipped to Resistance characterizes the period. A week where resistance spikes usually means the design wobbled or you were in unfamiliar territory.
Signal cards
- What shipped / What pushed back — which work fell on each side
- Files that came up again and again — if the same file keeps appearing in bug and error entries, that file is the problem
- Where the effort went — files that took the most work, overlaid with dependency fan-out from the code map. A file flagged Core hub is heavily edited and heavily depended on — the most dangerous place to touch
- Agent contributions — who did what, if you run several
"Files that came up again and again" is the fastest way to pick refactor candidates. Not by instinct, but by how many times that file actually cost you time.
The deferred-shortcut ledger
Leave a comment like this in code and the retro harvests it into one place:
// oculpm-defer: <ceiling>; <revisit trigger>- Ceiling — how far this shortcut holds (e.g. "up to 1000 items")
- Revisit trigger — when to look again (e.g. "when pagination lands")
A marker with no trigger is flagged No trigger, because a shortcut without a revisit condition quietly rots — write a trigger, or promote it to a planner item outright.
Generating artifacts
Turns collected entries into something shareable. Three formats:
| Kind | Use |
|---|---|
| Standup | What I did yesterday / what's next |
| PR description | Explaining this period's changes on a PR |
| Weekly status | A week summarized |
Two ways to produce them:
- Generate — uses the default AI provider and model from Settings (your API key, your billing)
- With Claude Code — hands it to a terminal Claude Code session. No API key and no extra billing — it runs on your subscription, and you watch it work in the terminal
Even without AI, the plain template gives you the bones. Generated artifacts show how many entries they came from and whether they were AI-written or templated.
Exporting
- Export
.md— the period's entries as a single markdown file - Copy to clipboard — paste the artifact straight out
- To Notion — creates a page and opens it in a new window
Eval trend
Record scores from a test suite and the retro draws a trend. With no records the card says it's empty — log scores with the run-evals skill and it starts filling in.
When to open it
- Friday afternoon — build the weekly status and check "files that came up again and again"
- Right before opening a PR — pull a PR description as your draft
- Starting a new sprint — pick what to repay from the deferred-shortcut ledger
Next steps
- How the raw material accumulates → Work Journal
- Promoting what you picked into a plan → Planner