retro

Retro

A retro that runs on memory rehashes opinions; one that runs on the ledger starts from what actually happened. The engine returns the facts for the window; render them as a short pack the user can bring to their own retro — including the numbers that didn't go well. No blame, no peers, no spin.

"${CLAUDE_PLUGIN_ROOT}/bin/waybill" mine --all      # catch-up metering first, always
"${CLAUDE_PLUGIN_ROOT}/bin/waybill" query report --from <iso> --to <iso>
"${CLAUDE_PLUGIN_ROOT}/bin/waybill" query manifest
"${CLAUDE_PLUGIN_ROOT}/bin/waybill" query untracked --from <iso> --to <iso>

Window: the sprint the user names; "last sprint" from config.baseline.window sprint length if known, else the last 14 days — state the window in the header either way.

Render (short sections, in this order)

Work from query report's .data:

1. Shipped — one line per item from .data.shipped (key, title, points, PR); then the totals line: points, PRs, tokens, cost if priced.
2. Estimates vs. reality — from .data.calibration: coverage first ("4 of 5 shipped items pre-registered — 80%"), then the positions: below = the work came in under the no-Claude range (the claimed saving held in full), within = partial, above = the item took longer than the no-Claude estimate — name those items and say it plainly; a retro that hides the above-range items is theater. If with_actuals lags pre_registered, one line: recording actual_hours at ship time is what makes next sprint's claims checkable.
3. Model mix — from .data.model_mix: tokens-per-point by dominant model ("opus-carried stories: 0.9M/pt over 4 stories; sonnet-carried: 0.4M/pt over 2"). Note mixed stories separately. Own history only; two rows with one story each is an anecdote, not a comparison — say so.
4. Waste & rework — .data.costs.waste (retried commands, repeated reads) and .data.costs.reopened_count, each with one factual line. Zero is worth stating ("nothing reopened").
5. Cache economics — from .data.spend_ledger.cache_savings: one line ("cache reads were 61% of volume and saved ~$118 vs. uncached list rates — derived at today's rates, covering 94% of cache volume"). Always label it derived; skip the dollars when saved_usd is null.
6. Still on the truck — from query manifest: open items with spend, flagging sitting rows ("PLAT-510: 1.2M tokens, idle 16 days"). From query untracked: the untracked share, with a one-line handoff to the salvage skill when it is notable.
7. Close — at most three factual process observations drawn ONLY from the numbers above (e.g. "estimates covered 80% of shipped work, up from nothing", "two items sat idle past the demurrage line"). No advice the numbers don't support; the user draws the conclusions — it's their retro.