Version. 0.1 · Date. 2026-08-06 · Author. Claude.ai (drafting) / Marvin Percival (approving).
Target. /Users/dunin7/loomworks-engine, branch main.
Baseline. Post CR-2026-178. CC confirms exact HEAD at Step 0.
Companion to. scoping-notes/loomworks-b12-scoping-note-v0_6.
Charter. Autonomous regime. Status. Ready for CC execution.
Recall is wholesale over a fifty-item window. The largest engagement anyone actually works in holds fifty-eight notes. Twenty-three engagements hold notes; two hundred thirty-five between them; eleven hold one or two each.
For every engagement that exists today, raising the constant closes this. No index, no search, no extension. The scoping note settled that.
What remains is the part the build list explicitly refuses to guess at: measure what the model's prompt actually holds against fifty-eight notes, and raise the limit to a figure someone measured rather than a round number that sounds safe.
Step 1 — measure (read-only). Against the largest real engagement (58 notes), determine what the assembled prompt actually costs in tokens, and how much headroom remains against the model's context limit after the rest of the prompt (system material, conversation history, the question itself) is accounted for. Report: tokens per note (mean and worst-case), total at 58, and the headroom.
Step 2 — choose and set. Raise the constant to a figure derived from Step 1's measurement, with stated headroom. CC states the figure and the reasoning in its report. The build list's own instruction governs: measured, not round-and-safe-sounding.
Step 3 — test. A test asserting the new constant is honored, and one asserting behavior at the boundary (an engagement at exactly the limit, and one over it, degrade predictably rather than truncating silently).
| Step | What | Mode |
|---|---|---|
| 0 | Pre-flight. | Auto |
| 1 | Measure against the 58-note engagement. Report figures. | Auto |
| 2 | Raise the constant to the measured figure; state reasoning. | Auto |
| 3 | Tests — constant honored; boundary behavior predictable. | Auto |
| 4 | Full suite; ruff 0; mypy at/below baseline. | Auto |
| A | Checkpoint. | Checkpoint |
| 5 | Tag cr-2026-179-b12-recall-window. Push. | Auto |
Execute CR-2026-179 v0.1 at ~/Downloads/cr-2026-179-loomworks-b12-
recall-window-v0_1.md. Confirm the CR number against the ledger first.
B-12: recall is wholesale over a 50-item window; the largest real
engagement holds 58 notes. Raising the constant closes this for
everything that exists today.
Step 1 is a measurement, and it's the point of the CR. Measure what the
prompt actually holds against the 58-note engagement — tokens per note,
total, and headroom against the model's limit after system material and
conversation history are accounted for. Report the figures.
Step 2 sets the constant FROM that measurement. The build list is
explicit: a measured figure, not a round number that sounds safe.
State your reasoning in the report.
Out of scope, and don't let it creep in: the universal commons (grows
with adoption, parked as its own question) and anything resembling
search or an index (ruled out by the scoping note).
Baseline: post CR-2026-178. Confirm current HEAD at Step 0.
DUNIN7 — Done In Seven LLC — Miami, Florida CR-2026-179 — B-12 recall window — v0.1 — 2026-08-06