Version. 0.29
Date. 2026-08-05
Charter. standing-notes/dunin7-standing-authorization-charter-v0_1.
Author. Claude Code (execution session). Operator: Marvin Percival.
Supersedes. v0.28 at record 200312b. All earlier versions stand as siblings.
Changes from v0.28. CR-2026-169 (B-11, the unstable answer) executed and CLOSED. Both target repositories merged, tagged, pushed; both first CI runs against the push succeeded.
change-requests/cr-2026-169-answer-sources-v0_1.md, execution session, grounded in inspection-briefs/loomworks-b11-step-0-findings-v0_1.md (v0.28's inspection). Two repositories: loomworks-engine (from main afe8507) and loomworks (from main c12a1ad). Step 0 baselines confirmed clean and matching the CR's stated HEADs before any change: engine pytest -q 3522 passed / 68 skipped; surface lint 0 / typecheck 0 / tests 696/696 / build clean.
Engine — main bd60cec, tag answer-sources-v0_1.
temperature=0 (responder.py) — the one LLM seam left at the SDK default.ConverseResponse.sources: list[AnswerSource] | None, attached from RouteResult.operation_data (system fact, never model-derived) — reuses the existing conduit recall_personal/show_held already use, no second mechanism._format_past_input states the retrieval boundary (computed, not left to the model) when the 50-row window is partial./operator/converse endpoint with a foreign-engagement assertion proving the sources field cannot contain what wasn't retrieved.
Surface — main 1245f1c, tag answer-sources-surface-v0_1.
sources rendered as a compact, read-only reference row (SourcesInline), mirroring the held-tray/matched-files pattern.item_id (both sides of the wire) — the engine's assertion_id is forbidden vocabulary on this surface's vocabulary-wall test; item_id is the term HeldItem/MatchedFileItem already use for the identical concept.sources via structured_data (the mechanism held_items uses) — the CR's §3.4 asks only that sources render, not that they survive a reload. Small, well-scoped follow-on if wanted.
Live pass (Checkpoint A), 3 real Anthropic calls: two identical ask_about_past_input responder calls came back byte-identical (direct demonstration of the D-1 fix); one partial-window call produced a real model reply plainly stating the retrieval boundary ("I'm only showing you the 2 most recent of 137 notes on record...").
Both Step 0 determinations recorded per the CR's Step 5 requirement:
CreditExhaustedError enforcement, a credit-integrity change outside this CR's scope. Recorded per acceptance gate item 9.
Implementation notes filed in both repositories (docs/phase-crs/cr-2026-169-implementation-notes-v0_1.md, each).
Steps 1 through 6 were committed directly to each repository's main rather than on a feature branch, so Checkpoint B's --no-ff branch-merge ceremony had no branch to merge — HEAD was tagged and pushed directly instead. This is a process deviation from the CR's Section 5 template (which names a working branch per the pattern CR-2026-167/168 used), not a scope or evidence-base divergence — C-3 does not apply; the CR's text, targets, and steps were followed exactly, only the branching mechanic was skipped. Every prior CR in this record used a feature branch; this session did not follow that convention and should have. No cleanup action taken — rewriting already-pushed history to retrofit a branch would be worse than the deviation itself. Flagged here for the Operator's awareness; future execution sessions should open a feature branch at Step 0 before the first commit, per the established pattern.
Both succeeded: engine run 31024698451 (success, ~17min, Engine test suite); surface run 31024711320 (success, Surface CI).
R-2 held — CR executed through its own checkpoints, per-step commits, no Operator confirmation needed at Checkpoint A (no anomaly). C-1 protocol followed at Checkpoint B: fetch + HEAD re-verify immediately before push, both repos confirmed not diverged from origin, clean index. R-4 held — one throwaway database at a time, dropped after each use; playground_dev/playground_test untouched. F-1 untouched — no deployment. No anomaly condition (§6) triggered.
None opened by this session's work. D-7's non-implementation is a recorded CR outcome (§2 above), not an open fork — the CR's own conditional language ("if it settles cleanly") already resolved it.
DUNIN7 — Done In Seven LLC — Miami, Florida DUNIN7 — status brief — v0.29 — 2026-08-05 Two identical calls, one pinned temperature, one identical answer — the item is closed.