b0dc30b parked working-tree state — resolution note — v0.1Version. 0.1 Date. 2026-07-19 Author. Claude Code on DUNIN7-M4. Status. Inspection complete. Recommendation only — nothing applied or discarded. Answers. Stage 2 build brief v0.1 §6.3, which required this commit be resolved SEPARATELY from the CR-2026-153 salvage decision.
b0dc30b was created at the start of the Stage 1 perimeter build to park two files that were sitting modified in the engine's working tree. It was committed deliberately without inspection, so Stage 1 could branch cleanly off main, and its message says exactly that.
It has now been inspected. It is not junk. It is a coherent, self-consistent improvement to the Companion's intent-classifier prompt, addressing a real ambiguity — and it is not on main, is untested, and would otherwise retire unexamined along with the CR-2026-153 transport it happens to sit beside.
Recommendation: keep the content, land it as its own change with tests. Do not retire it with CR-2026-153, and do not merge it as-is.
Two files, +15 / -2.
src/loomworks/orchestration/prompts/intent_classifier.md — the substantive part. It extends the add_knowledge intent description with a caveat and a contrast block:
add_knowledge ("Set the pH to 6.5" is a domain decision, not a command) — but the sentence's subject matter decides, not its grammatical shape.add_knowledge:recall_personalremember_about_meask_about_past_input
uv.lock — an ordinary dependency-lock update, carrying no intent.
The change targets a genuine failure mode: the classifier had guidance saying imperative shape does not disqualify add_knowledge, without the converse — that imperative shape does not qualify it either. A prompt that only argues one direction biases toward that direction. The contrast block supplies the missing side.
It also sits in the same area as previously recorded classifier problems (personal-memory misclassification, and a separately filed tune_setting false positive), which suggests it was written in response to observed behaviour rather than speculatively.
It is untested. Of the three contrast examples, only "Show me my notes" appears anywhere in the test suite; "Show birthdays" and "Parked on 13" appear in no test file. A prompt change with no test asserting the behaviour it intends is a change whose effect nobody can demonstrate — and prompt edits are exactly where silent regressions live, because nothing fails loudly.
Its provenance is thin. The commit message is honest that the content was uninspected. There is no record of which observed misclassification prompted it, or whether it was verified against one.
It is unrelated to what it sits beside. It rode cr-2026-153-tools-menu-engine by accident of timing, not by subject. Retiring that branch would retire this with it, which is why the build brief required it be handled separately.
b0dc30b's intent_classifier.md change onto its own branch off current main.recall_personal, remember_about_me, and ask_about_past_input respectively — and, as the control that makes them meaningful, that "Set the pH to 6.5" still classifies as add_knowledge. Without that control, a change that broke imperative add_knowledge entirely would pass.uv.lock half; let the lock regenerate normally.Not urgent. The classifier works today; this improves a known-imperfect edge. It is recorded so the work is not lost to a branch retirement, which is the only thing that would have happened otherwise.
DUNIN7 — Done In Seven LLC — Miami, Florida
Loomworks — b0dc30b parked working-tree state — resolution note — v0_1 — 2026-07-19