Loomworks — intent classifier, imperative-shape disambiguation — work item — v0.1

Version. 0.1 Date. 2026-07-19 Author. Claude Code on DUNIN7-M4, on Operator direction. Status. QUEUED. Not started. Explicitly outside Stage 2 scope (Operator ruling, 2026-07-19). Reads with. investigations/loomworks-b0dc30b-parked-state-resolution-v0_1 (the inspection this came from).

Plain-language summary

A prompt improvement to the Companion's intent classifier exists as an unmerged, untested commit. It is good work that would have been lost to a branch retirement. This item is to land it properly — with tests — or discard it on the evidence.

The work

1. Cherry-pick b0dc30b's src/loomworks/orchestration/prompts/intent_classifier.md change onto its own branch off current main in loomworks-engine. Drop the uv.lock half — it carries no intent and the lock regenerates normally.

1. Add classifier tests for the three contrast cases the change introduces:

| Utterance | Expected intent | |---|---| | "Show birthdays." | recall_personal | | "Parked on 13." | remember_about_me | | "Show me my notes." | ask_about_past_input |

1. Add the control that makes those tests mean something: "Set the pH to 6.5." must still classify as add_knowledge. The change's whole purpose is to stop imperative shape being read as a qualifier for add_knowledge — without this control, a change that broke imperative add_knowledge entirely would pass all three tests above and look like a success.

1. Land or discard on the result. If the tests pass with the change and fail without it, the change is demonstrated and lands. If they pass either way, the prompt edit is not doing what it claims and should be discarded rather than kept for looking sensible.

Why it is worth doing

The classifier had guidance that imperative phrasing does not disqualify add_knowledge ("Set the pH to 6.5" is a domain decision, not a command to execute) without the converse — that imperative phrasing does not qualify it either. A prompt that argues only one direction biases toward that direction. The change supplies the missing side.

It sits in the same area as previously recorded classifier problems, which suggests it was written in response to observed misclassification rather than speculatively — though no record of the prompting observation survives, which is part of why it needs tests rather than trust.

Why it is queued rather than done

It arrived by accident of timing, riding a branch about something else entirely, and Stage 2 is a relocation build. Landing an untested prompt change inside it would mean two unrelated things moving in one step. Prompt edits are also exactly where silent regressions live — nothing fails loudly — so this deserves its own change with its own evidence.

Not urgent. The classifier works today; this improves a known-imperfect edge.

DUNIN7 — Done In Seven LLC — Miami, Florida Loomworks — intent classifier, imperative-shape disambiguation — work item — v0_1 — 2026-07-19