[P0] ZARA-007 — Eliminate ambiguous intent precedence and decode Python skill terms correctly #8

New issue

Closed

opened 2026-07-18 19:19:10 +00:00 by lost-rob0t · 0 comments

lost-rob0t commented

2026-07-18 19:19:10 +00:00

(Migrated from github.com)

Problem

kb/intents.pl defines conflicting surfaces where clause order silently chooses behavior:

stop is media stop before conversation stop.
end is dictation stop before conversation stop.
search/find are web search before Python todo search.
hello is greet before the Python greeting skill.

Python skill intents are represented as compound Prolog terms (python(Skill)), while zara/wake.py only handles strings beginning with python(. Depending on PySWIP conversion, valid skills can bypass Python execution and fall into Prolog command execution.

Required fix

Define deterministic precedence based on conversation state and required slots, not source clause order.
Add an explicit typed result adapter from PySWIP values to Python intent objects.
Separate conversation controls, dictation controls, media controls, web search, and todo search.
Add missing-slot behavior for commands such as bare open, text, and schedule instead of accepting empty arguments.

Required tests

Table-driven resolver fixtures covering every duplicate surface in passive, active-conversation, and dictation states.
Tests for compound python(Skill) values using the actual PySWIP representation plus a mock adapter.
Add scripts/test-intents.sh that runs a stable corpus and diffs actual intent/args against a checked-in expected fixture.

Acceptance

end/stop do the state-appropriate action.
Todo search cannot be shadowed by web search.
Every Python skill route is executable from wake mode.
Missing required slots produce a pending/clarification result, not fake execution.

Branch

fix/zara-007-intent-precedence

Dependencies

ZARA-003, ZARA-006.

## Problem `kb/intents.pl` defines conflicting surfaces where clause order silently chooses behavior: - `stop` is media stop before conversation stop. - `end` is dictation stop before conversation stop. - `search`/`find` are web search before Python todo search. - `hello` is greet before the Python greeting skill. Python skill intents are represented as compound Prolog terms (`python(Skill)`), while `zara/wake.py` only handles strings beginning with `python(`. Depending on PySWIP conversion, valid skills can bypass Python execution and fall into Prolog command execution. ## Required fix - Define deterministic precedence based on conversation state and required slots, not source clause order. - Add an explicit typed result adapter from PySWIP values to Python intent objects. - Separate conversation controls, dictation controls, media controls, web search, and todo search. - Add missing-slot behavior for commands such as bare `open`, `text`, and `schedule` instead of accepting empty arguments. ## Required tests - Table-driven resolver fixtures covering every duplicate surface in passive, active-conversation, and dictation states. - Tests for compound `python(Skill)` values using the actual PySWIP representation plus a mock adapter. - Add `scripts/test-intents.sh` that runs a stable corpus and diffs actual intent/args against a checked-in expected fixture. ## Acceptance - `end`/`stop` do the state-appropriate action. - Todo search cannot be shadowed by web search. - Every Python skill route is executable from wake mode. - Missing required slots produce a pending/clarification result, not fake execution. ## Branch `fix/zara-007-intent-precedence` ## Dependencies ZARA-003, ZARA-006.