memtrace-intent-verification

name

memtrace-intent-verification

description

Verify whether a past decision actually held or was violated, and surface the arc of episodes that implemented that specific decision. Returns a Held | ViolatedAt | CannotProve verdict and the implementing arc from Cortex decision memory. Use when the user asks whether a decision held, was followed, or drifted, or before relying on a decision or reporting drift. Requires a decision_id (from memtrace-decision-recall); do NOT use for free-text decision lookup — use memtrace-decision-recall first. Do not assume a decision was followed; verify it.

Overview

A decision being recorded doesn't mean it was followed. verify_intent checks whether a decision held across its arc; get_arc returns the episodes that implemented it. Use these to confirm a rationale is still in force before relying on it, and to detect drift where the code quietly diverged from what was decided.

Full parameter spec for every Memtrace tool: references/mcp-parameters.md (bundled at the memtrace-skills plugin root).

Quick Reference

Tool	Purpose	Returns / FactStatus
verify_intent	Did the decision hold across its arc?	Held | ViolatedAt | CannotProve
get_arc	Arc of episodes implementing the decision	DeterministicallyDerived | CannotProve

Both take a decision_id (uint64), not free text. Get it from recall_decision (see memtrace-decision-recall) or from why_is_this_here (see memtrace-provenance). Do not invent ids.

Honesty contract: verdicts are deterministically defended. CannotProve means the decision is invisible or has no implementing episode — unknown, not "fine."

Steps

1. Get the decision_id

Run recall_decision(...) and take a decision_id from a hit, or get one from why_is_this_here(symbol_id). These tools do not accept names.

2. Verify it held

verify_intent(decision_id):

Verdict	Meaning	Action
Held	The decision held across its arc	Safe to rely on; honor it
ViolatedAt	The code drifted from the decision at a point	Surface the drift — it's either a bug or an unrecorded reversal; ask which
CannotProve	Decision invisible or no implementing episode	Treat as unknown; don't claim it held

3. Inspect what implemented it

get_arc(decision_id) → the episodes reachable by Produced/DerivedFrom edges. This is the implementation trail: where the decision actually landed (or where it should have).

4. Cross to the code graph for the "what"

The arc names episodes/symbols; to see the current code or its blast radius, hand those over to memtrace-first / memtrace-impact. Decision memory proves whether it held; the code graph shows the current state.

Decision Points

Situation	Action
Relying on a recalled decision for a choice	verify_intent first — only trust a Held verdict
Auditing whether the codebase follows its own decisions	verify_intent per decision; report every ViolatedAt
"Where/when was decision D implemented?"	get_arc(decision_id)
verify_intent returns ViolatedAt	Don't silently 'fix' it — it may be a deliberate (unrecorded) reversal; confirm with the user
verify_intent returns CannotProve	Treat as unknown; fall back to recall_decision + the code graph

Output

Call	Returns
verify_intent(decision_id)	Verdict: Held | ViolatedAt (the point where code drifted from the decision) | CannotProve (decision invisible or no implementing episode — unknown, not "fine")
get_arc(decision_id)	FactStatus DeterministicallyDerived | CannotProve, plus the implementing arc: the episodes reachable by Produced/DerivedFrom edges