Deferred deliberately by PR #2000, which added check I23 (context-budget directive to stop, summarize, or hand off) with no pre-scan pattern seeded. The row states the reason in its own terms; this item carries the trigger for revisiting it.
Why it was not seeded at authoring time
Unlike I8-e — which withholds seeding because no instance is attested — I23 has an attested, unfenced true positive: plugins/session-flow/skills/handoff/SKILL.md, whose own invocation triggers are "context heavy (check /context output)", "quality degrading", and a window-position threshold ("roughly beyond the final third of the window"). So this row is not waiting on an instance. It is waiting on calibration.
What needs calibrating
The threshold and window-position phrasings vary far more than the fixed shapes a row like I8-b matches. A pattern seeded before that variance is characterised would over-produce badly, and the blast radius is the reason it matters: a continuation skill can barely be model-invocable without naming a context trigger somewhere, so a loose pattern fires on every consumer's handoff, compaction, and continuation skill.
One of the attested hits is in a description: field, which is resident whenever the skill listing admits it rather than only when the body loads — a distinction I23 currently draws nowhere, and one a pattern would need to respect.
The open policy question the row deliberately leaves to a human
Given that blast radius, two coherent policies exist and PR #2000 did not choose between them:
- Continuation skills become user-invocable only, so their context triggers are not model-facing instructions; or
- Their trigger moves to the measured mechanism — which is what
context-guard already is.
Answering that first would tell a pattern what it is allowed to match, which is why the seeding is downstream of it rather than a mechanical follow-up.
Deferred deliberately by PR #2000, which added check I23 (context-budget directive to stop, summarize, or hand off) with no pre-scan pattern seeded. The row states the reason in its own terms; this item carries the trigger for revisiting it.
Why it was not seeded at authoring time
Unlike I8-e — which withholds seeding because no instance is attested — I23 has an attested, unfenced true positive:
plugins/session-flow/skills/handoff/SKILL.md, whose own invocation triggers are "context heavy (check/contextoutput)", "quality degrading", and a window-position threshold ("roughly beyond the final third of the window"). So this row is not waiting on an instance. It is waiting on calibration.What needs calibrating
The threshold and window-position phrasings vary far more than the fixed shapes a row like I8-b matches. A pattern seeded before that variance is characterised would over-produce badly, and the blast radius is the reason it matters: a continuation skill can barely be model-invocable without naming a context trigger somewhere, so a loose pattern fires on every consumer's handoff, compaction, and continuation skill.
One of the attested hits is in a
description:field, which is resident whenever the skill listing admits it rather than only when the body loads — a distinction I23 currently draws nowhere, and one a pattern would need to respect.The open policy question the row deliberately leaves to a human
Given that blast radius, two coherent policies exist and PR #2000 did not choose between them:
context-guardalready is.Answering that first would tell a pattern what it is allowed to match, which is why the seeding is downstream of it rather than a mechanical follow-up.