Skill discovery goes straight to a leaf and skips the repo-pointer hop #354

New issue

Open

opened 2026-08-26 02:30:38 +00:00 by coilyco-ops · 0 comments

coilyco-ops commented

2026-08-26 02:30:38 +00:00

Member

Axis C from #351, narrowed after reading context. The disk-location half of this axis has already been fixed. The selection half has not.

What happened

b88f60ea 2026-08-07T21:33:58Z, devrel seat. The agent diagnoses its own miss in the turn before the correction: it went to a leaf writing skill directly and never went up to the repo pointer, which it names as the intended discovery hop. The human then asks whether every repo skill should state its likely on-disk location, and separately asks why the agent read a raw corpus when the curated skill should have carried enough.
e2455c6e 2026-08-10T16:50:40Z, devrel seat. The agent produced drafts, and the human replies with a link to a skill it never opened.

Already fixed, do not re-do

The human's proposal in the first instance has landed. Shipped repo-* skills now open with their on-disk pointer, for example repo-agent-compose states the projects path in its first line. That half is closed.

What is still open

The failure that produced both instances is a leaf-first selection habit. In the first instance the agent had the pointer available and did not traverse to it. In the second the skill was not reached at all. A pointer that exists but is not hopped through fixes nothing, and the current evidence is one confirmed hop-skip plus one unexplained miss.

What would resolve it

Establish whether this is a selection defect or a traversal defect, because the fix differs:

selection - the skill was never in the composed set, which is a trace.json question and answerable directly once #350 lets a session name its bundle
traversal - the skill was present and the agent did not follow the pointer, which is a doctrine question

The second instance is currently unclassified between these two and cannot be settled without knowing what that session composed.

Acceptance condition: for each instance, the decision trace either shows the skill excluded, with a reason, or shows it selected, which makes it a traversal failure.

Limits

Two instances, both on the same seat, three days apart, both in writing-adjacent work. A single-seat single-domain cluster may be a skill-catalogue shape problem rather than a general one. Method and corpus in #351.

Axis C from #351, narrowed after reading context. The disk-location half of this axis has already been fixed. The selection half has not. ## What happened * `b88f60ea` 2026-08-07T21:33:58Z, devrel seat. The agent diagnoses its own miss in the turn before the correction: it went to a leaf writing skill directly and never went up to the repo pointer, which it names as the intended discovery hop. The human then asks whether every repo skill should state its likely on-disk location, and separately asks why the agent read a raw corpus when the curated skill should have carried enough. * `e2455c6e` 2026-08-10T16:50:40Z, devrel seat. The agent produced drafts, and the human replies with a link to a skill it never opened. ## Already fixed, do not re-do The human's proposal in the first instance has landed. Shipped `repo-*` skills now open with their on-disk pointer, for example `repo-agent-compose` states the projects path in its first line. That half is closed. ## What is still open The failure that produced both instances is a leaf-first selection habit. In the first instance the agent had the pointer available and did not traverse to it. In the second the skill was not reached at all. A pointer that exists but is not hopped through fixes nothing, and the current evidence is one confirmed hop-skip plus one unexplained miss. ## What would resolve it Establish whether this is a selection defect or a traversal defect, because the fix differs: * selection - the skill was never in the composed set, which is a `trace.json` question and answerable directly once #350 lets a session name its bundle * traversal - the skill was present and the agent did not follow the pointer, which is a doctrine question The second instance is currently unclassified between these two and cannot be settled without knowing what that session composed. Acceptance condition: for each instance, the decision trace either shows the skill excluded, with a reason, or shows it selected, which makes it a traversal failure. ## Limits Two instances, both on the same seat, three days apart, both in writing-adjacent work. A single-seat single-domain cluster may be a skill-catalogue shape problem rather than a general one. Method and corpus in #351.

coilyco-ops referenced this issue

2026-08-26 02:31:38 +00:00

Transcript mining round 1: six axes where the composed bundle did not bind, ranked #351

coilyco-ops referenced this issue

2026-08-26 02:45:37 +00:00

Eval prep: the measured gap on the seven-seat board, and every eval, prose, and content ticket folded into one ordering #357

coilyco-ops referenced this issue

2026-08-26 02:55:02 +00:00

Seats declare an incapacity they never tested, then hand the work back #352

coilyco-ops referenced this issue

2026-08-26 03:11:07 +00:00

Eval prep: the measured gap on the seven-seat board, and every eval, prose, and content ticket folded into one ordering #357