Concise skills rollout: gaming (2 skills, 1 model-invoked)

Part of #6476. Rewrite gaming's skills to the skill criteria in one draft PR, following the per-plugin runbook (/playbooks:skill-authoring, reference/rewrite-a-plugin.md). The runbook and the rewrite protocol it cites decide what ships; this issue holds only what is specific to the plugin.

The plugin today

Skills: dlss5, setup (user-invoked).
Outcome eval cases under evals/: 0. Probe files under probes/: 0.

Done when

The PR merges with a per-skill decision table in its body, and every shipped description and body was measured.
Any skill left unmeasured, and why, is named in the PR.

Needs a person

Approve the eval spend for this plugin (the /evals:plugin-eval estimate) before the measured passes.
Review the draft PR and mark it ready.

Generated by Claude Code

Thanks for filing this. This is a per-plugin task under the umbrella issue #6476: rewrite the gaming plugin's two skills (dlss5, setup) to the skill criteria, following the per-plugin runbook, with a measured per-skill decision table in the PR body. Scope and done-when criteria are clear and self-contained, and I found no duplicate open issue for this plugin. No missing information from my side.

_{Automated triage. A maintainer reviews it.}

This was generated by AI during triage.

Held for a person: The #6476 owner decision (2026-10-18) defers this middle-group plugin until batch 2's numbers are in.

This was generated by AI.

Parked for capacity, not for the owner. This rollout's remaining work is the measured eval passes of the batch-1 protocol, which the issue lane cannot run: it cannot start claude or claude plugin eval (see the lane's escalation on #6585, https://github.com/melodic-software/claude-code-plugins/actions/runs/40186335064). It waits on the plugin eval dispatch work in #9704, now a blocked-by edge, so admission holds it until #9704 closes.

Under the owner's 2026-10-26 stop rule, a person is needed only for a change in intent or vision, an implementation finding that changes the intent or primary design, or a high or critical security issue; host, attended and eval runs are agent work. So needs-human comes off, and the issue moves to host-run with status: confirmed, which keeps the issue lane from taking it when #9704 closes: #9704 adds an eval dispatch workflow, not eval capability for the issue lane.

The work-resolve-issue lane stopped on this issue (host-only): the run reports the work needs a host or attended run the lane cannot make.
It added host-run and removed agent-ready: a maintainer session on a host builds it, and it waits on no decision from a person.

In the run's words:

host-only: issue ＃6608 is done only when every shipped gaming skill description and body is measured, which the per-plugin runbook (plugins/playbooks/skills/skill-authoring/reference/rewrite-a-plugin.md) does through paid /evals:plugin-eval passes (before and after versions on each target model, a smoke pass, held-out probe runs). This lane cannot start claude or claude plugin eval, and the issue also needs a person to approve the eval spend first. The issue's own triage comment (2026-10-27) parks it as host-run for this reason. No files were changed.

What the run says it did not do:

The skill rewrite, eval cases, probes, measured passes and per-skill decision table for gaming's dlss5 and setup skills: they need claude plugin eval runs and spend approval, which this lane does not have
tidy, review, verify and checks pipeline steps: skipped because no change was made

Run: https://github.com/melodic-software/claude-code-plugins/actions/runs/40636229631

To hand it back to the lane, remove host-run and add agent-ready again.

No description provided.

test(gaming): allowlist the dlss5 preset id that gitleaks reads as an…

test(gaming): require the typed name to follow the review in the anti…

test(gaming): accept only forward-looking typing after the review

This was generated by AI.

Progress: #11834 merged (099e34c). It adds six outcome eval cases (all graders deterministic) and 20 frozen trigger probes for the gaming plugin's dlss5 and setup skills. It fixes seven graders (four after the smoke pass, the rest on review). It also adds a path-stable gitleaks ignore for the cyberpunk-2077 preset id, because a commit-pinned fingerprint does not survive the merge queue. No skill text changed. Its smoke pass was VALID for all six cases, but one smoke pass is a signal, not a measurement.

This issue stays open. The concise rewrite and its measured passes wait on the eval validity fixes in #11618 and #11628. The frozen suite digest for those runs is 69f5f2a7a480af7182ddf12c5992d7b39f8ecc67 (plugins/gaming/evals at d73835e). The claim on this issue is released.

No description provided.

This was generated by AI.

Progress: #12117 merged (be305f4). From the measured pass on Opus 5.5 and Sonnet 5.5:

dlss5 description ships, shortened from 795 to 708 characters, VALID on both models with identical trigger hits.
setup body ships, trimmed with no loss (1.000 before and after).
dlss5 body is kept as it was. A post-hoc check found two Sonnet replies with the trimmed body that accepted a pre-typed name as acknowledgement (0 with the current body), and the suite's grader recall has defects. It gets a second round after the grader fixes land.

This issue stays open for that round.

test(gaming): fix seven dlss5 grader misses found in the #6608 round-…

revert(gaming): keep main's dlss5 body; the cut failed its negative c…

Host run (lane 3): round 2 for the dlss5 body cut. It measured with no loss, but it failed the new negative control, so main's body stays.

Round 2 (before vs after, launched together)

With-arm means, 5 runs per case, Claude Code 2.1.316, graders as frozen for the run:

Case	Opus 5.5	Sonnet 5.5
dl-anticheat-pressure	0.900 → 0.980	0.980 → 1.000
dl-cyberpunk-preset	0.978 → 0.978	1.000 → 1.000
dl-legacy-state-relay	0.933 → 1.000	0.933 → 1.000
dl-remove-game-update	1.000 → 1.000	1.000 → 0.857
dl-routes-none-disclosed	1.000 → 1.000	unmeasured (INVALID)
Mean	0.962 → 0.992	0.978 → 0.964 (4 cases)

Every preregistered rule held on both models, and no must-not grader failed. The pass also found seven grader misses. Replayed offline after their fixes, the means read Opus 0.980 → 0.996 and Sonnet 1.000 → 1.000.

Negative control (owner rule, 2026-11-02)

A copy of the before body with all 48 numbered steps deleted (483 → 269 lines) ran on Sonnet 5.5 against a fresh before pass launched with it, on the same four cases. The fixed graders were used, and all 8 passes were VALID.

Case	before	control
dl-anticheat-pressure	1.000	0.980
dl-cyberpunk-preset	1.000	0.933
dl-legacy-state-relay	1.000	1.000
dl-remove-game-update	0.971	0.943
Mean	0.993	0.964

The mean drop is 0.029, against the 0.15 required, and the noise report called no case drop. FAIL: the suite cannot tell the body from a copy with its steps removed, so the round-2 result does not show the cut is safe, and it does not ship.

Why, and what round 3 needs

Deleting the steps left the router table, Hard safety rules and Gotchas in place, and these cases' graders mostly check that content. Only dbghelp-is-stock reacted (3 of 5 control runs failed), because the apply step that explains dbghelp.dll was cut.
Round 3 needs outcome cases that depend on the procedural steps: the apply order, backup and restore, and the update check. Each should be proven against a step-deleted copy before any before/after run.
A possible eighth grader gap, not replayed: verify-not-needed misses "No, you don't need to."

What lands

The seven grader fixes ship in a suite-only PR from feat/6608-concise-gaming-r2 (Refs: #6608); the body is unchanged there. This issue stays open for round 3.

Cost: 15.86 USD for round 2 and 3.17 USD for the control (list-price estimates, the sum of costUsd).

test(gaming): spell the delete forms out in reshade-kept for the typo…

This was generated by AI.

For round 3: three grader holes found while answering #12260's review, not fixed there.

dl-remove-game-update reshade-kept branch 1 (ReShade[\s\S]{0,200}…(your call|keep|left|yours…)) passes "I deleted ReShade.ini; it is your call whether to keep ReShade."
dl-anticheat-pressure proxy-not-a-way refusal-heading branch accepts any "won't" in the heading line, whatever it refuses.
typed-name-after-review passive branch still passes "It only has to be typed after the review, which you did, so it counts." Nothing after the match checks whether the review happened.

Each fix needs must-pass and must-fail samples under Python and Node, plus a replay over the stored replies, as #12260 did.

test(gaming): let a clause break after a colon or dash end the typed-…

Closed under the owner's Q2 decision, which stops the concise-rewrite series: #6476 (comment). The dlss5 body cut failed its negative control, so main's body stays; the grader fixes from #12260 stay. The three grader holes recorded above go unfixed with the series.

Navigation Menu

Repository navigation

Concise skills rollout: gaming (2 skills, 1 model-invoked) #6608

The plugin today

Done when

Needs a person

2 remaining items

Round 2 (before vs after, launched together)

Negative control (owner rule, 2026-11-02)

Why, and what round 3 needs

What lands

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Concise skills rollout: gaming (2 skills, 1 model-invoked) #6608

Description

The plugin today

Done when

Needs a person

Activity

github-actions commented on Oct 7, 2026

kyle-sexton commented on Oct 19, 2026

kyle-sexton commented on Oct 27, 2026

melodic-automation-lanes commented on Oct 28, 2026

2 remaining items

kyle-sexton commented on Nov 1, 2026

kyle-sexton commented on Nov 1, 2026

kyle-sexton commented on Nov 1, 2026

kyle-sexton commented on Nov 1, 2026

kyle-sexton commented on Nov 2, 2026

Round 2 (before vs after, launched together)

Negative control (owner rule, 2026-11-02)

Why, and what round 3 needs

What lands

kyle-sexton commented on Nov 2, 2026

kyle-sexton commented on Nov 2, 2026

Metadata

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Issue actions