スキルを探す

Run the improvement loop, led by the solve-wall share and the follow-up of each earned fail, with signal a rare confirming event: find the one link holding the climb, choose one simple domain-agnostic change against it, measure it as a matched arm, freeze a prediction, launch, read recorded bytes, decide the next move. Owns experiment selection and evidence reading; the product controller owns build, measure, climb, rebuild, claim and promotion. Also the Super Loop's Meta Agent: a second session that audits and steers the session driving the loop so it keeps moving (references/meta-agent.md).

日本語の概要は準備中です。原文の説明を表示しています。

Set up, reshape, title, publish, version or land an Anabasis pull request the way the operator wants it: where it goes on the stack, when to regroup a back-and-forth history into a few finished commits, how the body reads, who pushes, and when a version tag is attached. Use when asked to set up or append to a PR, regroup or squash its commits, rewrite its title or body, push it, cut a release for it, or merge a stack.

日本語の概要は準備中です。原文の説明を表示しています。

Measure and raise the precision of `bun run simplify`, the simplify census, shape by shape. Harvests every recorded census reading and not-slop answer, replays the current scans over history, has blind judges label a seeded sample, scores each shape with a Wilson interval, and tunes a scan's source until the sites it prints are ones a reader would act on (target 0.8). Use when asked how good the census is, why it flags the wrong things, or to tune a tree scan or a staged catcher.

日本語の概要は準備中です。原文の説明を表示しています。

Review the changed code for reuse, simplification, efficiency and altitude cleanups in this repository, then apply the fixes. Quality only — it does not hunt for bugs; use /code-review for that. The method, its scripts and its cases live in this directory, so the skill works without any user-level skill pack.

日本語の概要は準備中です。原文の説明を表示しています。

Move safely to the current stack head, append a stacked PR, or repair an authorised PR chain. Preserve local work, replay each child onto its actual parent as a linear range, publish batches through one push that gates every commit, and isolate a failed checkpoint. A context hop alone does not authorise remote stack repair.

日本語の概要は準備中です。原文の説明を表示しています。

Compare recurring implementation patterns, choose the best existing approach, teach it in the owning skill, and enforce reliable distinctions through existing lint tooling. Use when asked to generalise inconsistent code or turn a repeated anti-slop finding into a convention and fix its remaining occurrences.

日本語の概要は準備中です。原文の説明を表示しています。

Measure which test files carry the least evidence and decide which may be merged or removed. Use when the suite has grown, when a run is slow, when asked which tests matter, or before adding a test that may already exist. Ranks by marginal coverage, then adjudicates each shortlisted file by planting faults.

日本語の概要は準備中です。原文の説明を表示しています。

Compare one or several recorded source states to see whether changes improved Anabasis's climb. Build a state-by-condition census first; compare adjacent states only when prompt, model, command, project lineage, budget and runtime identities support it; read equal windows, earned failures and the mechanisms that fired. Use for 'did the fixes work', 'is this wave better', or before relaunching on a new source.

日本語の概要は準備中です。原文の説明を表示しています。

Plan a huge chunk of work (more than one agent session can hold) as a shared map of decision tickets on your issue tracker, and resolve them one at a time until the way to the destination is clear.

日本語の概要は準備中です。原文の説明を表示しています。

Investigate a live, stalled or completed Anabasis run and turn findings into an evidence-bound fix proposal. Also answers whether a campaign is climbing: how the climb is going, whether the batteries are getting harder, why a difficulty decision keeps repeating, how to climb faster, why the controller chose its action. Also the narrow read: reviewing one campaign run or recorded case through its outcomes, whether a live run is still producing useful evidence. Reads evidence; does not launch the next experiment or stop a run.

日本語の概要は準備中です。原文の説明を表示しています。

Invoke an already-built Built Harness on one or more queries, without building a new one. Use when the operator points at a domain bundle and asks to run a task, a family, or a query they wrote, through it.

日本語の概要は準備中です。原文の説明を表示しています。

Design, add, read or retire an Anabasis runtime safeguard: a bounded log-only sensor around a measured brittle decision. Use it for a post-run finding, not as a substitute for the owning product fix, verifier, score or launch gate.

日本語の概要は準備中です。原文の説明を表示しています。

Zip one recorded Harness Builder run for handover at a chosen depth (--light, --medium, --verbose). Use when the operator asks to zip, bundle, archive or send a run, its traces, its logs or "everything the model saw", or names a size cap. One script, three cumulative levels, a README and MANIFEST inside every zip.

日本語の概要は準備中です。原文の説明を表示しています。