solve
Single entry point for any coding task (change, feature, refactor, bug). Triages the task through the jev-api tools and routes it to the right track, loading the phase references it needs. Flags --headless (the decider agent answers every question) and --size=<S|M|L|XL>. Use when the user hands over a task to implement or fix.
- License 许可
- MIT
- 最近更新
- 2026-11-15
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- solve
- description
- Single entry point for any coding task (change, feature, refactor, bug). Triages the task through the jev-api tools and routes it to the right track, loading the phase references it needs. Flags --headless (the decider agent answers every question) and --size=<S|M|L|XL>. Use when the user hands over a task to implement or fix.
02 资源文件 13
- 参考文档debug.reference.md1.4 KB
- 参考文档decide.reference.md5.7 KB
- 参考文档design.reference.md1.6 KB
- 参考文档execute.reference.md1.2 KB
- 参考文档finish.reference.md1.2 KB
- 参考文档plan.reference.md1.7 KB
- 参考文档pr.reference.md1.9 KB
- 参考文档receive-review.reference.md1.4 KB
- 参考文档request-review.reference.md1.3 KB
- 参考文档spec.reference.md942 B
- 参考文档tdd.reference.md1.0 KB
- 参考文档verify.reference.md889 B
- 参考文档worktree.reference.md1.5 KB
03 SKILL.md 正文
solve
Entry point and the only skill of this plugin. The phases are reference files next to this one; open one (Read) only when its row in the decision matrix applies. Do not start work before the triage below.
jev-api first. In this skill and all its references, a decision that fits options, a rating or yes/no is put to the jev-api tools before own reasoning; own reasoning only on low confidence or a failed call. The catalogue is at the end of decide.
0. Arguments
Strip these flags from the task text before anything else; the rest is the task.
--headless: no question goes to the user. See Questions and --headless.--size=<value>:S,M,LorXL(case-insensitive). That value is the size; the size question of the triage is skipped. Any other value counts as no--size: the triage rates the size.--type=<value>:tiny,featureorbugfix(case-insensitive). That value is the track; the track question of the triage is skipped and its confidence rule does not apply. Any other value counts as no--type: the triage picks the track. The track-specific rules below (worktree prefix, flow, size flow) apply as for a triaged track. A--type=bugfixthat the root cause shows to be a feature still switches to the feature track, as in the bugfix track.
Questions and --headless
Every question to the user goes through AskUserQuestion, never as plain text. This holds for this file and for every reference.
Before an AskUserQuestion that concerns a file (design, spec, plan, review notes), first print the file's content to the user as a summary in bullet points, with the file's full absolute path. Then ask for approval (approve, change, abort); on change, fix the file and ask again until approved. The summary is skipped with --headless: the decider reads the file itself. Each reference only says what its summary covers.
With --headless, AskUserQuestion is never called. Instead, for each question spawn the decider agent (subagent_type: jev-solver:decider, the plugin-namespaced name; model and effort by size: S, M and L run with the agent's own opus/high, XL with model: fable, effort: high, passed in the Agent call; continue the same agent with SendMessage for later questions) and brief it with:
- the question and the options that
AskUserQuestionwould have shown (label and description, the recommended one first), - the task text, track, size, branch and worktree path,
- the absolute paths of
design.md,spec.md,plan.mdandreview.mdas far as they exist, and what the answer is used for.
It replies answer: (an option label, or free text if none fits) and reason:. Apply the answer as if the user had chosen it and note the decision in one line in the spec or plan. Approval gates (spec, discard) are answered by it too; discarding a branch, merging and other destructive or outward-facing steps beyond the normal flow are not taken headless: keep the branch.
1. Triage
Ask jev-api once, with the user's task text as state:
mcp__jev-api__batch
questions:
- type: choice
instructions: Which kind of task is this?
criteria:
tiny: only a single-file or configuration tweak (rename, typo, one-line edit) without significant influence on functionality or tests; any change to behaviour is feature
feature: new behaviour, a feature, a refactor or anything needing design decisions
bugfix: something is broken, a failing test, an error, unexpected behaviour
- type: score
instructions: How large is the work?
criteria: [S: a few lines in one place, M: one component, L: several functionalities, modules or subprojects, XL: cross-cutting or architectural]
Put the task-ambiguity question in the same call (noul: is the task text too ambiguous to start?; it counts only for the feature and bugfix tracks, and only a confident yes goes to the user via AskUserQuestion, headless: decider). With --size, leave out the score question; with --type, the track question; with --type=tiny, all questions (tiny needs neither size nor ambiguity check, so no call is made). With --type=feature or bugfix and a valid --size, only the ambiguity question remains. The score is fractional: round to the nearest level. Apply the confidence rule of decide to every answer; if it is low or the call fails (any error, no configuration), triage with your own reasoning using the same criteria and say so in one line.
Size floor
The triage under-rates tasks, so after rounding (and after the fallback to own reasoning) raise the size on the feature track:
- A functional change (behaviour of code, skills, hooks or tests; not documentation-only) is at least
M. - A change touching several functionalities, modules or subprojects (several plugins, packages or components) is at least
L. tinyonly for a single file or configuration edit without significant influence on functionality or tests; a tiny pick that changes behaviour or needs test changes is afeatureand takes sizeM(Lper the rule above) without a score.
The floor applies to a triaged size only: an explicit --size is never raised. The bugfix track keeps its triaged size.
2. Worktree and work branch
Always, for every track, right after triage and before any spec, plan, debugging or edit:
Follow worktree (branch name fix/<slug> for the bugfix track, feat/<slug> for tiny and feature; never nest; baseline), then report branch and worktree path in one line and continue. The same slug names the scratchpad directory below.
3. Decision matrix
| Situation | Reference |
|---|---|
| Choosing between defined options, rating risk or size | Decide |
| Work starts, or starts outside a worktree | Worktree |
| Feature, size L or XL, no approved design yet | Design |
| Feature, size XL, approved design, no spec yet | Spec |
| Feature, size M or larger, no approved plan yet | Plan |
| A plan (or a tiny change) is to be carried out | Execute |
| Writing any production code | Test first |
| Bug, failing test, error, unexpected behaviour | Debug |
| Before any claim of done, fixed or passing | Verify |
| A change, phase or finished task needs review | Request review |
| Review feedback arrived (reviewer, user, PR) | Receive review |
| First commit, every push, merge-ready at the end | Pull request |
| Implementation and review done | Finish |
4. Tracks
Below, a name in brackets is the matrix row to follow.
tiny
Implement directly. Then Verify and a light review: read your own diff once (git diff), nothing more. Then Finish. No spec, no plan.
feature
Flow by size, each "confirm" is an AskUserQuestion after a bullet-point summary of the file (headless: decider):
| Size | Flow |
|---|---|
| S | implement directly |
| M | Plan, confirm, implement |
| L | Design, confirm, Plan, confirm, implement |
| XL | Design, confirm, Spec, confirm, Plan, confirm, implement |
Then:
- Implement with Execute (with Test first for code); on S there is no plan, so execute works from the task text.
- Verify.
/simplifyonce on the finished work (any size), then Verify again.- Request review, depth by size (L and XL: the per-phase reviews and, on XL, the design and spec reviews; the final diff is covered by the
/code-reviewof the section below):
| Size | Review |
|---|---|
| S | read your own diff |
| M | one review of the final diff |
| L | a review after each plan phase, plus one of the final diff |
| XL | additionally a review of the design and of the spec before coding |
Findings are handled with Receive review. Then the final /code-review and Finish.
bugfix
On its fix/ work branch (step 2): Debug (root cause, failing test, one fix), Verify, Request review (S, M: read your own diff; L and above: one review of the final diff; findings via Receive review), the final /code-review, Finish. No spec or plan, unless the root cause shows the fix is really a feature (noul via decide): then switch to the feature track and say so.
Final code review (feature and bugfix, never tiny)
After the track's own reviews and before Finish, run /code-review <level> --fix on the branch, the level taken from the size: S low, M medium, L high, XL xhigh. Never max unless the user asks. Re-run Verify after it applied fixes.
Pull request
After the first commit on every track: push and open a draft PR; after every push: update the PR description (Pull request; no question about it). Commit and push per step or phase, not only at the end.
Done means merge-ready
The skill ends only when every change is committed and pushed on the work branch, the PR description is current, all merge conflicts with the base are resolved, the checks are green and the PR is merge-ready (Finish with Pull request does this). Report the PR URL.
Documents
Design, spec, plan and review notes live in the session scratchpad directory (named in the system prompt) under jev-solver/<task-slug>/ (design.md, spec.md, plan.md, review.md), never in the repository. Name them with their full absolute path when you mention them to the user (and to the decider).
Code exploration
In a graphmind project (graphmind which exits 0) start every exploration with /gm (/gm-fn for reading a symbol, /gm-impact for blast radius); grep only for string literals and config. Otherwise use the normal tools. Save decisions and conventions found along the way with graphmind memory add (project scope).
04 同仓库更多 Skills
universal-lint
User-invoked. Takes a free-text description of a set of files ("all files in src/", "files changed in PR 123", or an explicit list) and lints each of them exactly as this plugin's Write/Edit hook would — read-only, no autofix — printing one aggregated findings report. Self-contained: depends on no other plugin.
dprint
Runs dprint for the project. The first argument selects the action - init checks for the plugin's minimum dprint version, installs or updates dprint at user level after confirmation via AskUserQuestion, offers to create a dprint config when none exists and to add dprint plugins for file types the config does not cover; check [scope] runs dprint check; fmt [scope] runs dprint fmt and rewrites files; without an argument it runs fmt on the whole current directory. Use init when the user wants dprint set up, installed, updated, checked or its plugins completed; check when they want formatting verified with dprint; fmt only when they ask to format files with dprint. When invoking this skill yourself, always pass init, check or fmt as the first argument.
split-hooks-lib
Split the hooks module of a function-hooks plugin of this repository into a readable entry register.mts plus helper modules that the build bundles into one compressed plugin/hooks/lib.mjs, to keep the shipped size small. Use when the user asks to split a plugin's register.mts, to make a plugin smaller, to move a plugin to register.mjs plus lib.mjs, or when a hooks module has grown too large for one file.
__SKILL_NAME__
__SKILL_DESCRIPTION__ Use when the user asks to …
scaffold-plugin
Scaffold a new Claude Code plugin in this repository with the repo layout (<name>/package.json, plugin/, src/, tests/). Use when the user asks to create, add, scaffold or start a new plugin here.
refactor-plugin
Refactor one plugin of this repository end to end in a fresh worktree on its own branch cut from a freshly fetched origin/main - perf-worklist, then bughunt, then a merge-ready PR. Use when the user asks to refactor, overhaul or run the full performance and bug hunt pass on a plugin.