* fix(trash): keep install metadata when restoring from the trash Uninstall pruned a skill's, tracked repo's or agent's entry from the central .metadata.json, and only the files went to the trash, so a restore brought back a remote install as a local one: update and check no longer knew its source (#679). Uninstall now saves the entries it prunes, with their audit acceptances and target overrides, beside the trash entry, re-keyed to the trash name so they match where restore puts the item. uninstall.Restore and RestoreAgent merge them back after the files return, and every restore path (CLI, trash TUI, dashboard) goes through them; the dashboard passes its cached store so it does not go stale. The saved copy is taken with the same rule RemoveByNames uses and copies each entry, so legacy repo-group entries are kept and re-keying never touches the live store. In project mode uninstall also drops the skill's .gitignore rule and lock pin, so a skill restore reconciles the project skills afterwards. trash.Delete removes an entry with its saved metadata, and replaces the direct RemoveAll calls in delete, empty and the 7-day cleanup. Items trashed before this change have no saved metadata and still restore as local. * fix(trash): keep live records of another path when restoring metadata A legacy store keys a grouped skill by its basename, so the saved copy of frontend/foo was keyed "foo". If a different skill was installed as root-level foo while frontend/foo sat in the trash, restoring it replaced the root skill's record and lost its source. Extract now writes full source-relative keys (legacy repo-group entries keep theirs, since their Group is not their path), and Merge keeps a live entry whose key names a different path instead of overwriting it. Merge is shared by skill and agent restore, so both are covered.
11 KiB
CLI Development
Use when adding or changing Go CLI commands, flags, handlers, interactive TUIs, domain packages, or any mutating behavior.
Before Starting
- Search for the closest existing command pattern instead of guessing APIs from memory or documentation.
- List affected command, domain, test, schema, documentation, and Web API files.
- Define acceptance criteria before changing public behavior. Ask only when ambiguity would materially change scope or interfaces.
- Use TDD: create a reproducible failing test, then write the smallest implementation that passes it.
Command Layering
cmd/skillshare/<command>.go should contain only flag parsing, validation, and mode dispatch. When a handler approaches roughly 300 lines or mixes concerns, split it using existing suffixes:
| Suffix | Responsibility |
|---|---|
_handlers.go |
Core orchestration |
_project.go |
Project-mode behavior |
_render.go / _format.go |
Output rendering and formatting |
_prompt.go / _prompt_tui.go |
Decisions and prompts |
_tui.go |
Full-screen Bubble Tea UI |
_batch.go |
Batch orchestration |
_resolve.go |
Target/resource resolution |
_context.go |
Mode-specific context |
Put testable core logic in internal/<domain>/. Do not let the CLI, Web API, and UI implement separate versions of the same business logic.
Global and Project Mode
Most commands route through parseModeArgs() for global (-g) or project (-p) mode. Pass the command's value-taking option names and aliases as additional arguments (for example, parseModeArgs(args, "--name")) so their values cannot become mode flags. Do not include optional-value flags that reject a following flag, such as search --hub. The parser preserves -- and everything after it for the command parser. Before changing behavior, verify:
- whether the modes use different configuration, registry, source, or target paths;
- whether project mode needs a separate handler;
- whether their flag sets are actually the same;
- whether tests cover mode precedence, working directory, and missing configuration.
Output and Interaction
- Ask inline questions with
internal/ui/prompt.go(Select,MultiSelect,Confirm,Input, built onhuhwith theme colors); keep full-screen Bubble Tea components for browsing lists. Do not add another prompt framework. - Full-screen TUIs share the frame in
cmd/skillshare/tui_frame.go: a title line (skillshare <cmd> · scope · counts, tabs on the right), the list and its details side by side without divider lines, and one key line with the position on the right. Show only the common keys there and the rest under?; ask confirmations and take filter input on the key line, not on a separate screen.escclears the filter, then goes back, then quits. Action keys are lowercase with one meaning per letter:dremove,uupdate,ssync,eedit settings,nnew,iimport,ccollect,rrestore,tenable/disable,mmanual only,osort,!audit. Uppercase is only for whole-set actions such asDempty trash, always with a confirmation. Screenshot every changed TUI withscripts/screenshots/record-tui.sh. - A question asked without a terminal takes its default. Print each decision with the flag that changes it instead of failing or silently doing less.
- Confirm a single action (remove, delete, apply anyway) with
ui.ConfirmAction: esc answers no, and a decline printsui.Cancelled("removed"). Without a terminalui.Confirmreads one line from stdin (y/yes,n/no, anything else is the default), soecho y | skillshare …keeps working. - Print results with
internal/ui/rows.go:ui.Row(mark, label, value, ui.RowWidth(labels...))per item (✓/!/✗, orui.MarkNonefor plain information),ui.Sectionfor a bold block name,ui.Donefor the closing line with its duration,ui.Notefor dim detail andui.Nextfor up to three follow-up commands. Leave zero counts out. Integration tests match a row withAssertRowContains(t, label, value). - Preserve dispatch order: structured JSON → TUI when interactive and allowed → empty state → plain text.
- Structured-output stdout must remain machine-readable; progress, spinners, and diagnostics must not contaminate JSON.
- When adding or changing a flag, inspect
--help, completions, website command documentation, and tests. - Completions are five hand-written scripts (
cmd/skillshare/completion_{bash,zsh,fish,powershell,nushell}.go) that do not read the CLI's dispatch, so they drift silently. Any new or renamed command, subcommand, or flag must be added to all five in the same change; the literal-content check intests/integration/completion_test.goshould name it. Installed completions are a copy, so users only see changes after rerunningskillshare completion <shell> --install. - Noninteractive automation should use explicit flags. Never treat
--forceas a universal prompt bypass.
Mutating Behavior
Operations that change configuration, sources, targets, or managed files must:
- follow existing dry-run, backup, rollback, and conflict-handling patterns;
- write an oplog entry to
operations.logwith the operation, status, duration, and necessary arguments; - send security scan events to the audit log rather than the regular operation log;
- preserve path validation, scope checks, and ownership checks;
- use domain uninstall/remove flows for managed state rather than replacing them with filesystem deletion. Skills, tracked repos and agents go through
internal/uninstall(Preflight,Run,Agents): it owns the move-out check, the dirty check for tracked repos, the move to trash,.gitignorecleanup and the metadata prune (MetadataStore.RemoveByNames), keeping a copy of the pruned metadata beside the trash entry. Restore goes throughuninstall.Restore/RestoreAgent, which put that copy back, and permanent deletion throughtrash.Delete; never calltrash.Restoreoros.RemoveAllon a trash entry directly. The CLI and the dashboard handlers only resolve names and report its per-item results; - write the skills source through
internal/sourcefs, which refuses paths with a link component unless the operation's*sourcewalk.Followpolicy (built byconfig.SkillsWalkwhenfollow_source_linksis on) resolves a first-level link; such paths are rebased onto the link target with the same escape checks. The ratchet ininternal/sourcefs/ratchet_test.gofails on any new rawoswrite call until it gets a truthful reason intestdata/raw_writes.tsv; - create and remove first-level source links through
internal/sourcelink(Createprechecks withsourcewalk.Follow.Allow;Removeis theunlinkflow and moves only the link to trash;uninstall <link>is refused when the linked folder itself is a skill, andDiscardrolls back a link whose follow-up step failed); - read the skills source through
internal/sourcewalkand pass the samesourcewalk.Optionsthe operation's discovery used, so CLI, dashboard, and reconcile agree on which first-level links are followed. The ratchet ininternal/sourcewalk/guard_test.gofails on any new raw walk until it is listed inallowlist.json.
Tests
Integration tests use internal/testutil.NewSandbox(t) and RunCLI or RunCLIInDir:
func TestFeature_BasicCase(t *testing.T) {
sb := testutil.NewSandbox(t)
defer sb.Cleanup()
sb.CreateSkill("test-skill", map[string]string{
"SKILL.md": "---\nname: test-skill\n---\n# Content",
})
result := sb.RunCLI("command", "args...")
result.AssertSuccess(t)
result.AssertOutputContains(t, "expected output")
}
For a bug fix, prove that the test fails before the fix. Add an E2E runbook in ai_docs/tests/<slug>_runbook.md for new commands, install/uninstall/sync flows, security behavior, multi-step workflows, or OS/network/permission cases that integration tests cannot cover.
All command execution and isolation rules live in testing. Never run the CLI or tests on the host.
Web API
When a dashboard endpoint is needed:
- Use existing
writeJSONandwriteErrorhelpers ininternal/server/handler_<name>.go. Decode request bodies withdecodeJSONordecodeJSONWith(internal/server/json_body.go), which cap the body size and return 413 when it is too large; do not calljson.NewDecoder(r.Body)directly. - Register the method and route in
internal/server/server.go. - Handle scope explicitly through
s.IsProjectMode()or the existing guards. - Add handler tests and verify UI client types and query invalidation.
- Share an
internal/package when the CLI and API expose the same operation; do not shell out between them.
Sync is the main example: the CLI and the server both run per-target sync through internal/sync (SyncSkillTarget, RunAgentSync, RunExtraTargets). Both follow CLI semantics: every target runs, and a failed target is reported as a partial failure instead of stopping the loop. Change sync behavior in these runners, not in a command or handler loop.
Updating one tracked repo is another: both call update.TrackedRepo (internal/update), which checks for uncommitted changes, pulls into an isolated candidate (install.StageGitUpdate), audits it, publishes its exact approved commit, and refreshes metadata. Rejection never resets or changes the active checkout. It, install --track --update, and the in-place pull of a git-installed skill (install.handleUpdate) hold install.LockCheckout, so concurrent skillshare processes take turns on one checkout. A forced git-installed skill update fetches and then resets to upstream, so its local edits are gone once the fetch succeeds, even if the audit then rolls back. install --track --update shares this candidate mechanism. Candidates preserve sparse-checkout selection and audit metadata; unresolved or escaping symlinks fail closed. install.GitCheckoutRecovery shares partial-checkout cleanup between ordinary pulls and candidate publication, preserving pre-existing dirty paths. Candidate diffs are captured before publication so CLI confirmation does not depend on unpublished objects being available in the active checkout. The gate is fail-closed: a scan that cannot run is never overridden by force. Findings an earlier override accepted are always honored, and an override records the ones it lets through. Callers only choose inputs (Confirm) and word the result; change the sequence in the operation, not in a command or handler.
Completion Criteria
- The failing test now passes, along with proportionate neighboring tests.
- Handler split, dual-mode behavior, structured output, oplog, Web API, and all five shell completions were reviewed.
- Public behavior changes were synchronized after loading
documentation. - Commands were verified inside the devcontainer using
testingguidance.