Files
skillshare/wiki/cli-development.md
T
Willie 828ba93f27 fix(git): keep a caller's GIT_DIR out of skillshare's git commands (#854)
* fix(git): keep a caller's GIT_DIR out of skillshare's git commands

Git exports GIT_DIR, GIT_WORK_TREE, GIT_INDEX_FILE and related variables
to hooks that run in a linked worktree. Every git process skillshare
started inherited them, and those variables override the directory the
command runs in. Run from a post-checkout or post-merge hook, skillshare
therefore acted on the hook's repository instead of its own: install's
sparse clone turned on sparse checkout there (a later merge removed every
file outside the skill's path), the remote tree check's `git init --bare`
made it a bare repository, install's reconcile recorded its origin and
branch, and push staged and committed into it.

internal/gitenv builds git commands without the variables that
`git rev-parse --local-env-vars` lists for choosing a repository, work
tree, index, object store or config file. GIT_CONFIG_COUNT and
GIT_CONFIG_PARAMETERS stay, because users pass credentials through them
and install's token auth appends to the existing count. Every git call in
cmd/skillshare, internal/install, internal/git, internal/check and
internal/config now goes through it, and plugin runCommand starts native
CLIs from the same environment because they run git as well. The hooks
service keeps its own stricter filter.

Closes #852

* test(gitenv): fail on new git commands that bypass gitenv

Each git call site converted for #852 is safe only because it goes
through gitenv.Command, and about half have no test of their own. A
future exec.Command("git", ...) would silently bring the bug back, so a
syntactic guard now rejects one outside gitenv, hooks and testutil,
mirroring the sourcefs and sourcewalk ratchets.

The Environ test also checks a lowercase name, since Windows environment
names are case-insensitive, and compares against literal keys so that
removing one from repoVars fails it. The wiki records the rule.
2026-11-11 21:08:47 +08:00

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 on huh with 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. esc clears the filter, then goes back, then quits. Action keys are lowercase with one meaning per letter: d remove, u update, s sync, e edit settings, n new, i import, c collect, r restore, t enable/disable, m manual only, o sort, ! audit. Uppercase is only for whole-set actions such as D empty trash, always with a confirmation. Screenshot every changed TUI with scripts/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 prints ui.Cancelled("removed"). Without a terminal ui.Confirm reads one line from stdin (y/yes, n/no, anything else is the default), so echo y | skillshare … keeps working.
  • Print results with internal/ui/rows.go: ui.Row(mark, label, value, ui.RowWidth(labels...)) per item (✓/!/✗, or ui.MarkNone for plain information), ui.Section for a bold block name, ui.Done for the closing line with its duration, ui.Note for dim detail and ui.Next for up to three follow-up commands. Leave zero counts out. Integration tests match a row with AssertRowContains(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 in tests/integration/completion_test.go should name it. Installed completions are a copy, so users only see changes after rerunning skillshare completion <shell> --install.
  • Noninteractive automation should use explicit flags. Never treat --force as 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.log with 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, .gitignore cleanup and the metadata prune (MetadataStore.RemoveByNames), keeping a copy of the pruned metadata beside the trash entry. Restore goes through uninstall.Restore / RestoreAgent, which put that copy back, and permanent deletion through trash.Delete; never call trash.Restore or os.RemoveAll on 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.Follow policy (built by config.SkillsWalk when follow_source_links is on) resolves a first-level link; such paths are rebased onto the link target with the same escape checks. The ratchet in internal/sourcefs/ratchet_test.go fails on any new raw os write call until it gets a truthful reason in testdata/raw_writes.tsv;
  • create and remove first-level source links through internal/sourcelink (Create prechecks with sourcewalk.Follow.Allow; Remove is the unlink flow and moves only the link to trash; uninstall <link> is refused when the linked folder itself is a skill, and Discard rolls back a link whose follow-up step failed);
  • read the skills source through internal/sourcewalk and pass the same sourcewalk.Options the operation's discovery used, so CLI, dashboard, and reconcile agree on which first-level links are followed. The ratchet in internal/sourcewalk/guard_test.go fails on any new raw walk until it is listed in allowlist.json;
  • run git through internal/gitenv (Command, CommandContext, Environ), which drops GIT_DIR and the other repository variables a git hook exports, so skillshare started from a hook never works on the hook's repository (#852). TestNoRawGitCommands in internal/gitenv fails on any new exec.Command("git", ...) outside internal/gitenv, internal/hooks (its own sanitized environment) and internal/testutil.

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.

The Nix flake also runs the unit tests, inside the Nix sandbox, which has /bin/sh but no /bin/bash or /usr/bin/env. A change to go.sum also needs a new vendorHash; see release.

Web API

When a dashboard endpoint is needed:

  1. Use existing writeJSON and writeError helpers in internal/server/handler_<name>.go. Decode request bodies with decodeJSON or decodeJSONWith (internal/server/json_body.go), which cap the body size and return 413 when it is too large; do not call json.NewDecoder(r.Body) directly.
  2. Register the method and route in internal/server/server.go.
  3. Handle scope explicitly through s.IsProjectMode() or the existing guards.
  4. Add handler tests and verify UI client types and query invalidation.
  5. 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 testing guidance.