feat(REL-12752): adding TTY detection for auto-JSON output#660
feat(REL-12752): adding TTY detection for auto-JSON output#660
Conversation
JackDanger
left a comment
There was a problem hiding this comment.
Love that this follows the gh / glab convention — great precedent. This is the PR I spent the most time thinking about because it's the one most likely to bite someone.
1. This is a breaking change — should it bump the major version?
ldcli uses semver via release-please (v1→v2 was an OpenAPI spec change). Today, if a user has a script like:
ldcli flags list --project my-proj | grep "my-flag"…running in CI (no TTY), it gets plaintext and the grep works. After this PR, it silently gets JSON and the grep matches nothing. The script doesn't fail — it just returns no results. That's the worst kind of breakage: silent.
Research shows that existing LD consumers almost universally pass -o json explicitly (sandbox-data-generator, gonfalon dogfood-flags.sh, agent-prompts guide). So the blast radius is likely small. But "likely small" and "zero" are different things, and this is a public CLI distributed via Homebrew and npm. Worth discussing whether this warrants a v3 bump, or at minimum a deprecation warning period where non-TTY plaintext emits a stderr warning before flipping the default.
2. FORCE_TTY is checked via raw os.Getenv while isTerminal is injected
You went to the effort of making isTerminal a parameter for testability — nice. But FORCE_TTY is read directly from os.Getenv inside NewRootCommand, which means tests use os.Setenv/os.Unsetenv. These are process-global mutations that aren't safe with t.Parallel(). Consider adding FORCE_TTY to the injected function or to the envChecker interface from #659.
3. The LD_OUTPUT env var test implies a contract
The test LD_OUTPUT=plaintext overrides non-TTY default documents that LD_OUTPUT controls the output format. Is this intentional new behavior or pre-existing? If new, it should be in the README/docs. If pre-existing, good to have the test — but make sure this doesn't interact unexpectedly with --output flag and FORCE_TTY (three-way precedence: flag > env > TTY detection).
4. The isTerminal == nil → JSON default is a footgun
If any future caller of NewRootCommand forgets to pass isTerminal, they silently get JSON-only output. Consider making this required (no nil check) or at least panicking on nil in debug builds.
5. Consider a stderr notice during the transition
gh shipped this same change and it was smooth because people expect it. But gh had it from early on. ldcli has been plaintext-default for its entire life. A transitional approach: when non-TTY triggers auto-JSON, emit a one-line stderr notice like note: non-interactive session detected, defaulting to JSON output. Use --output plaintext to override. You can remove it after a release cycle.
When stdout is not a terminal, the default output format switches from plaintext to json. Explicit --output or --json flags still override. Follows the pattern used by GitHub CLI (gh) and GitLab CLI (glab). Adds a FORCE_TTY environment variable escape hatch.
Requirements
Note
Medium Risk
Changes the CLI’s default output format based on TTY detection, which can affect existing automation and downstream parsing when output is piped/CI. Behavior is gated by explicit flags/env vars and covered by new tests, reducing but not eliminating compatibility risk.
Overview
Changes default output behavior so when stdout is not a TTY the CLI defaults
--outputtojson(instead ofplaintext), with an opt-out viaFORCE_TTYand existing precedence for explicit--output,--json, andLD_OUTPUT.Updates help/docs text for
--outputto describe the new auto-default, wires TTY detection intoNewRootCommand, and expands tests to cover default selection, overrides, and improved error formatting/validation ininternal/output.Written by Cursor Bugbot for commit b07eac9. This will update automatically on new commits. Configure here.
Related Jira issue: REL-12752: TTY detection for auto-JSON output