[pradyb]

What / why

workflow approve --watch (#26) streams its events as NDJSON under --json. workflow view --watch had no equivalent — --watch and --json were mutually exclusive (--watch is an
interactive bubbletea view, incompatible with machine-readable output). Raised as #64 while
resolving #26.

With --json, --watch now streams each job/step status transition as compact single-line JSON
(NDJSON) until the run completes, instead of erroring:

{"kind":"job_started","job":"build"}
{"kind":"step_started","job":"build","step":"checkout"}
{"kind":"step_completed","job":"build","step":"checkout","conclusion":"success"}
{"kind":"job_completed","job":"build","conclusion":"success"}
{"kind":"run_done","conclusion":"success"}

Behaviour notes

• pkg/workflow/watch_view.go (WatchViewRun) polls GetWorkflowRunDetail and diffs each job's
  and step's status against what was already reported (viewWatchState/viewWatchProgress), so
  a poll that finds nothing new produces no events — a job/step is reported "started" the first
  time it's seen in_progress and "completed" the first time it's seen completed. A job/step
  that jumps straight from unseen to completed (skipping in_progress between polls) reports
  only job_completed/step_completed, no synthetic "started" event.
• Reuses the existing --interval flag (no new flags) and the same Sleep-injection pattern
  approve --watch (workflow approve: add --watch to decide each approval gate as it appears #26) established, so tests don't really wait between polls.
• Stops cleanly on cmd.Context() cancellation (Ctrl-C), reporting nothing further — same
  semantics as approve --watch.
• The plain interactive --watch (no --json) is unchanged.
• The now-obsolete TestViewCommand_WatchAndJSONRejected test was replaced with two end-to-end
  tests asserting the NDJSON shape and event ordering, plus a run-lookup-error case.

Testing

• pkg/workflow/watch_view_test.go: unit tests for the diff logic (started-then-completed,
  completed-without-observed-start, independently-keyed steps within a job) plus WatchViewRun
  end-to-end tests (polls until completion, already-completed run, context cancellation).
• cmd/workflow/workflow_test.go: TestViewCommand_WatchJSON_StreamsNDJSONUntilRunCompletes
  (asserts exact event kinds/order) and TestViewCommand_WatchJSON_RunLookupErrorSetsHasError.
• Mutation-tested the diff logic (removed the !p.completed guard) and confirmed the test suite
  catches the regression (a re-poll of the same completed job re-reports job_completed).
• gofmt -l ., go vet ./..., go build ./..., go test -race ./... all clean;
  ./scripts/check-coverage.sh reports 93.7% (floor 85%).
• Not verified live against a real GitHub Actions run in this PR, for the same token-permission
  reason noted on feat(workflow): add --watch to workflow approve for multi-gate runs (#26) #63/workflow approve: add --watch to decide each approval gate as it appears #26 — this reuses the already-live-tested GetWorkflowRunDetail path
  unchanged, just adding a new consumer of it.

Note on base branch

Stacked on feat/26-approve-watch (itself stacked on feat/1-keyring-token-storage, #62), per
the same "implement it stacked on the last one" request. Will need rebasing/retargeting once the
branches below it merge to develop.

Closes #64

https://claude.ai/code/session_01Wy6Fqe5usEojMuNcAjKvRD