* chore(release): pre-release P0/P1 cleanup
P0 fixes:
- delete src/analysis.ts (179 lines, 0 importers across src/clis/extension)
- remove dead OPENCLI_DIAGNOSTIC negative test assertion
- rename OPENCLI_BROWSER_TIMEOUT to OPENCLI_BROWSER_IDLE_TIMEOUT — the env
controls workspace lease idle release, not command runtime; old name was
misleading and undocumented (no fallback needed)
- add 'fill' to validate.ts KNOWN_STEP_NAMES so adapters using PR #1222's
fill pipeline step do not trip "unknown step name" warnings during validate
P1 fixes:
- BrowserConnect daemon-not-running hint: replace stale "make sure port is
available" with actionable "run opencli doctor / opencli daemon restart"
- TimeoutError hint: lead with --timeout flag, demote env var to secondary
* fix(validate): derive step allowlist from pipeline registry
@pr-monitor flagged the prior "add 'fill' to KNOWN_STEP_NAMES" fix as
treating only the symptom — two parallel hand-maintained lists will keep
drifting whenever a new pipeline step is registered.
Address the root cause: pipeline/registry.ts now exports
`getRegisteredStepNames()` and validate.ts builds KNOWN_STEP_NAMES from
that. Adding a step via `registerStep()` automatically allowlists it.
* test(validate): regression guard for pipeline step allowlist linkage
@pr-monitor follow-up: lock the validate ↔ pipeline registry linkage at
the test layer so future drift is caught immediately.
Changes:
- recompute KNOWN_STEP_NAMES per-call (was const at module load) so
steps registered after validate.ts import (plugins, dynamic registration)
are honoured
- add src/validate.test.ts with 3 cases:
1. every step name from getRegisteredStepNames() exists
2. an adapter using every currently registered step does not warn
3. a step registered at runtime is automatically allowlisted by
validate without any source change to validate.ts
* fix(capabilityRouting): add fill to BROWSER_ONLY_STEPS
Same double-list drift pattern as validate.ts KNOWN_STEP_NAMES (audit
follow-up flagged in this PR's evolution thread). The fill step was
registered in pipeline/registry.ts (PR #1222) but never added to the
browser-only allowlist in capabilityRouting.ts.
Concrete impact:
- shouldUseBrowserSession() didn't recognize a `[{ fill: ... }]` pipeline
as needing a browser, so PUBLIC adapters using fill could end up
without a page and crash inside stepFill at `page!.fillText(...)`
- pipeline/executor.ts's per-step retry policy (BROWSER_ONLY_STEPS gets
2 retries on transient errors, others get 0) skipped fill — losing
retry coverage on a DOM-touching step
Fix:
- add 'fill' to BROWSER_ONLY_STEPS
- add a documenting comment explaining BROWSER_ONLY_STEPS is the
browser-touching subset of registered steps (not the full set)
- export _validateBrowserOnlyStepsAgainstRegistry() so the test layer
catches the inverse drift (browser-only step that no longer exists)
- 3 new tests in capabilityRouting.test.ts:
* pipeline with fill routes to browser session
* BROWSER_ONLY_STEPS subset of registered step names
* fill is in both lists
This addresses @pr-monitor follow-up #3 (audit similar double-list
patterns) for the obvious in-scope candidate. Other candidates outside
this PR's scope: build-manifest serialization vs registry shape, error
code unions vs lint baselines.
* test(validate): use Strategy.PUBLIC enum instead of string cast in regression test
Self-review nit: `strategy: 'public' as never` worked but bypassed the
typed CliOptions union. Use `Strategy.PUBLIC` so the test exercises the
real public API.
jakevin