Skip to content

fix(cli)!: remove browser fork, add adapter override to hosted mode - #361

Merged
ankitranjan7 merged 2 commits into
mainfrom
fix/remove-browser-fork
Aug 19, 2026
Merged

fix(cli)!: remove browser fork, add adapter override to hosted mode#361
ankitranjan7 merged 2 commits into
mainfrom
fix/remove-browser-fork

Conversation

@ankitranjan7

@ankitranjan7 ankitranjan7 commented Aug 19, 2026

Copy link
Copy Markdown
Contributor

What

webcmd browser fork was an undocumented alias bound to the same handler as webcmd adapter override, and it leaked the alias in its own error message:

$ webcmd browser fork hackernews
Error: Usage: webcmd adapter override <site>/<command> (got "hackernews")

The caller is told the usage of a command they did not run. browser fork appears nowhere in docs/, skills/, README.md, or CHANGELOG.md — it was only ever reachable by guessing. It also sat in the browser namespace behind a require-existing session policy for what is a local file copy.

This removes it from the local CLI and from the hosted browser catalog.

adapter override in hosted mode (second commit)

browser fork turned out to be hosted's only fork path. adapter override was never routed in hosted mode — src/hosted/runner.ts matched only source, path, and help — and cloud's forkSystemAdapterCommand had exactly one caller: the fork browser action. Removing browser fork alone would have taken hosted's fork capability away with no replacement.

So the second commit routes adapter override <site>/<command> in hosted mode to a new POST /v1/adapters/override. After both commits the command means the same thing in either mode: fork an installed upstream command into a private copy you can edit, refusing commands that are not installed upstream and commands already forked.

Storage differs, so the output does:

local hosted
writes ~/.webcmd/clis/<site>/<cmd>.js + .base copy tenant-private adapter package
prints yours: / base: paths package: id + source: file
edit with the path from adapter path adapter source get|put
not upstream "not provided by an installed plugin" HOSTED_ADAPTER_FORK_SOURCE_INVALID
already forked "an override already exists at …" HOSTED_ADAPTER_FORK_ALREADY_EXISTS

Behaviour after

$ webcmd browser fork hackernews
error: 'webcmd browser fork' was removed. Use: webcmd adapter override <site>/<command>
# exit 2

$ webcmd browser click x
error: unknown command 'click'
# exit 2  (was 1)

Unknown browser subcommands now exit 2 (USAGE_ERROR) instead of 1.

Why fork stays in BROWSER_SUBCOMMAND_NAMES

rejectPositionalBrowserSessionArgv reads any unreserved token after browser as the retired positional-session grammar. Dropping fork from that set would answer webcmd browser fork hackernews with:

error: Browser sessions are root selectors. Use: webcmd --session fork browser hackernews

— worse than what we started with. Retired names stay listed so commander can report them properly. The set was already a superset of the declared subcommands (~40 names, 7 real); the comment now says that is deliberate instead of asking the next person to "keep in sync".

Also

Corrects the usage string in the adapter-shadow hint, which advertised adapter override <site> <command> (two positionals) for a command that takes one <site>/<command>.

Hosted impact

Hosted impact: C — webcmd-cloud change required
Trigger: 3.4 browser command surface (src/browser/command-catalog.ts)
Cloud files: src/default-adapters/source.ts:268, src/live-gates/cli-gates.ts:14
Why: cloud asserts the exact list ['bind','close','fork','init','run','snapshot','tabs','verify']
     and the count browserCommands: 8. Removing fork throws BrowserRunProtocolMismatchError.
Sequencing: land the cloud PR before `npm run bump:webcmd`.

The cloud side is agentrhq/webcmd-cloud#50, which serves POST /v1/adapters/override, drops the fork action, and re-pins the parity digests. Without it the hosted adapter override here 404s.

Order: merge this → bump PINNED_WEBCMD_SHA in cloud → merge webcmd-cloud#50.

Known window: between this release and the cloud deploy, hosted mode has no fork command — browser fork is gone and adapter override 404s until webcmd-cloud#50 ships. Accepted deliberately rather than splitting the PRs; keep the two merges close together.

Verification

  • npm run typecheck — clean
  • npm test — 5682 passed, 1 failed: src/doctor.test.ts:218 (profile alias rendering), pre-existing on main, unrelated to this change
  • npm run check:hosted-contract — passes; regenerated hosted-contract.json now lists 7 browser commands
  • npm run check:codex-plugin — passes

Cross-repo, per §7 of .github/hosted-impact-review.md: packed this branch, installed it into webcmd-cloud, ran cloud's full suite against it — 1189 passed, 0 failed.

Follow-ups from the same audit (not in this PR)

  • runCli() has no error handler, so adapter path, adapter source get|put, site fixture get, and session close <bad-id> dump raw Node stack traces
  • webcmd adapter path <site>/<cmd> fails for every command that is not already a local override — the documented private-fork workflow
  • Adapter error envelopes are always YAML, even under -f json
  • --help -f yaml is missing on site, skills, session, external

🤖 Generated with Claude Code

`webcmd browser fork` was an undocumented alias bound to the same handler as
`webcmd adapter override`. It leaked the alias immediately:

    $ webcmd browser fork hackernews
    Error: Usage: webcmd adapter override <site>/<command> (got "hackernews")

The error names a command the caller did not run, and nothing in docs/, skills/,
README, or CHANGELOG ever referenced `browser fork`. It also sat in the browser
namespace behind a `require-existing` session policy for what is a local file
copy.

Remove the local registration and the hosted catalog entry, so `adapter
override` is the single fork path in both modes.

`browser <retired>` now reports the replacement instead of a bare unknown
command, and unknown browser subcommands exit 2 (USAGE_ERROR) rather than 1:

    $ webcmd browser fork hackernews
    error: 'webcmd browser fork' was removed. Use: webcmd adapter override <site>/<command>

`fork` stays in BROWSER_SUBCOMMAND_NAMES: the preprocessor reads an unreserved
token after `browser` as a retired positional session id, so dropping it would
answer with "Browser sessions are root selectors" instead. The comment there now
says retired names are kept on purpose — the set was already a superset of the
declared subcommands.

Also corrects the usage string in the adapter-shadow hint, which showed
`adapter override <site> <command>` for a command that takes one
`<site>/<command>` positional.

BREAKING CHANGE: `webcmd browser fork` is removed. Use `webcmd adapter override
<site>/<command>`.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@github-actions

github-actions Bot commented Aug 19, 2026

Copy link
Copy Markdown
Contributor

🟠 Maintainer review suggested — low confidence

The automated review could not reach a fully supported conclusion.

Limitations

  • The automated review returned an invalid structured result.

This review is advisory and does not block merging.

Removing `browser fork` in the previous commit took away hosted mode's only
way to fork an adapter: `adapter override` was never routed in hosted mode
(runner.ts matched only `source`, `path`, and help), and cloud exposes no
override route — `forkSystemAdapterCommand` had exactly one caller, the
`fork` browser action.

Route `adapter override <site>/<command>` in hosted mode to the new
`POST /v1/adapters/override`, so the command means the same thing in both
modes: fork an installed upstream command into a private copy you can edit,
refusing commands that are not installed upstream and commands already
forked.

Storage differs, so the output does. Local prints the file it wrote and its
`.base` copy; hosted prints the private package id and the source file, and
points at `adapter source get|put` as the editing path — the hosted analogue
of editing the file at `adapter path`.

Requires the matching webcmd-cloud change; without it the route 404s.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@ankitranjan7 ankitranjan7 changed the title fix(cli)!: remove the duplicate browser fork subcommand fix(cli)!: remove browser fork, add adapter override to hosted mode Aug 19, 2026
@ankitranjan7
ankitranjan7 merged commit 3919f6f into main Aug 19, 2026
37 checks passed
Agnik47 added a commit to Agnik47/webcmd that referenced this pull request Aug 20, 2026
…q#361

agentrhq#361 landed on main after this branch was opened and already removed
`browser fork` for good (retirement redirect to `adapter override`).
Rebasing reintroduced this PR's now-stale fork re-registration, which
duplicated that removal and broke the retirement tests. Drop it and
update the grouping/tests/docs to match the current adapter authoring
surface (init, verify).
Agnik47 added a commit to Agnik47/webcmd that referenced this pull request Aug 20, 2026
…q#361

agentrhq#361 landed on main after this branch was opened and already removed
`browser fork` for good (retirement redirect to `adapter override`).
Rebasing reintroduced this PR's now-stale fork re-registration, which
duplicated that removal and broke the retirement tests. Drop it and
update the grouping/tests/docs to match the current adapter authoring
surface (init, verify).
Agnik47 added a commit to Agnik47/webcmd that referenced this pull request Aug 28, 2026
…gentrhq#317)

`browser --help` listed its session commands and its adapter-authoring commands
in one flat block, so the two unrelated surfaces read as one. Group them from
the shared catalog: `tabs`, `bind`, `run`, `snapshot`, and `close` under
"Browser session commands", `init` and `verify` under "Adapter authoring
commands". Local and hosted help draw from the same catalog, so they stay
aligned.

`browser fork` is also removed here. It ran the same action as the adapter
command and only existed as a hidden re-registration, which agentrhq#361 superseded;
`adapter override` gains a `fork` alias so the local spelling survives.

Rebased onto 0.7.10. Three upstream changes landed in the same lines and are
kept: the structured-help imports from agentrhq#425/agentrhq#428, `BROWSER_RUN_HELP_TEXT`, and
the `adapter override` description and hosted artifact-download paragraph. The
alias is additive to that description rather than replacing it.
@ankitranjan7
ankitranjan7 deleted the fix/remove-browser-fork branch September 1, 2026 12:41
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant