fix(subagent): start agents on windows from a stored launch command

[LivXue]

Summary

Stored Windows launch commands were parsed with POSIX rules, which removed backslashes from interpreter and launcher paths. Launch consumers now share platform-aware parsing and quoting, and bare program names are resolved once on the child's PATH.

• Preserve quoted Windows and POSIX paths in readiness checks. Literal or escaped quotes in preceding arguments no longer hide a missing launcher from discovery or stale-row detection.
• Share path substitution across discovery, registration, scaffold smoke checks, the generated installer, and all five shipped installers. Under older Raven versions without the shared resolver, installers retain plain-path compatibility but refuse unsupported whitespace in command paths before writing a row. A working directory passed separately remains supported.
• Quote Raven's own ACP executable and forwarded config path for the launch platform. Windows quoting preserves embedded quotes and trailing backslashes.
• Resolve programs on the child's PATH, in directory order with executables preferred within each directory. ACP launches can resolve batch shims for configured server commands; CLI and Kimi prompt arguments do not opt into that lookup.

Type

• Fix
• Feature
• Docs
• CI / tooling
• Refactor
• Other

Verification

Checked on Windows after rebasing onto the current integration branch. Commands use the existing uv environment with this checkout and its plugin sources on PYTHONPATH, and PYTHONUTF8=1.

• uv run --no-sync pytest -n 0 tests/test_utils_commands.py tests/test_subagent_vendored_agents.py tests/test_subagent_builtin_agents.py tests/test_subagent_registry.py tests/test_cli_agents_commands.py::test_the_generated_installer_on_older_raven_checks_paths_before_registering -q -x: 237 passed.
• Expanded comparison: uv run --no-sync pytest -n 0 tests/test_utils_commands.py tests/test_subagent_vendored_agents.py tests/test_subagent_builtin_agents.py tests/test_subagent_registry.py tests/test_cli_agents_commands.py -q --tb=no --junitxml $prReport, with $prReport set to a temporary XML file. Fixed tree: 298 passed, 1 skipped, 7 failed, 4 errors. The rebased PR before the two follow-up fixes: 249 passed, 1 skipped, 12 failed, 4 errors. Both runs use the platform-safe permission-test skip needed to collect the CLI module on Windows. Comparing every failing test ID, including setup errors, found no introduced failures.
• The remaining Windows failures cover existing command-string/path-separator assertions, POSIX permission behavior, unavailable os.fchmod, and insufficient privileges for symlink fixtures. They are not reported as passing tests. The complete repository suite and a live WebUI dispatch were not run for this revision.
• Replacing the readiness and installer fixes with the old implementations made 28 regression cases fail. An inline uv run --no-sync python - probe compared readiness and launch parsing for 3,110 generated quote cases across both platform branches; all agreed.
• $prPythonFiles = @(git diff --name-only origin/main...HEAD -- '*.py'), followed by uv run --no-sync ruff check @prPythonFiles and uv run --no-sync ruff format --check @prPythonFiles: all 24 changed Python files passed.
• uv run --no-sync lint-imports: 10 contracts kept, 0 broken.
• uv run --no-sync python scripts/check_source_language.py origin/main and uv run --no-sync python scripts/check_large_files.py origin/main...HEAD: passed.
• uv run --no-sync python scripts/check_commit_messages.py origin/main..HEAD and node node_modules/@commitlint/cli/cli.js --from origin/main --to HEAD --config commitlint.config.cjs: passed.
• git diff --check origin/main...HEAD: passed.

The pre-submit sweep covered the full branch diff, command producers and consumers, platform behavior, older-version fallbacks, error paths, architecture, repository conventions, and test strength. No additional blocker was found.

• Relevant tests pass locally
• Relevant lint / type checks pass locally
• User-facing docs or screenshots are updated when needed

No UI or documented command syntax changed; no documentation or screenshot update is needed.

Risk

• Security impact considered
• Backward compatibility considered
• Rollback path is clear for risky changes

Windows commands retain native separators and quoting. POSIX launches still use shlex. Readiness preserves foreign absolute-path spellings and checks launcher existence; it is not a general command validator. Host ACP commands are derived when the roster loads, so no stored-command migration is needed.

Older installers now reject command paths they cannot represent, with an upgrade or space-free-path remedy, before changing the roster. Plain command paths and a separately passed working directory remain supported.

On Windows, ACP batch shims execute configured server arguments through cmd.exe. Prompt-bearing CLI and Kimi launches retain their existing lookup restriction. A CLI agent installed only as an npm shim still needs an explicit executable path. Reverting the squash commit restores the prior behavior.

Related Issues

#880

[gloryfromca]

Blocking: Windows CLI executable resolution and the relevant test regressions must be fixed before merge.

I reviewed the full diff and the affected launch, probe, manifest-resolution, and compatibility call paths. I also checked the branch history and PR description; the repository rules in AGENTS.md/CLAUDE.md and CONTEXT-MAP.md; backward compatibility for stored CLI rows and cross-shaped commands; test changes for weakening/skips; and the inner-layer import direction. The three concrete issues are inline.

Verification:

• PYTHONPATH=.:plugins-dist/everos-memory:plugins-dist/design-engine:plugins-dist/ppt-engine uv run pytest -n 0 tests/test_utils_commands.py tests/test_subagent_vendored_agents.py::TestAWindowsShapedCommand::test_a_missing_windows_launcher_disables_the_row_with_the_token_named tests/test_subagent_vendored_agents.py::TestASpaceInTheResolvedPath::test_resolution_quotes_a_spaced_subagent_dir -q -> 11 passed, 2 failed.
• PYTHONPATH=.:plugins-dist/everos-memory:plugins-dist/design-engine:plugins-dist/ppt-engine uv run pytest -n 0 tests/test_subagent_probe.py -q -> 66 passed.

[gloryfromca]

Blocking: the three existing Windows launch findings remain unresolved.

This revision only reformats one assignment in vendored_agents.py; it does not change the CLI executable resolution, cross-shaped command parsing, or the Windows-parser test setup raised in the existing threads. I reviewed the delta and rechecked it against the full diff, affected callers, repository rules and architecture, backward compatibility, and test changes. No new finding was introduced by the formatting commit.

Focused verification on this head:
PYTHONPATH=.:plugins-dist/everos-memory:plugins-dist/design-engine:plugins-dist/ppt-engine uv run pytest -n 0 tests/test_utils_commands.py tests/test_subagent_vendored_agents.py::TestAWindowsShapedCommand::test_a_missing_windows_launcher_disables_the_row_with_the_token_named tests/test_subagent_vendored_agents.py::TestASpaceInTheResolvedPath::test_resolution_quotes_a_spaced_subagent_dir -q -> 11 passed, 2 failed, with the same two failures already reported.

[gloryfromca]

Blocking: the three open Windows launch findings remain unresolved.

The tree is identical to the previous revision; only the second commit message changed from style(subagent) to refactor(subagent). I rechecked the full diff, affected callers, repository rules and architecture, backward compatibility, and test changes. The commit-message correction introduces no new code finding, but it does not address the open findings.

Focused verification:
PYTHONPATH=.:plugins-dist/everos-memory:plugins-dist/design-engine:plugins-dist/ppt-engine uv run pytest -n 0 tests/test_utils_commands.py tests/test_subagent_vendored_agents.py::TestAWindowsShapedCommand::test_a_missing_windows_launcher_disables_the_row_with_the_token_named tests/test_subagent_vendored_agents.py::TestASpaceInTheResolvedPath::test_resolution_quotes_a_spaced_subagent_dir -q -> 11 passed, 2 failed, unchanged.

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

The two reproduced regressions are fixed, and I withdrew the remaining prior blocker because it was pre-existing and has an explicit-path workaround. I reviewed the new delta and full PR diff, affected launch/probe callers and history, AGENTS.md/CLAUDE.md/CONTEXT-MAP.md rules, backward compatibility, test-strength changes, and the inner-layer import direction. One narrow POSIX spaced-path readiness edge is marked inline as a follow-up; it does not meet the fourth-round blocking bar because it needs a spaced product root plus a missing launcher and can be escaped by relocating the product.

Verification:
PYTHONPATH=.:plugins-dist/everos-memory:plugins-dist/design-engine:plugins-dist/ppt-engine uv run pytest -n 0 tests/test_utils_commands.py tests/test_subagent_vendored_agents.py tests/test_subagent_probe.py -q -> 180 passed.

[0xKT]

Not a blocker. Four more findings from the same pass, none of which holds the merge

Severity of this one finding, not a verdict on the pull request. The blocking findings from this pass are review threads on the changed files; GitHub renders those collapsed, as a file name with no text.
button. The one that does is in the review thread on vendored_agents.py.

Gates on the merged tree: gates.sh ruff/lint-imports/commit/large/lang rc=0.

1 -- command_quote does not double a trailing backslash run, so its own round-trip is false

raven/utils/commands.py:122-133. The Windows arm wraps in double quotes and escapes only the
inner quote. The Win32 rule is that a run of backslashes immediately before the CLOSING quote
must be doubled, or that quote is consumed as an escaped one. Forcing only
raven.utils.commands.os.name to "nt" -- the same thing your own _as("nt") helper does at
tests/test_utils_commands.py:19-20:

C:\Program Files\node\npx.cmd  -> "C:\Program Files\node\npx.cmd"  -> round-trips   <- control
C:\Users\me\agents\            -> "C:\Users\me\agents\"            -> ValueError: unbalanced quotes
C:\Users\me\agents\\           -> "C:\Users\me\agents\\"           -> C:\Users\me\agents\   (one lost, SILENTLY)

The docstring at :131-132 states the property this breaks: "Producing this platform's
quoting is what makes a template round-trip back into the same argv it was built from."

Reachability, which is why this is not blocking: the only route is SUBAGENT_PYTHON
(vendored_agents.py:418, raw operator bytes with only .strip()). The other call site,
command_quote(str(folder)), cannot produce a trailing separator -- folder is
manifest.parent from a */subagent.json glob, always one level below root. A
SUBAGENT_PYTHON ending in \ names a directory and was never going to launch; the cost is
that the failure becomes "unbalanced quotes" instead of "no such file". The two-backslash arm
is the one worth fixing: it is silent.

2 -- the round-trip property is asserted by name, not by test

tests/test_utils_commands.py:61-65 is docstringed for the round trip and samples one value
with a space, no inner quote and no trailing backslash. Three mutations of command_quote
survive the whole file at 11 passed; the control (quote returns the value unquoted) fails 2,
including that very test, so the suite does reach it. The sharpest survivor escapes an inner
quote as "" -- the other escaping real CommandLineToArgvW understands -- while
_split_windows at :53-56 only toggles quote state on a bare ". Under that mutant the
module's own quoter and its own splitter disagree (C:\dir\a"b round-trips to C:\dir\ab)
and the test named for the round trip still passes.

3 -- the guard in _resolve_executable_windows is unasserted

raven/utils/commands.py:147-149, test at tests/test_utils_commands.py:82-87. The test
monkeypatches shutil.which to lambda exe: None, and the guarded line is
return shutil.which(exe) or exe -- so with which() pinned to None both branches return
exe and both assertions hold whether or not the guard ran. Deleting the guard entirely leaves
11 passed; the control (the function always returning exe) fails
test_windows_resolves_a_bare_extensionless_name, so the suite does reach the function. A
which() returning a wrong-but-truthy path instead of None kills it. Production behaviour on a
real Windows host is correct -- this is a test hole, not a defect.

4 -- _is_absolute_path's docstring states the constraint this PR exists to lift

raven/agent/subagent/vendored_agents.py:458-461: "Tokens come from str.split(), so a path
containing spaces arrives here as fragments. The command templates the manifests and each
install.py emit keep their paths space-free, and that constraint is cheaper than
re-tokenizing every stored row's command line." Both halves are false at head: the two callers
pass command_tokens output (:477 and :920), and the PR's own new test class is named
TestASpaceInTheResolvedPath. That paragraph is the stated safety argument for judging shape
on a whitespace split, so a later reader who trusts it reasons from the wrong tokeniser.

Stated, not filed

• The Windows arm the PR set out to fix does work. Probed both directions:
  "C:\Program Files\gone\python.exe" C:\gone\run.py keeps the quoted path as ONE token and
  _launcher_missing returns it (fail-closed), where the base str.split() gave
  ['"C:\Program', 'Files\gone\python.exe"', ...]. The regression in the thread is confined to
  the POSIX convention this PR introduced.
• The shlex.split sweep is complete for what it claims: all eight launch readers at the base
  are converted, and the survivors under raven/ (permissions/rules.py x4,
  rpc/methods/{command_dispatch,input,slash_routing}.py, cli/ops_connection_commands.py)
  read slash commands and permission patterns, not launch commands.
• _split_windows diverges from real CommandLineToArgvW in two documented places -- ""
  inside quotes yields nothing rather than a literal ", and an unbalanced quote raises where
  Windows takes the rest as one token. Neither is filed: command_argv's own text says it
  raises "the same class callers catch today", and in this codebase the real parser never sees
  the string. Worth a sentence in the docstring naming the narrowing.
• The new module's docstring uses "product", which CONTEXT.md's Discovered agent
  _Avoid_ list retires for prose. Checked and dropped: the surrounding modules use it the same
  way throughout, so this PR is consistent with its neighbours rather than introducing it.

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

The POSIX quoted-root defect is fixed on this revision. command_tokens now groups both quote forms, and the added end-to-end tests drive discover_product_rows over a spaced root with both a missing and present launcher, covering the fail-closed and ready outcomes. I reviewed the delta against the full PR, affected callers and history, repository rules and architecture, backward compatibility, and test strength. I found no new blocker. The previously noted bare-name Windows CLI .cmd behavior remains a pre-existing follow-up with an explicit-path workaround.

Verification:
PYTHONPATH=.:plugins-dist/everos-memory:plugins-dist/design-engine:plugins-dist/ppt-engine uv run pytest -n 0 tests/test_utils_commands.py tests/test_subagent_vendored_agents.py tests/test_subagent_probe.py -q -> 185 passed.

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

The shared resolver makes discovery, registration, and newly generated installers agree, and I reviewed the delta against the full PR, affected callers/history, repository rules and architecture, backward compatibility, and test strength. One incomplete consumer remains inline: the default scaffold smoke still whitespace-splits the now-quoted command. I am carrying it as a nonblocking follow-up under the late-round bar because --no-smoke provides a working path through creation/registration, although the default workflow and four existing CLI tests currently fail.

Verification:
PYTHONPATH=.:plugins-dist/everos-memory:plugins-dist/design-engine:plugins-dist/ppt-engine uv run pytest -n 0 tests/test_utils_commands.py tests/test_subagent_vendored_agents.py tests/test_subagent_probe.py tests/test_cli_agents_commands.py tests/test_cli_subagent_setup.py -q -> 298 passed, 4 failed, 1 skipped. The four failures are the whitespace-policy CLI cases named inline.

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

This revision updates the CLI assertions to the new quoted-path success contract, but the previously reported nonblocking smoke follow-up remains unresolved: the default smoke still raw-splits quoted commands. The spaced-home and --here cases therefore still write the scaffold and exit nonzero. The SUBAGENT_PYTHON case now also expects success from a deliberately nonexistent interpreter, which readiness correctly rejects. I reviewed the test-only delta against the production flow, affected callers, repository rules and architecture, backward compatibility, and test strength; no new finding beyond the already recorded follow-up emerged.

Verification:
PYTHONPATH=.:plugins-dist/everos-memory:plugins-dist/design-engine:plugins-dist/ppt-engine uv run pytest -n 0 tests/test_utils_commands.py tests/test_subagent_vendored_agents.py tests/test_subagent_probe.py tests/test_cli_agents_commands.py tests/test_cli_subagent_setup.py -q -> 299 passed, 3 failed, 1 skipped.

[gloryfromca]

Blocking: the new spaced-path tests must assert the discovery-only contract or request registration.

The production smoke fix is correct: it now parses the stored command with command_argv, and the spaced-root invocations reach exit code 0. I reviewed the delta against the full PR, affected CLI/discovery callers and history, repository rules and architecture, backward compatibility, and test strength. The one new finding is inline.

This meets the late-round blocking bar: this revision introduces the failing assertions; every relevant Linux test run reaches them; and the branch has no passing-test path until the tests either add --register or inspect the discovered row.

Verification:
PYTHONPATH=.:plugins-dist/everos-memory:plugins-dist/design-engine:plugins-dist/ppt-engine uv run pytest -n 0 tests/test_utils_commands.py tests/test_subagent_vendored_agents.py tests/test_subagent_probe.py tests/test_cli_agents_commands.py tests/test_cli_subagent_setup.py -q -> 299 passed, 3 failed, 1 skipped.

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

The registration-contract blocker is fixed: the three spaced-path tests now opt into registration before reading the pinned roster, and the production smoke path continues to parse stored commands through command_argv. I reviewed the delta and the full PR diff, relevant callers and history, repository rules and architecture constraints, backward compatibility, and whether the tests were weakened. The pre-existing bare-name Windows CLI .cmd resolution limitation remains a follow-up with an explicit executable path as a working escape hatch.

Verification: PYTHONPATH=.:plugins-dist/everos-memory:plugins-dist/design-engine:plugins-dist/ppt-engine uv run pytest -n 0 tests/test_utils_commands.py tests/test_subagent_vendored_agents.py tests/test_subagent_probe.py tests/test_cli_agents_commands.py tests/test_cli_subagent_setup.py -q (302 passed, 1 skipped); git diff --check github/main...HEAD (clean).

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

The new revision fixes the Windows single-quote failure: host_raven_acp_command() and its forwarded --config path now use the same platform quoter as the launch parser. The added tests cover sibling and PATH launchers, spaced and unspaced paths, both simulated platforms, and config forwarding through command_argv. Range-diff confirms the prior eight reviewed commits are unchanged rebases and this is the only substantive delta.

I covered the repository rules, full PR diff and new delta, callers and history, backward compatibility, architecture constraints, and whether tests were weakened. The pre-existing bare-name Windows CLI .cmd resolution limitation remains only a follow-up with an explicit executable path as a working escape hatch.

Verification: PYTHONPATH=.:plugins-dist/everos-memory:plugins-dist/design-engine:plugins-dist/ppt-engine uv run pytest -n 0 tests/test_utils_commands.py tests/test_subagent_builtin_agents.py tests/test_subagent_vendored_agents.py tests/test_subagent_probe.py tests/test_cli_agents_commands.py tests/test_cli_subagent_setup.py -q (342 passed, 1 skipped); git diff --check github/main...HEAD (clean).

[LivXue]

Answering the board note above, item by item, at 7ce8b10a6. The same push rebases onto main (c3d282130).

1. command_quote and a trailing backslash: fixed in fix(utils): double the backslashes a windows quote would swallow (e20b07836). Every run of backslashes in front of a quote, the closing quote included, is now doubled, so C:\Users\me\agents\ and C:\Users\me\agents\\ both split back as themselves.
2. The round trip asserted by name: test_a_quoted_windows_token_round_trips now runs 11 tokens (one and two trailing backslashes, a backslash before an inner quote, quotes at both ends, a UNC path, the empty string) at either end of the line, with a POSIX twin over 7. Your three survivors (an inner quote escaped as "", left bare, dropped) now fail 6 tests each, and backslash runs left single fail 4.
3. The unasserted guard: _resolve_executable_windows went with launch_argv (below), and the same guard in resolve_program is asserted with a which() that answers every name. Deleting the path guard, or letting an unstartable suffix through, fails test_on_windows_only_a_bare_startable_name_is_looked_up.
4. _is_absolute_path's docstring now says its tokens come from command_tokens (347b5e2b0).

From the stated-not-filed list, _split_windows' docstring now names its two narrowings from the real parser, neither of them a spelling command_quote produces.

#880's lookup is folded in: the ACP launch looks a bare program up once, on the child's PATH, directory by directory with .exe first (f4103d5c4). resolve_program takes batch_files. The ACP launch passes it, because its argv is configuration and an npm-installed server has only a .cmd; the cli launch and the kimi ask do not, because their argv carries the prompt. Each of the three call sites has its own test.

Also in this push:

• The shipped install.py files write through the quoting rule (604747147).
• Under a raven without raven.utils.commands (v0.2.0 through v0.2.4), the shipped and scaffold installers write the unquoted row instead of failing on the import (7e6dbe1c6 for the scaffold).
• Text that still described the removed whitespace refusal is updated (81677c5de).
• The author of fix(agent): quote host acp paths for the launch platform is corrected. It was committed on a machine whose git address resolves to an unrelated GitHub account. Its patch-id is unchanged.

Verification and the mutation check are in the description.

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

The nine follow-up commits are sound. Windows quoting now round-trips backslashes and embedded quotes; ACP program lookup uses the child's PATH and permits batch shims without extending that exposure to prompt-bearing CLI launches; shipped and generated installers use the shared resolver while retaining their older-Raven fallback. The newly posted resolutions of the POSIX probe and host-ACP quoting failures match the current implementation and tests.

I covered the repository rules, full diff and revision delta, callers and history, backward compatibility, architecture constraints, and whether tests were weakened. All threads I opened are resolved. The pre-existing bare-name Windows CLI .cmd limitation remains only a follow-up with an explicit executable path as a working escape hatch.

Verification: PYTHONPATH=.:plugins-dist/everos-memory:plugins-dist/design-engine:plugins-dist/ppt-engine uv run pytest -n 0 tests/test_utils_commands.py tests/test_subagent_builtin_agents.py tests/test_subagent_vendored_agents.py tests/test_subagent_probe.py tests/test_subagent_host_env.py tests/test_subagent_kimi_code.py tests/test_subagent_third_party.py tests/test_subagent_acp.py tests/test_cli_agents_commands.py tests/test_cli_subagent_setup.py -q (998 passed); git diff --check github/main...HEAD (clean).

[0xKT]

Not a blocker -- from the re-acceptance pass on head 7ce8b10 (merged onto main 289426c), over the fix commits that came after the first pass.

Severity of this one finding, not a verdict on the pull request. The blocking findings from this pass are review threads on the changed files; GitHub renders those collapsed, as a file name with no text.

1. A quote in a hand-written manifest still splits the judge from the spawn. _split_shape (raven/utils/commands.py:100) groups on both quote characters with no escape, while the spawn splits with shlex on POSIX and _split_windows on nt. A template with an escaped quote before the launcher -- {PYTHON} --name=O\'Brien {SUBAGENT_DIR}/run.py --acp, or the same with \" -- opens a quote the judge never closes. Its tokens are ['<py>', '--name=O\Brien <root>/run.py --acp'], so _launcher_missing has no absolute token to check and a folder whose run.py is gone reads ready, while the spawn reads ['<py>', "--name=O'Brien", '<root>/run.py', '--acp'] and fails at dispatch. A double-quoted "--name=O'Brien" agrees on both sides. On the nt branch a bare --name=O'Brien does the same (noted on the earlier thread). Only a user-edited manifest can carry this -- none of the shipped ones does -- and main's str.split judge disabled the same rows. The _split_shape docstring (commands.py:80-93) still says single quotes are literal and that an unclosed quote answers "the path is not there".
2. Under an older raven, the generated installer now registers a row that cannot launch. raven/templates/agents_scaffold/install.py:26-32 falls back to plain substitution when raven.utils.commands is missing, so a folder with a space registers a row that splits at the space, where the previous installer refused the folder and said why. The comment there says this is deliberate; a refusal under the older raven would at least tell the user what to do.

Checked on the merged tree: the machine gates pass. The acceptance pass also ran the full suite: nothing red because of this PR; the reds are the known environment-only ones plus one process-group timing test that also fails on main under load.

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

The two new fixes close the re-acceptance notes without weakening coverage. command_tokens now honors host escape rules so literal quotes cannot hide a following launcher, while preserving cross-shaped absolute paths for fail-closed readiness checks. Shipped and generated installers running under an older Raven now reject spaced command paths before registration, but still accept a spaced cwd when the command does not interpolate it.

I covered the repository rules, full diff and revision delta, callers and history, backward compatibility, architecture constraints, and test strength. All threads I opened remain resolved. The pre-existing bare-name Windows CLI .cmd limitation remains only a follow-up with an explicit executable path as a working escape hatch.

Verification: PYTHONPATH=.:plugins-dist/everos-memory:plugins-dist/design-engine:plugins-dist/ppt-engine uv run pytest -n 0 tests/test_utils_commands.py tests/test_subagent_builtin_agents.py tests/test_subagent_vendored_agents.py tests/test_subagent_probe.py tests/test_subagent_host_env.py tests/test_subagent_kimi_code.py tests/test_subagent_third_party.py tests/test_subagent_acp.py tests/test_cli_agents_commands.py tests/test_cli_subagent_setup.py -q (1042 passed); git diff --check github/main...HEAD (clean).

[0xKT]

Not a blocker -- three items from the re-acceptance pass on head 144669d (merged onto main 6126965). Nothing blocking was found, and the touched tests pass (262 passed across tests/test_utils_commands.py, tests/test_subagent_vendored_agents.py and tests/test_cli_agents_commands.py).

Severity of this one finding, not a verdict on the pull request. The blocking findings from this pass are review threads on the changed files; GitHub renders those collapsed, as a file name with no text.

1. The Windows launch split drops or merges an argument made only of backslashes or an escaped quote. In _split_windows, the backslash branch (raven/utils/commands.py:41-57) never sets token_started, so such a token is dropped at the end of the line or glued onto the next one. The judge's copy of the same rule in _split_shape does set it (:99-100), so readiness judges a different argv from the one that starts. Measured with os.name patched to nt, each row built by subprocess.list2cmdline(want):
   • "C:\Program Files\agent\agent.exe" --root \ acp launches as [..., '--root', '\acp'];
   • prog \ launches as ['prog'];
   • prog \\ b launches as ['prog', '\\b'].
     Rows that command_quote writes are always quoted, so only hand-written Windows rows reach this. Adding token_started = True as the first line of the backslash branch fixes it: a run of backslashes always puts at least one character into the token. With that line, a command_argv(list2cmdline(argv)) == argv round trip over 50000 random argvs fails 0 times (6413 on head), and the three test files above still give 262 passed. They pass unchanged on head too, so a round-trip row or two would pin it.
2. On POSIX, the readiness split ignores newline and CR separators that the launcher splits on. _split_shape separates only on space and tab (:135), while the launcher's shlex.split also separates on \n and \r. So python\n/opt/gone/run.py --acp launches as ['python', '/opt/gone/run.py', '--acp'], but the judge reads ['python\n/opt/gone/run.py', '--acp']. main's str.split judge agreed with the launcher. The input needs an escaped newline inside a command string, so it is rare.
3. The new module docstring says "a Windows-shaped product" (raven/utils/commands.py:14). CONTEXT.md:363 lists "product" under Avoid for a discovered agent and keeps it only in code-level product_* spellings.