GodopuandClaude Opus 5.5
50c8da8269
Review loop (--review-only, --all-reviewer): round 1 (@325b056) 4/4 PASS,
round 2 (@db7f4ca, HERDR_SESSION) 4/4 PASS. Full suite at 325b056 on
macOS / Python 3.9 / bash 3.2 / herdr 0.9.3 from an agent pane:
1845 passed / 0 failed / 56 skipped (18 test_deploy_freshness skips: no
nats-docker submodule in the fresh worktree); the panes of the real
servers (default, multi-agent-mux) and ~/.claude/projects were unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VCgu9Fa9iT7TbJUSzdq4Ff
6.7 KiB
Review 191f0ec3 — a test run from a pane no longer opens workspaces on that pane's server
Head 325b056057fcdd717d2c81017fc5ef57183cb4d0 (tree 7905896f005159d596e3872d0ee2c3c12061f09a) is two commits on base 8b0b4a07dc76a4915895bdde6170dc8eed9b5f74: 525df2b (the leak fix) and 325b056 (the generated-shim pin moves to 525df2b). Four files, +106/−5. git diff --check is clean. Snapshot porcelain empty. Control-plane main is still 8b0b4a07dc76a4915895bdde6170dc8eed9b5f74, porcelain empty.
Both measured causes are right, and the fix matches them.
The suite that left seven workspaces on multi-agent-mux ran with --basetemp .../pytest-pr56-shim. mock_herdr puts its bin/herdr under that basetemp. Both resolvers used to skip any PATH directory whose path contained -shim anywhere, so they walked past the mock and executed /opt/homebrew/bin/herdr. The real binary follows HERDR_SOCKET_PATH ahead of HOME, and mam_sandbox left the pane's socket in the environment.
_is_shim_path (lib.sh:113-125) still matches /.mam/shim/, /multi-agent-herdr-shim/, and /.agents/skills/.bin. The substring *-shim/* is gone. A second case matches only when the path with one trailing slash stripped is itself *-shim. _resolve_real_herdr_path (:127-143) skips a PATH directory on that test, then skips the candidate when _is_shim_path matches dirname of the canonical file (:136). It still prints the PATH entry, $dir/herdr. The generated shim's _resolve_real_herdr (:172-186) now requires ${dir%/} to fail the pattern *-shim, and it still rejects a directory containing .mam/shim.
mam_sandbox deletes HERDR_SOCKET_PATH, HERDR_PANE_ID, and HERDR_TAB_ID (tests/conftest.py:296-298).
Wrapper directories
A decoy-only probe sourced this snapshot's lib.sh with WORKSPACE_ROOT=/tmp/m191ws. PATH was the case directory, a stand-in, /usr/bin, and /bin. The real binary was not on it.
| PATH directory | _resolve_real_herdr_path |
generated shim |
|---|---|---|
.../pytest-x-shim/case0/bin |
that mock | that mock |
a directory whose own name is herdr-shim |
stand-in | stand-in |
.../multi-agent-herdr-shim/srv |
stand-in | that directory |
.../.agents/skills/.bin |
stand-in | that directory |
.../.mam/shim/extra |
stand-in | stand-in |
An ancestor named pytest-x-shim above bin/ is selected by both resolvers. A directory that is itself *-shim is skipped by both. .mam/shim and a child of it are skipped by both. L1 and L1b assert the same two outcomes inside the suite, with a decoy behind the mock, and both passed.
lib.sh still excludes /multi-agent-herdr-shim/ and /.agents/skills/.bin. The generated shim excludes a PATH entry whose own name ends in -shim, so the directory multi-agent-herdr-shim itself is skipped by both. A child of that directory, whose own name does not end in -shim, is excluded only by _resolve_real_herdr_path. README.md:102 and README.ko.md:84 still describe /tmp/multi-agent-herdr-shim/<HERDR_SERVER_NAME>/herdr. No script in this tree creates that path; _init_herdr_isolation writes $WORKSPACE_ROOT/.mam/shim/herdr. The .agents/skills/.bin gap on the generated shim is older than this commit: the previous generated check was the substring *"-shim"*, and .bin never matched it. Nothing in lib.sh prepends .agents/skills/.bin onto PATH.
What still reaches the real binary
A direct call of mam_sandbox, with probe values set first, left HERDR_SESSION and HERDR_ENV set and removed HERDR_SOCKET_PATH, HERDR_PANE_ID, HERDR_TAB_ID, HERDR_SESSION_NAME, HERDR_WORKSPACE_ID, and HERDR_CONFIG_PATH. HOME was the sandbox. L2's PANE_VARS (tests/test_herdr_real_server_leak.py:74-75) are the socket, the pane and tab ids, the workspace id, and HERDR_SESSION_NAME. The test passed. It does not include HERDR_SESSION.
This reviewer process inherited HERDR_SESSION from its pane. lib.sh:1411-1415 records that herdr 0.8.2 left HERDR_SESSION unset on a default-server pane and exported the socket to every pane. docs/contracts/issue-6/implementation_plan.md records the measurement that the real binary follows HERDR_SESSION, HERDR_SOCKET_PATH, and --session, and that HERDR_SESSION_NAME alone reaches the default socket. The calls that created the seven workspaces go through these resolvers, which now return the mock, so they do not reach the binary that reads HERDR_SESSION. A later call that misses the mock still inherits HERDR_SESSION.
The suite's intentional real-server tests already drop it. tests/test_issue45_restore.py lists HERDR_SESSION in SCOPE_ENV and builds the server env from what remains. tests/test_issue50_dispatch_herdr.py Box.env drops every HERDR_* name, and Box.herdr passes only HOME, PATH, and --session. tests/test_herdr_shim_contract.py::test_h10_real_herdr_schema_match calls herdr api schema with the inherited environment and does not use mam_sandbox. That test was already in the tree. I did not run it, and I did not run the real binary. api schema is a read. The workspace-creating path was the resolver.
HERDR_ENV also survived the fixture call. The measured contract names HERDR_SESSION, HERDR_SOCKET_PATH, and --session. I did not treat HERDR_ENV as a second route.
Shared control plane
test_s11_this_branch_does_not_change_the_generated_shim (tests/test_issue23_24_startup_path.py:294-315) compares the shim heredoc with git show 525df2b:.agents/skills/lib.sh, unless MAM_BASE_SHA overrides it. It passed, so the heredoc at 325b056 is the heredoc at 525df2b. The second commit changes the pin and the comment. Sourcing lib.sh rewrites $WORKSPACE_ROOT/.mam/shim/herdr. This review sourced it with WORKSPACE_ROOT pointed at /tmp/m191ws and at pytest sandboxes. The narrowing becomes the control plane's shim when this branch is merged and something sources lib.sh with that workspace root.
What was run
Control-plane venv, Python 3.9.6, pytest 8.4.2, PYTHONDONTWRITEBYTECODE=1, -p no:cacheprovider, --basetemp=/tmp/m191bt-leak2 (the path contains neither shim nor mam-review). Inherited HERDR_SOCKET_PATH, HERDR_PANE_ID, HERDR_TAB_ID, HERDR_WORKSPACE_ID, HERDR_SESSION_NAME, HERDR_SESSION, HERDR_SERVER_NAME, HERDR_CONFIG_PATH, HERDR_ENV, WORKSPACE_ROOT, MAM_ENV_FILE, HOME_DIR, CLAUDE_PROJECT_DIR, and CODEX_HOME were unset.
tests/test_herdr_real_server_leak.py and tests/test_issue23_24_startup_path.py::test_s11_this_branch_does_not_change_the_generated_shim: 4 passed in 1.50s. The session guards that travel with that run completed. The #55 agent-home guard did not fire.
User ~/.config/herdr/config.toml mtime stayed 1784456394. Snapshot porcelain stayed empty after the run.
[VERDICT: PASS]