[crazy-daddy]

Summary

Adds hooks/project_identity.py, an optional SessionStart hook that writes karma-project.json into the session's project dir (~/.claude/projects/<encoded>/):

{ "git_identity": "owner/repo", "subdir": "api", "project_path": "/home/me/code/repo/api", "written_by": "hook" }

The file records which repository a project dir belongs to, independent of where the repo lives on disk. Extra session roots (#115) use it to match the same project across machines whose paths differ. Karma writes this file itself for machines running it; the hook covers machines whose ~/.claude is synced to a Karma install but that don't run Karma.

Independent of #115 and its follow-up: no shared code, merges in any order. Without #115 the file is simply not read.

Behavior

• Stdlib only, Python 3.9+, always exits 0 with no output; each of its two git calls times out after 2 s
• Writes atomically (per-process tmp file + os.replace) and only when the identity changed
• Writes nothing when there is no origin remote, when the repo root can't be determined, or when the cwd lies outside the repo root, so a wrong subdir never merges a subproject into its repo root
• Only writes into a project dir inside a projects dir; reads stdin and git output as UTF-8 (non-ASCII paths on Windows)
• normalize_git_url is a copy of api/services/git_identity.py's; a test compiles the API's version (without importing the API package) and checks both agree

Docs

SETUP.md → Hook Configuration Reference → "Project Identity Sidecar (Multiple Machines)"

Testing

hooks/tests/test_project_identity.py: URL normalization, subdir computation (Windows paths, case, unrelated paths), end-to-end runs against a temp git repo (write, skip unchanged, no remote, non-ASCII path, non-project dir, missing dir), malformed stdin, API parity. The 2 failures in hooks/tests (test_ticket_branch_detector.py) also fail on current main on Windows.

[JayantDevkar]

Read through. Small, careful hook: silent exit 0 on every path, list-form git calls with timeouts, the encoded dir taken from transcript_path so it cannot disagree with Karma's encoding, atomic write with no mkdir. No blockers; three things to fix before it lands, and it lands together with #115.

Should fix

• normalize_git_url (hooks/project_identity.py:56-64, and the same logic in api/services/git_identity.py:71-86) collapses an Azure DevOps HTTPS URL https://dev.azure.com/org/proj/_git/repo to _git/repo, so unrelated Azure repos with the same name share one identity. Drop a _git segment before taking the last two parts, in both copies at once (the parity test requires them to agree), and add the case to the shared list.
• The parity test's case list (hooks/tests/test_project_identity.py:31-42) only has easy shapes. Add https with credentials, ssh://git@host:2222/o/r.git with a port, a trailing slash, an uppercase .GIT suffix, and Azure.
• SETUP.md:985-989, the "Removing a Hook" table has no row for project_identity.py.

Nits

• SETUP.md:979: give the snippet an explicit "timeout": 10 like the other hook entries (worst case here is about 4 s of git on a synchronous SessionStart).
• The hook never clears a stale sidecar if origin is later removed, and a change to the three key fields overwrites any extra fields. Fine, but say so in the doc, and confirm the four keys match what #115's indexer writes.

No CI covers hooks/, so none of the new tests have run anywhere yet; I'll add hooks/** to a workflow on main. Until then please run pytest hooks/tests on 3.9 before pushing the next revision.

[crazy-daddy]

Rebased onto main and pushed two commits: 22e4fb7 for the review, 309a449 for two defects a second pass found in the same function.

Should fix

• Azure DevOps: normalize_git_url drops a _git segment in both copies, but only in the position before the repo (parts[-2]), so a repo or owner literally named _git still parses. https://dev.azure.com/org/Proj/_git/Repo, https://org@dev.azure.com/org/Proj/_git/Repo, https://org.visualstudio.com/DefaultCollection/Proj/_git/Repo and git@ssh.dev.azure.com:v3/org/Proj/Repo all give proj/repo.
• Parity cases now include https with credentials, ssh://git@host.example.com:2222/Owner/Repo.git, trailing slash, .GIT, Azure https and ssh, and https://github.com/Owner/_git. The .GIT case needed a case-insensitive suffix strip in both copies.
• While adding the port case I found the scp regex matched scheme URLs with a port first (ssh://git@host:2222/Owner/Repo.git parsed as path 2222/Owner/Repo and only passed because the last two segments are taken). Scheme URLs now go to _URL_RE first; ssh://git@host:2222/repo.git gives None instead of 2222/repo. Both copies, parity test covers it.
• SETUP.md: "Removing a Hook" row for project_identity.py.

Nits

• Snippet has "timeout": 10; the prose says the two git calls self-limit to 2 s each so the entry is a safety net.
• Documented: the hook never clears a stale sidecar after origin is removed, and a change to any of the three key fields rewrites the file, dropping extra keys.
• Confirmed the four keys (git_identity, subdir, project_path, written_by) match what write_sidecar in Merge sessions from other machines via extra session roots (1/2: core) #115 writes; only written_by differs by value (hook vs karma-api).

Verified on Python 3.9.13: pytest hooks/tests/test_project_identity.py 39 passed. api/tests/test_git_identity.py 31 passed on the api venv. ruff check/format --check clean with the repo config. The two test_ticket_branch_detector.py failures in hooks/tests fail identically on main on this Windows box and are untouched here.

[JayantDevkar]

Re-reviewed the two commits against the code: every item in your reply is as described (Azure _git only at parts[-2], scheme URLs before scp, case-insensitive .git, parity cases, SETUP row, timeout note). Both copies of normalize_git_url are logically identical.

Pushed 031d65f on top with three small things from the second read, so you don't have to round-trip:

• The sidecar guard now requires the transcript's parent to resolve to <config dir>/projects (CLAUDE_CONFIG_DIR, else ~/.claude), not any directory whose parent happens to be named projects. New end-to-end test for the outside-config case.
• The temp-file assertion could never fail (the real name carries a pid suffix); it now globs .karma-project.json.tmp*.
• One-line comment on the ssh://git@host:2222/_git/repo case saying the _git drop is positional.

40 + 31 tests pass, ruff clean. Lands together with #115 once that one's finishing commit is in.

[JayantDevkar]

Approved with #115. Hook is 3.9-safe, every path exits 0 silently, both normalize_git_url copies are logically identical and the parity test covers the edge list. Merging after #115.