fix(mcp,ingest): _setup_claude parse-miss + stderr + atomic settings.json + timestamped backup#348
Open
Huntehhh wants to merge 3 commits into
Open
fix(mcp,ingest): _setup_claude parse-miss + stderr + atomic settings.json + timestamped backup#348Huntehhh wants to merge 3 commits into
Huntehhh wants to merge 3 commits into
Conversation
install.ps1 and install.sh invoked `truememory-mcp --setup` and `truememory-ingest install` directly. On Windows hosts running Defender's Attack-Surface-Reduction rule 01443614 in Block mode, those console- script shims (setuptools / uv trampolines with per-install unique hashes) silently fail the cloud-prevalence check and never launch. Switched both installers to `\$toolPython -m truememory.mcp_server` / `\$toolPython -m truememory.ingest.cli`, routing through the high- prevalence signed python.exe — invisible to ASR. This mirrors the invocation form `mcp_server._setup_claude` already writes into the Claude Code / Claude Desktop MCP config, so installer and runtime are now consistent. Added a Windows-ASR troubleshooting note to the install.ps1 footer and to the README's advanced-setup section so Block-mode users have the manual re-run command. Co-Authored-By: claude-opus-4-7 <wontreply@getfucked.ai>
…th migration Bundles install-script hardening and a one-time _setup_claude shim-path migration with the ASR fix so users with stale Claude MCP configs from prior installs get auto-migrated to the python -m form. install.ps1 / install.sh - Soften the early Die-on-missing-toolPython to a Warn so TRUEMEMORY_SKIP_SETUP=1 (and the model-download fallback) still work - install.sh: add `set -o pipefail` so mid-pipeline failures aren't silently swallowed (currently the script reports "installed successfully" on a broken install) - install.ps1: quote $PKG_SPEC so TRUEMEMORY_SOURCE paths with spaces don't get split by PowerShell's argument parser - install.ps1: surface `uv tool uninstall` exit codes > 1 (locked files, permissions) instead of swallowing them with `*> $null` - install.ps1 footer: add a Block-mode ASR hint with the python -m equivalent commands so users who copy-paste the bare-shim commands have an escape hatch truememory/mcp_server.py: _setup_claude - New _is_shim_path() detects setuptools / uv truememory-mcp shim command paths in existing Claude Code / Claude Desktop configs - When such a path is found, migrate the registration to the [python_path, "-m", "truememory.mcp_server"] form instead of preserving it (the previous "existing config preserved" branch left ASR-blocked shim paths in place — every Claude launch then tried to CreateProcess the blocked binary) - Migration is reported as "migrated from shim to python -m form" in the setup output so the change is visible - Route the two remaining stdout error prints in _setup_claude to stderr for MCP-stdio-protocol safety README.md - Step 4: clarify Windows "quit" — clicking X minimizes to system tray, MCP config only reloads at a full process launch - BibTeX citation version: 0.6.0 -> 0.6.8 CHANGELOG.md: full Unreleased / Fixed entries for all of the above Co-Authored-By: claude-opus-4-7 <wontreply@getfucked.ai>
…json + timestamped backup Four small, related polish fixes layered on top of PR buildingjoshbetter#346's ASR + shim- migration work: mcp_server.py:_setup_claude - Treat empty existing_cmd as a parse miss (preserve the existing registration) rather than as a stale entry (force-remove + re-add). Pre-fix: `claude mcp list` output that didn't match the parser regex fell through to `_path_exists("")` = False = stale → silently destroyed any working dev-venv path the user had set up. Now preserve and emit a stderr note explaining how to recover. - Same fix on the Claude Desktop branch: an entry with an empty / missing 'command' field is preserved so user-set env / args / cwd fields survive. - Route the remaining stdout prints in _setup_claude (success summary block + Claude Code failure path) to sys.stderr. Defense in depth for the MCP stdio protocol — --setup exits before mcp.run() so the current behavior is safe, but a future refactor calling _setup_claude mid-session would silently corrupt JSON-RPC on any bare print(). ingest/cli.py:_run_install - settings.json write via tmp + Path.replace() instead of direct write_text(). The old truncate+write pattern had a race window where a concurrent hook reader (SessionStart, Stop) saw a partial / empty file and crashed with JSONDecodeError. Atomic on POSIX, best-effort on Windows (much narrower window). Falls back to direct write on cross-volume rename failure. ingest/cli.py:_merge_claude_md - Timestamp the CLAUDE.md.bak filename — fixed `.with_suffix(".md.bak")` silently clobbered the previous backup on every re-install. Now `CLAUDE.md.bak.<unix-ts>` preserves the backup chain. Tests (tests/test_mcp_setup_hygiene.py — 6 new): - Parse-miss preserves Claude Code entry + diagnostic on stderr - Stale (non-existent) path replaces Claude Code entry - Empty command field preserves Claude Desktop entry (user env survives) - Stale path replaces Claude Desktop entry - Atomic settings.json write leaves no .tmp leftover - CLAUDE.md.bak timestamping produces distinct files across runs Co-Authored-By: claude-opus-4-7 <wontreply@getfucked.ai>
This was referenced May 17, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Four small polish fixes layered on top of #346's ASR + shim-migration work. All scoped to
_setup_claudeand_run_install; behaviour-preserving for the happy path, fixes real footguns on the unhappy paths.What changed
mcp_server.py:_setup_claudeParse-miss preserves existing registration (both Claude Code + Claude Desktop branches). The old code treated an empty
existing_cmdas stale and force-removed — which silently destroyed any working dev-venv path the user had set up ifclaude mcp list's output format ever drifted or the regex missed for any other reason. Now empty → preserve, with a one-line stderr note explaining how to recover. Symmetric fix on the Desktop side: an entry with an empty/missingcommandfield is preserved so any user-setenv/args/cwdsurvive.All
_setup_claudeprint()calls route tosys.stderr— defence in depth for the MCP stdio protocol.--setupreturns beforemcp.run()so the current behaviour is safe, but any future refactor that calls_setup_claudemid-session would silently corrupt the JSON-RPC transport on any bareprint(). User runs--setupinteractively, so stderr lands in their terminal anyway.ingest/cli.py:_run_installsettings.jsonwrite — the previous directwrite_text()truncated then wrote. A concurrent hook reader (SessionStart,Stop) landing inside that window parsed a partial/empty file and crashed withJSONDecodeError, silently breaking lifecycle hooks until the next install. Switched to tmp +Path.replace(): atomic on POSIX, best-effort on Windows (much narrower window). Falls back to direct write on cross-volume rename failure so the user never loses their config.ingest/cli.py:_merge_claude_mdCLAUDE.md.bak.<unix-ts>backup filename — the previous fixed.with_suffix(".md.bak")silently clobbered the prior backup on every re-install. Re-installs across--userarg swaps or upgrades destroyed the backup chain. Timestamping preserves it.Tests
tests/test_mcp_setup_hygiene.py— 6 new tests, all passing:commandfield preserves Claude Desktop entry (userenvsurvives)settings.jsonwrite leaves no.tmpleftover on diskCLAUDE.md.baktimestamping produces distinct files across runsTest plan
pytest tests/test_mcp_setup_hygiene.py -v— all 6 pass on Linux / macOS / Windowspython -m truememory.mcp_server --setup— verify "existing config preserved" (not "stale entry replaced") in the setup output--setup— verifyenvblock survives untouchedtruememory-ingest install: hook reader never seesJSONDecodeErrortruememory-ingest installtwice with differentCLAUDE.mdcontent — two distinct.bak.<ts>files existWhat I deliberately did NOT bundle
Saved as separate PRs in the multi-agent coordination plan:
start_new_session=TrueWindows guards,ps/pgrepcross-platform branches, AF_UNIX fallback) — agent-B's PR-2 scopetests/test_cli_help.pytests invoke the baretruememory-mcp.exe/truememory-ingest.exeshims directly and ASR-block on Windows Block-mode hosts)docs/setup-*.md,docs/cli.md,docs/guides/debugging.md— shipping next as a separate PRPre-merge note for maintainers
This is one of a series of focused PRs from the multi-agent TrueMemory hardening sweep. The recommended merge order (other agents' PRs included) is:
Order constraints affecting this PR specifically: #346 must land first (this PR's
_setup_claudechanges layer on top of #346's_is_shim_path()migration). Everything else can shift.Merge ordering
Depends on: #346 (agent-B installer ASR bypass — agent-C's parse-miss guard + stderr routing in
mcp_server.py:1280-1400layers on top of agent-B's_is_shim_path()shim-migration in the same region). Clean mechanical merge once #346 lands.Blocks: none.
Recommended sequence position: directly after #346 / #351 (PR-2b subprocess portability).