fix: diary wing routing for non-Projects layouts + empty-wing reads (#1145 bugs 1 & 2)

[jphein]

Addresses two of the three findings @igorls flagged on #1145 from the v3.3.3 smoke test. Batched into one PR because both are in the same file class and share the "LLM agent / non-~/Projects/ user defaults to empty string" failure mode. Bug 3 (Stop-hook auto-mine MEMPALACE_PALACE_PATH env propagation) lives in a different code path and is deferred to a follow-up PR pending local repro.

Bug 1 — _wing_from_transcript_path is ~/Projects/-centric

The heuristic only matched -Projects-<project> in the encoded transcript folder name. Users whose source lives outside ~/Projects/ — ~/dev/, ~/src/, ~/code/, or any other Linux convention — got transcript paths like ~/.claude/projects/-home-<user>-dev-<Parent>-<project>/*.jsonl with no -Projects- segment, fell through to the wing_sessions fallback, and lost the per-project diary scoping #659 was meant to deliver.

Fix: keep the explicit -Projects-<x> match first (preserves existing behavior / test coverage), then fall back to the last --delimited segment of the encoded folder name. That segment is the actual project directory name regardless of source layout — Claude Code's encoding puts it there for any path, not just under ~/Projects/.

Bug 2 — tool_diary_read didn't adopt #1097's empty-string pattern

#1097 fixed mempalace_search to treat wing="" / room="" as "no filter" (LLM agents frequently default optional string parameters to ""). tool_diary_read still defaulted an empty wing to wing_<agent_name>, silently hiding entries written to named wings — e.g., the per-project wings from #659 that the stop-hook now derives.

Fix: route wing through the _sanitize_optional_name() helper #1097 introduced so empty / whitespace / None all coerce to None and drop out of the query. An explicit non-empty wing still scopes to that wing. Build the where clause from a list of required conditions (room=diary, agent=<name>) and conditionally append the wing filter — matches the pattern tool_list_drawers already uses.

Tests

Six new tests, all green. Full suite: 1067 passed (this branch, from upstream/develop).

tests/test_hooks_cli.py — four _wing_from_transcript_path cases:

• test_wing_from_transcript_path_non_projects_layout — igor's specific scenario (~/dev/<Parent>/<project>)
• test_wing_from_transcript_path_src_layout — ~/src/<app>
• test_wing_from_transcript_path_code_layout — ~/code/<app>
• test_wing_from_transcript_path_bare_folder_no_delimiters_falls_back — genuinely unstructured parent → wing_sessions

tests/test_mcp_server.py::TestDiaryTools — two tool_diary_read cases:

• test_diary_read_empty_wing_returns_across_wings — wing="" returns entries from any wing (the bug as reported)
• test_diary_read_explicit_wing_still_filters — wing="wing_a" still scopes correctly; regression guard against overcorrecting the empty-string fix

Not in this PR

Bug 3 from #1145 (Stop-hook auto-mine subprocess ignores MEMPALACE_PALACE_PATH env) — env-propagation bug in a different subprocess-launch site, same file but different narrative/tests. Deferred to a follow-up PR once I've reproduced locally and confirmed root cause. Happy to bundle if reviewers prefer one PR; say the word.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Fixes two hook/diary routing issues triggered by non-~/Projects/ source layouts and LLM agents defaulting optional parameters to empty strings, improving per-project diary scoping and cross-wing diary reads.

Changes:

• Extend _wing_from_transcript_path() to derive a project wing for non-~/Projects/ transcript folder encodings.
• Update tool_diary_read() to normalize wing using _sanitize_optional_name() and conditionally apply the wing filter.
• Add unit tests covering non-Projects transcript layouts and empty-wing diary reads.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
mempalace/hooks_cli.py	Adds a second heuristic for deriving project wings from encoded transcript folder names outside ~/Projects/.
mempalace/mcp_server.py	Changes diary read filtering to treat empty/whitespace/None wing as “no filter” and builds the where clause conditionally.
tests/test_hooks_cli.py	Adds coverage for wing derivation across common non-Projects directory layouts plus a fallback case.
tests/test_mcp_server.py	Adds coverage ensuring wing="" reads across wings while explicit wing still scopes correctly.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

tool_diary_read now treats the default wing="" (i.e., callers omitting wing) as “no wing filter”, but the tool schema still documents “If omitted, reads from wing_{agent_name}”, and this is a behavioral change from the previous default-to-agent-wing behavior. If you want to support both behaviors (omit => agent default, explicit empty string/whitespace => no filter), consider using a sentinel default (e.g., wing=_UNSET) to distinguish “not provided” from "", and update the TOOLS["mempalace_diary_read"] schema description to match the actual semantics.

[Copilot]

With wing optional (and potentially unset/empty), tool_diary_read can now fetch across all wings for an agent, but it still does col.get(..., limit=10000) and then sorts/slices client-side. This risks silently truncating results (there’s an existing note elsewhere about a 10K get() limit) and makes the total field inaccurate when there are >10K diary entries. Consider paginating get() (offset loop) or fetching only metadatas first to identify the latest last_n entry IDs, then retrieving documents for just those IDs; this avoids large reads and removes the 10K correctness hazard.

[jphein]

Closing as duplicate — we filed #1146 and @igorls filed #1147 within four minutes of each other covering the same two bugs. Their PR is the right vehicle for 3.3.3:

• Cleaner primary regex — /\.claude/projects/-([^/]+) matches Claude Code's actual encoded-folder location as the primary strategy, with -Projects-<x> as the legacy fallback. That reflects the encoding rules more directly than my "-Projects- first, last-segment fallback" ordering.
• CHANGELOG updated — the 3.3.3 followup context is preserved in the release notes where it belongs.
• Docstring update on tool_diary_read is clearer about the "empty = no filter" contract.
• Tests cover the same ground; theirs have the macOS /Users/ layout case which mine didn't.

Thanks @igorls for the smoke test + quick turnaround on both the findings and the fix. Re the Bug 3 retraction — confirmed on my side: subprocess.Popen without env= inherits parent env, MempalaceConfig.palace_path reads MEMPALACE_PALACE_PATH / MEMPAL_PALACE_PATH before falling back to config-file default. No propagation gap. Good catch on the measurement artifact.

No action needed on my end.