fix: add wing param to diary_write/diary_read, derive from transcript path

[jphein]

Problem

All diary entries from the stop hook land in wing_session-hook regardless of which project the Claude Code session is in. This makes per-project diary search impossible — everything piles into one wing.

Solution

mcp_server.py

• Add optional wing parameter to tool_diary_write(). If provided, sanitizes it (lowercase, spaces→underscores) and uses it directly. If omitted, falls back to the existing wing_{agent_name} behavior.
• Add optional wing parameter to tool_diary_read() for filtering by target wing.
• Expose wing in the input_schema for both mempalace_diary_write and mempalace_diary_read in the TOOLS dict.

hooks_cli.py

• Add _wing_from_transcript_path(path) helper that extracts the project name from Claude Code transcript paths:
  • ~/.claude/projects/-home-jp-Projects-kiyo-xhci-fix/... → kiyo-xhci-fix
  • Falls back to "sessions" for unrecognized paths.
• In hook_stop, derive the project wing from the transcript path and append a wing=<project> hint to the block reason so Claude writes diary entries to the correct per-project wing.

Backwards Compatibility

Fully backwards compatible — wing is optional in both functions and the TOOLS schema. Omitting it produces identical behavior to before.

Test Changes

Updated test_stop_hook_blocks_at_interval to use startswith(STOP_BLOCK_REASON) since the block reason now includes a project wing suffix.

Test plan

• python -m pytest tests/ -q --ignore=tests/test_convo_miner.py — all 597 tests pass
• ruff check mempalace/mcp_server.py mempalace/hooks_cli.py — no issues
• Manual: tool_diary_write(agent_name="claude", entry="test", wing="my_project") stores in my_project/diary
• Manual: tool_diary_write(agent_name="claude", entry="test") still stores in wing_claude/diary (backwards compat)
• Manual: stop hook block reason includes wing=<project> for Claude Code sessions

[Copilot]

Pull request overview

This PR fixes stop-hook diary entries being written to a single agent wing by introducing an optional wing parameter for diary read/write and deriving a per-project wing from the Claude Code transcript path.

Changes:

• Added optional wing parameter to tool_diary_write() and tool_diary_read(), and exposed it via the MCP tool schemas.
• Added transcript-path parsing in the stop hook to append a wing=<project> hint to the block reason.
• Relaxed the stop-hook test assertion to account for the new reason suffix.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 5 comments.

File	Description
mempalace/mcp_server.py	Adds wing support to diary tools and updates MCP tool schemas accordingly.
mempalace/hooks_cli.py	Derives a per-project wing from transcript paths and includes it in stop-hook block messaging.
tests/test_hooks_cli.py	Updates assertion to tolerate the new stop-hook reason suffix.

---

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

[Copilot]

wing is currently only normalized via lower().replace(" ", "_") and is not passed through sanitize_name like other wing/room inputs (e.g. tool_add_drawer). This allows slashes, .., very long strings, etc., which can break IDs/metadata and bypass the repo’s path-traversal protections. Please validate/sanitize wing (and ideally the final computed wing string) using sanitize_name, and treat whitespace-only wing as “not provided”.

[jphein]

Fixed — replaced wing.lower().replace(" ", "_") with sanitize_name(wing) for proper path-traversal/length/character validation.

[Copilot]

tool_diary_read now accepts a user-provided wing but does not validate it (or agent_name) and doesn’t handle ValueError like tool_diary_write. For consistency and to avoid invalid metadata queries / unexpected behavior, sanitize agent_name and wing with the shared sanitizers and return a structured {success: False, error: ...} on validation failure.

[jphein]

Fixed — added sanitize_name() for both agent_name and wing in tool_diary_read(), plus try/except ValueError matching tool_diary_write()'s pattern.

[Copilot]

The tool_diary_read docstring still describes reading an agent’s personal diary only, but the function signature now supports overriding the target wing. Please update the docstring to mention the wing parameter and clarify how it changes the data source (agent-default vs explicit wing).

Suggested change

	Read an agent's recent diary entries. Returns the last N entries
	in chronological order — the agent's personal journal.
	Read recent diary entries for an agent.
	Returns the last N entries from the diary room in chronological order.
	If `wing` is omitted or empty, entries are read from the agent's default
	personal wing derived from `agent_name`. If `wing` is provided, entries
	are read from that explicit wing instead.

[jphein]

Fixed — updated docstring to document wing parameter and how it changes the data source.

[Copilot]

_wing_from_transcript_path only treats / as a path separator. On Windows (CI runs windows-latest), transcript paths are typically backslash-separated, so this will always fall back to "sessions" and won’t actually fix per-project diary wings. Consider supporting both separators (or use Path(...).parts / normalize separators) when extracting the project segment.

Suggested change

	match = re.search(r"-Projects-([^/]+?)(?:/|$)", transcript_path)
	normalized_path = transcript_path.replace("\\", "/")
	match = re.search(r"-Projects-([^/]+?)(?:/|$)", normalized_path)

[jphein]

Fixed — added path separator normalization (replace("\\\\", "/")) before regex match. Added dedicated tests for Windows backslash paths.

[Copilot]

The test now only checks reason.startswith(STOP_BLOCK_REASON), which can pass even if the new wing=... suffix is missing or malformed. Add assertions for the expected suffix (e.g. contains wing=sessions for non-Claude paths) and a new case covering a Claude Code transcript path so _wing_from_transcript_path() behavior is exercised.

Suggested change

	assert result["reason"].startswith(STOP_BLOCK_REASON)
	assert result["reason"].startswith(STOP_BLOCK_REASON)
	assert "wing=sessions" in result["reason"]
	def test_stop_hook_blocks_at_interval_claude_code_path(tmp_path):
	transcript = tmp_path / ".claude" / "projects" / "example" / "t.jsonl"
	transcript.parent.mkdir(parents=True, exist_ok=True)
	_write_transcript(
	transcript,
	[{"message": {"role": "user", "content": f"msg {i}"}} for i in range(SAVE_INTERVAL)],
	)
	result = _capture_hook_output(
	hook_stop,
	{"session_id": "test", "stop_hook_active": False, "transcript_path": str(transcript)},
	state_dir=tmp_path,
	)
	assert result["decision"] == "block"
	assert result["reason"].startswith(STOP_BLOCK_REASON)
	assert "wing=projects" in result["reason"]

[jphein]

Fixed — tightened assertion to check for wing=sessions suffix. Added 5 new tests: wing extraction from Claude Code transcript path, project extraction, fallback to sessions, Windows paths, and lowercasing.

[jphein]

Rebased onto current develop (post-#647 and #667). The wing param in diary_read now preserves #647's sanitize_name() validation and last_n clamping — both combined cleanly with the wing parameter logic.

[Qodo-Free-For-OSS]

Hi, tool_diary_read() ignores agent_name in its Chroma where filter when wing is provided, so it can return entries written by other agents into the same wing.

Severity: action required | Category: security

How to fix: Filter diary read by agent

Agent prompt to fix - you can give this to your LLM of choice:

Issue description

tool_diary_read() can return other agents’ diary entries when callers pass a custom wing, because the Chroma where filter does not include the agent metadata field.

Issue Context

tool_diary_write() stores diary entries with metadata including agent. tool_diary_read() requires agent_name but currently does not use it in the DB query, relying on default wing naming for isolation.

Fix Focus Areas

• mempalace/mcp_server.py[989-1045]
  • Update the where clause to include { "agent": agent_name } (and keep wing/room filters).

We noticed a couple of other issues in this PR as well — happy to share if helpful.

---

Found by Qodo code review

[jphein]

Friendly ping on review — adds an optional wing parameter to tool_diary_write / tool_diary_read and derives the project wing from the Claude Code transcript path when writing from the stop hook, so per-project sessions don't all collapse into the default session-hook wing. Clean against current develop.

[bensig]

Code review

Found 2 issues:

1. _wing_from_transcript_path() returns bare project names (e.g., "myproject") without the wing_ prefix, but all existing wings follow the wing_* convention (documented in AAAK_SPEC: "WINGS: wing_user, wing_agent, wing_team, wing_code, wing_myproject..."). Diary entries written via the stop hook land in wing "myproject" while diary_read defaults to "wing_claude" -- entries become unreachable by default reads.

mempalace/mempalace/hooks_cli.py

Lines 214 to 217 in ca5b357

	match = re.search(r"-Projects-([^/]+?)(?:/|$)", normalized)
	if match:
	return match.group(1).lower().replace(" ", "_")
	return "sessions"

2. tool_diary_read() does not include agent_name in the ChromaDB where filter when a custom wing is provided. The filter is {"$and": [{"wing": wing}, {"room": "diary"}]} with no {"agent": agent_name} clause. Any caller passing a custom wing can read diary entries written by other agents into that wing -- a cross-agent information leak. (Also flagged by Qudo, unresolved.)

mempalace/mempalace/mcp_server.py

Lines 1017 to 1019 in ca5b357

	)
	if not results["ids"]:

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[bensig]

@jphein small fix needed: _wing_from_transcript_path() returns bare project names (e.g. myproject) but existing wings all use the wing_ prefix (wing_myproject). Entries written via the stop hook will be unreachable by default diary_read calls. See review comment above for details.

[jphein]

Both fixed in 49704af:

1. _wing_from_transcript_path() now returns wing_<project> (matches AAAK_SPEC) and falls back to wing_sessions. 6 existing tests updated.
2. tool_diary_read() now includes {"agent": agent_name} in the $and filter — no cross-agent read access when a custom wing is shared.

All 51 hooks_cli + 62 mcp_server tests pass locally.