Repository navigation
Pathfinder v0.1.4
Pathfinder v0.1.4
Release notes for the changes since v0.1.3, reviewed on 2026-10-22.
Summary
- Scans of the same project now wait for one another across the UI, CLI, MCP,
and library API. - Queries accept bare method names and list candidates when a name is ambiguous.
- Deep dives and briefings identify a structural centre for modules without a
recognized entry point. - Descriptions record when they were written.
- The live graph saves 3D depth in the project database, and symbol cells use
fuller 3D layouts. - AI briefings separate test fixtures from runtime architecture and show named
CommonJS exports. - Core MCP orientation tools carry metadata requesting upfront loading in
supporting clients. - Opening the packaged executable without arguments starts the browser UI.
Scanning and queries
Concurrent scans
Scans started through the UI, CLI, MCP scan command, or lib/api.js now share a
per-project lock in .pathfinder/scan.lock. A scan waits while another scan of
that project runs; different projects can still scan independently. The CLI
reports when it is waiting. Locks are released on completion or failure, and
locks left by dead processes can be recovered.
Method names and module entries
A bare method name such as compact can resolve to a symbol labelled
Memory.compact. Exact labels and stable keys still work. Ambiguous names now
report candidate stable keys rather than only asking for a more precise target.
When a module has no recognized entry point, deep-dive summary/full views and
the AI brief use its most-connected symbol as a structural entry. Results label
the basis as most_connected; this is not proof of runtime invocation.
Diagnostics reserve the missing-entry finding for modules with no detected
symbol calls. Deep-dive cells also centre on a connected hub when the recognized
entry does not participate in the local call structure.
Descriptions and AI briefings
- Writing a node or edge description records
descriptionUpdatedAt; clearing
the description clears its date. Node description dates survive rescans. - The inspector displays description dates. The current-state report and AI
brief show dates beside the descriptions they include. Older descriptions
remain undated until rewritten. - Test and fixture paths are excluded from runtime module, route, service, and
connectivity summaries in the AI brief. A separate test-evidence section
keeps their presence visible; they remain in the graph and targeted queries. - Npm script listings identify the package file they came from.
- CommonJS export detection names public members from object assignments,
aliases, property assignments, and direct identifier exports instead of
reporting onlymodule.exports. Spreads and computed keys remain unresolved.
3D graph and placement
- Module depth is saved as
pos_zalongside X/Y in the project database. It now
survives switching between the CLI'slocalhostURL and the VS Code
extension's127.0.0.1URL. - Existing browser-stored depth is read as a migration fallback. Modules with
saved X/Y but no depth receive depth without moving their X/Y positions. - Reset Layout clears saved depth too. Position updates that omit Z preserve
the saved depth. - Symbol-cell call structures spread through a volume, with unconnected
symbols arranged around an outer shell. Deterministic depth offsets prevent
call trees from remaining flat during layout. - Planned modules created from the context menu retain the clicked position,
including depth in 3D, after reloading the live graph.
MCP integration
The server adds anthropic/alwaysLoad metadata to six read-only orientation
tools: brief, working set, callers, context, impact, and deep dive. This requests
upfront loading from clients that support that metadata; it does not establish
the behavior of every client. Other tools receive no such flag.
The MCP developer documentation explains this metadata and possible client
configuration overrides. Automatically writing a server-level alwaysLoad
setting from the integration toggle remains a proposal, not an implemented
feature.
Portable package and release workflow
- A packaged executable launched with no arguments now runs
serve --reuse,
opening the UI or reusing the current build. Explicit command-line commands
retain their behavior. Windows quick-start instructions include the.exe. - The Windows ZIP smoke test now extracts the archive, checks help, starts the
bundled server, and loads the project picker over HTTP. - The automated Windows build runs this smoke test before uploading its ZIP.
- The CLI's
--versionnow reads the package version instead of reporting
the hard-coded 0.1.0. - Since the v0.1.3 tag, the GitHub workflow has also changed dependency setup
and builds both macOS architectures in the macOS 14 job, replacing the
separate macOS 13 Intel job. This describes the workflow configuration;
it is not evidence of a successful hosted build.
Developer documentation and validation
- The known-issues tracker records resolved method lookups, entry reporting,
scan coordination, and description dates, with remaining scanner gaps kept
separate. - A new A/B benchmark protocol specifies fresh contexts, identical repository
states, and evidence collection for comparing exploration with and without
Pathfinder. No benchmark outcome is claimed here. - Added test coverage includes scan coordination and lock recovery, bare and
ambiguous method names, connectivity-based entries, CommonJS export names,
and Python scan/query behavior across repeated scans. - The full test suite passed on 2026-10-22: 89 tests, zero failures or skips.
- On 2026-10-22, the Windows package was rebuilt locally. The extracted ZIP
passed help and HTTP project-picker checks, and launching the executable
without arguments started the server. Other platforms were not rebuilt or
verified during this packaging fix.
Upgrade notes
Restart Pathfinder to load the changed code. The database adds the depth column
automatically when opened. Existing descriptions do not acquire historical
dates. Restart the AI client session to pick up changed MCP tool definitions.