Skip to content

ROADMAP.md

Latest commit

 

History

History
140 lines (117 loc) · 6.99 KB

ROADMAP.md

File metadata and controls

140 lines (117 loc) · 6.99 KB
Raw

Roadmap

Where Konnect is going, in the order the work actually depends on itself. No dates — items ship when they survive verification. Opening an issue is the best way to influence priority; several of the largest items below exist because a contributor measured something we hadn't.

This revision (post-v0.7.0) is built in large part on @neusse's improvement backlog and his benchmark work (discussions #239, #295, #224) — hands-on, evidence-first evaluations that found what CI could not. If you want to work on any item here, say so on its issue; design agreement first, then one focused PR.

1. Truth and write-safety (the standing doctrine)

Every response field derives from the result, never from the request. A verdict without evidence is INCOMPLETE, never approval. Fixtures come from real KiCAD output. These rules are enforced by tests (doc_tool_counts.rs, schema_parameter_usage.rs, schema_migrations.rs) and documented in docs/DEVELOPING_TOOLS.md; every new tool must satisfy them.

Open work in this theme:

  • Stale-board detection (#240) — after KiCAD dies, the next closed-board call cannot tell that from "KiCAD was never running." Lock files become supporting evidence, not sole authority. Depends on a document-answering IPC mock (#241) so refusal branches finally get shared test coverage.
  • Project and library-table targeting (#189) — ambiguity in the ancestor walk fails with a structured error instead of guessing an unrelated project.
  • Server lifecycle ownership (#103) and a non-mutating startup (#242) — a server process must not reinstall configuration as a side effect of launching.

2. Schematic correctness

  • Multi-unit symbols (#182) — the read half now flows through the shared ConnectivityIndex (#323); the mutation half (#273) lands next after a rebase onto the index.
  • Connectivity on move and delete (#120) — junction creation/pruning when things move. This is also the prerequisite for a real move_connected (#315), which currently refuses rather than pretending.
  • Remaining indentation-sensitive scans (#84) — replace the last literal-string probes with structural parsing.

3. Autorouting: a real Freerouting bridge

The single biggest capability gap. KiCad 10's CLI never had Specctra DSN/SES, so Konnect's old autoroute could only fail; it was removed rather than faked (#253).

The agreed direction is @neusse's design from #253: drive a standalone Freerouting JAR directly — DSN export, route, SES import, editor refresh — with discovery reporting engine found and bridge available as separate facts, and PCM-installed JARs discovered on every platform. Board identity validation and atomic SES import are requirements, not options. Design is agreed; implementation is open for whoever claims it on #253.

4. Client compatibility

The dynamic-toolset architecture (small starter kit, load_toolset on demand) assumes MCP clients either expose many tools or re-fetch after tools/list_changed. Two reports show real clients violating both assumptions: VS Code Copilot caps callable tools (#325) and a Linux client never re-fetched (#233).

Planned: a compact tool surface mode (a few generic call/help tools proxying the full catalog — credit @simachines' working prototype in #325), MCP resources/ exposure of the tool directory, and documenting eager_toolsets per client. The 200-tool catalog is a feature; falling over on real clients is not.

5. Workflow and skills layer

@neusse's benchmark-driven proposals in #295: an evidence hierarchy the bundled skills teach (KiCad's own DRC/ERC outrank Konnect's own summaries), single-owner rules for the live board, Freerouting-first PCB flow, and a physical pin-map requirement before custom-part creation. Gated on the tool surface stabilizing (v0.8.0); invited as skill-layer PRs.

6. Platform and KiCad-forward

  • KiCad 11 readiness (#257) — the SWIG pcbnew bindings the legacy ActionPlugin uses are removed in KiCad 11; the executable IPC plugin is the replacement path and needs its lifecycle/settings parity planned before KiCad 11 ships. The one item here with a deadline.
  • macOS signing/notarization (#131), then Homebrew (#154) once artifact identity is stable.
  • Per-user Windows installs (#254) — discovery must find %LOCALAPPDATA%\Programs\KiCad.
  • KiCAD PCM publication — submit to the official addon repository.

7. The quality flywheel

What keeps the rest honest:

  • End-to-end benchmarks against reality. @neusse's konnect-codex runs found silent corruption, fail-open verdicts, and contract gaps that 900+ unit tests missed. A re-run against each minor release is part of the release rhythm now, and the standing invitation is open.
  • Live-test honesty (#221) — fix the stale CI claims and the rotation-readback flake.
  • A real-KiCAD fixture corpus — boards and footprints saved by KiCAD itself (tabs, CRLF, ki_fp_filters, positionless properties), so no fixture can share the code's wrong assumption again.
  • The catalogue-wide guards stay: doc counts, schema-parameter usage, and schema migrations are all swept by CI on every PR.

Done (eras, not items)

  • v0.1–v0.3 — full toolset surface across schematic, PCB, library, export; three-platform PCM packaging; HTTP transport; hierarchical sheets.
  • v0.4–v0.5 — atomic update_pcb_from_schematic; footprint graphics authoring and board-side editing; client-scoped installs.
  • v0.6–v0.7 — the truth-and-safety arc: required-argument enforcement, the layer-crash fix, sync corruption fixed with post-apply read-back, DRC-evidence-gated review verdicts, honest autoroute removal, and the developer docs set (docs/).