Lataa Termpolis 1.20.0

🧭 A native code graph — plus the memory brain's next rung
v1.20.0 gives agents a structural map of your codebase and closes out the memory-hardening roadmap's "next rung." Everything is native, local, and native-free (no tree-sitter, no external services), test-first.
New: a native code graph (CodeGraph-like, built in — no dependency)
Agents no longer have to grep the same files over and over to understand structure. Termpolis now indexes your repo into a symbol + call graph and exposes it over MCP:
- code_explore — ask one question ("how does X reach Y", or just a symbol name) and get the relevant symbol's source plus its direct callers and callees.
- code_callers / code_callees — who uses this / what does this depend on.
- code_impact — the transitive blast radius of a change, so you see everything a shared function edit could break before you touch it.
- code_search — locate any symbol by name across the whole codebase.
It's a heuristic multi-language extractor (TypeScript/JavaScript, Python, Go, Rust, Java) over a pre-indexed, incrementally-updated local graph — built entirely in-house, secret-denylisted (never graphs .env/keys), and it (re)builds alongside the existing code index. Structural "what/where" now complements the memory brain's semantic "roughly where is X."
The memory brain's next rung
- Shard compaction — the append-only memory log no longer grows unbounded; a lossless, atomic, round-trip-proven compaction reclaims dead lines while preserving every live memory, deletion, and usage count (and it aborts rather than risk data it can't read).
- Honest embedder status — the status is now a truthful ready / unavailable / unprobed tri-state instead of implying "healthy" before the first embed, warmed by an off-thread probe that never stalls typing.
- Cross-agent conflict detection can now use a real NLI model (opt-in, off by default until it's benchmark-proven to beat the conservative heuristic) — catching contradictions like "use Postgres" vs "use MySQL" that have no negation word.
- Richer recall evaluation — order-sensitive nDCG and per-scenario slices, so future learning changes are judged on whether they actually help, not a single averaged number.
- Production burn-in log — a device-local, bounded record of anything the brain finds unusual (degraded init, corrupt shard lines, …), so real-world use leaves a reviewable trail.
Quality
Every piece ships with failing-test-first regression guards, including a strict lossless round-trip proof for compaction. Full suite green; lint + typecheck + build clean; coverage gate green across Windows, macOS, and Linux.
Full Changelog: https://github.com/codedev-david/termpolis/compare/v1.19.5...v1.20.0