Skip to content

CHANGELOG.md

Latest commit

 

History

History
119 lines (80 loc) · 9.03 KB

CHANGELOG.md

File metadata and controls

119 lines (80 loc) · 9.03 KB
Raw

Changelog

All notable changes to state-trace are documented here. The format roughly follows Keep a Changelog, and the project adheres to Semantic Versioning (with the usual pre-1.0 caveat that minor bumps may include breaking changes).

Unreleased

Changed

  • Clarified MCP install docs to mirror Geometra's per-client syntax, with simple Codex install first and the absolute-path command labeled as a PATH-safe fallback.

0.3.3 — 2026-04-25

Added

  • state-trace-mcp-config helper command, plus python -m state_trace.mcp_config, to generate project-scoped .mcp.json with an absolute installed state-trace-mcp command path. This makes Codex and other MCP clients more reliable when they launch servers with a minimal PATH.

Changed

  • README MCP install docs now separate official PyPI installs from editable source installs and lead with per-repo Codex .mcp.json setup.
  • server.json package metadata now matches the current 0.3.3 PyPI version.

0.3.2 — 2026-04-25

Fixed

  • iso-trace ingestion: O(N²) auto-link blowup. extraction._extract_edge_candidates ran a per-node Jaccard scan against every existing node, generating ~70 spurious depends_on/related_to edges per node when fed an iso-trace transcript. The trajectory-source carve-out that disables this for swe-agent and openhands was missing iso_trace. Verified on a real Claude Code session (179 turns, 977 nodes): edges 76,123 → 10,578 (-86%). A 1,467-turn session that previously OOM'd at 4 GB during ingestion now completes cleanly (6,411 nodes / 61k edges) and retrieval correctly surfaces the expected files.

Performance

  • retrieve_brief ~270× faster on large graphs. On the 6,411-node / 61,017-edge JobForgePersonal session: per-query wall time dropped from ~250 s to ~1 s warm-cache. Five behavior-preserving changes:
    • graph.find_edges uses networkx's indexed out_edges / in_edges when an endpoint is pinned, instead of a full O(E) scan. _supersession_bonus was calling it twice per scored node — ~615 M edge tuple comparisons per retrieve.
    • retrieval._build_causal_chains builds a lightweight nx.Graph topology instead of graph.to_undirected() (which deep-copies every Pydantic Node/Edge, ~17 s on a 61k-edge graph).
    • retrieval._seed_nodes memoizes tokenize_sequence(_document_text(node)) per (node.id, len(content)) — node BM25 tokens are deterministic, so recomputing on every retrieve was wasted work.
    • graph.namespace_view returns self when every node already lives in the requested namespace, instead of rebuilding the full graph (~1.5 s on a 6.4k-node single-namespace engine).
    • _artifact_bridge, _edge_semantic_bonus, _provenance_bonus, _supersession_bonus are now computed in single O(E) edge sweeps that update both endpoints in one pass, replacing 2×|scored| per-node neighbor walks. Cuts get_neighbors calls from ~5,800 to ~830 per retrieve.

All 52 tests pass. No public API changes.

0.3.1 — 2026-04-25

Added

  • mcp-name verification comment in README.md and server.json manifest at the repo root, in preparation for publishing to the MCP Registry. No code changes.
  • CHANGELOG.md (this file).

0.3.0 — 2026-04-24

Added

  • Module-to-path translator in retrieve_brief's lexical fallback. Dotted Python module references (astropy.modeling.separable_matrix) now resolve to file path candidates (astropy/modeling/separable.py), with conservative filtering on capitalized segments. Pushed SWE-bench-Verified A@1 0.216 → 0.254 and A@5 0.322 → 0.376 at n=500.
  • iso-trace ingestion adapter at state_trace/iso_trace_adapter.py. Reads @razroo/iso-trace's normalized Session → Turn → Event[] JSON and converts it to state-trace's agent_log format. Lets users seed working memory with accumulated Claude Code / Cursor / Codex / opencode session history. Two tests cover the pure transform and end-to-end ingestion.
  • GitHub / GitLab blob URL extraction in the lexical fallback. Issue text containing URLs like github.com/.../blob/main/astropy/io/ascii/html.py now contributes the embedded path to file candidates.

Changed

  • README headline benchmark: state_trace now leads every baseline (BM25 and Graphiti) with non-overlapping 95% CIs across both A@1 and A@5 at n=500.
  • README adds a Live solve-rate section reporting the n=20 docker-harness result (35% both arms — see "Honest finding" below).

Honest finding

  • Real solve-rate via swebench docker harness at n=20: state_trace 7/20, no_memory 7/20 (35% each). The two arms resolve different subsets (5 overlap, 2 unique each, union 9/20 = 45%). state_trace's localization wins do not translate to aggregate solve-rate wins when paired with Codex CLI as the downstream agent — Codex is already near-ceiling at file localization on its own. Published as a negative result rather than buried.

0.2.1 — 2026-04-23

Fixed

  • retrieve_brief.target_files no longer collapses to a single candidate. Three changes together:
    • target_files[:4] → target_files[:5].
    • Lexical fallback when fewer than 5 file nodes exist in the graph: mines candidates from the query, top-scored node issue_text metadata, and embedded GitHub blob URLs.
    • The fit_agent_brief budget loop's floor for target_files raised from 1 to 3, so the budget squeeze can't strip Artifact@5 candidates back down to a single file.
  • The pre-fix collapse made A@5 ≡ A@1 by construction. Fix moved n=500 A@1 from 0.150 → 0.216 and A@5 from 0.150 → 0.322.

0.2.0 — 2026-04-23

Added

  • MemoryEngine.current_state(session) — direct query for the live working-memory view (active task, latest observation/decision, non-stale active files). Filters out invalidated and superseded nodes.
  • MemoryEngine.failed_hypotheses(session) — direct query for "do not propose again" memories. Four sources: invalid_at set, status="error" without later validation, nodes pointed at by supersedes edges, and (added in this release) nodes with metadata.rejected_angle=True.
  • MCP tools for both: current_state, failed_hypotheses exposed via state-trace-mcp.
  • Graphiti-on-Kuzu backend in examples/swebench_verified_eval.py for an honest head-to-head at n=500 with deterministic stub embedder/reranker.
  • examples/swebench_verified_solve_rate.py scaffold for live solve-rate evaluation.
  • ARCHITECTURE.md explaining the networkx + SQLite design choice vs Neo4j/Kuzu.
  • BENCHMARKS.md translating every metric into real-world terms.
  • vs-Graphiti README section with 13-row head-to-head table backed by benchmark numbers or concrete API calls.

Fixed (dogfood-surfaced)

  • retrieve_brief no longer forces patch_file for non-file-oriented queries.
  • Top two evidence lines now get a wider character budget so load-bearing detail (numbers, paths) survives small-model compaction.
  • failed_hypotheses recognizes metadata.rejected_angle=True as a fourth failure source.
  • current_state.latest_observation sorts by actual recency (event_at, created_at, step_index) instead of step_index alone — fixed a bug where the oldest write surfaced as "latest" when multiple observations shared step_index=0.

Benchmarks

  • SWE-bench-Verified localization at n=500 (no_memory, BM25, state_trace, graphiti). state_trace beats Graphiti on Artifact@1 with non-overlapping CIs in this release; ties BM25 with overlapping CIs.

0.1.0 — 2026-04-23

Added

  • Initial PyPI release.
  • MemoryEngine core API: store, retrieve, retrieve_brief, record_action, record_observation, record_test_result, ingest_text, store_agent_log[_file], step.
  • GraphManager over networkx.MultiDiGraph with namespace-view filtering.
  • Hybrid retrieval (BM25-style lexical seed + edge-aware graph traversal + state-aware reranking).
  • Bounded capacity with decay, compression, and lifecycle retention.
  • WAL SQLite + FTS5 backend for durable working memory; auto-selected when storage_path ends in .db / .sqlite / .sqlite3.
  • Stdio MCP server (state-trace-mcp entry point) shipping in the [mcp] extra.
  • LangGraph and LlamaIndex framework adapters that don't import the host frameworks.
  • FastAPI surface with /store, /retrieve, /retrieve_brief, /graph endpoints.
  • Repo-local benchmark suite: held-out live, offline retrieval, Graphiti head-to-head, budgeted harness proxy, live harness replay, long-horizon memory pressure.