v0.1.11 — import-aware call links, gitignore support

Upgrading: existing indexes are rebuilt once automatically (index format 2).

Fixed

• Call links follow imports (Python). Cross-file calls were linked by name alone, so
  asyncio.run() could link to your own run, str() to a project str, and a method call to
  a field of the same name. Calls now follow the caller's imports (including package
  re-exports and function-local imports); builtins and third-party calls link to nothing; calls
  on objects link only when there is exactly one candidate in a file the caller imports. On a
  new cross-file benchmark, precision went from 0.68-0.96 to 1.00 on requests, flask, httpx
  and CodePrism, with recall equal or better. Other languages still resolve by name. (#33)
• Gitignored files are no longer indexed (vendored checkouts, build output). Uses git's own
  file list, with a plain-walk fallback outside git. Opt out with respect_gitignore = false.
  (#31)
• Ambiguous paths are reported, not "not indexed". A short path matching several files
  (utils.py) now returns the candidate paths and a hint. (#32)

Added

• benchmarks/run_call_precision.py: cross-file call accuracy against a Python ast oracle.

Install

pip install -U codeprism-ai==0.1.11

v0.1.10 — tools accept project-relative paths

Fixed

• Tools now accept project-relative file paths. Files are stored with absolute paths, so
  get_callers("src/app.py", ...), get_module_summary, get_context, search_symbol's
  project_path and the other file tools silently returned empty results for the relative
  paths agents normally use. Relative paths, ./, either separator and (on Windows) any case
  now resolve against the served project root; an ambiguous name (two utils.py) is never
  guessed. (#30)
• get_callers / get_callees report errors for an unindexed file or unknown symbol instead
  of an empty list, and include each caller's file path, not just its internal id. (#30)
• No duplicate file records: codeprism index . stored relative paths while the MCP server
  stored absolute ones. Paths are now always absolute; an old relative index heals itself on the
  next index run. (#30)

Install

pip install -U codeprism-ai==0.1.10

v0.1.9 — graph sync & scale, working Claude Code setup, Codex + AGENTS.md

Graph sync and scale fixes, a working Claude Code setup, Codex support, and
AGENTS.md as the shared agent guide.

Upgrading: existing indexes are rebuilt automatically on the next codeprism index
or server start (new index format version), which also clears stale rows left by the
re-index bugs fixed below. Re-run codeprism setup <agent> to move to the new config
locations and get AGENTS.md.

Added

• codeprism setup codex: registers [mcp_servers.codeprism] in .codex/config.toml
  (or ~/.codex/config.toml with --global). Comments and other settings are preserved.
• AGENTS.md is the shared guide. Every setup writes the CodePrism usage guide to
  AGENTS.md (read by Codex, Cursor, Windsurf, Zed, Copilot, ...). CLAUDE.md becomes a thin
  @AGENTS.md import. Continue.dev gets .continue/rules/codeprism.md. (#27)
• One user-level config for every project. codeprism serve without a path serves the
  project containing the current directory (nearest .git / .codeprism.toml) and builds or
  refreshes its index in the background at startup. setup claude|codex --global use it.
  get_graph_stats reports index_status. (#29)
• Parallel parsing. codeprism index and the MCP server parse in a process pool
  (--workers/-j, parse_workers config), and SQLite is tuned for bulk writes.
  Full index of a 2M LOC corpus: 89.6s → 42.2s. (#24)
• Index format version. Indexes from an older format are fully re-parsed once instead of
  being trusted by checksum. (#25)

Fixed

• Claude Code setup never actually registered the server. MCP entries were written to
  .claude/settings.json, which Claude Code doesn't read. Now .mcp.json (project, pre-approved
  in settings.local.json) or ~/.claude.json (--global); stale entries are migrated, and
  setup files go to --project instead of the current directory. (#26)
• Java, C, C++, Ruby and PHP files were never indexed, and the watcher ignored them (and
  .rs). All 10 languages are now indexed and watched by default. (#20)
• Re-indexing left stale data: renamed symbols lingered and edges were duplicated when line
  numbers shifted; --force didn't remove deleted files. (#21)
• update_file dropped callers from other files from the in-memory graph. (#22)
• Single-file updates did whole-repo work: 2.6s → ~55ms at 2M LOC. (#28)
• search_symbol kind filter was case-sensitive; kind="Function" returned nothing.
  Thanks @MilindLate. (#19, fixes #18)

Install

pip install -U codeprism-ai==0.1.9

v0.1.8 — Java, C/C++, Ruby, PHP parsers (10-language support)

What's new in v0.1.8

10-language parser support

CodePrism now indexes 10 languages out of the box:

Parser	Languages	New in this release
Java (L1)	.java	✅ Classes, interfaces, methods, annotations, generics
C / C++ (L2)	.c, .h, .cpp, .cc, .cxx, .hpp, .hh	✅ Functions, structs, classes, inheritance, namespaces
Ruby (L3)	.rb, .rake, .gemspec	✅ Modules, classes, instance/singleton methods, visibility
PHP (L4)	.php, .php5, .phtml	✅ Namespaces, classes, traits, interfaces, methods

Previously supported: Python, JavaScript, TypeScript, Go, Rust.

Documentation

• README Supported Languages table updated (all 10 languages)
• Mermaid benchmark charts added to README and docs/benchmark-results.md
• INTEGRATIONS.md now has a Supported Languages section with extensions

Install

pip install codeprism-ai==0.1.8

Full changelog

See docs/changelog.md.

v0.1.7 — Corpus expansion + get_dependencies fix

What's new

Added

• Benchmark corpus expansion: pallets/flask 3.0.3 and encode/httpx 0.27.2 (10 tasks each). Token reduction: flask 91.3%, httpx 93.0%. Overall average across 3 production codebases: 91%.

Fixed

• get_dependencies output format: was returning imported symbol names instead of source modules. Root cause: source_module stored in extra{} (excluded from SQLite). Fix: store in the signature field (persisted). Re-index required for updated output.
• Benchmark ground truths for requests_004, requests_006, requests_008: updated to match actual tool output.

Benchmark summary

Corpus	Baseline	CodePrism	Reduction
psf/requests v2.32.3	6,407 tokens	738 tokens	88.5%
pallets/flask 3.0.3	9,558 tokens	828 tokens	91.3%
encode/httpx 0.27.2	12,685 tokens	894 tokens	93.0%

Full results: docs/benchmark-results.md

v0.1.6 - Level 2 latency benchmark + get_dependencies fix

What's new in v0.1.6

Level 2 latency benchmark

• benchmarks/run_latency_benchmark.py: measures p50/p95/p99 query latency with warmup runs
• Engine held open across all tasks — times pure query latency, not indexing overhead
• Results on psf/requests v2.32.3 (20 reps, 3 warmup discarded):
  • get_context p50=1.2ms, p95=1.8ms
  • get_callers p50=1.2ms, p95=1.8ms
  • get_impact p50=2.0ms, p95=3.4ms
  • get_dependencies p50=3.8ms, p95=5.1ms (after fix)
  • Overall p50=1.9ms, p95=2.5ms

All queries under 5ms p95 — noise relative to LLM API latency (500-3000ms).

Performance fix: get_dependencies

• Was doing SELECT * FROM symbols (full table scan) on every call
• Fixed with SELECT DISTINCT name WHERE kind != 'import' (name-only projection)
• Latency dropped from 10.7ms to 3.8ms p50 on the requests corpus (64% faster)
• All 341 tests pass

v0.1.5 - Benchmark harness + cross-file call edge fix + docs

What's new in v0.1.5

Indexer fix

• Cross-file call edges now resolve correctly. BaseParser.resolve_intrafile_refs was consuming CALLS refs that matched import stubs, preventing them from reaching cross-file resolution. Fixed by skipping CALLS-to-import-stub wiring intra-file. Impact: run_payment -> compute_checksum and all similar cross-file calls now appear correctly in get_callers().

Level 1 benchmark harness

• Token reduction measurement via tiktoken (no API key required)
• LLM-as-judge accuracy eval via Ollama cloud (gpt-oss:120b)
• Fixture corpus (10 tasks) + psf/requests v2.32.3 corpus (10 tasks)
• CI workflow: runs on every push/PR, token-only, results as build artifact

Measured results

• 88.5% token reduction on psf/requests v2.32.3 (real-world, 6k-12k token files)
• Accuracy: baseline 0.81 / CodePrism 0.66 on requests corpus

Documentation

• docs/benchmark-results.md - full methodology and all run results
• docs/architecture.md - 5-layer system design and data flow
• docs/changelog.md - version history
• README: Benchmarks section + Documentation table

Run the benchmark yourself

pip install -e ".[bench]" tree-sitter-python
python -m benchmarks.setup_repos
python -m benchmarks.run_token_benchmark --tasks benchmarks/tasks/requests_tasks.json

v0.1.4 - Property-based tests (Phase B)

Phase B: Hypothesis Property-Based Tests

341 tests total (up from 324). Added 17 property-based tests using Hypothesis that assert invariants hold for any valid input.

Coverage

• SecurityScanner: status always in {PASS,WARN,BLOCK}; BLOCK iff BLOCK issue present; identical diff always PASS
• SecretsDetector: any api_key/password pattern fires; all findings carry fix_suggestion
• InjectionDetector: f-string SQL always flagged; severity always BLOCK or WARN
• WeakCryptoDetector: findings are always WARN severity
• GraphEngine: N unique symbols produce N nodes; removal makes node unreachable; edge roundtrip; caller/callee symmetry
• SearchMatch: score always in [0.0, 1.0]

v0.1.3 - Semantic search via embeddings (Phase A)

Phase A: Semantic Search

search_symbol now supports embedding-based semantic search.

New features

• codeprism index --embeddings: builds ChromaDB vector index alongside the graph
• index_project MCP tool: new embeddings param
• search_symbol finds symbols by meaning, not just name substring
• Falls back to substring search silently if embeddings not built
• enable_embeddings = true in .codeprism.toml wires semantic search into MCP server

Usage

pip install 'codeprism-ai[embeddings]'
codeprism index /path/to/project --embeddings

v0.1.2 — Spec compliance fixes

What's fixed

• search_symbol: added project_path param to scope results by directory prefix
• get_graph_stats: path filter now actually filters stats instead of being silently ignored
• get_dependencies: circular_deps[] is now populated via NetworkX cycle detection
• pyproject.toml: bumped version, fixed author email
• CODE_OF_CONDUCT / CONTRIBUTING: updated contact email