codedb 0.2.5830

codedb 0.2.5830

codedb 0.2.5830 is a performance and correctness release for steady-state MCP and core read-only tools. It makes common code-intelligence operations substantially faster while preserving retrieval results, ordering, ranking, security policy, telemetry behavior, and client-visible MCP responses.

Release: https://github.com/justrach/codedb/releases/tag/v0.2.5830

What changed in 0.2.5830

Faster indexed retrieval

• Exact symbol lookup now uses the existing symbol hash index directly. Exact-name requests no longer scan the broader symbol path. Prefix, pattern, fuzzy, kind-only, incomplete-index fallback, result-cap, and deterministic-ordering behavior remain unchanged.
• Tree, outline, exact-word, and fuzzy-file results use bounded, generation-validated caches. Mutations advance the Explorer generation before changed or removed content can be served. Tree entries are capped at 16 MiB; outline and word rendering use 32-entry, 16 MiB LRUs with a 4 MiB per-entry ceiling; fuzzy scoring keeps at most 32 query/limit entries.
• Repeated and deep reads reuse cached content hashes and newline offsets. After the first offset-table build, ranged extraction scales with the requested range instead of every byte before it.
• Context symbol bodies use zero-copy, offset-based extraction when content is resident, with the previous disk fallback retained.
• Generic and incremental indexing use rolling trigram construction. Each incoming byte is loaded and normalized once while preserving case folding, whitespace suppression, location masks, following-character masks, and final-boundary behavior.
• Compact JSON-RPC responses use a one-slice copy fast path. Any payload containing raw CR/LF bytes still takes the original sanitizing fallback, preserving one-response-per-line MCP framing.

Cache correctness and concurrency

• Cache reads are synchronized with watcher mutation paths.
• Content-cache identity uses non-reused entry generations rather than allocator addresses.
• Cached fuzzy results are copied before request-specific boosts, so callers cannot mutate shared scores.
• Add, replace, remove, reparse, and generation-change tests verify that stale content and rendered output cannot be reused.

No ranking formula, result cap, parser behavior, path-security policy, sensitive-file filtering, telemetry behavior, or MCP response schema changed in this release.

Zig 0.16.0 → pinned Zig 0.17

The release line moved from Zig 0.16.0 to the exact tested snapshot 0.17.0-dev.813+2153f8143 in 0.2.5829, immediately before this performance release. codedb 0.2.5830 inherits and validates that migration across native, Windows, and freestanding/WASM targets.

The migration includes:

• a pinned compiler/package minimum and release jobs that verify the compiler version;
• Zig 0.17 build-graph passthrough and lazy-path APIs;
• migration from removed repetition, sentinel-formatting, writer, I/O, and allocator APIs;
• target-appropriate atomics and host-operation guards for Windows and WASM;
• borrowed process-lifetime argv instead of another command-parser copy;
• a vendored, licensed Zig-0.17-compatible nanoregex dependency;
• internal protocol-neutral JSON helpers replacing an otherwise incompatible server-framework dependency.

The migration line initially measured 1.20× faster cold CLI search, 1.14× faster warm snapshot-backed CLI search, and a 3.217× exact-word geometric-mean speedup against the Zig 0.16 build. The final 0.2.5830 generic initial-index comparison then measured 136.0 ms → 131.0 ms, a 3.7% improvement, with 20/20 paired wins and identical hit counts across all 14 query rows.

The detailed migration inventory and reusable upgrade guide are in docs/zig-0.17-migration.md.

Measured performance

Final release gate: 20 real paired runs

The final source-attribution gate used:

• production baseline v0.2.5829 (dd36e94), plus pinned harness-only parity backport 24e89c7;
• executable candidate bc72bfc;
• the same fixed 21-file corpus copied from the immutable baseline;
• pinned Zig 0.17.0-dev.813+2153f8143 for both binaries;
• 20 counterbalanced AB/BA pairs.

Every sample passed commit/tree, clean-worktree, compiler, corpus, pair, order, and sequence provenance checks. Every parity-enabled tool matched across every measured iteration after normalizing only response duration. No parity-enabled tool crossed the paired regression threshold of more than 10% and more than 50 µs.

Largest in-process paired-median latency reductions:

Operation	Improvement
Exact symbol	81.72%
Tree	67.60%
Outline	59.79%
Exact word	37.16%
Bundle	27.73%
Hot-file lookup	17.81%
Context	11.89%
Read	4.81%
Find	3.27%
Edit	2.04%

End-to-end MCP round trips

A separate immutable-corpus transport benchmark—including JSON-RPC serialization, stdio transport, parsing, dispatch, envelope generation, escaping, response writing, and client receipt—measured 2.16× to 99.11× speedups across tree, outline, symbol, read, find, word, search, and bundle workloads. Normalized responses were byte-identical.

MCP operation	Baseline	0.2.5830 candidate	Speedup
Tree	343.8 µs	40.1 µs	8.56×
Outline	995.0 µs	148.9 µs	6.68×
Symbol	229.4 µs	20.9 µs	10.96×
Read	189.3 µs	53.9 µs	3.51×
Find	5.52 ms	409.2 µs	13.50×
Word	13.83 ms	215.0 µs	64.32×
Search	72.1 µs	33.4 µs	2.16×
Bundle	13.36 ms	134.7 µs	99.11×

Large synthetic edge cases improved 12.8×–109.7× for huge outlines, deep reads, large trees, fuzzy-file lookup, and exact-word output; exact symbol lookup improved 22.4×.

These are workload-specific measurements. The largest cache-backed gains are steady-state gains; first calls still construct cached representations. Full methodology, limits, raw evidence, and per-tool results:

• docs/performance-0.2.5830.md
• docs/bench-0.2.5830-paired-report.md
• attached bench-0.2.5830-paired-samples.tar.gz, SHA-256 104edc6875a9d121d2e5a4c1e69c12735f2c60128e48ae384aa8c6cfb950bc24

Token and deployment efficiency snapshot

At release time, the Codegraff dashboard reported:

• 19.7 billion tokens saved over the previous 30 days;
• 580,000 codedb operations over the previous 7 days;
• 63 µs p50 per codedb operation;
• approximately 47 tokens per outline and 14 tokens per lookup.

These are aggregate deployment/dashboard counters from activated installs, not a causal A/B measurement of 0.2.5830. They describe codedb's broader context-efficiency footprint; the version-specific performance claims above come from the immutable paired benchmarks.

Benchmark-gate hardening

The release also makes future performance claims harder to get wrong:

• baseline and candidate run from clean throwaway worktrees;
• the runner binds evidence to expected source commit and tree identities;
• both sides receive the same fixed corpus copied from the baseline;
• corpus fingerprints and compiler executable identities are recorded per sample;
• AB/BA order, side, pair number, and sequence are validated;
• full normalized JSON-RPC responses are hashed across every measured iteration;
• parity exemptions require both case metadata and an explicit comparator allowlist;
• duplicate tool records, missing pairs, dirty worktrees, and mismatched provenance fail closed;
• regressions use paired medians and deterministic bootstrap intervals rather than single-run minima.

Verification

• zig build test
• native ReleaseFast build
• Windows x86_64 ReleaseFast cross-build
• WASM build
• MCP E2E: 20/20 scenarios passed
• paired comparator tests: 12/12 passed
• final local release gate: 20/20 pairs passed
• GitHub paired benchmark workflow: passed
• git diff --check

Release assets

Published binaries:

• macOS arm64
• macOS x86_64
• Linux arm64
• Linux x86_64
• Windows x86_64

The shipped macOS arm64 binary is signed with hardened runtime and a secure timestamp. Apple notary service returned Accepted, and Gatekeeper reports source=Notarized Developer ID for the exact shipped binary.

• arm64 notarization submission: de536fc8-44ae-41dc-9826-b1f7847dcfe2

The x86_64 macOS binary is intentionally unsigned because post-link signing reproduces the Rosetta startup crash tracked in #504. The exact shipped unsigned binary was executed successfully under Rosetta as codedb 0.2.5830. A signed x86_64 candidate was notarized but failed that runtime check and was not shipped.

Install or update:

curl -fsSL https://codedb.codegraff.com/install.sh | bash

Or through npm:

npx -y codedeebee mcp

SHA-256

2b786d7404cf8ea79857dcd3a067d35493e67d3eb523ea1eb5810103dec048cd  codedb-darwin-arm64
514424656a3c632315c90d08ea00349ba879327d70d0e997c435136a74129fbd  codedb-linux-x86_64
abca6aceb29c8665c3ac18071b6e0260ebe8419de2d5aa0c1473f07e6284cf42  codedb-linux-arm64
efbff83e4482cd441dc3ef89871603e836f7e5d700381815e616803ed7b4a79a  codedb-darwin-x86_64
fbb106ca1cc084cb514994a937d1508f3a9785e0f44964fd5af9f09bf94f662e  codedb-windows-x86_64.exe

Verify downloads with the attached checksums.sha256.