Repository navigation
Changelog
All notable changes to Code-Index-MCP will be documented in this file.
The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.
[Unreleased]
[Unreleased]: # placeholder for post-rc5 work
[1.6.0] - 2026-11-09
Prepared minor release: the port to the mcp 2.x SDK (Code-Index-MCP#124,
with #128, #131 and #127 in #132 and its re-check fixes in #135), so
index-it-mcp co-installs with PMCP 3.x; the watcher sweep, incremental sync
and generation-retention fixes (Code-Index-MCP#114 and #111 in #129, follow-ups
in #130). This section does not imply publication; protected-main dispatch and
independent delivered-artifact acceptance are separate gates.
Changed
Port to the mcp 2.x SDK (#124)
-
The STDIO server runs on the
mcp2.x SDK. The requirement is now
mcp>=2.2,<2.4(locked at 2.3.0), soindex-it-mcpinstalls in the same
environment as PMCP 3.x (mcp>=2.2,<3) andphase-loop-runtime==0.7.30
with a cleanuv pip check.jsonschemais now a direct dependency. -
The cap below 2.4 is deliberate. The Tasks runtime depends on the SDK's
Server.middleware, which mcp marks provisional ("may change in a 2.x minor
release"). Each new mcp minor must be re-tested before the cap is widened.
At startup the server also checks the middleware contract it needs and fails
with an actionable message if an installed mcp lacks it. -
Clients see the same tools, input and output schemas, results, error codes
(except for methods the server does not serve, below) and readiness-aware
index_unavailableresponses. The port keeps two
contracts the 1.x SDK'scall_tooldecorator provided and the 2.x lowlevel
server dropped: arguments are validated against each tool'sinputSchema
(Input validation error: ...results), and a failing tool returns an
isErrorresult instead of a JSON-RPC error. -
The 2.x SDK removed its experimental Tasks runtime. A repo-owned runtime keeps
task-augmentedreindexandwrite_summaries, thetaskscapability, and
tasks/get,tasks/result,tasks/listandtasks/cancel, with the 1.x
results and error codes, on connections negotiated at a handshake-era
protocol revision (2025-11-25 and earlier). -
Clients that probe
server/discovernow negotiate protocol revision
2026-07-28 instead of falling back toinitializeat 2025-11-25. Claude
Code 2.1.x does this. On that revision, which has no Tasks,tools/list
omitsexecution, task metadata ontools/callis ignored (the call runs
synchronously),tasks/*answer-32601, and results addresultTypeand
aserverInfo_metastamp. The official Python SDK'sClientSession, the
1.x SDK client and the PMCP 3.x gateway negotiate 2025-11-25. -
Every tool's
outputSchemanow carries the top-level"type": "object"the
MCP schema requires and the 2.x SDK enforces. Each schema'soneOfbranches
were already objects, so the same results validate. -
Requests for methods the server does not serve now answer
-32601(Method
not found) with the method name indata, per JSON-RPC; the 1.x SDK
answered unknown methods with-32602. -
The PMCP pilot's client uses
streamable_http_clientwith anhttpx2
client, because the 2.x SDK removedstreamablehttp_client. -
Task retention is bounded. A task's
ttlis clamped to between one minute
and one hour, a task without one gets one hour, and the effective value is
returned. The one-minute floor means atasks/resultwoken by the completion
always finds the result, even forttl: 0. Expiry starts when a task
finishes, so a running task is never purged. At most 256 tasks are kept: the
oldest finished tasks are evicted first, but never a task that still has a
tasks/resultwaiter or that finished within the last minute. When every
slot holds a running, awaited or just-finished task, a new task is refused,
and the client gets that reason ("Too many tasks (256 running or awaiting
retrieval); retry after one finishes") rather than an opaque
"details": "ValueError". (#132 re-check) -
Concurrent
tasks/resultrequests for the same task are all answered when
the task finishes. Previously, as in the 1.x SDK runtime, a second request
left the first one waiting forever. Deleting or expiring a task now wakes its
waiters, which get-32602. -
On the POSIX STDIO transport, a line that is not a valid JSON-RPC message
gets a JSON-RPC error instead of no reply (the 2.x transport drops it
silently):-32700(Parse error,idnull) for invalid JSON, and-32600
(Invalid Request) for other JSON, echoing a string or integerid. A
malformed response (resultorerrorwith nomethod) is dropped with a
debug log instead, because responses are never answered and itsidbelongs
to the server's own requests. Blank lines are ignored. A warning logs only the
size and the error code. On Windows the SDK's own STDIO reader is used, and
malformed lines are still dropped silently. -
On protocol 2026-07-28, which has no server-to-client requests, summarization
no longer attempts MCP sampling for every chunk. It logs once and uses the
next summarization path.
Fixed
Watcher sweep and generation retention (#114)
- The watcher sweep counted files the indexer never stores (disabled-by-default
languages such as YAML, shell and JSON, and unsupported extensions) as
"created" drift. On this repository that was 112 files, so every sweep
(default 60 minutes) forced a full rebuild that could not clear the drift,
left readiness atmissing_provenanceduring each rebuild, and published a
new generation each time. The sweep, the full index walk and single-file
indexing now share one admission predicate
(mcp_server/indexing/admission.py,EnhancedDispatcher.index_admission_reason):
the same extension, plugin-enablement and exclusion rules. Right after a full
rebuild the sweep reports no drift. Rows a rebuild deliberately retains
(imported documents and history) are no longer reported as deleted. - Drift that survives a successful reconciliation is not reconciled again
until the tracked tree or the drift itself changes; the sweep logs it once.
A reconciliation that failed, was refused or raised is retried on the next
sweep, then with exponential backoff capped at 16 sweeps. - Bounded generation retention. Each publication and each failed rebuild is
recorded ingenerations/generation-history.json, and only recorded
generations can be pruned: after a publication, recorded generations beyond
the live one plus theMCP_INDEX_GENERATION_RETENTION(default 2, minimum 1)
most recently live previous generations, and recorded failed generations
beyond the newest one. Never removed: generations without a record
(everything from before the upgrade, or after a lost history file), any
generation a registry entry references, generations pinned with a
generations/<id>.keepfile, and any generation a reader may still hold (its
SQLite sidecars exist, or a process has it open). The semantic directory and
sidecars are removed before the database, and the history file is fsynced.
Quarantine copies and server-side Qdrant collections are not touched. - Sandboxed plugin workers kept per-file in-memory state keyed by each
rebuild's staging path, which grew by about 11 MB per rebuild of a 29-file
fixture. A staged build now releases the repository's plugin workers while
queries are still fenced; they restart on next use.
Incremental sync of non-indexable files (#111)
- Incremental sync failed with
Staged rebuild failed (RuntimeError)when a
commit touched a file in a language with no enabled plugin, such as a
workflow.yml. Incremental sync now asks the shared admission predicate,
as the full walk does: files it rejects are ignored, an indexed file that is
no longer admitted is deleted, and a rename to a non-indexable path (for
examplenotes.mdtonotes.txt) deletes the old row. The result matches a
full rebuild. A file the predicate admits whose plugin fails to resolve (for
example a sandbox worker that does not start) is a failure in both the
incremental and the full path, so the previous generation is kept. - A failed staged rebuild's stored
last_sync_errorand therepository sync
output now include the message of the rebuild's own refusal and a bounded
summary of the per-file errors (at most five, checkout and staging paths
removed). Text from other exceptions is still reported by type only. - The shared secret redaction used for these errors and rerank diagnostics now
also covers GitHub tokens (ghp_,gho_,ghu_,ghs_,ghr_,
github_pat_), AWS access key ids, Slack tokens, PEM private keys, URL
credentials and*secret*/*token*/*password*key-value pairs.
serverInfo.version (#128)
- The STDIO
initializereply'sserverInfo.versionis theindex-it-mcp
version (mcp_server.__version__), not themcplibrary version. The same
version is stamped on 2026-07-28 results.
Grammar prefetch retries (#131)
mcp-index grammars prefetchretries transient download failures from the
language pack's GitHub release assets: HTTP 5xx and 429, connection resets
and timeouts. It makes at most 5 attempts, with equal-jitter exponential
backoff (4 s doubling, capped at 60 s); no retry starts that would sleep past
a 120 s budget. Refused connections, DNS lookup failures and HTTP 408 are also
retried. Other HTTP 4xx such as 404, checksum mismatches and unrecognized
failures fail on the first attempt. Retries and the final error name only the URL (without credentials,
query or fragment) and the status. CI, Release Automation and the production
Dockerfile all prefetch through this command.
PMCP pilot grammar cache (#127)
scripts/v13_pmcp_pilot.pynow gives its fixtures and network-denied sandbox
workers a prefetched grammar cache.prepareprefetches into an explicit
cache, then binds it in the manifest:--grammar-cache, else the caller's
TREE_SITTER_LANGUAGE_PACK_CACHE_DIR, else<root>/grammar-cache. Every
mode checks the cache offline before it starts. The fixture, server,
registration and lifecycle environments get
TREE_SITTER_LANGUAGE_PACK_CACHE_DIR, plus a manifest URL that refuses
downloads, so offline mode'sreindexno longer fails with
grammar_unavailable.
Upgrade notes
mcp2.x is required. This release requiresmcp>=2.2,<2.4(locked at
2.3.0). Drop anymcp<2pin installed alongsideindex-it-mcp(the
workaround for1.4.1and1.5.0, and what1.5.1itself requires): it
makes the install unresolvable. Code that imports themcpSDK in the same
environment must use the 2.x API.- Co-install with PMCP 3.x.
index-it-mcpnow installs in the same
environment as PMCP 3.x (mcp>=2.2,<3) andphase-loop-runtime==0.7.30
with a cleanuv pip check.1.5.1could not share an environment with
PMCP 3.x. - Claude Code negotiates protocol 2026-07-28. Claude Code 2.1.x probes
server/discoverand negotiates 2026-07-28, which has no Tasks: task metadata
ontools/callis ignored, the call runs synchronously andtasks/*answer
-32601. The tools, their schemas and their results do not change.
Task-augmentedreindexandwrite_summariesremain available to clients
that negotiate 2025-11-25 or earlier (the official Python SDKClientSession,
the 1.x SDK client and the PMCP 3.x gateway). The official Python SDK 2.x has
no Tasks client; itsClientSessionvalidates everytools/callresult as a
CallToolResult, so task clients must speak the 2025-11-25 wire directly. - The
mcpcap moves one minor at a time. The cap below 2.4 is deliberate:
the Tasks runtime depends on the SDK's provisionalServer.middleware. Each
newmcpminor will be re-tested before the cap is widened to it. - Generation retention is automatic. After each publication only the live
generation, theMCP_INDEX_GENERATION_RETENTION(default 2) most recent
previous ones and the newest failed one are kept. Generations that exist
before the upgrade are never pruned automatically; remove them by hand once
they are no longer needed. Only generations published or failed after the
upgrade count towards retention. - After deploying, hosts that worked around #114 with
a longMCP_WATCHER_SWEEP_MINUTES(for example 1440) can return to the
default. - PMCP pilot manifests prepared before this change lack
grammar_cache. Run
prepareagain.
Known follow-ups
- #133: move the PMCP pilot from pmcp 2.7.3 to 3.x and
re-run offline and rehearsal acceptance. - #134: delete the unported legacy scripts
(patch_mcp_server.py,scripts/cliservers). - #121: plan a first semantic build and retire legacy
Qdrant collections. - #117: per-language degradation when a tree-sitter
grammar is unavailable.
[1.5.1] - 2026-11-08
Prepared corrective release for published 1.5.0 and 1.4.1: cap mcp below
2 so a plain install, with no separate mcp constraint, can start STDIO.
This section does not imply publication; protected-main dispatch and
independent delivered-artifact acceptance are separate gates.
Fixed
- Plain installs could not start STDIO.
1.5.0and1.4.1declare
mcp>=1.0.0with no upper bound, sopip install index-it-mcp==1.5.0
resolvesmcp2.x, andindex-it-mcp stdiofails at import with
ImportError: cannot import name 'McpError' from 'mcp.shared.exceptions'.
Release smoke and the PMCP pilot install underuv.lock(mcp1.27.0), so
neither caught it. The requirement is nowmcp>=1.27,<2; the locked version
stays 1.27.0.
Added
scripts/release_smoke.py --unconstrained, run bymake release-smokeand
so by the CI agent gate and Release Automation. It installs the built (or
delivered) wheel into a fresh Python 3.12 venv with no lock constraints, so
the newest versions the wheel allows resolve. It then checks
index-it-mcp --version, prefetches grammars into an explicit
TREE_SITTER_LANGUAGE_PACK_CACHE_DIR, and runs an MCP initialize and
tools/list exchange over raw JSON-RPC, then the lexical STDIO queries. It
fails on1.5.0's requirements (mcp2.3.0) and passes with the cap
(mcp1.30.0).
Upgrade notes
- Do not install
1.5.0or1.4.1without constrainingmcpbelow 2,
even asindex-it-mcp==1.5.0. Install1.5.1. Hosts that
must stay on1.4.1(for example glibc older than 2.34) add"mcp<2"to the
same install command. After1.5.1is published and accepted, the owner plans
to yank1.5.0and1.4.1. - No co-install with
mcp2.x. Environments that needmcp2.x, such as
PMCP 3.x, cannot install1.5.1in the same environment. Run it in its own
environment until the port lands. - No index, schema or configuration changes from
1.5.0; its upgrade notes
still apply when upgrading from1.4.x.
Known follow-ups
- #124: port the STDIO server to
mcp2.x and lift
the cap.
[1.5.0] - 2026-11-08
Prepared minor release for Code-Index-MCP#115 and Code-Index-MCP#119: the
treesitter-chunker 6 / tree-sitter 0.26 parser-stack upgrade and the
baml-bridge 0.23.0 alignment, so index-it-mcp can share one environment with
phase-loop-runtime 0.7.30. This section does not imply publication;
protected-main dispatch and independent delivered-artifact acceptance are
separate gates.
Changed
baml-bridgeand the BAML toolchain move from 0.20.1 to 0.23.0, matching
phase-loop-runtime0.7.29+ so both can share one environment.baml_sdk/is
regenerated (#119).- baml-bridge 0.21+ records every BAML call under
~/.baml/bteland uploads the
recordings to Boundary's cloud whenBOUNDARY_API_KEYor a saved Boundary
login exists. Recording and upload stay off by default: the generated SDK
bakesoff, so unset, blank and whitespace values of the new
MCP_BAML_TELEMETRYvariable mean off, whatever importsbaml_sdkfirst.
MCP_BAML_TELEMETRY=low|medium|highopts in. That uploads prompts containing
source code whenever a Boundary key or login is present. Upstream's
BAML_TELEMETRYis ignored. Seedocs/SUPPORT_MATRIX.md. - LLM calls now time out after 5 minutes in total, or 10 seconds to connect; they
had no limit before. On the single-file summarization path, a stalled provider
can take up to 5 minutes for the batch call and up to 5 minutes for each chunk
in the fallback. - Generated
FunctionSpec.client_id()is nowrequested_model(), returning the
bare model name. scripts/generate_baml_sdk.pyclearsbaml_sdk/before generating, so modules
a newer toolchain dropped don't survive. It and CI setDO_NOT_TRACK=1, which
stopsbaml-cliusage pings.- Parser stack moves to treesitter-chunker 6.x, tree-sitter 0.26.x and
tree-sitter-language-pack 1.20.x, matching the other Consiliency repositories
(#115). - The required agent gate runs on Python 3.12, 3.13 and 3.14.
Added
mcp-index grammars prefetchandmcp-index grammars checkmanage the
language-pack grammar cache. Language-pack 1.20 downloads grammars on first
use, so CI, container images and the devcontainer prefetch them.
--export-libs DIRwrites a shared library directory for
TREE_SITTER_LANGUAGE_PACK_LIBS_DIR, which images use because the pack only
accepts a per-user cache. Seedocs/SUPPORT_MATRIX.mdfor the cache location
and offline use.get_statusreports agrammarsblock (readyorgrammar_unavailable,
with the missing grammars and affected languages), and the server logs an
error at startup when registry grammars cannot load offline.
Fixed
- Pack grammar lookups use
csharpfor thec_sharplanguage; the public
language code is unchanged. - Sandboxed plugin workers keep the host grammar cache root and libs
directory after their environment is scrubbed, instead of starting from a
cold cache. - Sandboxed plugin workers warm the chunker parser for their language before
installing the filesystem guard. Chunker 6 reads installed package metadata
on the first python/javascript parser load, and the guard madechunk_text
fail, so the sandboxed Python plugin stored no chunks. - Network-denied sandbox workers never download grammars. The pack downloads
from native code the Python socket guard cannot see, so a cold or
rolled-over cache made workers fetch from GitHub and write into the host
cache. Workers now load only already-extracted grammars and fail the plugin
load closed otherwise. - A file whose grammar cannot load offline fails indexing with
grammar_unavailableinstead of being stored without chunks, so a reindex no
longer publishes a silently degraded generation. - Shutdown asks an in-flight staged rebuild to stop before it publishes, so the
shutdown watchdog no longer force-exits (exit code 1) a service whose
rebuild was still running when the signal arrived. grammars prefetch --export-libsreplaces each library atomically, refuses
symlinked or other-writable targets and directories that hold other files,
and never changes the mode of an existing directory.grammars checkfails
on an unsafeTREE_SITTER_LANGUAGE_PACK_LIBS_DIR.- The release smoke harness passes the grammar location through and keeps the
private grammar cache out of its world-writable cleanup sweep. - The devcontainer moves to Debian bookworm because the language-pack wheels
need glibc 2.34 or newer.
Upgrade notes
-
Full rebuild required. Chunk ids move to the
treesitter_chunk_id_v6
scheme. Indexes built with chunker 4.x reportscheme_mismatchand fail
closed until a fullreindex; they are never mixed with 6.x ids. Semantic
profiles fingerprint the chunker version, so existing vectors read as
profile-incompatible until they are rebuilt or re-embedded. -
6.0 drops erroneous keyword-only
lambdafragment chunks from Python and
changes CRLF handling inchunk_file. LF chunk ids are otherwise unchanged. -
Grammar prefetch is mandatory. Run
mcp-index grammars prefetchafter
installing, and again after every language-pack upgrade, while network
access is available. Sandboxed workers do not download a missing grammar:
without a prefetch every file whose plugin needs a grammar fails with
grammar_unavailable, and a reindex does not publish its generation. -
mcp-index grammars checknow fails whenTREE_SITTER_LANGUAGE_PACK_LIBS_DIR
is unsafe (a symlink, not owned by root or the running user, or group- or
other-writable);get_statusreports that case aslibs_dir_unsafe. -
Semantic vectors need re-embedding. After the full reindex, rebuild the
semantic profile's summaries and vectors; 4.x vectors are never reused for
6.x chunk ids. Budget the inference run before rebuilding large corpora. -
glibc 2.34 or newer. tree-sitter-language-pack 1.20 Linux wheels require
glibc 2.34+ (for example Debian 12, Ubuntu 22.04, RHEL 9). Hosts on older
glibc cannot install this release; stay on1.4.1there. -
Dependency floor.
tree-sitter>=0.26,<0.27,treesitter-chunker>=6,<7
andtree-sitter-language-pack>=1.20,<1.21. This letsindex-it-mcpshare
an environment with other tools that pin tree-sitter 0.26. -
baml-bridge 0.23.0.
baml-bridge==0.23.0replaces 0.20.1. Environments
that pinbaml-bridge==0.20.1cannot install this release; tools that pin
0.23.0, such asphase-loop-runtime0.7.30, now co-install. Code that imports
the checked-inbaml_sdk/gets the regenerated 0.23.0 SDK. -
BAML call recording and upload are off by default. Leave
MCP_BAML_TELEMETRYunset (oroff) to keep it that way. Setting
MCP_BAML_TELEMETRY=low|medium|highrecords calls under~/.bamland, when
BOUNDARY_API_KEYor a saved Boundary login exists, uploads prompts that
contain your source code to Boundary. The level is read at the first
import baml_sdk, so restart the service after changing it. -
Upstream
BAML_TELEMETRYis ignored. OnlyMCP_BAML_TELEMETRYcontrols
recording; existingBAML_TELEMETRYsettings have no effect. -
New LLM timeouts. Summarization calls now stop after 5 minutes in total or
10 seconds to connect. Size batch jobs for a worst case of 5 minutes for the
batch call plus 5 minutes per chunk on the single-file fallback path. -
client_id()renamed torequested_model(). Callers of the generated
FunctionSpec.client_id()must switch torequested_model(), which returns
the bare model name.
Known follow-ups
- #117: a missing grammar currently blocks the whole
repository's generation in every language. Per-language degradation is
planned; until then, prefetch on every host.
[1.4.1] - 2026-09-12
Prepared corrective release for Code-Index-MCP#97. This section does not imply publication;
protected-main dispatch and independent delivered-artifact acceptance are separate gates.
Fixed
- Installed wheel and non-root container startup, SQLite upgrade and MCP lifecycle checks.
- Repository state, query refusal and generation fencing across interrupted indexing.
- Path, credential, artifact and local-inference policy boundaries.
- SQLite/vector reconciliation, staged corpus provenance and duplicate vector mapping counts.
- Disabled watcher admission, bounded shutdown and repository selectors in the admin UI.
- Signed artifact metadata binds repository identity and archive checksum; restores use fresh output directories.
- Staged reindex cancellation and durable row counts, independent ignore rules, cleanup failures and multi-language search scope.
Verified
- Original pre-review pilot: installed PMCP 2.7.3, Inspector 2.6.0 and admin workflows on Linux/Python 3.12.
- Original bounded synthetic local-inference pilot with durable accounting and query/index contention; runtime repairs require fresh candidate-bound acceptance.
- Version-specific release checks and final review are tracked in
docs/operations/v13-release.md.
Limits
- Controlled rollout only; multi-repo and STDIO remain beta. No automatic fleet indexing.
- Other operating systems, Python 3.13+, broad retrieval quality and default reranking
require separate evidence. Provider immutable revisions were unreported in the pilot. - Archive-only signatures from earlier versions are rejected in enforce mode; prepare and sign canonical metadata as documented in
docs/security/attestation.md.
[1.4.0] — 2026-07-19
Minor feature release that lands the provider-neutral inference stack, the
tree-sitter chunker v4, and safer index rebuild/rollout gating on top of the
1.3.1 hardening surface.
Added
- Provider-neutral inference boundary so semantic/inference backends can be
swapped behind a stable interface without changing caller contracts (#74). - Tree-sitter chunker v4 adoption for more accurate and stable code chunking
across supported languages (#79). - Collection-resident, provenance-bound inference rollout gate (INFERLIVEGATE)
that only activates inference for verified, provenance-tracked collections
(#83). - Roadmap v12 charting the next feature and hardening milestones (#77) and an
inference-rollout operations runbook (#86).
Changed
- Chunk-identity hardening via the CHUNKERSAFE guard, atomic index rebuild, and
crash-ledger drain so interrupted rebuilds recover without corrupt or
duplicated chunks (#84). - Inference hygiene tightened end to end (INFERPOLISH) (#82).
- Four Lane A cleanups, including an explicit INDEX_LOCKED readiness state for
concurrent index operations (#85). - CI now derives the GHCR image namespace from the repository owner so release
images publish correctly after the organization transfer (#89). - Author standardization and CODEOWNERS added for deterministic review routing
(#87, #88).
Fixed
- Release container images are signed without crane, and the ownerless install
URLs in the release automation and install scripts were corrected (#92).
[1.3.1] — 2026-07-10 (Unpublished)
Fixed
- Enforced the FastAPI JWT boundary independently from the optional local
STDIO handshake and removed ambiguous repository-selector fallback. - Made staged index recovery atomic and fail closed, unified summarization
results, and bounded sandbox worker lifecycle and request handling. - Restored deterministic local quality gates, release-critical typing, and
hermetic test execution across the supported release matrix. - Split release preparation from protected-main publication and pinned every
external GitHub Action to an immutable upstream commit.
Release status
- Package, runtime, active documentation, installers, and local build artifacts
are prepared as1.3.1/v1.3.1. - No tag, GitHub release, container image, attestation, or PyPI package was
published by this hardening phase.
[1.3.0] — 2026-07-06
Added
- Added the public
mcp_server.clientPython API with typed local search,
symbol lookup, reindex, and status operations for programmatic use without
starting MCP STDIO or FastAPI. - Added source metadata filters for ordinary code, friction signals, and
historical GitHub issue context across the direct client, MCP tool handlers,
and FastAPI/search.
Changed
- Advanced the repo-owned package, runtime, release automation, installer, and
active public documentation surface from1.2.0/v1.2.0to1.3.0/
v1.3.0. - Preserved
index-it-mcpas the canonical Python distribution while keeping
code-index-mcplimited to repository, client-local server, and container
naming surfaces.
[1.2.0] — 2026-04-25
Changed (GAREL — stable surface preparation)
- Prepared the stable
1.2.0/v1.2.0repo-owned release surface across
package metadata, runtime metadata, release automation defaults, active
install docs, support docs, and Docker installer helpers. - Preserved
v1.2.0-rc8as historical prerelease soak evidence only while
routing the actual stable dispatch and GA release evidence to downstream
GADISP.
Known limitations
- This phase did not dispatch the stable release; GitHub Latest and Docker
latestremain stable-only outputs that are still verified inGADISP. - Beta, experimental, unsupported, and disabled-by-default support-tier limits
remain governed bydocs/SUPPORT_MATRIX.md; stable surface prep does not
broaden product topology or runtime claims.
[1.2.0-rc8] — 2026-04-25
Changed (GARECUT — post-remediation RC recut freeze)
- Advanced the active remediated RC recut contract to
1.2.0-rc8/
v1.2.0-rc8across package metadata, runtime metadata, release automation
defaults, and installer helpers without widening into GA channel changes. - Froze the rc8 recut evidence and final-decision handoff expectations so a
blocked or successful prerelease proof can route the next work either back
through renewed GAREL or into a GARECUT rerun without silently authorizing
GA.
Known limitations
v1.2.0-rc8remains a prerelease RC/public-alpha or beta channel artifact;- it does not authorize GA wording, GitHub Latest promotion, or stable Docker
latest; GitHub Latest is not the RC policy source. - GARECUT dispatch still requires a clean enough release worktree,
HEAD == origin/main, and unused local plus remotev1.2.0-rc8tags before
gh workflow runmay execute.
[1.2.0-rc7] — 2026-04-24
[1.2.0-rc6] — 2026-04-24
Changed (GARC — follow-up RC soak freeze)
- Advanced the active follow-up RC contract to
1.2.0-rc6/v1.2.0-rc6
across package metadata, runtime metadata, release automation defaults,
installer helpers, README status, and the active Docker/MCP configuration
docs. - Added explicit follow-up RC soak operator guidance and evidence reduction for
GARC so dispatch prerequisites, workflow observation fields, and blocked-stop
handling are frozen before any new release mutation.
Known limitations
v1.2.0-rc6remains a prerelease RC/public-alpha or beta channel artifact;
it does not authorize GA wording, GitHub Latest promotion, or stable Docker
latest; GitHub Latest is not the RC policy source.- GARC dispatch still requires a clean worktree plus fresh upstream
governance/E2E/operations evidence beforegh workflow runmay execute.
[1.2.0-rc5] — 2026-04-23
Fixed (P27-P33 — v3 multi-repo blocker closeout)
- Added fail-closed readiness handling for missing, empty, stale, building,
wrong-branch, and unsupported-worktree indexes. Query tools now return
index_unavailablewithsafe_fallback: "native_search"instead of
dispatching against an unsafe index. - Hardened the v3 repository model around many unrelated repositories on one
machine, one registered worktree per git common directory, and tracked/default
branch indexing only. - Added the P33 production multi-repo matrix for unrelated repo isolation,
same-repo worktree rejection, wrong-branch safety, stale/missing index
fallback, incremental repair, rename/delete/revert, force-push recovery, and
wrong-artifact rejection.
Changed (P34 — public alpha recut)
- Recut the public-alpha release contract to
1.2.0-rc5/v1.2.0-rc5across
package metadata, runtime metadata, release automation, README status,
Docker/MCP setup docs, anddocs/SUPPORT_MATRIX.md. - Aligned outside-developer docs, operator runbooks, and agent handoffs on the
readiness contract: indexed MCP results are authoritative only when readiness
isready; otherwise users should use native search or follow the returned
remediation. - Updated
docs/operations/deployment-runbook.mdto require P27-P33 gates,
make alpha-production-matrix, docs truth tests, release metadata tests, and
clean-checkout wheel/container smoke before public-alpha promotion.
Known limitations
- Same-repo sibling worktrees and non-default branch queries are not independent
indexed-search contexts in v3; they returnindex_unavailablewith
safe_fallback: "native_search"until a future release expands the topology
contract. - Language/runtime support remains scoped by
docs/SUPPORT_MATRIX.md, release
promotion remains gated bydocs/operations/deployment-runbook.md, and the
public-alpha go/no-go evidence remains linked from
docs/validation/private-alpha-decision.md.
[1.2.0-rc4] — 2026-04-22
Added (P26 — private alpha evidence and public alpha decision)
- Added the P26 private-alpha evidence contract:
docs/validation/private-alpha-decision.md,
docs/validation/private-alpha-evidence.schema.json, and the ignored
private-alpha-evidence/raw-output boundary. - Added
scripts/private_alpha_evidence.pyfor local private fixture runs across
python_repo,typescript_js_repo,mixed_docs_code_repo,
multi_repo_workspace, andlarge_ignored_vendor_repobefore public alpha. - Documented public-alpha release truth: supported install paths are
uv sync --locked/STDIO andghcr.io/viperjuice/code-index-mcp; language
support is scoped bydocs/SUPPORT_MATRIX.md; beta warnings and rollback
instructions remain indocs/operations/deployment-runbook.md.
[1.2.0-rc3] — 2026-04-21
Changed (P21 — release contract and dependency unification)
- Unified the P21 release contract across runtime metadata, package metadata,
README status, changelog, and release automation around1.2.0-rc3. - Removed active workflow and Docker dependency references to retired
requirements*.txtinputs in favor of the lockedpyproject.toml/uv.lock
dependency source. - Expanded release metadata and requirements-consolidation tests to guard the
P21 contract surfaces consumed by P22.
Added (P25 — blocking release gates and automation)
- Added explicit public alpha gate jobs for dependency sync, focused format/lint,
unit plus release smoke, integration smoke, Docker build/smoke, docs truth, and
required-gate aggregation. - Added release automation preflight refusal before version mutation, release
branch creation, artifact build, tag creation, GitHub release creation, PyPI
publish, or container publish. - Documented the operator gate checklist and private-alpha authenticated
attestation fallback for missingATTESTATION_GH_TOKENor repository
attestation settings.
[1.2.0-rc2] — 2026-04-20
Fixed (P20 — sandbox/OpenBLAS hotfix, observability bugs)
- (P20 SL-0) Raised
CapabilitySet.mem_mbdefault from 512 to 2048
so sandbox-default-on plugins that transitively import numpy/OpenBLAS
no longer abort onRLIMIT_AS. Closes IF-0-P20-3. - (P20 SL-2b) Prometheus
/metricsendpoint now exposes
mcp_tool_calls_total,mcp_rate_limit_sleeps_total, and
mcp_artifact_errors_by_class_total. Previously a registry split
meant the counters lived on the default REGISTRY while the exporter
served a private one. - (P20 SL-2b)
SecretRedactionResponseMiddlewareis now applied
on subprocess uvicorn boot. Previously it was registered in a startup
hook that ran after FastAPI sealed its middleware stack, so subprocess
gateways never redactedBearer ...tokens in HTTP responses.
Added (P20 — validation harness, runbook, observability smoke)
- (P20 SL-1) New
tests/integration/multi_repo/harness — pytest
fixture spawns ≥2 live uvicorn gateway subprocesses sharing one
registry; tests cover concurrency (200 entries), cross-repo delta
resolution, and dispatcher repo_id isolation. Closes IF-0-P20-1. - (P20 SL-2) New
tests/integration/obs/smoke — docker-compose
Prom+Loki stack proves JSON log parse rate, counter scrape, and
secret redaction over HTTP. Skips cleanly without docker. - (P20 SL-3) New
docs/operations/deployment-runbook.mdwith
measurable bake-pass criteria, bake windows, and rollback procedures
per stage (dev/staging/canary/full-prod). Closes IF-0-P20-2.
[1.2.0-rc1] — 2026-04-20
Added (P18 — Enforcement, Artifact Resilience & Ops)
- Route auth enforcement (IF-0-P18-4):
validate_production_config()body fills weak-JWT, CORS-wildcard, permissive rate-limit, and weak-admin-password checks.render_validation_errors_to_stderr()emits[FATAL]/[WARN]-prefixed listings. Gateway startup exits non-zero on any fatal validation error in production; WARN-continues in dev/test. Auth added to/search/capabilities,/graph/context,/graph/search. (mcp_server/gateway.py,mcp_server/config/validation.py,tests/security/test_route_auth_coverage.py,tests/security/test_startup_config_validation.py) - Artifact pipeline resilience: publisher try/finally rollback via
gh release deleteon downstream failure;artifact_downloaddelta-base-missing fallback to full artifact;_respect_rate_limitparsesRetry-After(int-seconds or HTTP-date, capped 300s) and distinguishes 403 (TerminalArtifactError) from 429 (backoff + retry); attestation preflight checksattestations:writescope. (mcp_server/artifacts/publisher.py,artifact_download.py,providers/github_actions.py,attestation.py) - Retention janitor CLI (IF-0-P18-2):
python -m mcp_server.cli retention prune --repo OWNER/REPO [--dry-run] [--older-than-days N] [--keep-latest-n N]. Protectsindex-latestandisLatest=truereleases. Newmcp_server/artifacts/retention.pypolicy module; newmcp_server/cli/retention_commands.pyclick group. - JSON logging: hand-rolled
JSONFormatterinmcp_server/core/logging.py; emits{"timestamp","level","name","message", ...extras}. Gated onMCP_ENVIRONMENT=productionorMCP_LOG_FORMAT=json. - Secret redaction middleware (IF-0-P18-1):
SecretRedactionResponseMiddlewarerewritesBearer \S+,JWT_SECRET_KEY=\S+,GITHUB_TOKEN=\S+to[REDACTED]in 4xx/5xx response bodies. Streaming bodies pass through unmodified (documented limitation). (mcp_server/security/security_middleware.py) - Prometheus counters:
mcp_rate_limit_sleeps_total(no labels) andmcp_artifact_errors_by_class_total{error_class}. Scrapeable on/metrics. (mcp_server/metrics/prometheus_exporter.py) - Attestation boot probe:
probe_gh_attestation_support()(lru_cache'd) runsgh attestation --help. Boot banner emitsATTESTATION_PREREQWARN if the subcommand is unavailable andMCP_ATTESTATION_MODE=enforce. (mcp_server/artifacts/attestation.py,mcp_server/cli/stdio_runner.py) - stdio_runner module surface (IF-0-P18-3):
initialize_services()andcall_tool(name, arguments)hoisted from_serve()closures to top-level async functions. Seven runtime state variables (dispatcher,plugin_manager,sqlite_store,initialization_error,_file_watcher,_indexing_thread,_fts_rebuild_thread) promoted to module globals. Restores the pre-P2B test surface on the post-P2B file layout. (mcp_server/cli/stdio_runner.py)
Changed (P18)
- Plugin sandbox default (behavioural change):
MCP_PLUGIN_SANDBOX_ENABLED=1is no longer required — sandboxing is now ON by default. SetMCP_PLUGIN_SANDBOX_DISABLE=1to opt out. Seedocs/security/sandbox.mdfor migration notes. (mcp_server/plugins/plugin_factory.py)
Fixed (P18 — P17 straggler burn-down)
- Full-suite failure count closed from 19 (P17 SL-4b close) to 4. The 17
tests/test_mcp_server_cli.pyfailures and 2tests/test_benchmarks.pySLO failures now pass; residual 4 are pre-existing integration tests gated onghCLIattestations:writescope and an unrelated P16 vocabulary signature mismatch. (mcp_server/cli/stdio_runner.py,mcp_server/benchmarks/benchmark_suite.py,mcp_server/plugins/sandboxed_plugin.py)
Added (P17 — Durability & Multi-Instance Safety)
- Registry flock protocol (IF-0-P17-1):
RepositoryRegistry.save()acquires
fcntl.flock(LOCK_EX)on a sibling.lockfile, then performs a read-merge-write
with an atomicrename(). Two concurrent processes callingsave()against the same
file no longer lose each other's writes. (mcp_server/storage/repository_registry.py,
tests/test_registry_concurrency.py) - Singleton reset wired:
reset_process_singletons()(P16 IF-0-P16-4 stub) is now
called at the top ofinitialize_stateless_services(). Repeat-init in the same
Python process yields fresh managers. (mcp_server/cli/bootstrap.py,
tests/test_singleton_reset.py) - Ref-poller edge-case handling (SL-2):
RefPollerdetects detached HEAD,
force-push, and branch-rename; each triggersenqueue_full_rescan(repo_id)to prevent
incremental indexing on a rewritten history. (mcp_server/watcher/ref_poller.py,
tests/test_ref_poller_edges.py) - Sweeper observability (SL-2):
WatcherSweeperexceptions emitWARNINGlog and
incrementmcp_watcher_sweep_errors_totalPrometheus counter.
(mcp_server/watcher/sweeper.py,tests/test_sweeper_observability.py) - MCP_MAX_FILE_SIZE_BYTES enforced (SL-2): Dispatcher walker (
index_directory)
now skips files larger thanMCP_MAX_FILE_SIZE_BYTES(default 10 MiB); oversize
files are logged and skipped without stalling the indexer.
(mcp_server/dispatcher/dispatcher_enhanced.py) - Checkpoint clears on clean exit (SL-2):
IncrementalIndexerclears the
checkpoint file even whenerrorsis non-empty on clean exit; a re-run starts fresh
rather than resuming from a stale error state.
(mcp_server/indexing/incremental_indexer.py) - ENOSPC → read-only store (SL-3):
SQLiteStorecatchesOSError(ENOSPC)on
commit(), transitions to read-only mode, and increments
mcp_storage_readonly_total. The server keeps serving reads without crashing.
(mcp_server/storage/sqlite_store.py,tests/test_disk_full.py) - Schema migration backup (SL-3):
SchemaMigrator.migrate_artifact()writes a
timestamped.backupcopy of the DB before any migration; rollback-on-failure
restores from backup. (mcp_server/storage/schema_migrator.py,
tests/test_schema_migration_backup.py) - UnknownSchemaVersionError in core/errors.py (SL-3):
UnknownSchemaVersionError
promoted fromschema_migrator.pytomcp_server/core/errors.py, extending the P16
IF-0-P16-1 taxonomy by one class. - RerankerFactory.create_default() (SL-4):
RerankerFactory.create_default()
returns a functional reranker; resolves the P14 carry-over.
(mcp_server/plugins/reranker_factory.py) - Ruff F401 clean (SL-4):
ruff F401count acrossmcp_server/dispatcher,
mcp_server/storage,mcp_server/watcheris 0. - Multi-instance runbook (SL-docs):
docs/operations/multi-instance.md— flock
protocol, singleton-reset semantics, DR/recovery procedure. - Vocabulary freeze (IF-0-P17-1): Registry flock + read-merge-write pattern is
frozen as a named interface for cross-phase reference. - Test-debt ledger (SL-docs):
docs/operations/known-test-debt.mdenumerates
residual failures after P17 close (56 total; 9 cross-repo coordinator carry-over +
47 pre-existing in other modules).
Added (P16 — Shared Vocabulary Preamble)
- Frozen error taxonomy (IF-0-P16-1): New error classes
TransientArtifactError,TerminalArtifactError, andSchemaMigrationErrorinmcp_server/core/errors.pyfor structured error handling in artifact operations. Stubs only — no consumer wiring. - Lazy env-var getters (IF-0-P16-2): Five new environment variable getters in
mcp_server/config/env_vars.py—get_max_file_size_bytes(),get_artifact_retention_count(),get_artifact_retention_days(),get_disk_readonly_threshold_mb(), andget_publish_rollback_enabled()— with sensible defaults. Stubs only — no consumer wiring. - ValidationError dataclass + validate_production_config (IF-0-P16-3): New
ValidationErrordataclass and keyword-onlyvalidate_production_config(environment: str = "production")function inmcp_server/config/validation.py. Stubs only — no consumer wiring. - reset_process_singletons (IF-0-P16-4): New utility function in
mcp_server/cli/bootstrap.pyfor test isolation and graceful restarts. Stubs only — no consumer wiring.
Added (P15 — Security Hardening)
- Plugin sandboxing (IF-0-P15-1): Plugins execute in isolated worker processes with JSON-line IPC, 30s timeout, and capability-based restrictions. New
mcp_server/sandbox/package with supervisor and adapter;SandboxedPluginwraps instances withCapabilitySet(filesystem, network, subprocess, env_read). (mcp_server/sandbox/supervisor.py,mcp_server/plugins/sandboxed_plugin.py,tests/security/test_plugin_sandbox.py) - Metrics endpoint auth (IF-0-P15-2):
/metricsrequiresAuthorization: Bearer <token>(app-level) or NetworkPolicy restriction (k8s-only). Wired viarequire_auth("metrics")middleware. (mcp_server/security/security_middleware.py,mcp_server/api/gateway.py) - Artifact attestation (IF-0-P15-3): Published artifacts are signed with
gh attestation signand verified withgh attestation verify(requiresGITHUB_TOKENwithattestations:writescope). Mode controlled byMCP_ATTESTATION_MODE(enforce/warn/skip). (mcp_server/artifacts/attestation.py,mcp_server/artifacts/publisher.py,mcp_server/artifacts/artifact_download.py) - Path traversal guard + token validator (IF-0-P15-4): Search results filtered through
PathTraversalGuard.normalize_and_check(path, allowed_roots)(controlled byMCP_ALLOWED_ROOTSenv). Startup validatesGITHUB_TOKENfor all five required scopes viaTokenValidator.validate_scopes()(soft-fail default; hard-fail withMCP_REQUIRE_TOKEN_SCOPES=1). Rate-limit backoff on cross-repo artifact fetches. (mcp_server/security/path_guard.py,mcp_server/security/token_validator.py)
Added (P14 — Multi-Repo Completeness + Schema Evolution)
- Reranker wiring (IF-0-P14-1):
CrossRepoCoordinator.__init__now accepts an injectedIReranker(mcp_server/indexer/reranker.py); falls back toRerankerFactory.create_default()with graceful degradation toNone. (mcp_server/dispatcher/cross_repo_coordinator.py,tests/test_cross_repo_reranker.py) - Dependency graph (IF-0-P14-2): New
mcp_server/dependency_graph/package withparsers.py,aggregator.py, and ecosystem parsers (Python/npm/Go/Cargo)._get_repository_dependenciesreturns resolved repo IDs. (mcp_server/dispatcher/cross_repo_coordinator.py) - Schema migrator (IF-0-P14-3): New
mcp_server/storage/schema_migrator.pywithSchemaMigrator,UnknownSchemaVersionError, andSchemaMigrationError. Artifact manifests carryschema_version;check_compatibilitygates downloads by version. (mcp_server/artifacts/artifact_download.py) - Auto-delta artifacts (IF-0-P14-4): New
mcp_server/artifacts/delta_policy.pywithDeltaPolicy/DeltaDecision. Publisher switches to delta mode when artifact exceedsMCP_ARTIFACT_FULL_SIZE_LIMIT(default 500 MB)._get_chunk_ids_for_pathacceptslimit/offsetpagination. (mcp_server/artifacts/publisher.py,mcp_server/artifacts/artifact_upload.py) - Watcher sweep + rename atomicity (IF-0-P14-5): New
mcp_server/watcher/sweeper.py::WatcherSweeperruns full-tree scans everyMCP_WATCHER_SWEEP_MINUTES(default 60) to recover dropped inotify/FSEvents events.move_filewrapped intwo_phase_commit; raisesIndexingErroron semantic failure. (mcp_server/dispatcher/dispatcher_enhanced.py,mcp_server/watcher_multi_repo.py)
[1.1.0] - 2026-04-14
Security
- #40 Removed hardcoded
JWT_SECRET_KEYandDEFAULT_ADMIN_PASSWORDdefaults from
gateway.py. Server now raisesRuntimeErrorat startup if either env var is unset,
preventing silent production deployments with known-weak credentials.
docker-compose.ymland.env.exampleupdated to use env var substitution with no
insecure literal fallbacks. (mcp_server/gateway.py,docker-compose.yml,.env.example)
Fixed
- #41
get_index_status()symbols_countfield now returns the actual number of
symbols stored for the file (was always0). AddedSQLiteStore.count_symbols_for_file()
and wired it intoIndexEngine.get_index_status().
(mcp_server/storage/sqlite_store.py,mcp_server/indexer/index_engine.py) - #41 Reference storage in
_store_parse_result()now callsget_symbol()+
store_reference()instead of silently passing; references are persisted in
symbol_referencesand queryable viaget_references().
(mcp_server/indexer/index_engine.py) - #42 Removed broken
curl .../install-mcp.shline from README (script does not
exist). Replaced silent TODO inwatcher_multi_repo._create_and_upload_artifact()with
an explicitlogger.warningdirecting users to the CI workflow for artifact upload.
(README.md,mcp_server/watcher_multi_repo.py) - Windows CI compatibility: forward-slash path normalization in
SQLiteStore.store_file(),
signal.SIGALRMguard inmcp_server_cli.py, double-quoted git commit messages in
tests,>= 0timing assertions for low-resolution Windows clocks.
(mcp_server/storage/sqlite_store.py,scripts/cli/mcp_server_cli.py,
tests/test_utilities.py,tests/test_repository_management.py,
tests/test_multi_repo_search.py,tests/test_dispatcher_advanced.py) - Benchmark scripts updated to current SQLiteStore/plugin API: replaced removed
store.add_symbol()withcreate_repository+store_file+store_symbol, and
replaced obsoleteplugin.extract_symbols()withplugin.indexFile()with
IndexSharddict-style access.
(benchmarks/symbol_lookup_benchmark.py,benchmarks/semantic_search_benchmark.py,
benchmarks/indexing_speed_benchmark.py)
Added
- Backstage catalog registration (
catalog-info.yaml) and standard repo layout
(docs/,mkdocs.yml). (04ea8ff) - Consiliency maintenance worker trigger workflow for scheduled index health checks.
(.github/workflows/,72f15a5)
Added
- Query-intent routing: symbol-pattern queries (
class Foo,def bar, CamelCase,
snake_case) route directly to the symbols table, bypassing BM25. Fixes retrieval for
class/function lookups that previously returned wrong files.
(mcp_server/dispatcher/query_intent.py,dispatcher_enhanced.py) FlashRankRerankerandCrossEncoderReranker(OSS, lazy-loaded); all three rerankers
share a unified sync interface. Configured viaRERANKER_TYPEenv var.
(mcp_server/dispatcher/reranker.py)- Matrix benchmark: 17-query suite covering all mode × embedding × reranker combinations.
(scripts/run_matrix_benchmark.py) - LLM chunk summarization:
ChunkWriter/LazyChunkWriter/ComprehensiveChunkWriter;
per-profile config viasummarization:block incode-index-mcp.profiles.yaml.
(mcp_server/indexing/summarization.py) - Initial project structure with plugin-based architecture
- FastAPI-based MCP server implementation
- Python plugin with Jedi integration for advanced code analysis
- Tree-sitter wrapper for language parsing
- Semantic indexer with Voyage AI integration
- C4 architecture diagrams using Structurizr DSL
- Comprehensive documentation structure
- Plugin base class for language extensions
- File watcher skeleton for real-time updates
- Cloud sync stub implementation
Changed
- Enhanced Python plugin with proper AST analysis
- Improved error handling in gateway endpoints
- Updated project documentation and README
Added
- MCP
InitializeResult.instructionsfield populated: every connecting client receives
search-first guidance at handshake time, over any transport (stdio, SSE, HTTP).
(scripts/cli/mcp_server_cli.py) last_indexedfield added tosearch_coderesults, surfacing the index timestamp so
agents can assess line-number freshness. (scripts/cli/mcp_server_cli.py,
mcp_server/storage/sqlite_store.py)FileWatchernow auto-starts insideinitialize_services()after the dispatcher is
ready, so the index stays current without any manual setup. Stopped automatically in
main()finallyblock on server exit. Only active when usingEnhancedDispatcher
(not BM25-onlySimpleDispatchermode). (scripts/cli/mcp_server_cli.py)- Auto-index on first run: when no index exists,
initialize_services()creates
.mcp-index/code_index.dband indexes the repository in a background thread. The MCP
server is immediately responsive; search results populate as indexing progresses.
SetMCP_AUTO_INDEX=falseto disable for large repos and run thereindextool
manually. (scripts/cli/mcp_server_cli.py) search_coderesponses now include"indexing_in_progress": truewhen the initial
background index is still running, preventing agents from concluding "no results"
prematurely while the index is being built. (scripts/cli/mcp_server_cli.py)FileWatcherstart deferred until after background initial index completes, eliminating
concurrent SQLite write contention between the bulkindex_directory()pass and the
watchdog event handler on first-run repos. (scripts/cli/mcp_server_cli.py).gitignoreentries for.mcp-index/code_index.db-shmand.mcp-index/code_index.db-wal
(SQLite WAL mode sidecar files) added toinstall.shand written automatically during
auto-init so WAL files never appear as untracked. (mcp-index-kit/install.sh,
scripts/cli/mcp_server_cli.py)code-index-mcp.profiles.yamldiscovery now searchesMCP_PROFILES_PATHenv var →
CWD → server package directory in that order. Previously the file was only found at CWD,
so semantic profiles silently vanished whencwdwas set to another repo. Fallback to the
server's own installation means semantic search and summarization work out-of-the-box in
any repo. (mcp_server/config/settings.py)${VAR:default}env-var substitution supported in allbase_urlfields of
code-index-mcp.profiles.yaml. SetVLLM_EMBEDDING_BASE_URLor
VLLM_SUMMARIZATION_BASE_URLto point at any server without editing the YAML.
(mcp_server/config/settings.py,code-index-mcp.profiles.yaml)- File-count guard before auto-index: if the target repo exceeds
MCP_AUTO_INDEX_MAX_FILES
(default 100 000), auto-indexing is skipped and a warning is logged. Prevents indefinite
startup I/O on very large monorepos. (scripts/cli/mcp_server_cli.py) - BM25 FTS auto-heal: on startup, if an existing index has tracked files but an empty
fts_codetable (e.g. after a migration or partial corruption), a background thread
callsrebuild_fts_code()automatically. Previously the empty state persisted until
a manualreindextool call. (scripts/cli/mcp_server_cli.py)
Fixed
scripts/cli/mcp_server_cli.pywas calling undefinedget_settings()on every tool
invocation, crashing the server before any tool could execute. Replaced with
Settings.from_environment(). (scripts/cli/mcp_server_cli.py)LazyChunkWriter.__init__was missing thesummarization_configparameter despite
the base classChunkWriteraccepting it, causing aTypeErroron every server
startup. (mcp_server/indexing/summarization.py)- File watcher
trigger_reindex()now callsremove_file()beforeindex_file(),
preventing stale chunks from persisting when a file edit changes chunk boundaries.
(mcp_server/watcher.py) EnhancedDispatcher.remove_file()now evicts_file_cachebefore SQLite deletion,
preventing a skip-on-reindex race where an unchanged-hash file would be left absent
from the index after a forced removal. (mcp_server/dispatcher/dispatcher_enhanced.py)SQLiteStore.remove_file()now deletessymbol_trigramsbeforesymbols, fixing a
FK constraint failure that silently left file rows in the DB on removal.
(mcp_server/storage/sqlite_store.py)SQLiteStore.remove_file()now clearsfts_codeentries regardless of whether
file_idwas stored as an integer, absolute path, or relative path string.
(mcp_server/storage/sqlite_store.py)SQLiteStore.search_bm25()now includeslast_modifiedin result rows so the
last_indexedfield is non-null in search results. (mcp_server/storage/sqlite_store.py)- Semantic rebuild (
_build_semantic_baseline) now usesIgnorePatternManager, respecting
.gitignoreand.mcp-index-ignore. Previously indexedtest_workspace/(20,803 Django
fixture files), causing 7-hour rebuilds. (mcp_server/cli/index_management.py) - Default ignore patterns expanded:
test_workspace/,test_repos/,vendor/,
third_party/,baml_client/,*_pb2.py,*_pb2_grpc.py.
(mcp_server/core/ignore_patterns.py) code-index-mcp.profiles.yamlbase_urlnow takes precedence overOPENAI_API_BASE
env default (was reversed, routing oss_high embeddings to wrong endpoint).
(mcp_server/config/settings.py)- Corrected
oss_highvLLMbase_urlfrom127.0.0.1:8001toai:8001.
(code-index-mcp.profiles.yaml) - Resolved merge conflicts in treesitter_wrapper.py
- Fixed Python plugin implementation issues
- Tree-sitter byte-offset/char-offset mismatch in Python plugin: symbol names were corrupted in files
containing multi-byte characters (e.g. em-dashes in docstrings) before the first class/function
definition. Fixed by usingnode.text.decode("utf-8")inplugin.pyandplugin_semantic.py. - BM25/FTS path was bypassed when
semantic=Truebut no Qdrant indexer was initialized. Fixed in
dispatcher_enhanced.pyso the FTS path runs whenever the semantic indexer is absent. - Python plugin
_preindex()now excludeshtmlcov,.venv,venv,node_modules,__pycache__,
.git,dist,builddirectories to prevent junk results from polluting in-memory indexes.
Security
- Added input validation for file paths
- Implemented basic security checks in API endpoints
- Added secret detection patterns
[0.1.0] - TBD
Planned Features
- Complete implementation of C, C++, JavaScript, Dart, and HTML/CSS plugins
- SQLite-based local storage with FTS5
- Full file system watcher implementation
- Performance optimizations and caching
- Authentication and authorization system
- Comprehensive test suite
- Production deployment guide
Version History
Pre-release Development
- Project inception and architecture design
- Core infrastructure setup
- Initial plugin system implementation