v0.1.2

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.

[Unreleased]

[0.1.2] — 2026-05-31

Removed

• Dead mcp.enabled config block. .codegraph.yml carried an
  mcp.enabled: false default that no code path ever read — confusing
  next to the active register_mcp: true. Removed from CodegraphConfig;
  the model is set to silently ignore the field so existing 0.1.x configs
  still load without error. New configs no longer emit it.

Added

• codegraph clean subcommand + auto-prune of .codegraph/explore/.
  codegraph serve generated a pyvis HTML page per explored function
  and never pruned them — long-lived projects could grow this cache to
  hundreds of MB. New module codegraph/cache_prune.py implements LRU
  eviction by mtime (prune_cache_to_size(path, max_size_mb)), and the
  new codegraph clean subcommand exposes it with --max-size-mb 50
  (default) and --all (wipe everything). codegraph serve runs the
  prune on startup and prints the current cache size.
• Auto-populate ignore patterns from detected ecosystem during init.
  The free-text "Extra ignore patterns" prompt was a footgun — a user
  once typed y thinking it was yes/no and got ignore: [- y] in
  their YAML. Init now sniffs the repo for pyproject.toml,
  package.json (with React Native detection), go.mod, Cargo.toml,
  pom.xml, build.gradle, and pre-fills the matching ignore patterns
  (__pycache__/, .venv/, node_modules/, ios/Pods/,
  android/build/, vendor/, target/, build/, …). The user
  confirms the auto-detected list with a single yes/no; the free-text
  extras prompt is still there for project-specific paths but no longer
  the only path to safety.
• codegraph init writes agent guidance to CLAUDE.md and AGENTS.md.
  After a fresh install on a project, coding agents (Claude Code,
  Cursor, Codex, …) defaulted to grep even with polycodegraph's MCP
  server registered, because they had no hint that polycodegraph was
  available. Init now prompts (default: yes) and writes a short
  "## polycodegraph" section telling the agent which MCP tool to use
  for each kind of structural query (find-symbol, callers, callees,
  blast-radius, dataflow-trace, untested, dead-code). Mirrors the
  .mcp.json behaviour: creates the file when absent, appends when
  present without the section, leaves alone when the section is
  already there. Both files are written so Codex / GitHub Copilot CLI
  (which read AGENTS.md) get the hint too.

Changed

• Edge classification: external vs unresolved-local. The
  unresolved_edges metric used to lump together imports of external
  packages (fastapi, os, react) — which static analysis without a
  venv/site-packages parse cannot resolve and isn't supposed to — with
  edges that pointed at in-repo symbols but couldn't be matched (rename
  drift, dynamic import). One number, two very different meanings.
  compute_metrics now also returns external_edges (root module not
  in the repo) and unresolved_local_edges (root is in-repo, symbol
  missed). The unresolved_edges total is kept as the sum for
  back-compat. Markdown report + MCP metrics tool both expose the
  split. (Surfaced by a user adding polycodegraph to a FastAPI project
  who saw "1,423 unresolved edges" and asked whether the tool was
  broken; almost all were external imports.)

Fixed

• codegraph init now writes a robust .mcp.json — uses the absolute
  path to the codegraph binary (resolved from the running interpreter's
  bin dir, then shutil.which, then bare fallback), passes an explicit
  --db .codegraph/graph.db flag, and sets cwd to the repo root. The
  0.1.1 fix wrote the file but used a bare codegraph command with no
  --db and no cwd, which forced users to hand-edit the file when
  their MCP client's $PATH or working directory didn't line up. Old
  pre-0.1.2 default entries are migrated forward in place; user-customised
  entries are left alone. (Surfaced when a user added polycodegraph to a
  FastAPI project — Claude reported needing to fix all three.)

• Type-annotation-only references no longer count as dead code. The
  Python parser now emits a reference edge from a function/method to
  each capitalized type name appearing in its parameter and return type
  annotations. Edges go through the existing unresolved::<TypeName>
  resolver pipeline, so they rewrite to real CLASS ids when the type
  is defined in the repo. The flagship case is FastAPI Pydantic
  request-body models — def create_defect(body: RetroDefect): ... was
  flagged dead because the only reference to RetroDefect came through
  the type annotation; that loop is now closed. Stdlib typing
  scaffolding (Optional, Annotated, Union, …) and FastAPI
  dependency markers (Body, Depends, …) are blocklisted so the
  graph stays clean. Framework-agnostic: any function with type
  annotations benefits, not just route handlers.

• MCP server returns an actionable error when the graph isn't built.
  Previously, every tool would happily return an empty result against
  a missing .codegraph/graph.db. Claude / Cursor paraphrased that as
  "workspace is empty," sending users down the wrong rabbit hole. Now
  the dispatcher catches the missing-db case and returns a structured
  {error: "graph_not_built", message: "polycodegraph: no graph found at <path>. Run \codegraph build` in the repo root first.", db_path:
  ...}. New GraphNotBuiltErrorexposed fromcodegraph.mcp_server.server` for callers that want to handle it
  explicitly.

• workspace_state MCP tool now distinguishes "not configured" from
  "no repos". When ~/.codegraph/workspace.yml didn't exist, the
  tool returned {workspace_size: 0, repos: []} — indistinguishable
  from a configured-but-empty workspace. LLMs paraphrased this as
  "everything is broken" and cascaded into wrong-answer territory.
  Result now carries a status field
  (workspace_not_configured / workspace_empty / ok) with an
  actionable message: "local-graph MCP tools work independently — this
  only affects cross-repo tools. Run codegraph workspace init …".

• semantic_search and hybrid_search MCP tools distinguish
  "not enabled" from "not built". Both used to return a bare
  {error: <stringified-exception>} when called without the optional
  embeddings index. LLMs read this as catastrophic. Now both tools
  return {status, message}:

  • embeddings_not_enabled — the optional codegraph.embed import
    failed (dep not installed). Message: install with
    pip install polycodegraph[embed].
  • embeddings_not_built — module imports fine but no index for the
    repo. Message: run codegraph embed (~140 MB, ~30s).

  Both messages spell out that the structural query tools
  (find_symbol, callers, callees, blast_radius, dataflow_trace) work
  without embeddings, so the LLM doesn't escalate the partial degrade
  into "everything is broken".

• Bench harness: repomapper venv now bootstraps on Python 3.14.
  aider-chat (transitive dep of repomapper) failed to install because
  the bundled pip didn't provision setuptools.build_meta quickly
  enough; build aborted with ModuleNotFoundError: No module named 'setuptools.build_meta'. bench.runners._venv.ensure_venv gained a
  build_deps=[...] kwarg that installs build-time deps in a separate
  pip install before the main one; the repomapper runner sets it to
  ["setuptools", "wheel"]. Bench-side only — does not affect end
  users.

[0.1.1] — 2026-05-24

Fixed

• codegraph init now actually writes .mcp.json when the user accepts
  the "Register MCP server" prompt. Previously the prompt was labelled
  "Phase 3 implementation" and only stored register_mcp: true in
  .codegraph.yml while doing nothing on disk — which meant the README's
  "Claude Code and Cursor auto-pick up polycodegraph" claim was broken in
  0.1.0. Init now writes a project-local .mcp.json, merges into an
  existing file when present, and preserves any user-customised codegraph
  entry it already finds. The prompt default also flipped from False to
  True. (Issue surfaced when a user ran pipx install polycodegraph on
  a fresh project and Cursor saw no MCP server.)

Upgrading from 0.1.0

If you ran codegraph init on 0.1.0 and answered yes to the MCP
register prompt, your .codegraph.yml says register_mcp: true but no
.mcp.json was ever written. Re-run codegraph init after upgrading
— it will re-prompt (defaults are pre-filled from your existing config)
and now actually write the file. Your existing graph + config are
preserved.

Post-launch-sprint additions (still pre-release)

Go language support — .go files parse + analyze cleanly

The Go ecosystem is the single biggest community pull outside Python/JS,
and a code graph that spans services + tooling + libraries was the
obvious next move. v1 of the Go extractor lives at
codegraph/parsers/go.py and produces the same node/edge shape as the
Python and TypeScript parsers, so every existing tool (analyze,
workspace_blast_radius, MCP queries) works on Go code immediately.

• New module codegraph/parsers/go.py — tree-sitter Go grammar via
  tree-sitter-go>=0.23 (added to pyproject.toml).
• Node kinds emitted: MODULE (one per .go file, qualname = the
  package X name so files in the same package share an addressable
  namespace), FUNCTION (top-level func), METHOD (qualname
  module.ReceiverType.Name, with receiver + pointer-ness in metadata),
  CLASS (the closest analog to Go's type X struct {...} and
  type X interface {...}), TEST (*_test.go files).
• Edge kinds emitted: `DEFINED_IN...

v0.1.1

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.

[Unreleased]

[0.1.1] — 2026-05-24

Fixed

• codegraph init now actually writes .mcp.json when the user accepts
  the "Register MCP server" prompt. Previously the prompt was labelled
  "Phase 3 implementation" and only stored register_mcp: true in
  .codegraph.yml while doing nothing on disk — which meant the README's
  "Claude Code and Cursor auto-pick up polycodegraph" claim was broken in
  0.1.0. Init now writes a project-local .mcp.json, merges into an
  existing file when present, and preserves any user-customised codegraph
  entry it already finds. The prompt default also flipped from False to
  True. (Issue surfaced when a user ran pipx install polycodegraph on
  a fresh project and Cursor saw no MCP server.)

Upgrading from 0.1.0

If you ran codegraph init on 0.1.0 and answered yes to the MCP
register prompt, your .codegraph.yml says register_mcp: true but no
.mcp.json was ever written. Re-run codegraph init after upgrading
— it will re-prompt (defaults are pre-filled from your existing config)
and now actually write the file. Your existing graph + config are
preserved.

Post-launch-sprint additions (still pre-release)

Go language support — .go files parse + analyze cleanly

The Go ecosystem is the single biggest community pull outside Python/JS,
and a code graph that spans services + tooling + libraries was the
obvious next move. v1 of the Go extractor lives at
codegraph/parsers/go.py and produces the same node/edge shape as the
Python and TypeScript parsers, so every existing tool (analyze,
workspace_blast_radius, MCP queries) works on Go code immediately.

• New module codegraph/parsers/go.py — tree-sitter Go grammar via
  tree-sitter-go>=0.23 (added to pyproject.toml).
• Node kinds emitted: MODULE (one per .go file, qualname = the
  package X name so files in the same package share an addressable
  namespace), FUNCTION (top-level func), METHOD (qualname
  module.ReceiverType.Name, with receiver + pointer-ness in metadata),
  CLASS (the closest analog to Go's type X struct {...} and
  type X interface {...}), TEST (*_test.go files).
• Edge kinds emitted: DEFINED_IN, IMPORTS (per-import-path,
  preserves aliases as metadata), CALLS (bare names like Foo,
  dotted-package like fmt.Println, receiver-method like g.Greet),
  INHERITS (Go composition — embedded fields in a struct become
  INHERITS edges to the embedded type).
• Registry hookup: builder.py imports codegraph.parsers.go at
  module load so the @register_extractor decorator fires. init's
  language-detection map now includes go: [".go"].
• Smoke test against spf13/cobra: 36 files parsed, 715 Go nodes
  (427 functions + 168 methods + 17 structs/interfaces + 17 test files
  • 19 modules), 2541 CALLS edges, 190 IMPORTS, 0 errors, 0.25s.
• 19 new parser tests at tests/test_go_parser.py covering every
  emitted node/edge shape plus edge cases (empty file, missing file,
  test-file detection, aliased imports, pointer receivers).

Limitations the parser is honest about up front:

• Generic type parameters land as text in metadata; not in qualnames.
• Interface satisfaction (does *Foo implement Bar?) is not detected
  by the parser — that needs a whole-package pass and belongs in the
  resolver layer, not the extractor.
• init() and main() get no special treatment yet.

Cross-repo workspace mode — treat N repos as one mental unit

Solo devs and small teams routinely switch between 3–10 repos a day. Until
now codegraph was strictly single-repo — every command resolved against
the .codegraph/graph.db in the current working directory. The new
codegraph workspace subcommand group lets you register N repos once and
then query them as a union.

• New CLI subcommands (codegraph/cli.py):
  • codegraph workspace init — create ~/.codegraph/workspace.yml
  • codegraph workspace add <repo> — register a repo (validates that the
    path exists and is a directory; idempotent)
  • codegraph workspace remove <repo> — deregister (the per-repo
    .codegraph/graph.db is untouched, only the membership)
  • codegraph workspace list — table of registered repos and whether
    each has a built graph
  • codegraph workspace status — per-repo git state: branch, dirty file
    count, last commit subject + timestamp, graph freshness, error notes
  • codegraph workspace sync [--only <name>] — rebuild the graph for
    every registered repo (or just one)
• New module at codegraph/workspace/:
  • config.py — WorkspaceConfig Pydantic model + YAML load/save with
    CODEGRAPH_WORKSPACE_FILE env-var override (useful for tests +
    per-shell workspaces)
  • operations.py — pure functions returning JSON-safe dicts so the CLI
    and MCP server share one implementation
• 3 new MCP tools (codegraph/mcp_server/server.py):
  • workspace_state() — git + graph state for every registered repo
  • workspace_diff_since(ref="main") — cross-repo file changes since a
    git ref; combines committed (<ref>...HEAD) and uncommitted (HEAD)
  • workspace_blast_radius(symbol, depth=None) — unioned blast radius
    across all repos; reuses the existing single-repo blast_radius() so
    behaviour matches exactly per-repo
• Pollution prevention — workspace MCP tools self-load their per-repo
  graphs from the workspace config; the MCP _call_tool dispatch
  bypasses the usual _load_graph() for workspace_* names so the
  server doesn't materialize a stray .codegraph/graph.db in the cwd
  when the user only calls workspace tools.
• Conflict handling — when the same memory id exists on both sides
  with different bodies at the same updated_at, the conflicting remote
  copy is saved as <id>.conflict-<timestamp>.md next to the local
  file for manual merging. Sync never silently clobbers.
• 39 new tests under tests/test_workspace_{config,ops,cli,mcp}.py
  covering config round-trip, real-tmp git-repo state probes,
  CliRunner-based CLI integration, and direct MCP handler invocation.
  Full suite remains green at 576 passing (was 537).
• README updated with a new "Cross-repo Workspace mode" collapsible
  block under CLI subcommands; MCP tools table now lists 18 total (was 15).

v0.3 unified trace — Architecture view shows the real chain

• Per-handler dataflow field on the HLD payload (PR #22). Each route
  entry now carries a dataflow.hops[] array shaped to the v0.3 contract,
  so the dashboard can render the actual frontend → backend → DB chain
  without re-running analysis client-side.
• Phase 4 of the Learn Mode lifecycle modal renders real hops (PR #23).
  The previously-generic "project-specific data layer" phase now shows
  the actual handler → service → repo → model chain across sequence,
  pipeline, and diagram modes. Empty / low-confidence states render
  gracefully with a "no trace data" panel.
• Argument-flow propagation — pick a parameter, watch it travel (PRs #24,
  #25). Per-hop arg_flow: dict[str, str | None] mapping starting keys to
  their locally-renamed names. Frontend renders a chip-strip param picker
  and animates the selected param along the swimlane (sequence) or the
  bezier path (diagram, via SMIL animateMotion). Rename annotations
  ((was userId)) surface where the local name diverges. Snake_case ↔
  camelCase ↔ PascalCase all normalise to the same key.
• ROUTE entry hop args backfilled from handler params (PR #26). The
  trace walker can't supply args at the entry hop (no incoming CALLS edge),
  so URL-template handlers (e.g. /api/users/{user_id}) used to produce
  empty arg_flow. Backfilled from the handler node's DF0 metadata.params
  (skipping self / cls). Closes the launch demo's hero shot.

Cleanup + analyzer hardening

• Examples directory excluded from dead-code / untested analysis
  (PR #19) — examples/cross-stack-demo is documentation, not call-graph-
  traceable code.
• Unused _propagate_class_role_to_members helper deleted (PR #19).
• # pragma: codegraph-public-api analyzer support (PR #21). Functions
  and classes preceded by the pragma comment (Python # or TypeScript
  // style, with optional # codegraph: public-api alias) are exempted
  from find_dead_code(). Applied to all 15 known intentional public-API
  methods (EmbeddingStore facade, SQLiteGraphStore.upsert_node /
  vacuum, all to_dict / as_dict serializers, _register.decorator
  closure). Self-graph dead-code count: 15 → 0, honestly.

Docs & developer experience

• README refresh (PR #20). 15 MCP tools listed; DF1–DF4 + Architecture
  view documented in the headline feature list; "What it doesn't do yet"
  rewritten to drop already-shipped items; numeric stats current.
• SESSION_HANDOFF.md (PR #20) rewritten as a self-contained briefing
  for the next session.
• PLAN_V0_3_UNIFIED_TRACE.md (PR #20) — concrete spec for the unified
  trace work; now marked shipped as of PRs #22–#26.

Cross-stack data-flow tracing (DF0 → DF4)

• DF0 — function signatures + per-call-site arguments. Python and
  TypeScript parsers capture parameter lists, return-type annotations,
  and the literal text of each call-site argument and kwarg.
• DF1 — HTTP routes + SQL data access. FastAPI / Flask / aiohttp
  ROUTE edges; SQLAlchemy READS_FROM / WRITES_TO edges including
  self.session.X repository-pattern detection.
• DF1.5 — role classification. Functions and classes are tagged
  HANDLER / SERVICE / COMPONENT / REPO based on framework patterns.
• **DF2 — frontend FETCH_CALL extractio...

v0.1.0

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.

[Unreleased]

0.1.0 is in-progress / pre-release. The README, CLI, MCP server, and 3D
dashboard reflect the current state of main, but the package has not yet
been pushed to PyPI. Items below describe what has shipped on main and
will roll into the eventual 0.1.0 tag. Install from source for now
(pip install -e .).

Post-launch-sprint additions (still pre-release)

Go language support — .go files parse + analyze cleanly

The Go ecosystem is the single biggest community pull outside Python/JS,
and a code graph that spans services + tooling + libraries was the
obvious next move. v1 of the Go extractor lives at
codegraph/parsers/go.py and produces the same node/edge shape as the
Python and TypeScript parsers, so every existing tool (analyze,
workspace_blast_radius, MCP queries) works on Go code immediately.

• New module codegraph/parsers/go.py — tree-sitter Go grammar via
  tree-sitter-go>=0.23 (added to pyproject.toml).
• Node kinds emitted: MODULE (one per .go file, qualname = the
  package X name so files in the same package share an addressable
  namespace), FUNCTION (top-level func), METHOD (qualname
  module.ReceiverType.Name, with receiver + pointer-ness in metadata),
  CLASS (the closest analog to Go's type X struct {...} and
  type X interface {...}), TEST (*_test.go files).
• Edge kinds emitted: DEFINED_IN, IMPORTS (per-import-path,
  preserves aliases as metadata), CALLS (bare names like Foo,
  dotted-package like fmt.Println, receiver-method like g.Greet),
  INHERITS (Go composition — embedded fields in a struct become
  INHERITS edges to the embedded type).
• Registry hookup: builder.py imports codegraph.parsers.go at
  module load so the @register_extractor decorator fires. init's
  language-detection map now includes go: [".go"].
• Smoke test against spf13/cobra: 36 files parsed, 715 Go nodes
  (427 functions + 168 methods + 17 structs/interfaces + 17 test files
  • 19 modules), 2541 CALLS edges, 190 IMPORTS, 0 errors, 0.25s.
• 19 new parser tests at tests/test_go_parser.py covering every
  emitted node/edge shape plus edge cases (empty file, missing file,
  test-file detection, aliased imports, pointer receivers).

Limitations the parser is honest about up front:

• Generic type parameters land as text in metadata; not in qualnames.
• Interface satisfaction (does *Foo implement Bar?) is not detected
  by the parser — that needs a whole-package pass and belongs in the
  resolver layer, not the extractor.
• init() and main() get no special treatment yet.

Cross-repo workspace mode — treat N repos as one mental unit

Solo devs and small teams routinely switch between 3–10 repos a day. Until
now codegraph was strictly single-repo — every command resolved against
the .codegraph/graph.db in the current working directory. The new
codegraph workspace subcommand group lets you register N repos once and
then query them as a union.

• New CLI subcommands (codegraph/cli.py):
  • codegraph workspace init — create ~/.codegraph/workspace.yml
  • codegraph workspace add <repo> — register a repo (validates that the
    path exists and is a directory; idempotent)
  • codegraph workspace remove <repo> — deregister (the per-repo
    .codegraph/graph.db is untouched, only the membership)
  • codegraph workspace list — table of registered repos and whether
    each has a built graph
  • codegraph workspace status — per-repo git state: branch, dirty file
    count, last commit subject + timestamp, graph freshness, error notes
  • codegraph workspace sync [--only <name>] — rebuild the graph for
    every registered repo (or just one)
• New module at codegraph/workspace/:
  • config.py — WorkspaceConfig Pydantic model + YAML load/save with
    CODEGRAPH_WORKSPACE_FILE env-var override (useful for tests +
    per-shell workspaces)
  • operations.py — pure functions returning JSON-safe dicts so the CLI
    and MCP server share one implementation
• 3 new MCP tools (codegraph/mcp_server/server.py):
  • workspace_state() — git + graph state for every registered repo
  • workspace_diff_since(ref="main") — cross-repo file changes since a
    git ref; combines committed (<ref>...HEAD) and uncommitted (HEAD)
  • workspace_blast_radius(symbol, depth=None) — unioned blast radius
    across all repos; reuses the existing single-repo blast_radius() so
    behaviour matches exactly per-repo
• Pollution prevention — workspace MCP tools self-load their per-repo
  graphs from the workspace config; the MCP _call_tool dispatch
  bypasses the usual _load_graph() for workspace_* names so the
  server doesn't materialize a stray .codegraph/graph.db in the cwd
  when the user only calls workspace tools.
• Conflict handling — when the same memory id exists on both sides
  with different bodies at the same updated_at, the conflicting remote
  copy is saved as <id>.conflict-<timestamp>.md next to the local
  file for manual merging. Sync never silently clobbers.
• 39 new tests under tests/test_workspace_{config,ops,cli,mcp}.py
  covering config round-trip, real-tmp git-repo state probes,
  CliRunner-based CLI integration, and direct MCP handler invocation.
  Full suite remains green at 576 passing (was 537).
• README updated with a new "Cross-repo Workspace mode" collapsible
  block under CLI subcommands; MCP tools table now lists 18 total (was 15).

v0.3 unified trace — Architecture view shows the real chain

• Per-handler dataflow field on the HLD payload (PR #22). Each route
  entry now carries a dataflow.hops[] array shaped to the v0.3 contract,
  so the dashboard can render the actual frontend → backend → DB chain
  without re-running analysis client-side.
• Phase 4 of the Learn Mode lifecycle modal renders real hops (PR #23).
  The previously-generic "project-specific data layer" phase now shows
  the actual handler → service → repo → model chain across sequence,
  pipeline, and diagram modes. Empty / low-confidence states render
  gracefully with a "no trace data" panel.
• Argument-flow propagation — pick a parameter, watch it travel (PRs #24,
  #25). Per-hop arg_flow: dict[str, str | None] mapping starting keys to
  their locally-renamed names. Frontend renders a chip-strip param picker
  and animates the selected param along the swimlane (sequence) or the
  bezier path (diagram, via SMIL animateMotion). Rename annotations
  ((was userId)) surface where the local name diverges. Snake_case ↔
  camelCase ↔ PascalCase all normalise to the same key.
• ROUTE entry hop args backfilled from handler params (PR #26). The
  trace walker can't supply args at the entry hop (no incoming CALLS edge),
  so URL-template handlers (e.g. /api/users/{user_id}) used to produce
  empty arg_flow. Backfilled from the handler node's DF0 metadata.params
  (skipping self / cls). Closes the launch demo's hero shot.

Cleanup + analyzer hardening

• Examples directory excluded from dead-code / untested analysis
  (PR #19) — examples/cross-stack-demo is documentation, not call-graph-
  traceable code.
• Unused _propagate_class_role_to_members helper deleted (PR #19).
• # pragma: codegraph-public-api analyzer support (PR #21). Functions
  and classes preceded by the pragma comment (Python # or TypeScript
  // style, with optional # codegraph: public-api alias) are exempted
  from find_dead_code(). Applied to all 15 known intentional public-API
  methods (EmbeddingStore facade, SQLiteGraphStore.upsert_node /
  vacuum, all to_dict / as_dict serializers, _register.decorator
  closure). Self-graph dead-code count: 15 → 0, honestly.

Docs & developer experience

• README refresh (PR #20). 15 MCP tools listed; DF1–DF4 + Architecture
  view documented in the headline feature list; "What it doesn't do yet"
  rewritten to drop already-shipped items; numeric stats current.
• SESSION_HANDOFF.md (PR #20) rewritten as a self-contained briefing
  for the next session.
• PLAN_V0_3_UNIFIED_TRACE.md (PR #20) — concrete spec for the unified
  trace work; now marked shipped as of PRs #22–#26.

Cross-stack data-flow tracing (DF0 → DF4)

• DF0 — function signatures + per-call-site arguments. Python and
  TypeScript parsers capture parameter lists, return-type annotations,
  and the literal text of each call-site argument and kwarg.
• DF1 — HTTP routes + SQL data access. FastAPI / Flask / aiohttp
  ROUTE edges; SQLAlchemy READS_FROM / WRITES_TO edges including
  self.session.X repository-pattern detection.
• DF1.5 — role classification. Functions and classes are tagged
  HANDLER / SERVICE / COMPONENT / REPO based on framework patterns.
• DF2 — frontend FETCH_CALL extraction. fetch / axios / useSWR /
  useQuery / api-client patterns emit FETCH_CALL edges with method,
  url, and body-key metadata.
• DF3 — URL stitcher (match_route). Stitches frontend FETCH_CALL
  URLs to backend ROUTE handlers with placeholder normalisation
  (/users/{id}, ${id}, :id, numeric segments) and a body-key
  overlap bonus.
• DF4 — codegraph dataflow trace. Walks the call graph + cross-layer
  edges and emits an ordered DataFlow. Available as a CLI subcommand and
  the MCP dataflow_trace tool.

v0.3 embedding layer

• codegraph/embed/ — chunker + embedder + LanceDB / JSON store.
• codegraph embed CLI. Default model nomic-ai/CodeRankEmbed
  (Apache 2.0, ~140 MB, 768-dim).
• MCP tools semantic_search and hybrid_search. Hybrid reranks...