• Date: 2026-06-09 (manual subprocess execution)
• CGC version: 0.4.16 — tested on PyPI (pip install codegraphcontext) and local editable (pip install -e . from this repo, uncommitted working-tree changes)
• Python: 3.12.3
• OS: Linux 6.8.0-124-generic
• Method: Subprocess-only per E2E plan. No source modifications. Tests run from /tmp with isolated HOME unless testing repo-local config bleed.
• Harness note: Initial automated harness (/tmp/cgc_e2e_harness.py) produced many false positives (truncated Rich output, missing bundle import -y, config bleed when cwd inside CGC repo). Findings below are manually verified.

Verdict (PyPI): Core indexing works, but call-chain analysis and golden parity regress vs repo goldens. Config bleed when cwd is inside a repo with .codegraphcontext/.

Verdict (local editable): Significantly better — call chains work, 20/20 golden parity exact match, cypher deprecation warning present. Several PyPI bugs appear fixed in the working tree but not yet released. Remaining local issues: delete orphans, per-repo FalkorDB default, API /health 404, CLI help gaps.

Golden baseline for Python sample: 482 nodes, 621 edges in tests/fixtures/goldens/sample_project/metadata.json; PyPI measured 482 nodes, 625 edges (0.6% edge delta — within tolerance).

• Severity: High
• Category: UX / Config
• Backend(s): All
• Repro steps:

  export HOME=$(mktemp -d)
  cgc config db neo4j
  cgc config set NEO4J_URI bolt://127.0.0.1:19999
  cd /path/to/repo/with/.codegraphcontext   # e.g. CGC repo itself
  cgc config show   # NEO4J_URI reverts to bolt://localhost:7687 from repo .env

• Expected: User's ~/.codegraphcontext/.env values take precedence, or repo-local overrides are clearly documented and scoped to per-repo mode only.
• Actual: Repo-local .codegraphcontext/.env merges and wins for DEFAULT_DATABASE, NEO4J_URI, etc. even in global mode with a fresh HOME. Index/find connect to the repo's FalkorDB graph (53k+ functions in dev tree) instead of the isolated test database.
• Impact: New users cloning CGC (or any repo with .codegraphcontext/) get silently redirected to the repo's database. E2E tests, CI, and multi-project workflows produce wrong results without obvious error.

• Severity: High
• Category: Accuracy
• Backend(s): All (parser / call resolver)
• Repro steps:

  export HOME=$(mktemp -d) && cd /tmp
  cgc config db falkordb && cgc context mode global
  cgc index --force tests/fixtures/sample_projects/sample_project
  cgc query "MATCH (a)-[r:CALLS]->(b) WHERE a.path CONTAINS 'function_chains' RETURN a.name, b.name, r.line_number"

  Fixture: function_chains.py line 5: result = f1(f2(f3(10)))
• Expected: CALLS edges f1→f2, f2→f3 (or equivalent nested-call representation).
• Actual: Only module-level edges: <module>→f1, <module>→f2, <module>→f3 (all line 5). Zero Function→Function CALLS for this pattern.
• Impact: cgc analyze chain f1 f3 returns "No call chain found". Docs/quickstart use this fixture as the canonical call-chain demo.

• Severity: High
• Category: Accuracy
• Backend(s): falkordb, kuzudb, ladybugdb (verified)
• Repro steps: After indexing sample_project: cgc analyze chain f1 f3
• Expected: Chain f1 → f2 → f3
• Actual: No call chain found between 'f1' and 'f3' within depth 5 (exit 0)
• Impact: Primary onboarding example broken for nested module-level calls.

• Severity: Medium
• Category: Accuracy / UX
• Backend(s): All
• Repro steps: Index sample_project → cgc analyze callers f2
• Expected: Caller f1 (per intuitive reading of f1(f2(f3(10)))).
• Actual: Single caller <module> at function_chains.py (technically correct given BUG-002, but contradicts user/docs mental model).
• Impact: Misleading caller analysis for module-level call expressions.

• Severity: Medium
• Category: Accuracy / Data integrity
• Backend(s): falkordb (verified)
• Repro steps:

  cgc index --force sample_project_c   # 83 nodes
  cgc delete sample_project_c
  cgc query "MATCH (n) RETURN count(n)"   # returns 9, not 0

• Expected: All nodes belonging to deleted repository removed.
• Actual: 9 orphan nodes remain (Parameters, Variables without paths — e.g. x, y, int v).
• Impact: Repeated index/delete cycles accumulate stale nodes; inflates stats and can cause golden parity drift in batch sweeps.

• Severity: Medium
• Category: UX
• Backend(s): All (MCP)
• Repro steps: MCP tools/call → delete_repository with {"path": "../../../etc"}
• Expected: Clean validation error (path not found / not indexed).
• Actual: Tool execution error: Failed to delete repository: 'NoneType' object has no attribute 'replace'
• Impact: Scary internal error instead of actionable message; suggests missing null guard on repository path resolution.

• Severity: Medium
• Category: UX / Config
• Backend(s): kuzudb (verified)
• Repro steps:

  cgc config db kuzudb && cgc context mode per-repo
  cd sample_project_go && cgc index --force .
  # Observe: "Context: Per-repo local mode (Database: falkordb)"

• Expected: Per-repo context inherits configured backend (kuzudb) or prompts user.
• Actual: Auto-created .codegraphcontext/ defaults to falkordb regardless of global cgc config db kuzudb.
• Impact: User selects KuzuDB globally but per-repo mode silently uses FalkorDB Lite.

• Severity: Low
• Category: UX / Docs
• Backend(s): N/A (API)
• Repro steps: cgc api start --port 8020 → curl http://127.0.0.1:8020/health
• Expected: Health check endpoint (common convention).
• Actual: HTTP 404. Working endpoint: GET /api/v1/status → {"status":"ok","message":"Connected",...}
• Impact: Load balancers / k8s probes using /health fail unless docs specify /api/v1/status.

• Severity: Low
• Category: UX / Docs
• Backend(s): All
• Repro steps: cgc cypher 'MATCH (f:Function) RETURN f.name LIMIT 1'
• Expected: Deprecation warning directing users to cgc query.
• Actual: Executes silently, returns results (exit 0).
• Impact: Deprecated command persists undetected.

• Severity: Low
• Category: Docs
• Backend(s): N/A
• Repro steps: cgc index --help — check --database enum
• Expected: Includes nornic, falkordb-remote per config_manager.py.
• Actual: Help text omits nornic and falkordb-remote.
• Impact: Users cannot discover remote backends from CLI help.

• Severity: Low
• Category: UX
• Backend(s): All
• Repro steps: cgc bundle import file.cgc --clear (no TTY)
• Expected: Non-interactive failure with clear message, or require -y in docs prominently.
• Actual: Hangs at Are you sure you want to continue? [y/N]: until timeout.
• Impact: CI/scripts hang; --yes/-y exists but is easy to miss.

• Severity: Medium
• Category: Accuracy / Release parity
• Backend(s): falkordb (PyPI 0.4.16)
• Repro steps: Fresh isolated index per language; compare MATCH (n) count to tests/fixtures/goldens/*/metadata.json.
• Expected: Node counts within 10% of goldens (same generator PYv0.4.16).
• Actual (fresh single-repo index):

• Note: Edge counts match exactly for all tested languages; node inflation is predominantly extra Parameter and Variable nodes. Suggests PyPI 0.4.16 indexes more node types than goldens captured, or goldens were generated from a different build artifact.
• Impact: Golden-based CI on PyPI install will fail for 11/20 languages on node count despite correct edge topology.

• CLI --database help missing nornic, falkordb-remote (BUG-010)
• MCP tool count: docs reference 21 tools; runtime exposes 25
• Quickstart call-chain example (f1→f2→f3) does not work with current call resolver (BUG-002/003)
• API health probe should document /api/v1/status not /health
• NEO4J_USER is not a valid cgc config set key; use NEO4J_USERNAME (discovered during exit-code testing)
• Config precedence: local .codegraphcontext/.env at cwd overrides global — underdocumented for global-mode users (BUG-001)

*Edge counts inferred from shared indexing pipeline; node counts verified via Cypher.

Local install: /tmp/cgc-local-venv via pip install -e . from commit dce24ed + 9 uncommitted files (cli_helpers.py, database_falkordb.py, database_kuzu.py, falkor_worker.py, api/router.py, scip_indexer.py, etc.).

The working tree contains fixes not in the published PyPI 0.4.16 wheel:

1. Nested/module-level call resolution (f1→f2→f3 CALLS edges)
2. Exact golden parity (edge count 621 vs PyPI's 625 on sample_project)
3. Cypher deprecation warning
4. Improved config / DB-init error handling (no silent repo-graph bleed)

Action: Cut a new PyPI release (0.4.17+) to ship these fixes to new users installing via pip install codegraphcontext.

1. BUG-001: Document config precedence clearly; consider not loading repo-local .env in global mode unless --context or per-repo mode is active.
2. BUG-002/003: Fix call resolver for nested call expressions at module scope, or update docs to use intra-function call examples.
3. BUG-005: Cascade-delete orphan Parameter/Variable nodes when removing a repository.
4. BUG-012: Regenerate language goldens from PyPI 0.4.16 wheel, or pin golden generator to exact release artifact hash.

Report generated by manual E2E bug hunt session 2026-06-09. Local editable comparison added same day.