Skip to content

Implementation validation — 2026-09-05–06

Validated locally on macOS with Python 3.12.12:

  • 2,066 existing regression tests passed, with 5 skips and 1 expected failure, covering code intelligence, extracted context modules, and the MCP tool audit.
  • 24 standalone product tests passed, including workspace separation, file freshness, local/shared knowledge, repository authorization, real stdio/HTTP MCP, resource access, response budgets, idle workspace eviction, remote installation validation, and concurrent AST hydration.
  • A CI rerun exposed a project-summary race while background hydration added AST entries. Analysis now reads a cache snapshot. A deterministic regression adds a file from another thread during summary iteration, verifies that the current response stays consistent, and verifies that the next response includes the newly indexed file.
  • All 20 schema migrations applied to a fresh PostgreSQL 16/pgvector container. The 21 product tests present at that validation point passed against PostgreSQL; the subsequently added repository-source validation test passed with the full 22-test suite. The disposable container was stopped and removed.
  • The wheel installed into an isolated environment without the attocode distribution. An import guard rejected agent, swarm, TUI, Anthropic, and OpenAI SDK imports; navigation and bundled rules still worked. The clean installation resolved MCP 1.29.1; the main regression environment used MCP 1.26.0.
  • Generated Codex, Claude Code, and Cursor configurations passed doctor, each opening a real stdio MCP session and querying the selected workspace. The fixture path contained spaces. User-level client configurations were not changed.
  • Actual Codex CLI 0.153.4 and Claude Code 2.1.212 sessions called the standalone wheel's symbols tool, returning a fixture function name that was absent from the prompt. Codex recovered from an initial invalid search_symbols call; Claude discovered the tool through ToolSearch. The sessions used temporary project configurations and existing account authentication. Cursor desktop 3.17.8 is installed, but its headless CLI is unavailable; an actual Cursor agent session remains unverified.
  • The dashboard production build passed. The existing bundle-size warning remains.
  • The Docker image built on Linux ARM64 and passed its runtime smoke test: health endpoint, dashboard HTML, MCP authentication, and real code navigation. Embedding dependencies are opt-in; the default image installs no CUDA packages. The disposable smoke-test container was removed.
  • The legacy agent CLI help command passed. Original agent documentation and compatibility import paths remain available.
  • New product modules pass the scoped Ruff check; all standalone Python files compile. The extracted legacy code still has existing style findings under a whole-package Ruff run.

The synthetic 10,000-source-file run discovered 10,001 files including project configuration and recovered all 9,999 expected dependency edges. Cold bootstrap took 7.32 s, a focused lookup 246 ms, and warm bootstrap 6.68 s. Both bootstrap responses contained 1,052 tokens under a 2,000-token budget. Raw results and workload caveats are in packages/code-intel/evals/results/.

The first GitHub Intelligence run passed on Linux for commit 09aaeca: Python 3.12 and 3.13 each applied all migrations, passed the complete 23-test product suite against PostgreSQL, passed the navigation and catalog gates, and verified an isolated wheel. The dashboard and Docker runtime jobs also passed.

The GitHub pypi-intelligence environment is configured to accept only intel-v* tags. Publishing requires a matching PyPI trusted publisher and passes through the complete reusable intelligence workflow. No package has been published or service deployed at this validation point. See the release procedure for the account setup and verification steps.