Skip to main content

Local agentic control plane for gpt2giga

Project description

gpt2giga-harness

Local agentic control plane for the gpt2giga compatibility gateway. The distribution provides the giga and gpt2giga-harness commands, the gpt2giga_harness Python namespace, a durable local worker, and the packaged Project Cockpit web UI.

Alpha status: 0.0.x storage, API, adapter, and automation contracts may change between releases. Use supervised local workflows first. The package metadata in the current checkout is the source of truth for the exact gateway dependency.

Install the current preview

Until the standalone Harness release is available in your package index, run it from a source checkout:

git clone https://github.com/ai-forever/gpt2giga.git
cd gpt2giga
git switch feature/harness_enrichment
uv sync --all-packages --all-extras --dev
source .venv/bin/activate
giga doctor
giga ui

Keep that environment active when you cd to the project you want to manage. On Windows PowerShell, activate .venv\Scripts\Activate.ps1 instead.

After publication, the installation will be:

uv tool install --prerelease allow gpt2giga-harness
giga doctor
giga ui

The current gpt2giga-harness==0.0.1a3 metadata depends exactly on gpt2giga==0.2.3a2. Installing only gpt2giga never adds Harness commands or the gpt2giga_harness namespace.

First local run

From a disposable project directory:

giga init
giga doctor .
giga doctor . --json
giga ui

The doctor reports ready, degraded, and blocked checks for the local proxy and routes, configured models, adapter CLI versions, workspace and Git, the durable worker, Harness-managed homes, and managed MCP snapshots. JSON evidence is redacted and omits absolute workspace paths; every degraded or blocked first-run prerequisite includes a safe remediation command. Runtime worker state is inspected read-only.

For a credential-free tour, copy the first-run demo. It keeps runtime state inside the disposable copy and verifies giga init, the local Echo read path, and the generated smoke eval without starting a proxy or an external agent CLI.

For the first model-backed north-star flow, copy the issue-to-reviewed-patch example. It packages reviewed agent profiles, a durable isolated-worktree workflow, and a post-apply eval while keeping apply, commit, push, and hosted writes explicit.

For unattended compatibility evidence, copy the nightly compatibility guardian. It packages a pinned Codex/Claude/Gemini eval, exact baseline dimensions, a read-only triage workflow, and a durable schedule that runs without the UI and raises Attention only after a tested contract regresses.

For parallel evidence-backed review, copy the cross-harness review team. It fans out read-only explorer, security, tests, and maintainability roles across Codex, Claude, and Gemini, then synthesizes retained child artifacts without introducing a shared writable workspace.

Open http://127.0.0.1:8091/. giga ui starts a durable local worker when needed; pass --no-start-worker when another supervisor owns that lifecycle. The default loopback binding is intentional. Remote binding requires explicit authentication and remote-access opt-in.

After the first run starts, Work reveals a compact Run → Evidence → Approval / worktree → Reuse / automation path. Once the run reaches a terminal state, Open evidence deep-links to that exact run's retained trace in Runs Center; prompts, responses, and workspace paths are not copied into the transition summary. For a retained isolated patch, Review worktree opens that run's Diff inspector, where apply and approval remain separate explicit operator actions. A successful run with no pending worktree review enables Reuse run, which opens the exact run's existing provenance/promotion inspector. Promotion preview/apply and later scheduling remain separate explicit actions.

Useful orientation commands:

giga config path
giga project info
giga harness list
giga harness inspect codex-cli --json
giga session list
giga native list
giga worker status

Built-in adapters

  • Direct Chat through the local gateway;
  • Codex CLI;
  • Claude Code;
  • Gemini CLI;
  • Echo for deterministic smoke tests.

External CLIs are discovered on PATH. Non-standard absolute paths belong in the user-owned ~/.gpt2giga/harness/config.toml:

[executables]
"codex-cli" = "/custom/bin/codex"
"claude-code" = "/custom/bin/claude"
"gemini-cli" = "C:\\Users\\me\\bin\\gemini.cmd"

Manage these values without editing TOML directly:

giga config set executables.codex-cli /custom/bin/codex
giga config unset executables.codex-cli
giga harness inspect codex-cli --json

Configured paths take precedence over PATH, must be absolute, and are checked for executable capabilities before a native or headless run starts.

Storage and safety boundaries

  • User-level coordination state: ~/.gpt2giga/harness/;
  • project configuration and references: <project>/.giga/;
  • vendor-owned Codex, Claude, and Gemini homes: read or imported only through adapter-specific contracts; Harness must not rewrite them;
  • edit runs: isolated worktrees with lease and policy checks;
  • risky native spawn, remote access, apply, and secret operations: explicit approvals.

The control plane stores redacted metadata by default. Content capture and secret materialization are opt-in and should have explicit access, retention, and cleanup policies.

Plugin contract

Third-party adapters register through the gpt2giga.harnesses entry-point group, while import targets live outside the gateway namespace:

[project.entry-points."gpt2giga.harnesses"]
my-harness = "my_package.my_harness:MyHarness"

Use giga harness scaffold, giga harness validate, and giga harness inspect to develop and diagnose an adapter.

Upgrade from the combined prerelease

Uninstall both old distributions before installing the split packages. Do not delete ~/.gpt2giga/harness/ or project .giga/ state during the package migration:

uv tool uninstall gpt2giga
uv tool uninstall gpt2giga-harness
uv tool install --prerelease allow gpt2giga-harness
giga doctor

Documentation

Package verification

From the repository root:

uv sync --all-packages --all-extras --dev
uv run pytest tests/harness -n 4 -q
uv build --package gpt2giga-harness --no-sources

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

gpt2giga_harness-0.0.1a3.tar.gz (489.5 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

gpt2giga_harness-0.0.1a3-py3-none-any.whl (547.2 kB view details)

Uploaded Python 3

File details

Details for the file gpt2giga_harness-0.0.1a3.tar.gz.

File metadata

  • Download URL: gpt2giga_harness-0.0.1a3.tar.gz
  • Upload date:
  • Size: 489.5 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: uv/0.11.28 {"installer":{"name":"uv","version":"0.11.28","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}

File hashes

Hashes for gpt2giga_harness-0.0.1a3.tar.gz
Algorithm Hash digest
SHA256 939c9ec1a01fc52c0eeae33dd5d6f90f9cf329d03fde9e50f7a2536f820a8688
MD5 ffe5c1e6ed9d247134df3ea2d7b91959
BLAKE2b-256 749e99892cea513e1b0c96d4ac144bc1f334faa027e25419d2391270266d7a76

See more details on using hashes here.

File details

Details for the file gpt2giga_harness-0.0.1a3-py3-none-any.whl.

File metadata

  • Download URL: gpt2giga_harness-0.0.1a3-py3-none-any.whl
  • Upload date:
  • Size: 547.2 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: uv/0.11.28 {"installer":{"name":"uv","version":"0.11.28","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}

File hashes

Hashes for gpt2giga_harness-0.0.1a3-py3-none-any.whl
Algorithm Hash digest
SHA256 491be800dea90e0afbf10674705c4500843e28997f74c76fef119f137cfa0218
MD5 7f26334670488ce0d705e59ed8889968
BLAKE2b-256 3578243e8153749985037444595a7df16c610edd9b14cab7cc9d036292140402

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page