# Host support

This matrix describes evidence from the exact versions tested for this release candidate. It is not a claim of universal compatibility. Every adapter reports capabilities from a live probe and must degrade when a capability disappears.

## Qualification levels

- **Runtime verified**: the shared conformance suite passed and an isolated one-shot host invocation passed without changing managed files.
- **Contract verified**: the shared conformance suite passed, but no model prompt was executed in that host during this qualification.
- **Degraded export**: deterministic instruction files are produced for manual transfer. Host hooks are not implied.

## Tested matrix

| Host | Exact tested version | Qualification | Important limits |
| --- | --- | --- | --- |
| Codex CLI | 0.144.5 | Contract verified | Version and tool discovery probes passed in a synthetic `CODEX_HOME`. The common five-case suite passed. A model prompt was not executed because the isolated home had no auth and copying live credentials was forbidden. Lifecycle hooks without probe evidence are reported as degraded. |
| Aider + Ollama | Aider 0.86.2; Ollama 0.32.0; `ollama_chat/qwen2.5-coder:7b` | Runtime verified | The common five-case suite and a synthetic one-shot prompt passed. Managed files remained byte-identical and no Git repository was created. Aider has no Agent OS compact/stop hook contract, embedded memory provider, or embedded graph provider. |
| Generic filesystem/instruction export | `filesystem-contract-v1` | Degraded export | Deterministic `AGENT_OS.md` and capability JSON are produced. Session, prompt, and resume are manual. Lifecycle hooks are unsupported and tool discovery is unknown. |

## Capability summary

| Capability | Codex CLI 0.144.5 | Aider 0.86.2 + Ollama 0.32.0 | Generic contract v1 |
| --- | --- | --- | --- |
| `session_start` | Degraded without discovered hook | Supported while executable is available | Degraded when export target is writable |
| `prompt_submit` | Degraded without discovered hook | Supported while executable, loopback endpoint, and configured model are available | Degraded manual transfer |
| `resume` | Degraded without discovered hook | Degraded, adapter does not manage history | Degraded manual transfer |
| `pre_compact`, `post_compact`, `stop` | Supported only with discovered hooks; otherwise degraded | Unsupported | Unsupported |
| `tool_discovery` | Supported in the tested probe | Unsupported | Unknown |
| `memory`, `graph` | Provider-dependent; unsupported in this qualification | Unsupported | Unsupported |
| `external_actions` | Unsupported until core approval is supplied | Unsupported until core approval is supplied | Unsupported until core approval is supplied |

## Evidence boundary

Generated conformance reports are deliberately excluded from source distribution. The exact local
qualification report hashes are bound by the external release-candidate manifest, not copied into
this source document. A fresh Aider run produces a fresh raw-log digest even when the bounded
capability and runtime assertions are unchanged.

The Aider probe stored raw model output only in its synthetic task workspace. The public-shaped report contains the result flags and a raw-log hash, not the output itself. The harness verifies the complete synthetic project tree after execution, but it is profile isolation, not an operating-system sandbox.

## Reproducing the non-Codex proof

Install Aider in an isolated environment, use an already available local Ollama model, and run `tools/run_aider_conformance.py` with an empty synthetic workspace. The harness accepts only a loopback HTTP endpoint, rehomes profile and temp directories, removes credential-like environment variables, disables Git and auto-commits, and fails if managed files change.

The adapter follows Aider's documented [`ollama_chat/<model>` convention](https://aider.chat/docs/llms/ollama.html) and [configuration options](https://aider.chat/docs/config/options.html). The exact package used here is documented on the [official `aider-chat` PyPI page](https://pypi.org/project/aider-chat/).