2026-11-05 16:22:12 +08:00

79 lines
3.6 KiB
Markdown

# Linux setup
Follow the installation and preparation sequence in [README.md](../../README.md#quick-start).
## Environment
Use Linux x86-64, Python 3.11, Git, and Docker Engine accessible to the current
user. The host environment runs the framework; the runtime Dockerfile supplies
tool-side Python, Node, Go, and Rust dependencies. Install Harbor in the separate
Python 3.12+ environment described in [BENCHMARKS.md](BENCHMARKS.md).
The reference embedding deployment needs an NVIDIA GPU supporting BF16 and may
run on a separate host. The Transformers service does not require Docker or
vLLM; the alternative container service requires NVIDIA Container Toolkit.
TB tasks requesting GPUs also require
NVIDIA Container Toolkit on the task host. CPU, RAM, GPU, and image requirements
come from each task's `task.toml`. Reserve additional capacity for concurrency
and image storage.
## Git checkout
For a ZIP download, initialize Git at the extracted **repository root** (the
directory containing `README.md` and the `execorchestra/` subdirectory), before
creating local configuration:
```bash
git -c init.templateDir= init -b execorchestra
git config --local core.autocrlf false
git config --local user.name Anonymous
git config --local user.email anonymous@example.invalid
git config --local commit.gpgsign false
git add .
git commit -m 'Initialize ExecOrchestra source'
cd execorchestra
```
The framework records the checkout identity. Commit intended source changes
before preparation. `.gitattributes` preserves prompt and resource bytes.
## Configuration
| File | Purpose |
| --- | --- |
| `.env` | Model credentials and reachable service/proxy URLs. Export it in every preparation or execution shell. |
| `sgar_mvp/config.json` | Generated by `init`; execution and recovery settings. |
| `sgar_mvp/config/embedding_release.json` | Generated by `init`; set the endpoint, API model ID, and tokenizer path. |
| `sgar_mvp/config/control_role_policy.json` | Control models and reasoning settings. |
| `sgar_mvp/config/candidate_pool.json` | The 57 configured research candidates: 10 models, 38 tools, 4 skills, and 5 agents. |
| `sgar_mvp/config/model_selection.json` | Registered model identities used by health and control stages. |
| `sgar_mvp/config/model_cost_policy.json` | Cost monitoring policy. |
| `sgar_mvp/config/retrieval_policy.json` | Generated by `init` and bound to the new index during preparation. |
The Composer selects execution models for agents. Model IDs, resource membership,
prompt bytes, and policy settings define the reference configuration.
[SERVICES.md](SERVICES.md) gives the required interfaces and deployment commands.
## Preparation outputs
Preparation creates the runtime image lock, current model/tool/agent readiness
records, control probe receipt, and `Pool/index_meta/`. Task-container tools are
checked in their benchmark environment. Full-pool readiness stops if a required
resource is unavailable; `readiness --allow-unavailable` explicitly selects a
reduced pool.
If the endpoint, source, resources, or configuration changes, refresh the
affected checks and rebuild the index. To rebuild, move `Pool/index_meta` to a
private backup, then repeat `readiness`, `index`, and `check`.
## Execution and accounting
For task files, pass `--input NAME=PATH` and declare the corresponding public
input root. Use a new run directory. `python -m sgar_mvp.main --help` lists the
invocation options.
The default cost policy is `monitor`; deadlines and recovery limits are
independent. Accounting covers control, execution, and recovery, using the
configured input, output, and cache prices. Benchmark commands and scoring
outputs are in [BENCHMARKS.md](BENCHMARKS.md).