Tags: hexo-ai/sia
Tags
feat: add focus argument for RL-based tuning and integrate RL Integra… …tion Guide and orchestration (#24) * feat: add focus argument for RL-based tuning and integrate RL Integration Guide and orchestration * Merge main into weights: complete weights/RL mode integration * fix: update prompt formatting and reference handling in prompts.py * fix: update prompt formatting and reference handling in prompts.py * docs: clarify training sandbox options and weights mode execution details in configuration.md * remove unwanted file * docs: update training Guide in prompts.py for clarity and consistency * chore: bump version to 0.5.0 and add claude-agent-sdk dependency * feat: enhance setup_run_directory with focus argument for conditional package installation * lint fix * lint fix * test fix.
feat: web visualizer for runs/ directory (sia web) (#19) * feat: web visualizer for runs/ directory (sia web) Add a FastAPI dashboard to browse local runs: trajectories, target-agent code, prompts, improvement plans, logs, and per-generation scores. - sia/web/runs.py: dependency-light data layer over runs/run_N/gen_M/ with path-traversal hardening (pydantic models, pure/testable functions). - sia/web/server.py: FastAPI app factory + serve()/serve_in_background(). - sia/web/static/index.html: dependency-free SPA (run list, accuracy chart, per-gen tabs, trajectory chat viewer, Python syntax highlighting, markdown-rendered prompts). Diff-based routing to avoid re-render flicker. - CLI: add `sia web` subcommand; `sia run` auto-starts a live dashboard in a background thread (--no-web/--web-port/--web-host). `sia --task ...` still works via a default `run` subcommand. - pyproject: `web` extra (fastapi, uvicorn); ship web/static as package-data. - tests: data-layer, API, and CLI-interface coverage. * docs: simplify inline comments, drop refactor-history notes * docs: document run/web subcommands + evaluation; fix welcome banner - README/configuration/walkthrough: switch examples to `sia run`, document the `sia web` visualizer and the live dashboard auto-started during runs. - README: add an Evaluation section (how per-generation scoring feeds the loop) linking EVALUATION_GUIDE.md. - orchestrator: print the welcome banner before arg parsing again, so it shows on `sia` / `sia run` (regressed when the subcommand dispatch moved it after parse_args). * fix: make web dashboard a core dependency; bump to 0.3.0 The live run dashboard is started by default on every `sia` run (orchestrator.main -> serve_in_background, unguarded), so fastapi / uvicorn / pydantic are not optional — a default `pip install sia-agent` would crash with ModuleNotFoundError on a normal run, and CI's `.[dev]` test job failed collecting tests/test_web.py for the same reason. - Move fastapi + uvicorn into core dependencies, add pydantic explicitly (runs.py imports BaseModel directly, not just via fastapi). - Drop the now-obsolete `web` optional extra and its install docs. - Replace importorskip("fastapi.testclient") with a plain import so a missing dep fails loudly instead of silently skipping. - Bump version 0.2.1 -> 0.3.0 for the web visualizer feature. * test: add httpx to dev deps for fastapi TestClient fastapi's TestClient (starlette.testclient) needs an httpx client, which fastapi does not pull in itself, so tests/test_web.py::test_api_endpoints failed in CI with "requires the httpx2 package". starlette 1.2.x prefers httpx2 but falls back to plain httpx, so the standard httpx satisfies it without depending on the non-standard httpx2 fork. --------- Co-authored-by: Selvam Palanimalai <selvam.palanimalai@gmail.com>