Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 20 additions & 6 deletions .github/AGENTS.md
Original file line number Diff line number Diff line change
Expand Up @@ -27,15 +27,29 @@ events so correcting a declaration can rerun CI.

## Installed-Wheel Acceptance

The acceptance script installs a supplied wheel offline into a temporary venv
with system dependencies available, and verifies that `pt_snap_cli` resolves
inside that venv. It clears source/catalog/focus environment overrides and uses
temporary HOME/CWD for helper installation and import-created focus.
The acceptance script audits a supplied wheel against source and validates the
reviewed fixture manifest before installing or importing. It installs offline
into a temporary venv with system dependencies available by default. Supplying
`--wheelhouse` instead installs dependencies offline into a venv without system
packages and verifies their installed paths too. It verifies that `pt_snap_cli`
resolves inside the venv, clears source/catalog/focus/resource-limit overrides,
disables user site packages, and uses temporary HOME/CWD for skill installation
and import-created focus.

It then checks help, bundled catalog/install status, capabilities, trusted fixture
import/reuse, overview and a peak query. This is executable CLI acceptance, not a
live model evaluation. The script accepts only the expected committed
`snapshot_with_empty_cache.pkl` path; run fixture provenance verification first.
live model evaluation. `--events-per-slice 4096` adds native-v2 and compatibility-v1
sharded imports/reuse, directory/manifest discovery, installed helper/peak skill
guidance, and real peak/cross-shard event parity with the standalone CLI. It
requires at least two physical slices on the selected device and verifies that
explicit dataset operations do not change focus. The script accepts only the
expected committed `snapshot_with_empty_cache.pkl` path, rejects an unhydrated
pointer, and runs the provenance gate itself; also run fixture provenance tests
before acceptance. Its output explicitly distinguishes offline CLI execution,
stored-baseline grading (not run here), and live-agent execution (not run).

Harness changes use `pytest tests/test_clean_wheel_agent_acceptance.py` plus an
actual audited-wheel invocation; mocked guard tests alone are not acceptance.

## Change Together

Expand Down
Loading
Loading