[Pepps233]

Summary

Add post-merge behavior testing for the exact packaged Diffex CLI in fresh Daytona sandboxes. Deterministic npm/Bun scenarios exercise the installed CLI, real TUI, tools and sessions; a separate bounded OpenAI smoke verifies the live transport.

Automatic main-branch runs remain disabled until GitHub Actions qualification passes. This PR does not merge itself, enable spending, change release gates or revert failed merges.

Reasoning

• Reuse the existing source/model freeze, offline build, packaging and checksum pipeline so acceptance tests consume reviewed package bytes outside the workspace.
• Control inference and built-in auth transport for deterministic coverage without replacing application sessions, tools or UI implementations.
• Keep deterministic behavior, live-provider availability, configuration, evidence integrity and Daytona cleanup outcomes distinguishable. Keep the harness independent of Harbor and Terminal-Bench.
• Protect secrets explicitly: this repository's Actions artifacts are public-readable to authenticated users with repository read access, so they are not confidential storage.

Changes

• Add an exact-SHA push: main / manual-dispatch workflow, independent deterministic/live jobs, aggregate validation and 30-day artifact retention, without path filters or cross-run cancellation.
• Extract shared installation helpers while preserving existing artifact acceptance coverage; add print, streaming terminal, read/write/edit/bash, cancellation/recovery, restart/resume and reviewed 80×24 / 120×40 / resize scenarios.
• Add a toolchain-only snapshot definition: Node 22.23.2, npm 11.16.0, Bun 1.3.11, Python, tmux, ripgrep and fd; 2 vCPU / 4 GiB RAM / 8 GiB disk and a 45-minute wall-clock sandbox TTL.
• Add internal run, recover and prepare-snapshot commands with recorded ownership, ambiguous-creation handling, incremental evidence, complete hash/JUnit validation and independently confirmed deletion. Pin the standalone Python Daytona SDK to 0.226.1.
• Add a fail-closed all-files publication guard. Only a completed, verified sanitized copy can upload; credential findings fail the job, authentication files/symlinks are excluded, and hidden evidence members remain retained. Credentials are not embedded in commands, snapshots, fixtures, commits or reports.
• Document provisioning, qualification and actual real-service CLI results, including incompatibilities discovered and fixed during qualification.

Testing

• npm run check — passed.
• ./test.sh — full credential-isolated non-E2E suite passed.
• python -m unittest discover -s scripts/daytona-behaviors -p 'test_*.py' in the pinned SDK environment — 59 passed.
• node --test scripts/artifact-acceptance.test.mjs scripts/daytona-behaviors-workflow.test.mjs scripts/daytona-behaviors/fixture.test.mjs scripts/daytona-behaviors/live-guard.test.mjs — 29 passed.
• Fresh standalone 0.4.0 packages installed locally with npm and Bun, lifecycle scripts disabled — deterministic installed scenarios and terminal baselines passed.
• Real Daytona qualification — 13 deterministic scenarios and 3 live scenarios passed against the same retained candidate and corrected bundle. Final live smoke used 4 requests, 98 output tokens and approximately $0.00205 in reported provider usage.
• Intentional baseline failure, active SIGTERM, runner-loss recovery and two overlapping CLI runs behaved as expected; a fresh independent audit confirmed deletion of all 10 qualification sandboxes.
• Final local publication scan — 955 files, zero exact matches for either real credential. Committed patch history was separately scanned before pushing; no credentials or local artifacts are included.
• git diff --check origin/main...HEAD — passed; terminal baseline EOF rows are explicitly preserved as screen geometry.

This pull request description is AI-generated.

[github-advanced-security]

CodeQL found more than 20 potential problems in the proposed changes. Check the Files changed tab for more details.