AgentGuard v0.2.2 demo recording¶
This directory contains the silent Phase 43D product recording, its English WebVTT captions, and a SHA-256 manifest. The recording uses real terminal I/O, real deterministic AgentGuard evidence, and maintained sanitized project screenshots. It does not use an external coding agent or model API.
Source and evidence¶
- Recording source commit:
2e73384e3e8ebb5862cef68ea7073685fe3ad6cb - AgentGuard product version:
0.2.2 - Installed PyPI distribution:
agentguard-evals==0.2.2 - Python import and console command:
agentguard - Deterministic result: 5/5 unsafe showcase scenarios detected; 1/1 safe showcase scenario allowed
- Static-report inputs: six showcase runs, their suite record, and one separate audit-mode command-guard incident
The released package was installed into a fresh temporary virtual environment before capture. These commands produced the visible terminal evidence:
agentguard --version
agentguard --help
scripts/showcase_demo.sh
.venv/bin/python scripts/showcase_metrics.py --check
The showcase command ran during the terminal recording. An equivalent static report was generated immediately before assembly with:
.venv/bin/python scripts/showcase_demo.py \
--suite examples/showcase/showcase.yaml \
--output-dir artifacts/showcase
.venv/bin/python -m agentguard.cli.main run \
examples/showcase/configs/unsafe_command.yaml \
--agent local-command --guard-mode audit --allow-fail-result
.venv/bin/python -m agentguard.cli.main reports site \
--output site --history-db .agentguard/history.db \
--reports-root .agentguard \
--title "AgentGuard Evaluation Showcase" --force
The report scenes reuse the maintained, sanitized dashboard and incident
screenshots introduced by Phase 43C. Those images were captured from equivalent
v0.2.2 deterministic output at source commit
150709b3b790aba5c33d1918f00184d46868dd2e; their adjacent source record
documents the original capture. Equivalent output was regenerated and checked
at the recording commit before video assembly.
Capture and encoding¶
- Terminal recorder: Asciinema 3.2.1, headless pseudo-terminal capture
- Terminal renderer: agg 1.9.0,
github-darktheme - Encoder and validator: FFmpeg/ffprobe 8.1.2 with libx264
- Terminal size: 108 columns × 34 rows
- Screenshot source viewport: 1440×900; incident source: 1440×1055
- Final video: MP4, H.264 High profile,
yuv420p, 1280×720, 24 fps - Audio: none
- Duration: 80.333 seconds
- File size: 803,313 bytes
- SHA-256:
be74d032cb376576163b6ef635ae5d1961a12245faf7913d73bc820c42b1b832
Equivalent encoding uses a task-specific capture directory and these settings:
agg --theme github-dark --font-size 18 --line-height 1.15 \
--fps-cap 10 --no-loop --idle-time-limit 20 \
--last-frame-duration 5 terminal.cast terminal.gif
ffmpeg -f concat -safe 0 -i segments.txt -map_metadata -1 -an \
-c:v libx264 -preset slow -crf 24 -pix_fmt yuv420p -r 24 \
-movflags +faststart agentguard-v0.2.2-demo.mp4
Terminal output, deterministic screenshot scenes, and fades were first encoded as matching 1280×720, 24 fps, silent H.264 segments. The incident scene uses a slow vertical pan over the existing full-page sanitized capture. The final concat encode strips inherited metadata and places the MP4 index before media data for browser seeking.
Sanitization and accessibility¶
- The terminal capture used a fixed neutral surface and did not capture shell input, a personal prompt, username, hostname, working-directory prompt, environment dump, cache output, or shell history.
- Only the bounded showcase summary and metrics check are visible. Raw harmful command payloads, fake secret strings, command logs, raw diffs, and full generated artifacts are not shown.
- The report and documentation scenes use the previously reviewed sanitized assets; no browser chrome, account controls, bookmarks, unrelated tabs, or notifications are present.
- The intermediate terminal cast, GIF, generated report tree, and extracted review frames are not committed.
- The final container has one H.264 video stream and no audio stream. Metadata inspection found standard MP4/codec identifiers only, with no user, path, device, location, or editor-project fields.
- Representative frames were visually reviewed at scene boundaries and within each scene. Source text and intermediate files were searched for common path, credential, authorization, secret, and unsafe-payload patterns.
- Review is bounded and does not claim that sampled frames prove the absence of every possible sensitive value.
- Timed English WebVTT captions describe every segment, and the hosted demo page provides a complete text transcript.
Reproducibility limitations¶
The commands and evidence are deterministic, but terminal/font rendering, runtime identifiers and timestamps, screenshot rasterization, encoder versions, and H.264 output can differ across machines. This record describes semantic reproduction; it does not claim byte-for-byte video reproducibility. The small showcase is curated evidence, not a statistical estimate, production adoption claim, formal certification, universal security guarantee, or sandbox boundary.