AGENT SKILL

verify-transcript

Use when validating transcript CLI behavior, locating its verification features, or running the paid YouTube end-to-end check.

License 许可
MIT
所属仓库
pascalandy/skills
最近更新
2026-11-15
Stars
★ 0
Forks
1

01 前言 Front Matter

kind
dev
name
verify-transcript
api-key
deepgram
description
Use when validating transcript CLI behavior, locating its verification features, or running the paid YouTube end-to-end check.

02 资源文件 10

03 SKILL.md 正文

Verify transcript

Drive transcript only through scripts/transcript.py. The verifier finds that public CLI in either the categorized authoring/ source layout or the flat applied skill layout.

Launch

This is a short-lived CLI. It does not start a server or retain a process.

VERIFY_DIR="<verify-transcript-skill-dir>"
uv run "$VERIFY_DIR/scripts/verify_transcript.py" --help

Resolve VERIFY_DIR from this skill's directory. Do not depend on the caller's working directory.

Doctor

Check the verifier before a run when the layout or local tools look wrong.

uv run "$VERIFY_DIR/scripts/verify_transcript.py" doctor

doctor checks uv, all seven Feature Map pages, and the adjacent transcript public script. It answers {"ok":true}, or one error per failed check, and -v lists every check. It does not call Deepgram, a summary model, YouTube, or Zoom.

Drive

Run all free checks with one command.

uv run "$VERIFY_DIR/scripts/verify_transcript.py" verify

The default run covers help and version output, structured recovery, transcript-only planning, discovery, configuration, diagnostics, YouTube dry-run planning, and Zoom source and summary planning. A valid diagnostic response can pass even when the machine is not ready. Read each diagnostic feature's readiness_ok observation.

Search the Feature Map before a focused check. Copy exact selectable values from features[].id. Matching Markdown pages are in documents[].

uv run "$VERIFY_DIR/scripts/verify_transcript.py" features zoom
uv run "$VERIFY_DIR/scripts/verify_transcript.py" features

Read features/README.md for the index. Each feature page names its public command, expected state, and proof.

The paid end-to-end check runs the canonical test video with the sonnet profile. The transcript README's Test videos rule pre-authorizes it, sets its limits, and says where it runs.

uv run "$VERIFY_DIR/scripts/verify_transcript.py" verify \
  --feature youtube.real-summary \
  --allow-paid

Paid work requires both the paid feature ID and --allow-paid. The flag alone does not select paid work. --all also selects the paid feature and therefore requires --allow-paid.

Evidence

Every command answers in one JSON line. A verification answers file, the run's result.json, which records the verdict and evidence_dir; a failed feature is also one entry of errors, with the command that reruns it. A run stopped by an interrupt or a failed write answers files instead, the evidence it saved so far. -v prints each feature's verdict as it ends. The default evidence root remains ${XDG_STATE_HOME:-~/.local/state}/eval-transcript/runs. Runs retained before the skill rename stay in the same history.

Each case retains:

  • The exact public command with no secret values
  • Captured stdout, stderr, exit code, duration, and normalized parsed output
  • The feature verdict and observations
  • An artifact manifest with names, sizes, and SHA-256 hashes for the paid flow

The final result.json records the selected layout, aggregate counts, paid authorization, and cleanup result. A successful paid check records the artifact manifest before cleanup. A failed paid check retains its command and process evidence, then cleanup removes any scratch output.

Cleanup

The verifier creates one marked directory under the system temporary directory. A finally block removes that directory after success, failure, timeout, or interruption.

Cleanup refuses a path outside the system temporary directory, a name without the current run ID, or a directory without the current marker. Cleanup never receives the evidence path. Retained evidence survives every normal teardown.

Do not delete evidence as part of verification. If the user asks to remove old evidence, inspect the exact run directories first and use a recoverable deletion method.

Helpers

After changing Deepgram uploads, run this bounded verification loop:

  1. Run the verifier doctor above with -v and set TRANSCRIPT_DIR to the directory its transcript_skill line names, whether source or applied
  2. Run the upload contract tests below, then the verifier helper tests
  3. Run the default free verification and inspect its verdict and evidence
  4. Run the youtube.real-summary feature above and require its audio_upload evidence
uv run --with pytest --with httpx --with rich pytest \
  "$TRANSCRIPT_DIR/scripts/tests/test_transcript.py" -k TestDeepgramContract -q

These deterministic tests check bounded upload chunks, intact audio, rejection of success before the full payload is sent, timeout byte reporting, and the workflow deadline. A successful real run alone does not prove behavior on a slow or interrupted connection.

Run the focused helper tests after changing this verifier.

uv run --with pytest pytest "$VERIFY_DIR/scripts/tests" -q

The tests cover layout discovery, exact feature discovery, public-surface ownership, output streams, paid selection, command confinement, process capture, cleanup, Feature Map coverage, and artifact validation. They stub external paid boundaries. The paid command above is the real end-to-end proof.

04 同仓库更多 Skills