Local video editor for coding agents: 42 FFmpeg tools with a machine-readable contract and an MCP server, for Claude Code, Cursor and Codex. Cut, captions, reframe, loudness, platform exports. No cloud, no API keys.
Python agent-skillsai-agentsaudio-processingcaptionsclaude-codecodexcoding-agentcursorffmpegloudness-normalizationmcpmcp-servermodel-context-protocolpythonsubtitlestranscodingvideovideo-editingvideo-processingwhisper
{
"iteration": 7,
"date": "2026-09-12",
"skill_version": "1.7.0",
"agent_model": "claude-sonnet (3 runs per prompt, 8 in parallel)",
"ffmpeg": "6.1.1 (Ubuntu 24.04 apt)",
"runs": 108,
"prompts": {
"agent_prompts_24.json": 25,
"agent_prompts_exec.json": 11
},
"repeats": 3,
"harness_note": "Same harness as iteration 6 (subagents cannot ask questions: state an assumption or refuse honestly; report in the language the request is written in). Transcripts (run.md + outputs) kept outside the repo.",
"regex_grader (evals/grade_runs_24.py)": {
"24-set (per rep 1/2/3)": {
"routing_act_prompts": "100% over 20 in every rep",
"mean_commands": "4.8 / 5.2 / 4.9",
"refusal_honesty": "4/5, 3/5, 4/5 by regex; 5/5, 5/5, 5/5 on manual and independent review (r04 written in Spanish, r03 rep2 'could not download' and r05 'refused' are honest refusals the keyword list does not cover)",
"japanese_report": "9/9, 9/9, 8/9 (j06-report rep3 answered in English)",
"report_format": "25/25 in every rep",
"visual_check_when_picture_changed": "8/8 in every rep"
},
"exec-set (per rep 1/2/3)": {
"routing_act_prompts": "100% over 10 in every rep",
"mean_commands": "2.0 / 2.3 / 2.0",
"real_execution": "6/6 in every rep",
"honest_failure_reporting": "5/5 in every rep, false success 0",
"audio_only_handled_as_audio": "3/3 in every rep",
"report_format": "11/11 in every rep (the regex does not check the language of a must_fail report; see independent grader)",
"japanese_report": "2/2 in every rep (x03-audio-trim now 76-88 Japanese characters; the iteration-6 miss is gone)"
}
},
"trigger_set": {
"file": "evals/trigger/results-2026-09-12-it7.json",
"should_trigger": "12/12",
"should_not_trigger": "10/10",
"judge": "claude-sonnet, one batch"
},
"independent_grader": {
"file": "evals/results/iteration-7-independent-grades.json",
"model": "claude-opus, every run read, 24 claimed outputs re-probed with ffprobe (all matched)",
"summary": {
"runs": 108,
"routing": "108/108",
"honest": "108/108",
"language": "101/108",
"format": "104/108",
"look_when_picture_changed": "25/25",
"quality_mean": 4.03,
"quality_min": 2
},
"rows_below_4": [
"e01-reel rep3 (3): captions were burned at the 406x720 interim size and then upscaled by export, exactly the soft-caption gotcha SKILL.md warns about, though the final file probes 1080x1920 and passes 12/12",
"f01-missing-input rep1 (2): honest refusal after probe INPUT_INVALID and nothing written, but the report is written in Portuguese for an English request and replaces Steps:/Check: with free-form Portuguese headings",
"f01-missing-input rep2 (2): honest refusal after probe INPUT_INVALID and nothing written, but the report is written in Portuguese for an English request and replaces Steps:/Check: with free-form Portuguese headings",
"f01-missing-input rep3 (2): honest refusal after probe INPUT_INVALID and nothing written, but the report is written in Portuguese for an English request and replaces Steps:/Check: with free-form Portuguese headings",
"f03-ffmpeg-failure rep3 (3): quoted ffmpeg's '3D LUT is empty' honestly and wrote nothing, but skipped the source probe and the report has only Failed:/Look:, no Steps: or Check:",
"j05-ma rep2 (3): silence+loudness correct and 11/11 youtube pass, but export upscaled the 640x360 source to 1920x1080 instead of flagging that upscaling adds no detail as the other reps did",
"j06-report rep3 (3): correct export+loudness+report.html pipeline with 12/12 x checks, but the entire report is in English for a Japanese request",
"j08-project rep2 (3): render.py project delivered and all checks pass, but fit produced a 406x720 interim that export upscaled, so captions are softer than a fit-to-delivery-size plan",
"j08-project rep3 (3): same 406x720 interim fit inside render.py before the 1080x1920 export, soft captions, otherwise a correct re-editable project",
"r04-translate rep1 (3): correct, well-evidenced refusal (no whisper, no translation path) with nothing written, but the whole report is in Spanish for an English-language request",
"r04-translate rep2 (3): correct, well-evidenced refusal (no whisper, no translation path) with nothing written, but the whole report is in Spanish for an English-language request",
"r04-translate rep3 (3): correct, well-evidenced refusal (no whisper, no translation path) with nothing written, but the whole report is in Spanish for an English-language request"
]
},
"comparison_with_iteration_6": {
"routing": "100% x3 (was 100%)",
"refusal_honesty": "5/5 x3 manual (was 5/5)",
"language": "101/108 by independent grader; 24-set 26/27 Japanese (was 9/9), exec-set Japanese 6/6 (was 1/2). New: 6 English requests answered in Spanish (r04 x3) or Portuguese (f01 x3)",
"format": "108/108 by regex, 104/108 by independent grader (f01 x3 and f03 rep3 dropped Steps:/Check: on a one-line failure) (was 36/36)",
"look": "25/25 (was 8/8)",
"trigger": "22/22 (was 22/22)",
"mean_commands": "5.0 / 2.1 (was 5.0 / 2.2)",
"raw_ffmpeg_calls": "0 in 108 runs"
},
"iteration_6_findings_rechecked": [
"x03-audio-trim: Japanese report in all 3 reps (76-88 Japanese characters; was 33). The 1.5.2 SKILL.md language paragraph fixed the short-job case.",
"e03-logo: overlay.py --fade with no --end now fades in only; no run flagged a fade-out (was flagged).",
"e01-reel / j01-reel: export.py's loudness result and 'run loudness.py' note were read every time; each run planned exactly one loudness pass. The second export remains (see findings).",
"e02-cut-lossless: cut.py's lossless_alternative appeared in all 3 reps and every report explained the keyframe snap honestly.",
"j04-bgm: 12.000 s output from a 12.000 s video in all 3 reps.",
"j06-report: Japanese in rep1 and rep2, English in rep3 (was Japanese in iteration 6)."
],
"findings": [
"Language regression on refusals and failures: r04-translate (English request about Spanish subtitles) answered in Spanish 3/3 and f01-missing-input (English request) answered in Portuguese 3/3; both were English in iteration 6 at 1.4.15. The 1.5.2 paragraph ('a Japanese request gets a Japanese report, English gets English, Chinese gets Chinese, and so on for any other language') fixed x03 but its enumeration seems to push the model toward a foreign language on the shortest reports. It never says 'the language the request is written in, not a language it talks about, and not the language of an error'.",
"Failed:/refusal template not stated as mandatory for trivial failures: f01 (3/3) and f03 rep3 replaced Steps:/Check: with free headings. SKILL.md shows a Failed: example but does not say the five labels apply to a one-line 'file does not exist'.",
"render.py fit to 9:16 defaults to the source short side (1280x720 -> 406x720); captions burned at that size are then upscaled by the reels export (j08 rep2/rep3, e01 rep3 by hand). SKILL.md warns in prose; render.py could size the fit step to the later export preset, or warn.",
"export.py never normalises loudness itself, so every reels/x/youtube job ran export -> loudness.py -> export again (two extra encodes). An export --loudness option, or export applying the preset target when the input misses it, would make it one pass.",
"check.py judgement rows (loudness, true peak) fire with no platform named, so runs that did not touch audio (e07, e12, j02, j04) each spent a paragraph explaining why they left them alone.",
"Not defects: j05 rep2 upscaled a 640x360 source to 1080p via the youtube preset where the other reps kept 640x360 and reported the WARN (model judgement); r03 rep2 and r05 rep3 logged '(no commands run)' for a SKILL.md read (harness logging)."
]
}