[UditAkhourii]

This is a curated version of #3 (by @sagniKdas53). It keeps the genuinely useful features, removes the problematic material, and fixes the confirmed bugs. All 74 tests pass locally on Python 3.10 install of ./cli[test], including the repo's original suites that #3 broke.

✅ What was kept (the good parts, credited to the original author)

• OpenRouter multi-model provider — --provider openrouter / auto-detection via OPENROUTER_API_KEY, native video upload for Gemini models, frame-sampling fallback for image-vision models (Qwen VL, Pixtral, Llama 3.2 Vision, GPT-4o), stdlib urllib fallback when requests isn't installed. Endpoints verified clean: keys go only to openrouter.ai (overridable via OPENROUTER_BASE_URL) or the Google GenAI SDK.
• Temporal chunking & parallel processing — --chunk-duration N --parallel M for long videos, with per-chunk timestamp offsetting and section-aware merging.
• Per-asset cost & token tracking — cost:, prompt_tokens:, output_tokens: sidecar headers.
• Domain modes — --mode screencast|meeting|demo|presentation prompt tuning, ## Environment & Tools / ## Meeting Details / ## Action Items & Decisions sections, SPEC/SKILL updates.
• Model alias table and cdaf models command — including the Gemini 3.7 Flash and Qwen 3.x flash aliases from the original PR (see correction below) — plus probe_duration_seconds() and the new test suite.

🗑️ What was stripped, and why

1. Undisclosed NSFW mode. PR This PR extends CDAF with multi-model OpenRouter support, parallelized temporal chunking for long/large video assets, automated generation cost tracking, and intelligent domain modes for screencasts, meetings, and product demos. #3's description never mentioned it, but the diff added an "Adult & NSFW Curation Mode" to the README, normative adult-indexing MUSTs to SPEC.md §2.5, nsfw-* model aliases, an nsfw: header key, and an --nsfw flag that set BLOCK_NONE on all four Gemini safety categories (dangerous content, hate speech, and harassment included). Routing "uncensored" traffic through Gemini with all filters off likely violates Google's Generative AI Prohibited Use Policy and risks the user's API account. Removed entirely; if wanted, it should be its own explicitly-scoped proposal.
2. Nonexistent model IDs ✏️ CORRECTION (retracted). The original review flagged gemini-3.7-flash, google/gemini-3.7-flash, qwen/qwen3.8-flash, and qwen/qwen3.5-flash-02-23 as nonexistent. That was wrong — the review was made against stale (pre-August-2026) model knowledge. All four are real: Gemini 3.7 Flash went GA on the Gemini API on 2026-08-13 and is live on OpenRouter, and both Qwen flash models are listed on OpenRouter. The original PR's aliases (flash-3.7, gemini-3.7, or-flash-3.7, qwen3) and pricing have been restored with published rates, and flash-3.7 is back as the top cdaf models recommendation. The one model-ID fix retained is the Qwen 2.5 VL OpenRouter slug (qwen/qwen-2.5-vl-72b-instruct → qwen/qwen2.5-vl-72b-instruct). Apologies to @sagniKdas53 for the wrong call here.
3. Import-time .env sweep. Merely importing cdaf.generate read and injected every key from .env files in the CWD, inside site-packages, and ~/.env — so a forgotten key in an unrelated file could be silently picked up and spent. Now: CWD .env only, loaded by the CLI (not at import), and existing env vars always win.
4. Whole-video base64 fallback. With ffmpeg absent, the frame extractor "fell back" to base64-encoding the entire raw video in memory and attaching it as an image_url — an OOM/garbage-request path that existed only to make the PR's own mocked tests pass. Replaced with a clear error, plus a 45 MB guard on inline OpenRouter uploads.
5. .gitignore tests/ entry — unanchored, it would have silently untracked any future test file (including cli/tests/). Removed; the video-extension ignores were kept.

🐛 Bugs in #3 that are fixed here

• CI was red on This PR extends CDAF with multi-model OpenRouter support, parallelized temporal chunking for long/large video assets, automated generation cost tracking, and intelligent domain modes for screencasts, meetings, and product demos. #3 (6 failures): the new tests needed requests, which CI never installed, and two existing test_local.py tests asserting the old provider tuple/error message were never updated. Fixed: CI installs ./cli[test], test_local.py updated, and the pytest→unittest rewrite of test_sidecar.py reverted. A further Windows-only failure (tempfile.NamedTemporaryFile keeps the file locked, breaking _video_data_url's re-open) is fixed with a TemporaryDirectory-based fixture.
• Multi-timestamp corruption: offset_body_timestamps replaced every timestamp on a line with the first match's offset values ([00:10.0-00:20.0] A then [00:30.0-00:40.0] B → both ranges became [01:50.0-02:00.0]). Each match is now offset individually; regression tests added.
• Invalid timestamps at minute boundaries: format_timestamp(59.96) produced 00:60.0 and never rolled into hours, so chunked videos over an hour mixed MM:SS.d and HH:MM:SS.d styles. Fixed with tenths-first rounding and a use_hours flag threaded through the chunk path; regression tests added.
• Unhandled network errors: requests timeouts/connection failures escaped as raw tracebacks, crashing cdaf generate mid-batch. Now wrapped in GenerationError so the batch loop prints FAILED <video> and continues.
• ffmpeg portability & quality: chunk cutting used input-side -to (ffmpeg ≥ 4.0 only) — now portable -ss/-t; Gemini chunks no longer get crushed to 640p/5fps (which destroyed the on-screen text that --mode screencast exists to capture); sub-second trailing chunks from probe rounding noise fold into the previous chunk instead of failing the whole video.
• Removed features restored: This PR extends CDAF with multi-model OpenRouter support, parallelized temporal chunking for long/large video assets, automated generation cost tracking, and intelligent domain modes for screencasts, meetings, and product demos. #3 deleted cdaf --version (which CI's smoke check runs!) and the _iter_videos warnings for skipped/invalid paths. Both restored.
• Cost honesty: unknown models now omit the cost: header instead of inventing a $0.20/$0.80 price; rates use published pricing, with the table documented as approximate.

⚠️ Known limitations (inherited, documented rather than fixed)

• merge_chunk_bodies concatenates chunk summaries rather than synthesizing one (N chunks → N paragraphs under ## Summary).
• Gemini cost estimates count candidates_token_count only (thinking tokens excluded), so cost: can underestimate.
• MODEL_PRICING is a static table and will drift (e.g. Gemini 3.7 Flash's current rates are promotional through 2026-12-31).

Test Plan

• python -m pytest cli/tests -q → 74 passed (includes the repo's original test_local.py / test_sidecar.py suites This PR extends CDAF with multi-model OpenRouter support, parallelized temporal chunking for long/large video assets, automated generation cost tracking, and intelligent domain modes for screencasts, meetings, and product demos. #3 broke)
• New regression tests for multi-timestamp offsetting, minute-boundary rollover, use_hours, unknown-model cost, and the restored 3.7/qwen3 aliases
• cdaf --version, cdaf models, and the example-sidecar validation (CI's exact smoke checks) run clean
• Model IDs verified against live Gemini API / OpenRouter listings (2026-08-28); endpoints limited to openrouter.ai + Google GenAI SDK; zero NSFW references in the final diff

Supersedes #3.

🤖 Generated with Claude Code

https://claude.ai/code/session_011jFGsfg58qPVdiSnwmWWXZ