llms-full.txt@docs git:20260930.581add6A

Immutable. This exact content never changes and is served at /api/v1/blob/2b3a8034dc00c717.

# Kinocut: extended agent guide

This is the repository's detailed companion to ../llms.txt. Load that index
first and retrieve this guide or a specific reference when needed. These files
are discovery aids; neither promises search ranking or universal crawler support.
Availability at kinocut.dev is a separate website publication check.

## Identity and release truth

Kinocut is an Apache-2.0 MCP server, Python client and CLI for guarded media
workflows. KyaniteLabs publishes it. The CLI is kino. The former mcp-video name,
compatibility import and shim bridge to Kinocut rather than describing a competing
product. Preserve documented compatibility instead of deleting historical names.

Canonical release facts: public_claims.json. Current operational status:
status/NOW.md. Package, repository, website and directory publication are separate
observations. A source edit or local test does not prove an external release.

## Setup and host integration

Install Python 3.11+ and FFmpeg/ffprobe using the appropriate supported platform
setup. Install uv for the uvx examples, or use pip and the installed kino command.
Install kinocut, then run kino doctor. Optional model packages,
weights, Node-based creation tools and providers have their own prerequisites.

Explicit MCP stdio launch:

```bash
uvx --from kinocut kino --mcp
```

Claude Code:

```bash
claude mcp add kinocut -- uvx --from kinocut kino --mcp
```

Gemini CLI and other clients: INTEGRATIONS.md. Use the client's supported MCP
configuration and retain workspace trust and tool confirmations. Kinocut also
defaults to MCP without a subcommand; --mcp states the intended mode explicitly.
MCP support does not imply support for every host's plugin package format.

## Select the smallest sufficient operation

Use deterministic code for path validation, exact times, FFmpeg construction,
codec selection, hashing, source confinement, budgets and release authority.
Use specialized models only where perception is required and the installed
backend provides real evidence. Use generative prose for requested, grounded
editorial work; preserve source names, numbers, negations and qualifications.

Discover tools with search_tools using the task phrase or exact name. Query
the connected server's MCP tools/list for its exact JSON schemas. That installed
surface is authoritative; TOOLS.md is the readable reference. Avoid loading or
duplicating all schemas for every task. Optional tool registration does not mean
the tool's model or provider is installed or executable.

For Python, start with Client().info(absolute_input_path). For an agent workflow,
inspect metadata, select an operation, validate its parameters, render, inspect
the resulting artifact and retain the receipt. Follow WORKFLOWS.md for the
operation allowlist and durable resume semantics; do not invent workflow names
from engine functions or tool descriptions.

## Paths and execution boundaries

Use explicit absolute input and output paths for plain-file APIs when the host's
working directory is uncertain. Relative plain-file paths use process cwd.
Workflow assets resolve against the job-spec directory, with confinement rules.
Compositor sources and masks resolve against their declared spec directory.
Do not guess an agent root or change cwd to bypass these checks.

An output must not alias an input, including resolved symlink and hardlink aliases.
Missing dependencies and unsupported filters are structured failures. A suggested
alternative is advice, not execution: choosing another codec changes the request
and must be intentional. Do not silently install dependencies or substitute a
delivery codec. Check supported local capabilities before scheduling a render.

## Existing media workflows

Long media: use storyboard sampling, scene detection and bounded longform
transcription before a heavy edit. Sampled frames are not a complete viewing
record. A transcript backend is not interchangeable with semantic video analysis.

Audio: AUDIO_MIXING.md describes single-encode timed mixing, plain-file ducking
and looping. New development-only Client additions must be checked against the
installed version. A governed audio-bed workflow also needs its immutable project
snapshot and approval contracts. Loudness measurements do not prove listening
quality, and silence detection does not prove editorial cut correctness.

Recipes: use the existing examples/workflows/captioned-vertical-short job and
WORKFLOWS.md. Build on supported job operations rather than inventing a second
recipe format or promising a person's distinctive editing style.

Creation: Hyperframes and local Revideo have separate runtime requirements.
Custom executable scenes are trusted code. CPU/GPU availability alone does not
prove encoder usability, decoded fidelity, portability or a fixed speedup.

## Evidence, receipts and human decisions

QUALITY_EVIDENCE.md distinguishes measured, unavailable and not-evaluated
evidence. A synthetic waveform is labelled synthetic. Native signal depth and
range matter. Missing visual/voice inference must not become a positive score.
Chronological image-change proxies do not establish semantic coherence.

VIDEO_RECEIPT.md defines receipt families and schema discriminators. Receipts
record artifacts, calls and findings; they do not automatically sign off an edit.
Use kino review-ui OUTPUT_DIR for the existing review surface. Consult
AI_VIDEO_REVIEW_AND_SALVAGE.md and VIDEO_RECEIPT.md for exact-artifact review and
decision contracts. An HTML review surface does not establish completed human
viewing or permission to publish. Verify the actual receipt and artifact hashes.

PROJECTSTORE_LIFECYCLE.md documents CAS repair and cancellation. Availability
must be supported by integrity/lifecycle evidence. A stop request does not prove
that a process has stopped; retain runner identity until shutdown is verified.

## Privacy and cost

FFmpeg editing can run locally without required cloud keys. Kinocut has no
software license fee. Local hardware, storage and electricity still cost money;
optional model providers and the host agent can have separate usage charges.

Prompts and tool results can reach the host's model provider. Results may contain
paths, transcripts or frames. Review the selected backend and exact data sent;
local execution is not a zero-egress guarantee. Do not upload private source
material merely to obtain tool discovery or advertise unsupported privacy claims.

## Source references

- INTEGRATIONS.md: host-specific setup and primary-source integration references.
- INSTALL.md: runtime prerequisites and platform setup.
- TOOLS.md and CLI_REFERENCE.md: readable public surfaces.
- WORKFLOWS.md: executable job schema and resume contracts.
- QUALITY_EVIDENCE.md: measurement domains and unavailable inference.
- VIDEO_RECEIPT.md and HUMAN_GATES.md: receipt families and exact-artifact review.
- PROJECTSTORE_LIFECYCLE.md: recovery, integrity and cancellation.
- faq.md: human-readable questions and answers.
- research/external-ai-audits/2026-09-30/REPORT.md: checked Google/Gemini proposals.