v0.24.0

Highlights

• Agent video jobs (experimental). tools/video_job.py turns a scene list into a narrated, captioned, QC-checked MP4 as a background job any agent can submit and poll. Cost is estimated and capped before any spend, one job runs at a time, and receipts land in a spend ledger. tools/video_mcp.py offers the same calls over MCP. The contract is in docs/agent-video-jobs.md.
• Submit-time guards. A job is refused before spending when a clip prompt asks for a screen, names an idea instead of a filmable thing, or is one word, or when narration only repeats the headline. Each guard comes from a real failed take.
• concept-explainer-short renders 16:9 as well as 9:16, and crossfades its scene joins.
• skills/openclaw-video-toolkit is now skills/video-toolkit, for any agent runtime. Update any path pointing at the old folder.
• product-demo text is at least 28px and passes WCAG AA contrast.

All commits

• Release v0.24.0: agent video jobs (experimental) — video_job.py and video_mcp.py turn a scene list into a QC-checked MP4 with spend caps and submit-time guards; concept-explainer-short renders 16:9 (d194ae2)
• FIX: the agent-path test gives its clip a three-word prompt, which the one-word guard from #147 now requires (dabd84d)
• FEAT: video_job.py gives every terse clip prompt a live-action footage style — a bare "filing cabinet drawers sliding open" came back as a cartoon (#149) (dadd610)
• FEAT: video_job.py refuses a clip prompt that names an idea instead of a thing — "scale animation" came back as a cartoon cat (#148) (499f12f)
• FIX: video_job.py guards from take two — spoken narration, ledger settlement, one-word clip prompts (#147) (fdc111c)
• FEAT: video_job.py refuses a clip prompt that asks for a screen before any spend — the video model draws fake software with garbled text (#146) (b229317)
• FIX: video_job.py fixes from the first showcase run — diagram cards, no stray people, sentence-boundary background prompts (#145) (0bf7e9f)
• FIX: concept-explainer-short crossfades scene joins instead of dipping through black (4f16594)
• FEAT: video jobs look finished by default — Lugh's voice, animated title and headlines over generated backgrounds, a music bed, and a contact sheet (2726ba8)
• FEAT: video_job.py takes scenes as flat --scene strings, so an agent never has to write nested JSON (f2647da)
• FIX: video_job.py reads a spec from @path and repairs the JSON slips open models make (' escapes, trailing commas, Python literals) (6fe35a2)
• DOCS: how video_job.py runs under longarm's run_tool (inline spec, short captions, files/ output, reservation receipt, ETA) (d08e134)
• FEAT: video_job.py runs under an agent's run_tool — inline --spec-json, a short caption form, --output into files/, a reservation receipt and an ETA (e88e3e8)
• DOCS: every ideogram scene in a video job carries a JSON caption, with a fill-in template agents can copy (f8c3cef)
• DOCS: agent video jobs — registry entries, docs/agent-video-jobs.md, a CLAUDE.md paragraph and the README link (3978841)
• FEAT: tools/video_mcp.py serves video_estimate/submit/status/result/list_presets over MCP for agents without a shell (85d7d9b)
• FEAT: tools/video_job.py turns a scene spec into a QC-checked MP4 as an async job any agent can poll (5c4df2a)
• DOCS: skills/video-toolkit replaces the stale OpenClaw-only skill, for any agent runtime (51b9ada)
• FEAT: concept-explainer-short renders 16:9 from config.json, not only 9:16 (995b4fd)
• FIX: product-demo template text is at least 28px and passes WCAG AA contrast (9251168)

Full Changelog: v0.23.0...v0.24.0