[WaterrrForever]

Skill wording from Wenbo's end-to-end reverse test on 0.9.8: https://www.heygenverse.com/a/reverse-style-sampling-findings-45a4fd9e-fb0a-4484-b12e-9289afa6868f

references/start.md

• F13, sign-in: the agent runs auth login itself instead of asking the person to type ! npx hyperframes auth login. Where no browser opens, it gives the printed link and code. It never signs in with an API key to get past Pro: that signs in a HeyGen app account, which has no HyperFrames plan. The person-facing line says their HyperFrames account is their HeyGen account (part of F14).
• F1, paywall and top-up: the person types nothing. The agent runs subscribe_command / billing top-up --needed <total> itself and puts the page in its message as a link (reverse.upgrade_url, the refusal's url from fix(cli): two CLI fixes from Wenbo's reverse test (paywall link, plan label) #7099, or the link in fix on an older CLI). Both examples lose their ! command blocks.
• F2, plan: reverse start's free/pro is final. auth status is read only for a null plan, and only its hyperframes field, never user.subscription.
• F12, old CLI: a CLI without reverse start used to be worked around, which is how 0.8.152 got through. Outside a plugin it now means an old install shadowing the release: run the session as npx -y hyperframes@latest and tell the person once.
• F9, hand-off (new section): last diff with --sbs, open the side by side, start preview, say they can keep editing there, pass on the Framey line, then skill_completed.

SKILL.md: one pointer to the hand-off section. To stay under the 5,120-byte budget (main was at 5,118), I dropped "Small score differences are not worth another round.", which repeats the "then stop" before it.

references/animated.md (F8): cut a generated shot's guide inside a dissolve, never across it (a guide holding dissolve frames came back as the neighbouring shot), and retime every generated shot with data-media-start / data-playback-rate (HeyGen does camera moves late; this lifted Wenbo's mean SSIM from 0.707 to 0.746).

Not verified live: an agent-run auth login inside Claude Code on 0.9.9. From source it opens the browser (isAgentDriven), but nobody has walked it yet. node scripts/check-core-skills.mjs passes.

[jrusso1020]

Reviewed at 9539c67ef4ae516a5025df3fc9af62c2ac051444.

Code/content reads clear, with three small wording corrections below. This head currently conflicts with main only in generated skills-manifest.json (confirmed with git merge-tree --write-tree HEAD origin/main), so this is a COMMENT review; approval needs the resolved head.

The plan instructions distinguish reverse's known plan from an unknown plan and avoid using the HeyGen subscription to bypass Pro. The refusal URL contract matches #7099's 66d1793a4 producer and serializer; the documented fallback to the link in fix also works on the older CLI. reverse start, rendered-video diff --sbs, and preview/conditional Framey guidance exist. The latest-CLI override explicitly excludes plugins, preserving their pinned launcher.

Nits:

• skills/reverse-any-video/references/start.md:13–14: use auth status --json when instructing the agent to read the named hyperframes field. Without it, the CLI prints a readable HyperFrames row rather than that JSON field.
• start.md:33–35: default auth login refuses remote/headless terminals with a --device hint; it does not automatically print a device code. Say to run auth login --device on that refusal and relay its complete link, or link plus code when separately printed. A local detected agent can use browser OAuth; a failed browser-open there yields a manual loopback URL, not a device code. Both this head and #7099 implement that distinction.
• skills/reverse-any-video/references/animated.md:28: clarify “cut the guide inside a dissolve” as trimming away blended edge frames and keeping the stable part of the shot. The following warning about retaining dissolve frames currently makes the instruction ambiguous.

Validation: bun run lint:skills passes, including command/path and size checks; bun run --cwd packages/cli gen:skills-manifest --check reports nine skills in sync. Full four-file content/diff and relevant CLI contracts were inspected. A fresh independent pass checked those contracts against both pinned trees and source-derived remote/local login guard cases, which I independently reran against both trees (node review-artifacts/7101/login-guards.mjs). No live agent-run login, purchase, generated-shot timing, browser preview or complete reverse run was exercised. The reported SSIM improvement remains the author's evidence, not my measurement. No runtime source changed, so no full build/runtime test run was needed for this review.

Need: yes · Size: bounded four-file documentation change · Reuse: existing CLI, plugin and hand-off contracts · Simpler: already minimal · Structure: fine
Risk: low — wording only; live sign-in/payment hand-off remains unverified.

Verdict: COMMENT — content clear with nits; resolve conflicts before approval
Reasoning: No verified Important content defect found. GitHub currently reports this exact head CONFLICTING/DIRTY; merge main, regenerate the skill manifest and return the resolved head for approval.

— hf-review

[WaterrrForever]

Thanks @jrusso1020, all three nits are in 302c0dd, on top of a merge of main (c54e856, manifest regenerated, gen:skills-manifest --check in sync).

• auth status --json when reading the hyperframes field.
• Sign-in: when the browser doesn't open, give the link it prints. When it refuses as a remote or headless terminal, run auth login --device and give its link, and the code when shown.
• Dissolves: trim the guide to the shot's own frames, leaving out the blended frames at either end.

Main's #7031 changed the same section, so two lines now sit beside its rule. upgrade_url is given as a link only when it isn't null. The "on a CLI without reverse start, run billing subscribe" line now applies only to a plugin's pinned CLI, since outside a plugin the old-CLI case switches to npx -y hyperframes@latest. lint:skills and check-core-skills pass.

[jrusso1020]

Reviewed at 302c0dd1755f5eef436d6e37ccd4dac19f504f93, delta from my reviewed 9539c67ef.

All three wording nits are fixed: the plan-field instruction uses auth status --json, headless refusal explicitly switches to auth login --device, and the guide excludes blended dissolve edge frames. The main merge preserves #7031's null-purchase semantics: a null upgrade URL is not relayed, and a null subscribe command stops rather than inventing a purchase path (skills/reverse-any-video/references/start.md:14,33–36,43–51; references/animated.md:28). The standalone latest override and plugin-only legacy fallback remain separate.

Validation: bun run lint:skills passes command/path and size checks; bun run --cwd packages/cli gen:skills-manifest --check reports nine skills in sync. I read the complete documentation delta and its merged reverse producer/refusal contract. Fresh independent source-derived controls exercise 18 reverse cases, six login guard cases, pinned-plugin invocation and exact nine-skill hash/file-count parity; I inspected and reran them with node review-artifacts/7101/r2/adversarial-offline.mjs. These execute extracted pure functions with fixture dependencies, not a packaged CLI or service walk.

Limits: no live Claude Code login, device approval, purchase/top-up, generation/retiming, browser preview or complete reverse run. The earlier field/URL and hand-off contract review still applies; no new runtime implementation was authored by this PR.

Need: yes · Size: bounded documentation re-review · Reuse: existing reverse/auth/plugin contracts and manifest generator · Simpler: already minimal · Structure: fine
Risk: low — wording and generated manifest, with live hand-offs explicitly unverified.

Verdict: APPROVE
Reasoning: The prior wording issues and manifest conflict are resolved; the merged null-purchase instructions match the actual CLI contract, with no new Important finding.

— hf-review

[jerrai-bot-heygen]

Reviewed 302c0dd against the installed-plugin CLI and purchase-consent contract. The normal path opens a page and leaves checkout to the person; it does not instruct the agent to complete a purchase. The new null-upgrade-link path and standalone @latest rule are also scoped correctly. One plugin-specific fallback still offers paid checkout when the pinned CLI cannot run Reverse at all; please resolve the inline finding before approval. Hosted exact-head skills and bootstrap jobs executed successfully; other green labels include skipped wrappers. No local run or merge. — Jerrai

[jerrai-bot-heygen]

This fallback can send someone to pay for Pro precisely when their installed plugin's pinned CLI does not implement reverse start, the command this skill needs to start/resume the actual run. A bare npx hyperframes also conflicts with the plugin execution rule that all examples use the pinned launcher, while this plain subscribe has no reverse feature/session id to credit the purchase to this run. Please make this an incompatibility/upgrade-the-plugin stop (or show an executable, version-pinned reverse path), rather than opening checkout for a run that cannot proceed. Test the plugin-without-reverse start branch.

[WaterrrForever]

Thanks Jerrai, fixed in 780a434. The plugin fallback is gone. Inside a plugin, a pinned CLI without reverse start now stops before anything is paid for, with one line saying the HyperFrames plugin needs updating to reverse a video. No billing subscribe and no bare npx. The rule sits beside the standalone old-CLI rule (npx -y hyperframes@latest), so each case has one path.

On testing the branch: this is skill text, with no runtime for a test to hold. lint:skills, check-core-skills and the manifest check pass, and nothing in the skills still says to subscribe on a CLI without reverse start (grepped).

[jerrai-bot-heygen]

Re-reviewed 780a434 against my 302c0dd plugin-paywall hold. A plugin whose pinned CLI lacks reverse start now stops before any payment, tells the person to update the plugin, and no longer runs a bare npx subscribe. Standalone-only @latest, normal nullable upgrade-link behavior, and human-completed checkout remain scoped as reviewed. The reverse skill manifest hash was updated for the changed file. Earlier head 302c0dd had executed green skills/bootstrap tests; at this new head the hosted check jobs were still starting at review time, so I do not call its tests green yet. Approval is source-quality only, not a merge action.

— Jerrai