The publish job's checkout persisted its contents:write token in
.git/config, where every later step, including the package install and
build, could read it. Check out without persisting credentials and give
the token only to "Create release tag", the one step that pushes. It
authenticates its own git calls through step-scoped GIT_CONFIG_* env and
masks the encoded value. The gh steps keep their own GH_TOKEN.
The workflow test now covers the checkout flag, which steps may receive
the token and under what name, the auth header the tag push actually
carries, and that nothing is written to .git/config.
— Rames Jusso
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(core): Safari keeps a small file's authored preload and never holds a clip already playing
* fix(core): Safari still holds a long file's later clips that were parsed before the runtime
The sweep sets preload none on them again and gives a clip its authored preload back once its file turns out not to be long.
* fix(core): Safari leaves the src of a long file's clip that is already playing or seeking
Its file's size can arrive after the clip started; writing an in-point src then restarted its load and sent it back to the in-point.
* fix(studio): check Host on the dev server's API routes and honor preview allowed hosts
The Studio Vite dev server now applies the same Host rule as the CLI
preview server to the routes its plugin serves (/api/*,
/__hyperframes_config), registered ahead of the plugin's other
middleware, and also accepts the IP-literal and configured hosts Vite
allows by default. HYPERFRAMES_PREVIEW_ALLOWED_HOSTS is now fed into
Vite's server.allowedHosts, so a host listed there loads the Studio page
as well as passing the API's checks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(studio): accept the *.localhost and leading-dot hosts Vite serves on the API routes too
The dev server's API Host check was stricter than Vite's own rule, so a page
Vite served on studio.localhost or under a leading-dot allowed host got a 403
from every /api call. The check now applies Vite's rule alongside the CLI's.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* ci(publish): build the Lambda handler zip in a read-only job and verify its hash before release
The handler zip build runs dependency install and bundling code. It moves out of the
publish job into its own job with `contents: read`, no secrets and no persisted git
credentials. The publish job downloads the artifact, checks its sha256 against the
build job's output, then pins and attaches it as before. A retried run still reuses
the zip already on the release.
Refs #7485
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(ci): cover the read-only Lambda zip build job and the pin step's hash check
The publish workflow test still looked for the old "Build Lambda handler zip"
step. It now checks the build job's read-only permissions and the publish job's
dependency on it, and runs the pin step for the retry, matching-hash,
mismatched-hash and empty-hash cases.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(studio-server): thumbnail origin from the server, sandbox previews from other sites
The thumbnail route builds the headless browser's URL from the server's own
listening origin (a new optional `serverOrigin` on the adapter, wired in the CLI
preview server and the Vite dev host) instead of the request's Host header.
Hosts that report no origin fall back to a loopback Host only.
Preview responses (bundled, sub-composition and project files) opened from
another site carry `Content-Security-Policy: sandbox allow-scripts` and
`Vary: Sec-Fetch-Site`. Studio's same-origin iframe, browser automation and
non-browser clients are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(studio): build the thumbnail origin from a hostname bind name
With HYPERFRAMES_PREVIEW_HOST set to a hostname, the listening address is
the IP it resolved to, and the Host check accepts the bind name, not that
IP. The thumbnail browser then loaded a 403 page. Use the bind name when it
is a hostname; IP and wildcard binds keep the listening address.
— Rames Jusso
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* style(studio): drop the serverOrigin comment that put it over the package's comment share
The hostname-bind rule is stated by its test name in serverOrigin.test.ts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(producer): a render uses the CLI's own GSAP, so the default CDN's file needs no network
* fix(cli): the GSAP vendor lookup runs as ESM from source, so render starts again
* chore(producer): raise the unpacked size budget by the bundled-script module's 3 KB
A project-local node_modules/.bin/hyperframes can shadow plain npx hyperframes,
but npm install -g never updates it. Classify the shadowing copy like the
running one: global installs keep their manager's command, project-local
installs get the project's own manager (from its lockfile), and unknown
layouts stay quiet. The command pins the release the npx copy is running.
* perf(core): Safari clips open at their in-point and are fetched ahead
* perf(core): Safari opens a long file's clips at their in-point and leaves short files as they were
* perf(core): Safari holds a long file's later clips with no source until each one is due
* perf(core): Safari plays a long file's sound from its element instead of fetching the whole file
* perf(core): Safari holds a later clip of a long file even when WebKit began reading it at parse
* style(core): fit the long-file comments to the comment budget and the formatter
* refactor(core): one proxy trigger path, and the producer budget at its measured size
* perf(core): Safari opens a later clip WebKit read at parse at its in-point rather than holding it
* build: the producer size budget at its measured 3162274 bytes
* refactor(core): name the condition that holds a long file's later clip
* fix(reverse): give agents the sign-in reason to say, and catch stale CLIs that know reverse
- `reverse start`'s signed-out `next` now carries a sentence the agent can say as is
(why HeyGen, and that the HyperFrames account is the HeyGen account), asks for it at
the start of the sign-in message, and on remote terminals has the agent send the
device link and code before it waits.
- references/start.md: same ordering, an example that fits both login routes, and no
links to or quotes from the instructions themselves.
- An old global install that knows `reverse start` but predates `next` was never caught:
agent runs get no update notice and no auto-install, so it stays frozen. The skill now
also treats `_meta.updateAvailable: true` from `reverse start` as an old install and
runs the rest with `npx -y hyperframes@latest`; every CLI that knows `reverse start`
already reports that field.
* fix(reverse): send remote terminals straight to device sign-in, reason in every link message
- `reverse start` now knows when no browser can open (the check `auth login` uses, moved
to attendedTerminal): `login_command` is `auth login --device` and `next` asks for one
message with the reason, the link and the code. The two-step route (a refused browser
login, then a bare follow-up with the code) is where the reason got lost.
- Every message that gives the sign-in link starts with the reason; agents say they are
waiting rather than quoting the skill, and give no file paths.
- The out-of-date line for an old global install goes in the next message to the person.
* fix(reverse): flag an older global install that shadows npx, and pin the run's release
- installerDetection: `shadowingInstall` finds the older global `hyperframes` that a plain
`npx hyperframes` runs instead of this npx copy (first on PATH outside the npx cache), with
the command that moves it. `detectInstaller` takes the entry to classify.
- `reverse start` run through npx opens `next` with that line as words to say, signed in or
not. Skill prose alone did not get agents to tell the person.
- Skill: a signed-out `reverse start` with no `reverse.next` is also an old install, which
covers a cold update cache; the run pins the release `@latest` resolved to.
* fix(aws-lambda): the handler ZIP installs its runtime from a committed lockfile with scripts off
build-zip.ts ran `npm install --no-package-lock` in the publish job, so
transitive versions floated at release time and any dependency's install
script ran beside the release credentials. The ZIP now runs `npm ci
--ignore-scripts` from runtime-lock/<source>/package-lock.json (versions
and integrity for every package), and the build fails if that lock
disagrees with bun.lock instead of falling back to a range or "latest".
`bun packages/aws-lambda/scripts/lock-runtime.ts` regenerates it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(aws-lambda): check only direct runtime versions against bun.lock; Renovate refreshes the rest
The runtime lock's transitive tree is npm's own resolution, not bun.lock's (modern-tar
0.8.6 there vs 0.8.4 in bun.lock), so the parity claim now covers puppeteer-core and
@sparticuz/chromium only. Renovate's weekly lockFileMaintenance refreshes the transitive
tree; its direct-version bumps are off for these dirs because those follow bun.lock.
— Rames Jusso
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(aws-lambda): keep the maintainer-only lock-runtime script out of the package
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(studio-server): project history lists files without holding the host's event loop
* test(studio-server): the history walk test counts the hold up to the sweep's end
* fix(studio-server): history rewrites video-state.json only when it changed
* refactor(studio-server): names carry what the history walk comments said
The Cloud Run render image ran bun's remote install script, fetched
chrome-headless-shell through an unpinned npx package, and built FROM a mutable
tag. The base image is now pinned by digest; chrome-headless-shell and bun are
downloaded directly and checked against pinned sha256s before use.
— Rames Jusso
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(studio-server): reject cross-origin state-changing requests
Non-GET routes in createStudioApi answer 403 unless the shared
sameOriginPost predicate (lifted into studio-server's ./request-identity
subpath) accepts the request: the Host must be a server identity, a
present Sec-Fetch-Site must be same-origin, and a present Origin must
match the effective host and scheme. Hosts behind Codespaces or tunnels
allow their public name with HYPERFRAMES_PREVIEW_ALLOWED_HOSTS, read by
`hyperframes preview` and the Studio dev server.
Squashed from #6128 (42574a7, 110a300, 89011e1, 41c6626, 27193bb) onto
current main.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): the preview server answers only the hosts it is reached by
Every route of `hyperframes preview` (the forwarded studio API, the SSE
stream, /__hyperframes_config and the Studio page) now answers only
loopback names, the HYPERFRAMES_PREVIEW_HOST bind, or a host listed in
HYPERFRAMES_PREVIEW_ALLOWED_HOSTS, the same list the write guard uses.
Only the Host header counts here, since a page can set X-Forwarded-Host
on its own requests and a read carries no Origin to compare it with.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Dextheking1 <Dextheking1@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(studio-server): buildSubCompositionHtml reads only compositions inside the project
The helper joined projectDir with compPath and read the result, trusting
every caller to have confined the path first. Studio's own preview route
does; an embedder that forgets would serve any readable file. The helper
now resolves compPath with core's resolveWithinProject and returns null
for a path outside the project, the same answer as a missing file.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(studio-server): keep subComposition.ts's comment share flat
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(studio-server): the sub-composition head reads only an index.html inside the project
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(cli): label GPT image prices as estimates and show the charged credits
GPT Image generations on a HyperFrames plan are moving to fal's reported
vendor cost, so the pre-run figure is an estimate. The --plan basis for
gpt-image-2.5, 2.6 and 2.6-flare now reads "about N credits (charged at
vendor cost)". Nano Banana is unchanged.
When the /v3/images job carries charged.amount, generate reports it as
charged_credits and records it in the media ledger's provenance. Without
it, nothing new is printed. --max-credits is unchanged.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(cli): keep 10% headroom under --max-credits for vendor-priced images
gpt-image-2.5, 2.6 and 2.6-flare are charged at the vendor's cost, so
--max-credits now refuses one when its estimate plus 10%, rounded up to a
whole credit, is over the limit. A 16:9 gpt-image-2.5 image (about 40
credits) is refused at --max-credits 40 and passes at 44. Other models keep
the plain estimate check. The refusal and the --max-credits help say so, and
the help notes this is not a hard ceiling.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): count only HyperFrames-credit charges, and keep headroom under the run's limit
Review fixes for the vendor-cost image work:
- charged_credits is reported only when the job's charged.pool is
hyperframes_credits. A charge to api_credits or any other pool is in
that pool's own unit, so an API-key caller no longer sees it as HF
credits.
- runJob logs that charged amount, not the estimate, on a job's completion
row in heygen-jobs.jsonl, so later credit-limit checks see the real total.
Without a HyperFrames-credit charge the estimate is kept.
- The run's credit limit (.hyperframes/credit-limit.json) uses the same 10%
headroom rule as --max-credits for gpt-image-2.5, 2.6 and 2.6-flare,
through one shared limitCheck helper, with the same refusal wording.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* docs(media-use): say the credit limit is not a hard ceiling for GPT Image
A GPT Image call is admitted at its estimate plus 10% but charged the
vendor's cost, so it can still go past a run's credit limit. The media-use
skill, its HeyGen reference, the --max-credits help and the limit doc
strings now say so, and that a finished call's reported charge counts
toward the limit afterwards. Other models keep the "refuses what would go
past it" wording. No logic change.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): hold a running GPT Image call at its headroom under the project limit
A reserved or still-running vendor-priced call was logged at its plain
estimate, so two concurrent calls could each pass the 10% headroom check
and together go over. Log the headroom amount until HeyGen reports the
charge.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(studio-server): history waits out a file a save has emptied, and rereads one written mid-copy
* style(studio-server): state the history copy's null result in the present tense
* style(studio-server): wrap the put doc comment
* style(studio-server): shorten the put doc comment to one line
* fix(studio-server): a file still changing as history opens starts at its first whole copy
* fix(producer): refuse a too-slow render only when its typical frame says it cannot finish
* fix(producer): keep the calibration estimate fields self-describing
* test(engine): split the frame extractor tests so each file stays under the flake gate budget
* test(engine): cut the comments the frame extractor test split carried over
* docs(studio-app): the Reference read is paid by Pro or a saved HeyGen API key
ReferenceAnalysis readGate: a saved usable key pays the read; without one
the plan gate opens the plans page (feature video_analysis). The page said
only the API key paid.
* docs(studio-app): a saved key comes first, and a used-up pro allowance opens plan and usage
* docs(studio-app): pro credits pay for the reference read first, a saved key when the plan can't
* fix(engine): a clip shows a new source frame on every output frame it covers
* refactor(engine): shorten the frame count doc to two lines
* style(engine): collapse the single-pass ffmpeg call the formatter wants on one line
* docs(studio-app): cloud exports come with the plan, not only an api key
The export page said cloud renders are always paid per minute from a HeyGen API balance. The app
now includes 5 cloud exports a month on Free and unlimited on Pro (plans.ts, cloudFit.mjs).
* docs(studio-app): export destination and api-key billing wording
* fix(engine): short clips from Apple-encoded videos keep their last frame
* refactor(engine): name the end-of-source hold and frame bound instead of commenting them
* fix(engine): frame ranges past a source's end stay empty with the held final frame
* test(engine): move the final-frame test into its own file
* test(engine): nightly motion blur run over generated compositions
A seeded generator of text-free compositions combines 19 features (nested
canvas, overflow, clip path, scroll container, callbacks, a last-child rule)
and checks the two pixel invariants on auto: outside the trail at two holds
and at a moving time, inside the trail at a moving time. The runtime lane
runs seeds 1 to 3; motion-blur-nightly.yml runs 200 a day on the Chrome 153
pin and on stable, and opens one issue per failing seed.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(engine): the nightly records every failure, ignores noisy references, caps its issues
A crash, a timeout or a launch failure is recorded before the job goes red; a
failing check whose blur-off reference disagrees with itself is reported as
unstable instead of failing; the report job moves into a tested script that
fetches open titles once, caps first, and opens one systemic issue above 5 seeds.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(engine): the nightly report resists injected failure text and keeps a crashed leg's record
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* fix(engine): nightly report escapes backslashes, test strips tags to a fixpoint
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
* docs(mcp): current ChatGPT setup steps, and the hosted render's real output
OpenAI moved custom MCP servers to Plugins > + > Add custom MCP server,
on Business, Enterprise and Edu only. The hosted render takes no format
or fps (render_video has a TODO to expose them), so the page no longer
suggests asking for 30 fps, WebM or MOV there.
* docs(mcp): connect from the claude directory, and name hyperframes in the first prompt
HyperFrames by HeyGen is listed in Claude's connector directory. In a Claude.ai test on 6 Nov
2026, a request without "Use HyperFrames" made Claude write its own video code instead of
calling the connector.
* docs(mcp): where to switch the connector on in a chat
* fix(cli): templates pull unpacks under .hyperframes so a watched project can take the template
* fix(cli): templates pull's staging comment fits on one line
* docs(changelog): weekly digest 2026-11-02-2026-11-09
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* docs(changelog): embed the Nov 2-9 weekly video
* docs(changelog): drop the 2k alias and name fixes as described above
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): await the player seek in seam-gate
#7378 made __player.seek() asynchronous (it resolves once the preview
picture is ready). seam-gate.mjs seeked and read in the same tick, so
every sample measured t=0: on a correct film each seam failed with
exit-visible op 0.00 and exit-moving 0 px/s, and probe found no movers.
Make the page-side seek/sample/scan helpers async and evaluate them with
awaitPromise, and note the false-failure signature in the reference doc.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* docs(skills): await the player seek in changelog-video spot checks
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
createSvgContactSheet built each path with path.join and then took the
label with split("/"), so on Windows the label was the whole path and
the 60-character cell limit cut the file name off. Use basename, as the
scroll and snapshot sheets in the same file already do.
* fix(cli): staging folders beside a project path no longer fail past 260 characters on Windows
mkdtemp does not get the long-path prefix that mkdir does, so every staging folder
made beside a project or output path failed once its full path passed 260 characters.
They now come from one helper, createStagingDir in @hyperframes/core/atomic-file, and a
scan test fails on a new mkdtemp beside a path.
* fix(cli): give mkdtemp the long-path form instead of replacing it
CodeQL treats mkdtemp as the safe way to make a folder under the system temp dir and
flagged every file written into a folder from the mkdir-based helper. createStagingDir
now calls mkdtemp with the long-path form of the path on Windows, so both hold.
* fix(cli): leave non-Windows paths alone when making a staging folder
The long-path form is now chosen by the shape of the path, not process.platform,
so tests that pretend to be Windows keep working and the comment share does not rise.
* fix(core): a staging folder keeps a Windows path that already names its namespace
* fix(core): a staging folder keeps a Windows device path as the caller gave it
* fix(core): the staging folder helper's comment fits the line limit
* fix(cli): no macOS developer-tools prompt from git, xcrun or cc checks on a Mac without the tools
* fix(cli): treat Xcode.app as installed tools and only take executables from PATH
* test(cli): give skills tests a developer folder so darwin cases run on any host
* fix(cli): read macOS tool paths the same way on every host
Follow-ups from review on #7443 and #7444:
- The quote hint returned the first run of words that existed, so a
"Final Cut/" folder hid "Final Cut export.mp4". It now names the longest.
- "--compare=my clip.mp4" got no hint; the value after = now starts a run.
- The telemetry section lists the skill a reverse or templates command ran
for, and the HeyGen session id a recorded reverse run carries.
* feat(skills): route a style outside the templates instead of listing them
A request to make a new video in some film's style reached style-sampling,
which only knows its five templates and answered any other style with that
list. Now it says the style is not one of them and offers the two routes
that fit: rebuild a short video and give it the person's content
(reverse-any-video), or a new film on their topic that takes the style's
grammar (look.md, "From a reference"). The hyperframes routing table points
the same request at look.md.
* fix(skills): keep the reverse-any-video Chinese trigger to videos
「照着做一个」 matched any "make one like this" request. 「照着这个视频做一个」
keeps it to a video, as review on #7445 suggested.
* fix(skills): ask for an example video when a style is only named
Both routes for a style outside the templates work from a video, so a
style the person only names needs one example first (review on #7450).
* fix(core): detect template entries structurally
isTemplateEntry only matched a raw file that starts with <template>, so the
full-document sub-composition shape (<!doctype html><html><body><template>
<div data-composition-id>...</template></body></html>) was not recognised and
resolveEntryHtmlPath compiled the raw file, whose root sits in an inert
template: no timeline, no duration, "Composition has zero duration".
A template entry is now a document whose <template> holds the composition
root with no live composition root outside it. Detection and extraction share
one lookup (locateSceneRoot), the same one readSceneRoot uses.
— Jerrai
* fix(core): drop comments from the template lookup
The comment ratchet allows no net rise in a file's comment share; the PR
description carries the explanation.
— Jerrai
* fix(core): find the scene template among several templates
locateSceneRoot read only the first <template>, so a scene template that
follows another template was treated as not a template entry. Scan every
template and take the first whose content holds a composition root.
* fix(core): drop the added comment (comment ratchet)
* chore: raise the inspector unpacked size budget by 0.5%
The inspector bundles core; the multi-template scene lookup added a few
hundred bytes and put it 10 bytes over the 1% tolerance.
* fix(core): refuse a file with several scene templates, naming them
* chore: trim the inspector size budget bump to what is needed
* fix(core): render the first scene template and lint the ambiguity instead of failing
A file with several scene templates and no live root rendered its first one
before; locateSceneRoot threw and stopped such projects rendering. It returns
the first again, and the multiple_scene_templates lint warning names each.
Co-authored-by: Jerrai <jerrai@heygen.com>
* chore(core): drop the locateSceneRoot comment line the comment ratchet counts
Co-authored-by: Jerrai <jerrai@heygen.com>
* fix(core): one shared scene-template pick for the compiler, the inliner and the loader
A scene template after a card template was detected by locateSceneRoot but
mounted from the first template of any kind, so it rendered blank.
findSceneTemplate picks the first template holding a composition root and
locateSceneRoot, resolveSubCompositionContent and the loader all use it.
Co-authored-by: Jerrai <jerrai@heygen.com>
* feat(cli): reverse and templates report their own runs; commands say which skills are installed
Gaps in what skill telemetry can answer for /reverse-any-video and
/style-sampling:
- The skill's beacon fires only when the agent runs it, and agents often
skip it. `reverse start` and `templates pull` now send `skill_invoked`
with `source: command`; beacons carry `source: beacon`. A person counts
once.
- style-sampling renders were never tagged. A pull marks the project the
episode is built in (only a folder that is already a project).
- Installs of an on-demand skill were invisible. `cli_command` carries
`skills_installed` and `skills_outdated`, the names from the cached
background skills check, so "never triggered" and "not installed" can
be told apart.
- `auth status` signed out exits 1 with a bare "Command failed", which
reads like a login failure. It now carries `NotSignedIn`, and an auth
error its own cause.
And a step that gets lost: a reverse started signed out often goes on
without ever trying to sign in. `reverse start` signed out now prints
`login_command` and a `next` line in the JSON the agent reads, not only
in the skill's reference.
* refactor(cli): keep comment share at the base in the reverse telemetry files
* fix(cli): command runs get their own event, counted once per piece of work
Review on the first cut: `skill_invoked` from the commands doubled the
#7176 beacon's single observation, and it fired before any work.
- `reverse start` and `templates pull` send `skill_command_ran`, not
`skill_invoked`; the beacon is back to its #7176 shape.
- `reverse start` sends it after the probe, with `signed_in` and the
reverse `session_id` that keys a retry; a video it cannot probe sends
nothing.
- `templates pull` sends it only when it downloads; a re-pull or a
refused pull sends nothing.
- A skills check that finds no install forgets the installed names.
- `auth status` failures carry only a code, never the auth error's text.