Skip to content

Release v1.5.0 and v1.6.0: image classes, derived codedApp.name, tenant capacity check - #4

Merged
ThisIsTJ merged 6 commits into
mainfrom
feat/tenant-capacity-check
Aug 27, 2026
Merged

ThisIsTJ merged 6 commits into
mainfrom
feat/tenant-capacity-check

Conversation

@ThisIsTJ

Copy link
Copy Markdown
Collaborator

Brings main up to date with two releases that are already tagged and already
consumed by all five workshop repos. main is currently at v1.4.0 while every
workshop pins v1.6.0, and the floating v1 tag — which every repo's CI resolves
for validate.yml — also points here.

v1.5.0 (c0ef42d)

  • Consolidate the image wrapper classes into .workshop-image-crop.
  • Enforce the 32-character Coded App name limit.
  • Derive codedApp.name from productSlug and tenant name.

v1.6.0 (3647969)

  • Make check:tenant-access actually run, and report capacity.
  • Gate the tenant check on the tenant, not the target name.

Please merge with a merge commit, not squash

v1.5.0, v1.6.0 and v1 all point at commits on this branch. A squash
rewrites them, which would leave every tag pointing at objects absent from
main's history. A merge commit keeps them reachable and matches how v1.4.0
sits on a merge commit.

feat/consolidate-image-classes is fully contained here and can be deleted
after this merges.

🤖 Generated with Claude Code

ThisIsTJ and others added 6 commits August 24, 2026 14:04
.workshop-component-visual was .workshop-screenshot without the card
chrome -- same margin, same img reset, no border, background or shadow.
Two classes for one job, differing only in whether the figure looked
like the rest of the site. The diagram it wrapped now uses
.workshop-screenshot and gets the standard treatment.

.workshop-new-chat-crop becomes .workshop-image-crop. The crop itself
cannot be merged away: it is a fixed 420x170 viewport over a
deliberately oversized 760px image, which is a different job from
displaying a figure at natural size. But naming it after one specific
screenshot guaranteed the next workshop would write its own copy rather
than reuse it, which is how the duplication started.

Consumers must repin to get this. A repo whose markup names
.workshop-image-crop while pinned to v1.4.0 or earlier renders the image
uncropped at full width, because the class simply is not in the
stylesheet it installed.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`uip codedapp deploy` refuses a name over 32 characters. `publish` does
not: it accepts the name, creates the package, and only the subsequent
deploy fails, leaving an orphan package in Orchestrator. That is how a
37-character name reached a real tenant before anything complained.

Three layers, because the schema alone cannot be trusted -- nothing
validates config against it at runtime, it only helps editors:

  - validateTarget throws, and every command path calls it, so doctor,
    build:target and codedapp check all fail before a tenant is touched
  - init refuses a product slug that would blow the budget, printing the
    arithmetic for each standard pattern rather than just a limit
  - the schema gains maxLength: 32 so an editor flags it while typing

The slug is what has to shrink. `staging-<slug>-workshop` spends 17
characters on fixed text, so 15 is the largest slug that fits every
standard pattern. Abbreviate the slug, never the prefix, since the
prefix is what participants recognise across workshops --
`comms-mining`, not `communications-mining`.

Note case-management is already at exactly 32 for its staging form, so
the headroom here is one character, not a comfortable margin.

Verified all 14 targets across the five content repos still validate.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A target that omits codedApp.name now gets <productSlug>-<tenant>,
normalised to lowercase hyphens: Bellevue_July2026_20260727 becomes
bellevue-july2026-20260727. Adding a new deployment target becomes a
matter of naming the tenant, with nothing to hand-compute.

Derived at config load, not stored. Docusaurus compiles baseUrl into
every asset path and the app is served at /<name>/, so a build and a
deploy that disagree produce a site where every asset 404s. Both paths
call getTarget(), so both see one value and cannot drift. Nothing is
written to .env for the same reason: that is the one place the name could
diverge from the build.

An explicit codedApp.name always wins. That is the escape hatch for a
tenant whose name will not fit once the slug is prepended -- six of the
twelve deployable targets today, all of them dated training tenants.
Nothing is truncated automatically: coding-agents-testcodingagents-
20260629 and its 20260707 sibling front-truncate to the same 32
characters, so truncation would silently point two deliveries at one app.

Adds a config-wide uniqueness check, which derivation is what makes
necessary. `local` and a hosted target normally point at the SAME tenant
-- the difference is how the site is served, not which tenant it talks to
-- so both deriving would produce one name and one deploy would replace
the other's app. Per-target validation cannot see that.

No behaviour change for the five content repos: every target names itself
explicitly, so derivation is inert. Verified all 17 still resolve and a
real target build still passes.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The check had never executed once since the flat-scripts migration: it
imported './workshop-target.mjs' but the file lives in src/config/, so
every invocation died with ERR_MODULE_NOT_FOUND. It was referenced by the
schema and reported on by doctor while being entirely dead, which read as
coverage. `node --check` in CI cannot catch that, so kit-ci gains a step
that resolves every relative import; it was mutation-tested against this
exact bug.

Wiring it up would also have failed the build. docusaurus-target invoked
it through the throwing run(), so the crash above would have broken any
deploy for a target setting requiresTenantAccess -- the opposite of "never
blocks a build". Advisory steps now go through runAdvisory(), which reports
and continues: shipping a workshop site must not depend on a reachable
tenant.

On top of the permission checks, the check now reports capacity, which is
the half that matters on the morning of a delivery:

- AI Units and Agent Units. Headroom is derived, not read: Available is
  null for consumables and Allocated is what the org handed to tenants.
  The two licence surfaces disagree -- on growthuipath the summary
  reported 3,675,583 AI Units consumed against the same 5,000,000 pool
  that the per-tenant rows summed to 1,796,083. The larger figure wins,
  because trusting the smaller one overstates headroom 2.4x in the one
  check meant to stop a workshop running out. The gap is printed.

- IXP quotas, via `re get quotas`. These are a Reinfer surface and uip
  cannot see them at all, so limits are discovered rather than declared:
  sources, datasets, buckets, comments, extraction_predictions and the
  rest arrive with real hard_limit and current_max_usage. A target
  declares only the per-participant need.

A shortfall reports how short and what to raise the limit to -- current
usage plus the full need plus a margin, not the gap, since existing usage
does not go away. Context resolution follows the order `pulse` already
uses (RE_CONTEXT, then UiPath_<org>_<tenant>, then an endpoint match) so
the two commands cannot disagree about which context belongs to a target.

Missing local tooling is a SKIP, not a warning, counted separately with a
note on how to enable it. `re` installs separately from uip and most
machines will not have it; a check that warns on every build is a check
nobody reads.

validate-config now reads the overlay-merged config, so config/targets/*
targets are validated at all -- they never were, despite the file's own
comment promising every target. It also rejects unknown capacity keys,
because aiUnitsPerParticipants would otherwise fall back to 0 and switch
the threshold off while still reporting green.

The lessons from ca-tenant-access-wiring.patch are recorded in
agents/base.md so they reach every content repo.
The UiPath Labs convention <tenant>-Participants / <tenant>-Facilitators is
confirmed correct, but it holds only for the uipathlabstraining tenants that
host real deliveries. The earlier README wording implied the convention itself
was unreliable, which was the wrong conclusion from the right observation: the
mismatch was found on growthuipath/Workshops, a testing tenant.

That distinction matters for wiring the remaining repos. A target named
"growth" addresses a real training tenant in workshop-ixp but the shared
testing tenant in the other three, so switching the check on by target name
would put "group not found" on every build for names that were never expected
to exist there.
@ThisIsTJ
ThisIsTJ merged commit e1f610b into main Aug 27, 2026
1 check passed
@ThisIsTJ
ThisIsTJ deleted the feat/tenant-capacity-check branch August 27, 2026 18:41
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant