Skip to content

feat(uipath-bundles): one preview skill for knowledge and skill bundles (cli#6015) - #4647

Merged
alexenica merged 25 commits into
mainfrom
refactor/kb-generic-bundles-api
Nov 13, 2026
Merged

alexenica merged 25 commits into
mainfrom
refactor/kb-generic-bundles-api

Conversation

@alexenica

@alexenica alexenica commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

Pairs with UiPath/cli#6015, which moves uip or kb content commands onto Orchestrator's generic /api/bundles routes and adds the preview command group uip or skill-bundles (alias skill-bundle).

What this PR does

One preview skill, uipath-bundles, replaces uipath-knowledge-bundles and uipath-skill-bundles. It covers every kind of Orchestrator bundle: knowledge bundles through uip or kb, skill bundles through uip or skill-bundles, with room for later kinds. It is preview like both skills it replaces. Both command groups register only on a prerelease uip build (isPreviewBuild in orchestrator-tool tool.ts), and only on tenants with that kind's feature on (Features.KnowledgeBundles / Features.SkillBundles). Step 0 checks both, and a miss means stopping.

Why one skill is safe

The two kinds differ where it matters most: a knowledge bundle's later versions come only from merged change proposals, while a skill bundle publishes them directly. A mixed-up agent still gets a clear refusal, never an unreviewed knowledge-bundle version:

  • The server refuses a direct later-version publish on a knowledge bundle. api/bundles grants BundleAccess.Publish to no knowledge-bundle caller and answers 400 (Orchestrator src/Bundles/AGENTS.md).
  • cli#6015's kind guard makes kb refuse a skill-bundle key, and skill-bundles refuse a knowledge-bundle key, before any content is read. The refusals are This key is not a knowledge bundle. and This key is not a skill bundle., with invalid_argument.

Shape

  • Kind first. The kind comes from what the user or task names, or from which group's get/list returns the record; get has no Kind field. A key never goes to the other group (Critical Rule 1).
  • Shared, written once: Step 0 (both groups preview-only; per-kind feature refusal; the content-404 + get check), folders, keys and paging, workspaces and the .okf marker (status/sync), version 1 and create --input recovery (RecordCreated / PublicationConfirmed / Data.Key, versions first, never re-create), path rules, and one failure table with a Kind column.
  • Per kind:
    • --destination and failed downloads: skill bundles need a new path; knowledge bundles may download over a directory, with Removed/Kept.
    • Workspace inference: knowledge bundles only.
    • Later versions: the full change-proposal workflow for knowledge bundles; publish with --yes only after the user approves DeletedPaths for skill bundles.
    • Tags: skill bundles only. share and build: knowledge bundles only.
    • Each group's exact command table and permissions.
  • Every message, ErrorCode and field the two skills quoted, verified against cli#6015 up to c73d26f88, is kept.

Rename mechanics (CONTRIBUTING § Renaming or removing a skill)

  • The skill: skills/uipath-bundles/ replaces both folders. Git shows the new SKILL.md as an add and two deletes rather than a rename, because it merges two files and is mostly rewritten.
  • Registries: assets/skill-status.json has one uipath-bundles entry (preview); skills.sh.json (Platform & Operations), the README table and one CODEOWNERS line (same owners) follow. The Studio Web ratchet has one entry, at fewer Node-only mentions than the two old skills combined. The planner's OPERATE_SKILLS has one name.
  • Tests: all tasks move under tests/tasks/uipath-bundles/, keeping their leaf directories, so the reference.directory: .. wiring and _shared/sb_calls.py keep resolving. Task IDs become skill-bundles-kb-* / skill-bundles-sb-*, tagged uipath-bundles.
    • Smoke: conflict_recovery, paging_a_long_history, status_before_proposing, content-404-permission-stop, immutable-tag-stop, publish-full-workspace.
    • E2e: e2e-publish-tag-read.
  • Activation: tests/tasks/activation/uipath-bundles.jsonl holds 37 positives (the 22 skill-bundle prompts plus 15 new knowledge-bundle ones), with the criterion renamed. The near-domain negatives are kept.

E2E task prepared but not executed. e2e-publish-tag-read runs a live lifecycle on a unique scratch bundle: create and v1 from a seed, download, edit, publish v2, and a mutable stable tag. It is graded on what the server holds afterwards, read back through uip, with a pre_run preflight that aborts as ERROR without the CLI build or the feature, and best-effort cleanup. It is skip: true because the eval tenant has Features.SkillBundles off (cli#6015 comment).

Planner tests

For the planner and check-genome code owners (@RaduAna-Maria, @roberts-cliff, @danmunteanu1, @vlad-coteanu-uipath), here is why this PR touches your files:

  • skills/uipath-planner/scripts/check-genome.mjs (be05906): the bundle skills' entry in OPERATE_SKILLS, now one name, uipath-bundles. test_skill_lists_cover_the_shipped_skills (the uipath-genome checker unit tests job) requires every shipped skill there, and failed without it. Nothing else in the planner changes.
  • tests/tasks/uipath-planner/smoke_skill_activation.yaml (c265a2f, its own commit): editing check-genome.mjs makes the smoke workflow run the whole planner suite, and this task failed deterministically, on main too (fix(uipath-planner): escape a hidden BOM in sdd-fingerprint #4638, run 43209115399).
  • skill-planner-genome-from-sdd is not changed. It failed here twice at 0.60 and passed on fix(uipath-planner): escape a hidden BOM in sdd-fingerprint #4638 at 1.00. In the run artifact, the agent wrote a process genome with its one component in a subfolder, instead of the single component genome the task and genome-format-guide.md § Two Levels expect. Main's check-genome.mjs gives the identical verdict on those files, so it is the agent's choice, not this PR.
    • Possible nudge for the owners: genome-from-sdd-guide.md rule 3 ends with "the process genome's Deployment lists the SDD's project names" in a rule that otherwise makes a dispatcher and performer one component. That may steer some runs toward a process genome.

Validation

  • npm run skills:validate, check-skills-sh.py, check-skill-status.py (31 skills), check-skill-links.mjs, check-sensitive-content.py, hooks/validate-skill-descriptions.sh (description 874 chars; with when_to_use, 1,428 of 1,536), Studio Web ratchet, check-task-host-paths.py and check-task-doc-citations.py: OK. The genome checker's skill lists cover every shipped skill. YAML and JSONL parse.
  • check-skill-verbs.py: 0 blocking. The or skill-bundles mentions are soft-stale (valid or prefix; not in the catalog snapshot until the CLI ships).
  • Graders, run from a staged copy of tests/tasks/uipath-bundles/: every mode exits 0 on a valid log and 2 on a missing log. The local controls also pass:
    • 32 call-log cases: aliases, option forms, last value wins, wrong bundle, ordering.
    • 28 re-login cases.
    • The fake-uip e2e helper cases.

Merging ahead of the CLI release: the skill documents commands that exist only on a prerelease build with cli#6015. Until then its Step 0 finds unknown command and stops, as designed.

🤖 Generated with Claude Code

…e as a permission stop

UiPath/cli#6015 moves `uip or kb` content commands onto Orchestrator's
generic /api/bundles routes, which hide a bundle kind the caller cannot
access. A caller with the bundle record but no knowledge-bundle content
permission, or a tenant with skill bundles but not knowledge bundles, now
gets 404 `BundleDto does not exist.` where it used to get a 403. Tell the
agent to confirm with `kb get`, then stop and report, instead of
concluding the bundle is gone and re-creating it. Add a smoke task that
grades that path.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@github-actions

github-actions Bot commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

Claude finished @alexenica's task in 48s —— View job


Summary

Updates uipath-knowledge-bundles for UiPath/cli#6015, where content commands move to the generic /api/bundles routes. On those routes, a bundle the caller can see but has no content permission for answers 404 BundleDto does not exist.. The skill now tells the agent to run kb get before deciding the bundle is gone. A new smoke task grades that the agent stops instead of re-creating the bundle, deleting it, logging in again or retrying.

Change-by-Change Review

1. skills/uipath-knowledge-bundles/SKILL.md: Critical Rule 5 (line 19)

Low. The guidance is correct, but the second sentence is hard to parse: "A 403, a feature refusal, and a 404 BundleDto does not exist. on a bundle get still shows all need an administrator." The phrase "on a bundle get still shows" can be read as "a 404 returned by get", which is the opposite of the meaning. This is a Critical Rule, so the wording should be unambiguous. Suggested:

A 403, a feature refusal, and a content-command 404 BundleDto does not exist. for a bundle that get still returns all need an administrator.

Fix this →

2. SKILL.md Step 0: new 404 bullet (line 36)

Low. Two small gaps:

  • The check command uip or kb get <bundle-key> has no folder flag. Critical Rule 2 requires a folder on every command, and the prose says "in the same folder". Make the command copy-paste ready: uip or kb get <bundle-key> --folder-path <p>, the same form as the table at line 76. Agents copy commands literally.
  • The get-fails-too branch says "find its key with list". If the cause is a wrong folder, list in that same folder will not show the bundle either, and Rule 2 says there is no tenant-wide listing. Spell it out: "ask the user which folder the bundle is in, then run list there". Otherwise the agent may loop on list in the wrong folder.

Fix this →

3. SKILL.md "What the output means" (lines 120–124)

OK. It gives one line and then points to Step 0, so it bridges rather than repeats. "Two" → "Three" is updated.

4. tests/tasks/uipath-knowledge-bundles/smoke/content_404_is_a_permission_stop.yaml

OK. Notes:

  • The tags include the required skill, tier, mode:diagnose and lifecycle:discover. This is the only kb task with a lifecycle:* tag; the three sibling tasks lack it, which predates this PR. task_id follows skill-<domain>-<capability>. Weights fit the guidance, and every criterion grades actual CLI calls, not what the agent says it did.
  • The prompt is minimal and does not hint at the 404.
  • The --help catch-all is placed last, with a comment explaining why. Good.
  • The "Checked the bundle record with kb get after the 404" criterion counts get calls but does not check their order. An agent that runs get before download also passes. This is acceptable for a smoke task, but the description claims more than the criterion checks. Either drop "after the 404" or accept it as-is.
  • The download max of 1 forbids a retry, which matches the skill. An agent that retries once with --destination changed also fails. That is fine, since the skill says not to retry.

What's Missing

  • Nothing blocking. The sibling smoke tasks (conflict_recovery, paging_a_long_history, status_before_proposing) lack a lifecycle:* tag. This predates the PR and is out of scope here, but worth a follow-up.
  • No flavor overrides exist for this skill, so no flavor review was needed. No CODEOWNERS or registry change is needed, because no skill folder was added or renamed.

Area Ratings

Area Status Notes
Frontmatter OK Unchanged
E2E Tests OK New smoke task covers the new behavior
Skill Body OK (Low nits) Rule 5 wording; Step 0 get command lacks folder flag
References & Assets OK None touched
Repo Hygiene OK Scoped to one skill; no secrets or paths; the GUIDs are obvious placeholders

Issues for Manual Review

  • I can't check the 404 envelope against the real post-#6015 CLI output: ErrorCode: not_found, Instructions: "HTTP 404: BundleDto does not exist.", Retry: "RetryWillNotFix", exit 1. Please confirm against the cli#6015 edge-case matrix.
  • Is the "upload step of change-proposal create/update" really behind the generic routes? The PR body says objects/upload moved, so this looks right.

Conclusion

Approve with minor nits. The guidance prevents a costly mistake: re-creating a bundle that is only hidden, which destroys its history. The test grades behavior correctly. The two Low items (Rule 5 wording, the folder flag on the get check) are worth a quick fix but are not blocking.

@github-actions

github-actions Bot commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

Claude finished @alexenica's task in 53s —— View job


Coder-eval task lint (advisory)

1 task YAML changed. Verdicts: 0 Critical, 0 High, 1 Medium, 0 Low, 0 OK.

Rubric: .claude/commands/lint-task.md. This check is advisory and never blocks merge.

Evidence of passing run

  • ❌ High: the PR body says "Smoke task run: by the smoke-skills check on this PR". That hands the run off to CI. It does not say the task has passed. Once smoke-skills is green, please edit the description to add a line like: Ran skill-kb-content-404-is-a-permission-stop-smoke and it passed.

Per-task lint

tests/tasks/uipath-knowledge-bundles/smoke/content_404_is_a_permission_stop.yaml, verdict: Medium

Issues:

  • [Medium] Meaningful coverage: nothing grades the diagnosis itself (L94–140). An agent that runs get and then tells the user "the bundle was deleted" or "try again later" passes every criterion. The task's point is to report missing permission or feature.
  • [Low] Could pass for the wrong reason: kb get (L115–120) has no ordering, so a get run as a preflight before the download also passes, even though the description says "after the 404".
  • [Low] Could pass for the wrong reason: the retry guard covers only download (L133–140). Retrying through the other mocked content verbs (file L50, versions L58, archive L66) goes ungraded, so an agent can loop on kb file and still pass.
  • [Low] Could pass for the wrong reason: the login guard (L122–129) probably also forbids read-only uip login status. That would fail an agent that only checked whether it was logged in. Critical Rule 5 forbids re-running uip login, not inspecting the session.

Suggested fixes:

  • Add a final-response check: a file_contains/response_contains-style check, or a light llm_judge alongside the deterministic guards, that requires permission|feature|administrator and excludes deleted|does not exist|re-create.
  • If cli_called supports ordering (after: / sequence), use it so or kb get must follow or kb download. Otherwise, edit the description so it doesn't promise "after".
  • Add max_count: 1 guards on or kb file, or kb versions and or kb archive, or one combined content-verb guard, so retries through sibling verbs fail.
  • Narrow the login guard to the mutating verb (e.g. login with no subcommand) if cli_called can tell it apart from login status.

CLI verb reachability: not run here, because scripts/check-cli-verbs.py needed approval in this runner. The criteria use cli_called + verb: rather than command_pattern, and every verb they name (or kb get|create|delete|download, login) is one the skill documents. Near-duplicate: none. The siblings (conflict_recovery, paging_a_long_history, status_before_proposing) share the scaffold but test different operations. Prompt: minimal and states only the goal. Good.

Within-PR duplicates

  • No duplicate clusters detected (only one task added).

Conclusion

  • ⚠ 1 task has issues, max severity Medium. Separately, no passing-run claim yet (High). Advisory only, not blocking merge.

alexenica and others added 3 commits November 12, 2026 14:53
…nd full command

The Studio Web ratchet caps Node-only command mentions per skill, and the
Commands table already spells out `uip or kb get`. Refer to it by verb,
as the rest of the skill does.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A failed `get` does not prove the bundle is absent: authorization,
feature, network and server failures pass through it, and a tenant with
knowledge bundles off fails `get` too. Branch on its envelope: success
means missing content permission, `not_found` means ask which folder,
anything else is reported as it came. Soften Rule 5 to "does not prove
the bundle is missing".

The smoke task moves to its own leaf directory. Help calls get their own
response and are excluded from every count; the positive criteria are
pinned to the prompt's key, folder and destination; failure mocks exit 1
like the CLI. Checked against coder-eval's argv matcher.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ontent retries

The login guard counted read-only `login status` / `login which`; read
the call log and fail only a re-login (bare `login`, `login --flags`,
`login refresh`). Add a guard so an agent cannot loop over `file`,
`versions` and `archive` after the 404.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@alexenica

Copy link
Copy Markdown
Contributor Author

Reply to the Claude review (written against d7c998f):

  1. Rule 5 wording. Fixed in 21c1930. The rule now says a content 404 "does not prove the bundle is missing" and points to the get check in Step 0.
  2. Folder flag on the get command. Not an issue any more. bf1bba7 removed the second full uip or kb get command: the Studio Web ratchet (tests/scripts/studioweb-browser-commands.json, ceiling or kb get: 1) caps Node-only command mentions, and the Commands table already gives uip or kb get <bundle-key> --folder-path <p>. Step 0 now says "run get once, on the same key in the same folder", the same way the skill names verbs elsewhere.
  3. The get-fails branch loops on list in the wrong folder. Fixed in 21c1930. The check now branches on get's envelope: on success, report the missing content permission; on not_found, ask the user which folder holds the bundle before listing any other folder; on any other failure, stop and report it, since it does not prove the bundle is missing.
  4. "after the 404" is not checked. Fixed in 21c1930. The wording is gone (cli_called has no ordering facet), and the get criterion is now pinned to the prompt's key and --folder-path Shared.
  5. Manual review: the envelope. Checked against the CLI. classifiedFailure (packages/common/src/classified-failure.ts) builds Message = the command's failure text and Instructions = the server detail. classifyHttpStatus maps 404 to not_found / RetryWillNotFix. The cli#6015 edge-case matrix shows Error downloading version HTTP 404: BundleDto does not exist., exit 1. The upload step of change-proposal create/update goes to /api/bundles/{id}/objects/probe and PUT …/objects/{hash}, as the cli#6015 live smoke trace shows.

@alexenica

Copy link
Copy Markdown
Contributor Author

Reply to the task lint (written against d7c998f; the task now lives at tests/tasks/uipath-knowledge-bundles/content-404-permission-stop/, in its own leaf directory per .claude/rules/test-writing.md):

  • [High] Evidence of a passing run. I'll update the PR body with the smoke-skills result once it finishes on the final head.
  • [Medium] Nothing grades the diagnosis. Not adding this, by design. .claude/rules/test-writing.md § Success Criteria says to grade side effects and never agent monologue. Every wrong path here (re-create, delete, re-login, retry, skipping the record check) is a CLI call, and each one is now graded.
  • [Low] get order. Fixed in 21c1930. The "after the 404" claim is dropped, and the criterion is pinned to the prompt's key and folder.
  • [Low] Retries through file/versions/archive. Fixed in 0855c18. A combined guard allows at most one of those after the download.
  • [Low] The login guard catches login status. Fixed in 0855c18. A run_command over cli_mocks/calls.jsonl now fails only a re-login: bare login, login --<flags>, or login refresh. Read-only login status and login which pass.
  • Also fixed in 21c1930: --help calls get their own mock response and are excluded from every count, and the download criterion is pinned to the key, folder and ./kb destination. Checked with coder-eval's argv_match.

alexenica and others added 4 commits November 12, 2026 15:14
…ogin, guard every download

The call log keeps raw argv, but the CLI strips its global options before
parsing, so `uip --output json login refresh` passed the re-login guard
and `uip login --output json status` failed it. Move the check into
check_no_relogin.py, which strips the global options from
`stripGlobalOptions` first.

The pinned download criterion missed a retry to another destination.
Add a max-one guard over every non-help `kb download`, and accept the
`-d` spelling of `--destination`.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…g the subcommand

The re-login check matched a subcommand name anywhere in argv, so an
option value read as a subcommand: `login --tenant status` (a real
login) passed and `login --tenant refresh status` (read-only) failed.
Consume `uip login`'s own options and their values (long, short and
`=` forms, from cli commands/login.ts) and take the first positional,
as Commander does. Fail with exit 2 on a missing log, malformed JSON,
or a record without a list `argv`, instead of passing or throwing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
UiPath/cli#6015 adds `uip or skill-bundles` (alias `skill-bundle`), a
separate command group for skill bundles: folder-scoped, versioned file
trees of agent skills that publish later versions directly and carry
version tags, unlike knowledge bundles, whose later versions come only
from reviewed change proposals. The two lifecycles contradict each other
on the rule that matters most, so they get separate skills with
redirects both ways.

The skill covers preflight, addressing by name or GUID, the folder rule,
download/read by number, tag, digest or latest, publishing a full
workspace (a missing base file is a deletion), WorkspaceUpdated after a
publish, conflicts, immutable tags, and the content-404 branch on
`get`. It documents no `share` and no `build`, which the CLI does not
offer. Two smoke tasks grade the immutable-tag stop and publishing the
whole workspace rather than a partial directory.

Registers the skill in skill-status.json (preview), skills.sh.json,
CODEOWNERS and the README table, records its Node-only Studio Web
mentions as ratchet ceilings, and adds a skill-bundles redirect to the
knowledge-bundles description.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…KILL.md

A SKILL.md anywhere in the repo can be picked up as a skill by repo-wide
linting; the fixture only needs file content, not skill discovery.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@alexenica alexenica changed the title fix(uipath-knowledge-bundles): treat a content 404 on a visible bundle as a permission stop feat(uipath-skill-bundles): add the skill-bundles skill; handle content 404 in knowledge bundles Nov 12, 2026
@github-actions

github-actions Bot commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

skill-lint: 0 new errors, 1 new warning

Rule Location Message
warning SEC-BROAD-BASH skills/uipath-bundles/SKILL.md:5 `allowed-tools` grants unrestricted Bash
What these rules check
  • SEC-BROAD-BASH: allowed-tools should scope Bash to specific commands instead of granting all of it. allowed-tools pre-approves tools while the skill is active. Bare Bash or Bash(*) lets the agent run any command without a permission prompt.

Only what this PR adds is listed, compared with the commit it branched from; issues already on the base branch are not. Updated on every push.

The genome checker's skill lists must cover every shipped skill. Skill
bundles operate a platform resource and build no artifact, so the skill
joins OPERATE_SKILLS next to uipath-knowledge-bundles.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
alexenica and others added 2 commits November 12, 2026 15:59
…base rules

cli#6015 settled `tags set <tag> --bundle <bundle>` and `tags delete
<tag> --bundle <bundle>`: the tag is positional, the bundle an option,
as KB proposals nest under --bundle-key. `--base-version` now has to
match a workspace's marker, `create --input` refuses a directory that
is already a workspace, a plain directory never gets a marker, and
publish reports DeletedPaths. Update the commands, the publishing
rules, the failure table and both task mocks; the immutable-tag task's
other-name guard reads the positional tag from the call log.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
skill-lint EVAL-DATASET-SYNC requires every skill in the activation
suite. Add 22 intent-based positives (publish, download, read, tags,
status/sync, conflicts) and three adversarial negatives near the domain
(authoring a SKILL.md, a JS bundle release, installing a skills plugin),
plus the skill_triggered criterion with stop_early.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
alexenica and others added 3 commits November 12, 2026 16:11
Every remote content command, the first publish included, needs
SkillBundleContent.View for the kind check; a write also needs the
matching Create, Edit or Delete content grant. A name is resolved
through the record, which needs SkillBundles.View; a key skips it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e binding, e2e

Skill: hidden base files are kept by default and need --include-hidden
on both status and publish; list's HasMore is exact (lookahead) and the
windowed lists report Total; only this group guards the bundle kind;
whole uppercase placeholders.

Smoke tasks: mocks and guards accept the `skill-bundle` alias; graders
parse argv the way Commander does (attached -iDIR, --opt=value, last
value wins) after stripping global options, bind publish and tags to
finance-skills or its key, require status on the published directory
and tags list before tags set, in log order. Every value option is
declared, since the matcher otherwise splits -ichanges into -h. The
task tells the agent to run uip from the task root, because the log
records no working directory. Help mocks list no `build`.

E2E: add a live publish/tag/read lifecycle task on a unique scratch
bundle, graded on what the server holds afterwards, with a preflight
that aborts as ERROR without the CLI build or the feature, and
best-effort cleanup. skip: true until the eval tenant has
Features.SkillBundles on; it has not been run.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
record_cli rules and cli_called match ANY recorded value of a flag,
while Commander keeps the last one. Drop the canned refusals keyed on a
partial --input and the folder predicate on the publish count; the
Python grader now also requires the final --folder-path to be Shared,
so a repeated --folder-path ending elsewhere fails. The versions mocks
answer each window of their two rows consistently.

E2E: `intact` compares v1 and the newest untouched files byte for byte
with the protected reference seed, and the newest invoice file must be
the seed plus the rule. Every mode, cleanup included, acts only on a
name of the generated sb-e2e-<12 hex> shape. Still skip: true; not run.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@alexenica

Copy link
Copy Markdown
Contributor Author

Reply to skill-lint:

  • EVAL-DATASET-SYNC (error): uipath-skill-bundles missing from the activation suite. Fixed in 80b8da0. Added tests/tasks/activation/uipath-skill-bundles.jsonl (22 positives), three adversarial negatives in negative.jsonl, and the skill_triggered criterion with stop_early.
  • UIP-EVAL-COVERAGE: no e2e task. Fixed in df34c2d. Added tests/tasks/uipath-skill-bundles/e2e-publish-tag-read, a live create, publish, tag and read lifecycle on a unique scratch bundle, graded on what the server holds afterwards. It carries skip: true with its re-enable condition, and has not been executed: the eval tenant has Features.SkillBundles off, and the commands ship with UiPath/cli#6015. Its pre_run preflight aborts as ERROR without the CLI build or the feature, so it cannot score on an unprovisioned tenant.
  • XSKILL-DESC-OVERLAP (0.47 with uipath-knowledge-bundles): not changing it. Both skills operate Orchestrator bundles, so "bundle", "version", "download" and "Orchestrator" are their real domain words. The descriptions differ on the distinguishing signals: uip or skill-bundles, agent skills and tags here; uip or kb, OKF documents and change proposals there. Each redirects to the other with →. The new activation rows (22 positives) are what measures the confusion, per the rule's own guidance.
  • SEC-BROAD-BASH: not changing it. This matches uipath-knowledge-bundles and the other CLI-driven skills. The workflow needs more than uip: copying an edit into a downloaded workspace, inspecting files, and writing outside the workspace. Scoping Bash to uip alone would make the documented publish workflow prompt on every step, or fail.

…ation smoke

skill-planner-smoke-skill-activation asks the agent to write the verbs
the skill defines and grades them against skills/uipath-planner/SKILL.md
§ Verbs. 9e33121 (#4330, PILOT-9071) wrote that assertion with `build`,
but the planner's verbs at that commit were design, decompose, author,
extract and edit: `build` belongs to uipath-genome (its SKILL.md § Verbs).
a4ccda0 (#4379) then added `check`. So an agent that copies the table
correctly scored 4/5 = 0.85 and failed every run, on main too (#4638,
run 43209115399).

Align only the includes list to the current table: design, decompose,
author, extract, edit, check. Weight, pass_threshold, tags, prompt and
the planner skill are unchanged. Checked with coder-eval's own
file_contains code: the current verbs pass at 1.00; the old list with
build (0.67), either missing verb (0.83) and build alone (0.00) fail.
The CI agent's real paths.txt from run 43238788631 scores 6/6.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
alexenica and others added 5 commits November 12, 2026 17:21
UiPath/cli#6015 (c925ae86d) makes `kb create --input` and
`skill-bundles create --input` report a failure after the record was
created as "Bundle '<key>' was created, but its initial publication was
not confirmed." with Data { Key, Name, RecordCreated: true,
PublicationConfirmed: false }. PublicationConfirmed: false means the CLI
got no confirmed commit, not that the bundle is empty.

Both skills now say: run `versions` on Data.Key in the same folder and
tenant first; v1 listed means done, none means `publish` that key;
never repeat `create`, never delete the record. In the KB skill a 409
from `create` is a taken name, and a 409 from `publish` while finishing
such a create means v1 landed, not that a change proposal is needed.
KB also notes that a bad --input is refused before any record exists.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ontent yet"

Review of 9e79ba5 against cli#6015 c925ae86d: the partial-create
failure keeps the underlying HTTP status, so `create --input` can answer
409 with Data.RecordCreated: true; the "409 on create, pick another
name" advice now applies only without that marker, and RecordCreated
takes precedence whatever the status. An unpublished bundle's
`versions` answers 404 "Bundle '<key>' has no content yet."
(BundleStore.cs), not an empty list: the resume `publish` now follows
only that answer (or a successful empty list); any other `versions`
failure stops under Step 0 rather than being read as "empty".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
cli#6015 8b41163b0 makes Data.Key optional on the partial-create failure:
when the server returned no key after the record POST, the message names
the bundle by name and Data carries only Name, RecordCreated: true and
PublicationConfirmed: false, and the Instructions say to list by name
first. The recovery Instructions now also print the real folder, tenant,
input and message, shell-quoted.

Both skills: without Data.Key, first `list` the same folder with
--search set to Data.Name and take the exact match's key, then the
versions check on that key; and point at the printed commands.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tes files

cli#6015 61d98dc6a: `skill-bundles publish` from a directory without a
marker, with --base-version above 0, that would delete base paths now
stops before any upload or commit: ErrorCode confirmation_required,
RetryWillNotFix, exit 1, Data { Key, BaseVersion, DeletedPaths }, and
proceeds only with -y/--yes. Workspace publishes are unchanged.

The skill shows the user every DeletedPaths entry and asks before
re-running with --yes, never adds --yes on its own, and prefers a
downloaded workspace for edits. Adds [--yes] to the publish row and a
failure-table row.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s state

cli#6015 a1db4cbef (head 1cb482ffe): every KB content command now checks
the bundle kind first and refuses a skill-bundle key with
invalid_argument "This key is not a knowledge bundle. ..." before any
content call; the KB skill names it and routes to uipath-skill-bundles.
`skill-bundles download` now stages its extraction, so a failure leaves
the destination as it was with no marker, and a destination filled
meanwhile is refused ("The destination is no longer empty. Download into
a new path."); the skill-bundles Reading section says so.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
alexenica and others added 4 commits November 12, 2026 18:04
Review of 662d472: in the KB skill the wrong-kind bullet sat between
the content-404 bullet and its nested Success / not_found / other
branches, detaching them; it now follows them. In the skill-bundles
skill, "run it again" covered refusals, the content 404 and a wrong-kind
key, which fail before anything is written; the retry now applies only
to a download that failed while writing, once its cause is fixed, and
those failures keep the Step 0 handling.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ted guidance

- Move the argv layer the two skill-bundle graders copied (global-option
  stripping, Commander-style option parsing, call-log reading, mode dispatch)
  into tests/tasks/uipath-skill-bundles/_shared/sb_calls.py; both tasks now
  stage the group folder as their reference (directory: ..), like the
  uipath-maestro-bpmn and uipath-insights tasks.
- Drop the unused `tags delete` option table from check_immutable_tag.py.
- check_no_relogin.py: login_subcommand returns the token or None instead of
  a redundant (found, token) pair.
- live.py: derive UNTOUCHED from SEED_FILES, drop unused exit codes, one
  preview helper for failure messages.
- uipath-skill-bundles: the failure-table row for an unconfirmed create points
  to the Publishing recovery instead of restating it.
- uipath-knowledge-bundles: drop the third "worth recognizing" bullet, which
  repeated Rule 5 and Step 0.

No grading change: the old and new graders give identical exit codes and
output on 182 synthetic call logs (all modes, malformed and missing logs,
global and short options, aliases, help renders).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…xist

cli#6015 c73d26f88: `skill-bundles download` now refuses any existing
--destination, even an empty directory, before contacting Orchestrator
(invalid_argument, "The destination already exists.", "Pass
--destination with a new path that does not exist. Do not create it
first."), and never removes an existing path. The skill says to pass a
path that does not exist and not create it first, retries a failed write
into a new path, drops the "no longer empty" message the CLI removed,
and adds the refusal to the failure table. kb download is unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s into one

Replace uipath-knowledge-bundles and uipath-skill-bundles with one preview
skill, uipath-bundles, covering every kind of Orchestrator bundle: knowledge
bundles through `uip or kb` and skill bundles through `uip or skill-bundles`,
with room for later kinds.

A mixed-up agent cannot do harm: the server refuses a direct later-version
publish on a knowledge bundle (api/bundles grants BundleAccess.Publish to no
knowledge-bundle caller and answers 400), and cli#6015's kind guard makes each
command group refuse the other kind's key before reading content.

The skill identifies the kind first (from what the user or task names, or
from whose `get`/`list` returns the record; `get` has no Kind field), then
shares one Step 0 (both groups preview-only, per-kind feature refusal), one set
of folder/key/paging rules, workspace and marker rules, version 1 and the
create --input recovery, and one failure table. What differs is kept per
kind: the destination rule and failed-download handling, workspace
inference, later versions (change proposals vs. publish with --yes only after
the user approves DeletedPaths), tags, share and build. Every quoted message,
ErrorCode and field from the two skills is kept.

Registries move with it: skill-status.json (preview), skills.sh.json,
CODEOWNERS, the README table, the Studio Web ratchet, and the planner's
OPERATE_SKILLS. Tests move under tests/tasks/uipath-bundles/ with their leaf
directories, so the `..` reference and _shared/sb_calls.py keep resolving;
task ids become skill-bundles-kb-* / skill-bundles-sb-*. The activation set
becomes uipath-bundles.jsonl, with knowledge-bundle positives added.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@alexenica alexenica changed the title feat(uipath-skill-bundles): add the skill-bundles skill; handle content 404 in knowledge bundles feat(uipath-bundles): one preview skill for knowledge and skill bundles (cli#6015) Nov 12, 2026
@github-actions

Copy link
Copy Markdown
Contributor

⚠️ Smoke test coverage gap — the following changed skills have no tests under tests/tasks/:

  • uipath-knowledge-bundles

Consider adding smoke tests before merging.

Review of 1ab078c against cli#6015 c73d26f88:
- "Drop the key" on a workspace mismatch is knowledge-bundle inference
  only; skill bundles keep the named bundle and folder and download it
  into its own new directory.
- local_permission_denied with Context.path is the knowledge-bundle
  marker refusal; skill bundles answer "The workspace marker is invalid
  or unreadable. Download a fresh workspace before publishing."
  (publish), "No valid workspace marker. Run 'uip or skill-bundles
  download' first." (status) and "Pass a workspace downloaded from this
  skill bundle." (sync).
- The kb download server_error advice applies to its two archive
  messages (mismatched bundle/version, unreadable shipped marker); any
  other server_error is reported as it came.
- Rule 9 again says an immutable tag can never be moved or deleted, by
  anyone.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@alexenica

Copy link
Copy Markdown
Contributor Author

Reply to skill-lint at 45a9115:

  • SEC-BROAD-BASH on skills/uipath-bundles/SKILL.md:5: not changing it. The reason is the one given for the two skills this replaces. The documented workflows need more than uip: copying an edit into a downloaded workspace, inspecting and removing drifted files before status and publish, and writing output outside the workspace. Scoping Bash to uip alone would make every one of those steps prompt or fail. The previous XSKILL-DESC-OVERLAP warning no longer applies, because the two bundle skills are now one.

@alexenica
alexenica merged commit ffb978f into main Nov 13, 2026
52 checks passed
@alexenica
alexenica deleted the refactor/kb-generic-bundles-api branch November 13, 2026 10:29
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.

4 participants