Repository navigation
Update CI, docs and repo scripts for the genome split (6/6) [PILOT-9071] - #4379
Conversation
|
Claude finished @RaduAna-Maria's task in 3m 27s —— View job SummaryThis PR is the last of the six stacked genome PRs. It resolves contradictions in the planner guides and templates: package versions, gap markers, deploy routing, HITL routing and the Platform constraints row. It gives each repeated rule one home. It also splits Most of the consolidation is clean. Nothing still links to the deleted Change-by-Change Review1.
|
| Area | Status | Notes |
|---|---|---|
| Frontmatter | OK | Only change: allowed-tools gains WebSearch (justified); description untouched |
| E2E Tests | Issue | Renames are clean; no coder-eval run for agent-facing changes |
| Skill Body | OK | SKILL.md trimmed to routing; the removed rules still have a home |
| References & Assets | Issue | Items 1–4 (Medium), 5–7 (Low) |
| Repo Hygiene | OK | CODEOWNERS entry for the deleted file removed; no secrets or personal paths; no cross-skill file reads |
Issues for Manual Review
- Whether the decompose Q2 default (
interactive) is the intended product decision. It conflicts with the(recommended)option and withdesign's Autonomous recommendation. - Changing the task_ids breaks the run history for these tasks on the codereval dashboard (the PR flags this).
- Whether
uipath-api-workflowreally always deploys throughuipath-solution(plan-guideStep 4) under the new "the deploy skill follows the packaging" rule.
Conclusion
Request changes (minor). The consolidation is sound and removes real contradictions. Fix items 1–4 before merge; three are one-line pointer or table moves, and item 2 needs a decision. Items 5–7 are optional polish. Please also add a coder-eval run for the decompose/plan smoke tasks.
|
Claude finished @RaduAna-Maria's task in 1m 33s —— View job Coder-eval task lint (advisory)This PR changes 11 task YAMLs (5 renamed, 6 with edits to comments only). Verdicts: 0 Critical, 0 High, 2 Medium, 3 Low, 6 OK. Rubric: .claude/commands/lint-task.md. This check is advisory and never blocks merge. Evidence of passing run
Per-task lint
|
skill-lint: 0 new errors, 3 new warnings
What these rules check
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. |
50de7c8 to
de270bd
Compare
d16cfb8 to
99a33c2
Compare
039379d to
3e893de
Compare
danmunteanu1
left a comment
There was a problem hiding this comment.
🟡 Commented · 11 blockers · 20 non-blockers
Re-review at 3e893de1d. The stack was rewritten with the builder named uipath-genome again. The rename is pure: at c90f2ebed the tree equals the reviewed 039379d5a apart from the name, the SKILL.md heading and one task id. 69a7b710a moves the refactor questions into extraction and adds the Decided: project chain; 3e893de1d drops the legacy-name handling.
Fixed
- The builder is
uipath-genomethroughout. Thanks. Decided:sub-projects: the chain names the projects the build creates (the queues between them are new, below).- What leaves
editon a designed genome has one home, the edit guide's § Any genome. - A design change on a designed genome ends in
edit's "The SDD was revised", notauthor. - The SDD hand-over sentence stated twice in the planner's SKILL.md.
- Live application has a Recommended answer for every genome.
- The builder's copy of the Decision Tree is gone; the rule file's other copies stay open.
Still open (threads unchanged)
- A retired skill name stops the build (builder
build-guide.md:9). - The builder, not the genome, decides which behaviour the rebuild follows (
genome-reading-guide.md:154). - Rule 5 against the decided values (
configuration-questions-guide.md:52). - The #4506 port: finding a workflow through the earlier build's hand-off (
execution-guide.md:68). - The #4519 port: choices the source never faced (
configuration-questions-guide.md:50). - A child SDD listed by bare filename is not fingerprinted (
sdd-fingerprint.mjs:20). - Required Windows smoke: it runs
--tags smoke, and neither builder task taggedwindows(build_component, the newdecided_chain_projects) carriessmoke. - A stale or hand-edited genome builds unchecked (builder
execution-guide.md:25).
New
- 🔴 A chain of three projects has no queue design; the PR's fixture holds two queues for one unit.
- 🔴 A genome extracted for documentation is refactored on its source's products, with no target question.
- 🔴 A migration asks Fold, and its build ignores the answer.
- 🔸 Three non-blockers inline.
Checks at 3e893de1d: links, flavors, descriptions and skill status pass. skills:test fails the same 21 tests as main. The planner and builder helper tests pass except the known Windows-only fingerprint test. CI had not started the pull-request workflows on this head when this was written. Pack UiPath/migration-source-frameworks#110 needs the builder name back and now conflicts with its main (review there).
| ## Refactor questions | ||
|
|
||
| Asked after the genome questions and before the split questions ([configuration-questions-guide.md](configuration-questions-guide.md)): the target of each component decides which split and scaffolding questions apply. Options are prepared from the genome — its steps, Transactional Shape and Handoffs — and from the Source Map's deployed-state and runtime rows (§ Provenance): | ||
| Asked after the genome questions and before the split questions ([configuration-questions-guide.md](configuration-questions-guide.md)). The targets are the genome's: each component is built on the product and project its Components and Build With rows name, and those targets decide which split and scaffolding questions apply. Options are prepared from the genome — its steps, Transactional Shape and Handoffs — and from the Source Map's deployed-state rows (§ Provenance): |
There was a problem hiding this comment.
🔴 Blocker · A genome extracted for documentation is refactored on its source's products, with no target question
Retire, Fold and Target are now asked only by the planner's extract Step 6c, and only for a refactor or migration: "a documentation-only extraction asks none". This guide no longer asks them. The usual path for a UiPath source is to extract it for review and decide to refactor later: the user points at that genome and asks to refactor it, which reaches this guide. That genome's Build With rows record the source's own products. The refactor then rebuilds a Windows project as Windows and a connector-only process as an RPA process, and never offers the planner's § Refactor Targets hosts. On main, every refactor build asked Target per component.
Fix: a refactor build of a genome with no Requested change row first goes to uipath-planner edit to ask the refactor questions (Step 6c), then builds; the planner's edit guide gets that row.
There was a problem hiding this comment.
e9efd78c7 closes the gap for an attended extraction by asking Retire, Fold and Target in every extraction. That has two costs:
- An unattended extraction records the Recommended answers as the user's decisions. Step 6c takes every Recommended option and "The
Requested changerow records every answer as the user's decision". The build then answers them from the genome and never asks (refactor-guide.mdrule 2). Example: an unattended "document this solution" folds a reader into its writer (Fold is Recommended). A refactor build months later then drops the reader's trigger and handoff, and nobody chose that. - A documentation request pays for a rebuild it did not ask for: two or three questions per component plus the delivery-model question. Its Build With rows can also name a recommended target instead of the product the automation runs on today.
Fix: keep Step 6c for a refactor or migration request. A refactor build of a genome with no Requested change row goes to uipath-planner edit for Step 6c first, as in the original fix. If every genome should carry a target design, record an unattended run's answers as (default) in the row, and have the build ask any (default) topic.
There was a problem hiding this comment.
Reproduced at bcf20b12a, on the PR's own extract_component fixture (InvoiceIntake, a Windows project):
- "Document what the project does … unattended" wrote
Requested change | — (documentation request) | Target: InvoiceIntake → RPA process (Portable) — refactor answers, unattended Recommended default. - An attended "Build the automation described in
invoice-intake-genome.md" then treated it as decided, 2 of 2 runs. The opening question read "rebuilt as a Portable RPA process", the build plan read "Target: one RPA process, Portable … (from the refactor answers: the genome's Build With row)", and both runs created the project with"targetFramework": "Portable". The runtime was never asked.
An attended refactor request ("Refactor it to current UiPath practice and build…") also skipped Step 6c: no Retire, Fold, Target or delivery-model question, and the row records "Target: … Portable (the refactor answers; unattended run, Recommended option)" in a session that could ask. An attended solution extraction did ask Target and Fold (1 run each).
| return [[], false]; | ||
| } | ||
|
|
||
| const DECIDED_CHAIN = /:\s+([^:]*→[^:]*|[A-Za-z][\w.-]*)\s*$/; |
There was a problem hiding this comment.
🔸 Non-blocker · A chain written with -> or commas passes strict, and the build makes the default two projects
Probed at 3e893de1d on the PR's fixture: B, one invoice — the SDD §10: InvoiceDispatcher -> InvoicePerformer and …: InvoiceDispatcher, InvoicePerformer and InvoiceReporter both pass strict. The builder then finds no chain and names the projects <Component>.Dispatcher and <Component>.Performer.
Fix: on a B line, fail a list after the source's last : (->, ,, and) that is not →-separated.
There was a problem hiding this comment.
Fixed for names that start with a capital and hold no space (e9efd78c7). Two forms still pass strict and build the default two projects. Both were probed on the PR's fixture:
…: Invoice Dispatcher -> Invoice Performer -> Invoice Reporter…: invoiceDispatcher, invoicePerformer and invoiceReporter
DECIDED_LIST matches only [A-Z][\w.-]* names.
Fix: match any name text between the separators (no : and no →), so these fail too.
3e893de to
f85a765
Compare
danmunteanu1
left a comment
There was a problem hiding this comment.
🟡 Commented · 13 blockers · 21 non-blockers
Re-review at f85a7650b: the stack rebased onto main ec2a62b26, and f85a7650b ports what main brought in.
Verified
- The port: every line #4613, #4581, #4529, #4534, #4463 and #4533 added to the genome and planner skills is at this head, a few in the stack's wording. #4529's case Build handoff row stays out, since the stack dropped that row before.
- Links, flavors, descriptions and skill status pass;
skills:testfails the same 18 tests as mainec2a62b26. - "Run skill smoke tests" passes now. "Run RPA skill smoke tests (Windows)" still fails with "No task.json files found" (the open
build_component.yaml:9thread).
New
- 🔴 The required activation gate fails under
uipath-genome, on the same three rows in both runs (inline onactivation_gate.py:81). I could not reproduce it in Claude Code itself, so I read it as the gate's harness, not routing; the gate still has to pass. - 🔴 The required README status check fails: the status row sits where
uipath-build-from-genomesorted (inline on #4327). - 🔸 #4327 alone fails Test Helpers (
test_genome_check.pyreads the planner'sskill-mapping-guide.md, which arrives in #4328) and the package build; both pass at this head. That is fine if the stack merges as one, not if #4327 merges first.
Still open: the 11 blockers of the previous re-review, threads unchanged.
danmunteanu1
left a comment
There was a problem hiding this comment.
🟡 Commented · 3 blockers · 4 non-blockers
Re-review at bcf20b12a, covering e9efd78c7, fc0dd41b5 and bcf20b12a on top of f85a7650b.
Verified
- Links, flavors, descriptions and skill status with
--check-readmepass.skills:testfails the same 18 tests as mainec2a62b26. The planner's 238 Python tests pass on Windows, the CRLF ones included. - The activation gate passes under
uipath-genome, measured under that name (thread resolved, with the cause confirmed). - Every file and section the copies table names exists on both sides. Main's rules 7, 14 and 17 are back as rules 12 to 14.
- Strict fails a chain written with
->or as a comma list. Two forms still pass (thread kept open). - "Run RPA skill smoke tests (Windows)" passes with the new
smoke_buildtask; that thread is resolved.
Fixed: 27 threads, resolved.
Still open
- 🔴 A stale or hand-edited genome still builds unchecked (builder
execution-guide.md, posted 11-10). That thread has no reply yet. - 🔴 The retired-name rule now builds
uipath-coded-agentsthroughuipath-rpa(the retired-name thread on #4327). - 🔴 Every extraction now asks the refactor questions, and an unattended run records the defaults as the user's decisions (the
refactor-guide.md:20thread). - 🔸 The case SDD missing on disk, and the examples no read plan opens: the rebuttals stand.
New
- 🔸 § Source values restates the migration preflight's disagreement rule (inline).
- The floor for this skill's gate leg is now 50%. At the gate's context size every plugin skill is listed by name only, so the leg guards little until the gate scores descriptions. That is a change to the gate, not to this PR.
| 2. A check whose failure the source logs as a pass or ignores is a decision or an optional wait, not a hard check; a check that passes when its target is absent asserts absence. | ||
| 3. A value bound to a built-in (the run date, the user's folder) is computed at run time, never taken from a same-named project variable's stored default. | ||
|
|
||
| Where the genome and the source disagree, the build follows the source, with the exceptions that step lists. Each disagreement is a genome deviation in the group report — step, source object and line, genome says, source says — and the genome is left to a `uipath-planner` edit. |
There was a problem hiding this comment.
🔸 Non-blocker · Restates the migration preflight's disagreement rule
source-migration-guide.md § Migration preflight step 4 already says it: "Where the genome and the source disagree, the kind of disagreement decides what the build follows; each is recorded as a genome deviation in the group report, and the genome is left to a uipath-planner edit." This copy says "follows the source" with exceptions, so the two can drift apart.
Fix:
| Where the genome and the source disagree, the build follows the source, with the exceptions that step lists. Each disagreement is a genome deviation in the group report — step, source object and line, genome says, source says — and the genome is left to a `uipath-planner` edit. | |
| Where the genome and the source disagree: [source-migration-guide.md § Migration preflight](source-migration-guide.md) step 4. |
|
Down to what we'd want before merging the stack:
The other open threads (retired |
6e0c850 to
71a9d7e
Compare
The planner and genome checker pytest jobs (Node for check-genome.mjs), the smoke and activation-gate scripts, the Studio Web flavor contract test and the genome docs follow the new skill layout. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…store the live-healing guide The checker's own tests move under tests/tasks/uipath-genome, so the builder's owners review them; the two cross-skill contract tests keep their place with owner lines naming both teams, and the fingerprint test runs in the job that sets up node. The required job keeps its name, with a comment saying why. .claude/rules/genome.md replaces the deleted uipath-genome/CLAUDE.md for both skills: subject homes, the rules that still apply, and the files a format, file-kind, building-skill or fingerprint change must touch together. docs/guides/live-healing-pass.md comes back from main as the guide for supervising a live build, linked from the builder's README row. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 65 dated from runs under its first name. Three runs on the stack top scored 13/13, 13/13 and 12/13; 90 is the nearest 5% below the lowest. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…enome as written The planner writes every genome, so it now checks each one: author, extract and edit end with the strict check and fix every FAIL (at most three rounds), and a new check verb reports findings. Whenever an SDD changes, the planner updates its genome and checks it in the same run, and strict fails a Designed in: line whose fingerprint no longer matches, so a stale genome never reaches a build. uipath-genome ships no script: it reads the facts its read plans branch on straight from the genome, with no context call, FAIL sorting or fingerprint comparison, and routes check or validate requests to the planner. Tests, the check_genome task, activation rows, CI paths, CODEOWNERS and the genome rule file follow the script. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…cided: project chain Refactor and migration extractions now ask Retire, Fold and Target per component, with the planner's own Decision Tree and delivery-model rule, and write the answers as the genome's target design; the builder builds the targets the genome names and keeps only run modes, run record, connections and defect rulings, dropping its copy of the Decision Tree. A B flow's Decided: line can name its project chain (B, <unit> — <source>: <Project> → <Project> …); the planner writes it from the SDD, the strict check validates it, and the build creates exactly those projects in that order. author no longer asks whether the UI application is reachable; the build asks. Execution autonomy is defined once per skill, the designed-genome edit list has one home, "the build" and "the lead agent" replace execution, executor and orchestrator, the builder trims its Shape Rules and repeats, and the fingerprint code lives once in sdd-fingerprint.mjs. New tasks cover an SDD revise updating its genome and a Decided chain reaching the build. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…gain uipath-genome is the current builder name, so the checker no longer treats it as a legacy builder (uipath-genome-builder stays one), and the builder and planner stop saying an older genome names it. An older preamble is the one asking for Execute mode. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…changes into the split Main #4613 names which side a bare "platform" means; the sentences now live in the planner's extraction and format guides and the builder's execution, refactor and migration guides. The BPMN template gets #4581's builder-SDK HITL and project-structure blocks with their Studio Web overrides, and the skill table's BPMN row matches main in multi-skill-patterns-guide.md, where the stack keeps the table. The case template gets #4529's validation-row block and #4534's stage-link ToC, and the case data guide states #4463's "any" escalation, so the twin parity test holds. author now runs only when no genome exists yet; an existing one is updated through edit. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rgets and CI Builder: a retired skill name builds through uipath-rpa instead of stopping the build; the genome's Rebuild follows cell decides a defect's behaviour; a decided-values table answers the SDD's and the request's decisions under their own labels; building into a project finds renamed workflows through the earlier hand-off; the source's values never answer a choice the source never faced; every build applies Retired and Folded rows; a refactor's capability probe runs before the plan fixes the shape. Planner: every extraction of a UiPath source asks Retire, Fold and Target, so every genome carries a target design; one statement of what leaves edit for design on a designed genome; a Decided: chain of N projects uses N - 1 queues and may come from the refactor answers; a solution root links each child SDD, and the fingerprint also finds bare filenames; Word PDDs fall back to PDF or pasted text. Checks and CI: strict fails a project chain not written with arrows; the checker drops the last legacy builder name; a Windows smoke task builds a small genome; the README row sorts where main has it; genome.md lists the copies both skills keep and restores main's rules 7, 14 and 17, with owners; the activation gate can keep its rows' transcripts and uploads them when a leg fails. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…'s verb does The builder's Source values keeps only what changes the build; each Running in Parts keeps the core both verbs need and drops what only the other verb uses. The rule file's copies table says what each side keeps. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two runs on its 11 rows scored 8/11 and 7/11: at the gate's context size the listing carries names only, and on implement, execute and run-the-migration prompts the agent reads the genome before invoking the skill. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
71a9d7e to
0da9096
Compare
Issue
PILOT-9071
Summary
uipath-genometouipath-build-from-genomeuipath-genomedocsTesting (stack top)
maintoo: maestro-case CLI snapshot test; Studio Web ceilings for agents / solution / test🤖 Generated with Claude Code