Skip to content

feat(uipath-agents): HTTP Request built-in tool for inline agents in a Flow - #4752

Open
ionut-mihalache-uipath wants to merge 1 commit into
mainfrom
docs/agents-http-request-tool
Open

ionut-mihalache-uipath wants to merge 1 commit into
mainfrom
docs/agents-http-request-tool

Conversation

@ionut-mihalache-uipath

Copy link
Copy Markdown

What

Adds the built-in HTTP Request tool (type: "internal", properties.toolType: "http-request") to the uipath-agents skill. The skill had no walkthrough for it.

Changes

  • New capabilities/built-in-tools/http-request.md: when to use it, the resource shape, the six arguments and how the runtime treats them, pinning arguments, output, limits and errors.
  • built-in-tools.md: an Argument Sources section for argumentProperties (the skill had none), the flow node type per tool in the Tool Registry, and the Flow SDK case in Critical Rule 5.
  • Registry and routing rows (SKILL.md, lowcode.md, agent-definition.md, inline-in-flow.md, guardrails.md).
  • New integration task tests/tasks/uipath-agents/lowcode/http_request_tool/.

Decisions a reviewer cannot get from the diff

  • The tool is for inline agents in a Flow only. Agent Builder has no UI for it on a standalone agent. uip agent validate accepts the resource on a standalone agent, so the skill states the rule itself: never author it there, tell the user, and offer the inline-in-flow route.
  • Builder-SDK Flow projects do not hand-author the resource. compile writes it and deletes a hand-written one, so the skill hands the tool to the uipath-maestro-flow skill. The resource shape and walkthrough here are for a JSON-authored .flow.
  • The flow node suffix is not the toolType (http-request is httprequest, deep-rag is summarize). Critical Rule 5 said builtin.<toolType>; the Tool Registry now lists the exact node per tool. The same rule cited Critical Rule 16 for the hand-off to the flow skill; the rule it meant is 15.
  • load-attachments: the planning docs said only four toolType values pass validation and recommended load-attachments. It passes validation but has no runtime implementation, so the docs now say not to use it.
  • The task is a guard: it fails when a standalone agent ends up carrying the tool, and passes when the agent declines or builds an inline agent instead.

Verification

Ran:

  • The documented resource.json, with the documented pinned arguments, through uip agent refresh and uip agent validate --inline-in-flow on an inline agent inside a real Flow project (uip 1.205.0).
  • The task's checker three ways: an inline agent carrying the tool passes, a standalone agent carrying it fails, an empty workspace passes.
  • Argument handling, defaults, the public-host check and redirect limit were read from the runtime tool's source, and the input/output schema was compared with the Flow canvas node's schema.
  • npm run skills:validate, skills:check-links, scripts/check-sensitive-content.py, scripts/check-skill-status.py: pass.
  • npm run skills:test: 471 pass, 1 fail. The failure is the built studioweb flavor mentions no new command the browser bundle cannot run (uipath-coded-apps), which fails the same way on an untouched main.

Not run:

  • coder-eval on the new task, so there is no passing-run claim yet.
  • skill-lint (not available locally).
  • No live HTTP call through a deployed agent.

🤖 Generated with Claude Code

https://claude.ai/code/session_0136ibA39NTYe4XHq5Z2Ch32

…a Flow

The agents skill had no walkthrough for the built-in HTTP Request tool
(`type: "internal"`, `toolType: "http-request"`).

- http-request.md: the tool is for inline agents in a Flow only. Agent Builder
  has no UI for it, so the skill never authors it on a standalone agent, even
  though `uip agent validate` accepts the resource there; it tells the user and
  offers the inline-in-flow route instead.
- Builder-SDK Flow projects declare the tool through the uipath-maestro-flow
  skill (`compile` owns the resource); the resource shape and walkthrough here
  cover a JSON-authored `.flow`.
- built-in-tools.md: an Argument Sources section for `argumentProperties`, the
  flow node type per tool in the Tool Registry (the node suffix is not the
  `toolType`), and the Flow SDK case in Critical Rule 5, which now cites the
  right rule.
- The planning docs said only four `toolType` values pass validation and
  recommended `load-attachments`, which has no runtime implementation.
- New integration task `http_request_tool`: fails when a standalone agent
  carries the tool.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0136ibA39NTYe4XHq5Z2Ch32
@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

Claude finished @ionut-mihalache-uipath's task in 1m 27s —— View job


Summary

Adds an HTTP Request built-in tool walkthrough (http-request.md) to uipath-agents, scoped to inline agents in a Flow only. Also adds an argumentProperties (Argument Sources) section, a per-tool flow-node column in the Tool Registry, a fix to the Critical Rule 5 citation (16 → 15), a load-attachments correction, and a negative-guard test task.

Change-by-Change Review

1. references/lowcode/capabilities/built-in-tools/http-request.md (new)

OK. Well structured. It says where the tool applies (inline only), when to use an IS tool instead (credentials), which files to author for Builder-SDK vs JSON-authored flows, which fields may change, runtime behaviour for each argument, limits and errors, and a numbered walkthrough with stop conditions ("If uip maestro flow registry get does not return that node, stop"). CLI commands use --output json, placeholders follow <UPPER_SNAKE>, and every link resolves.

  • Low: §Gotchas line 169 repeats Argument Sources rule 1 from built-in-tools.md, and line 170 repeats rule 3. Both already link to that section, so these two bullets could become a one-line pointer.

2. built-in-tools.md

OK. Critical Rule 5 now cites the right rule (CR15, "Never invoke other skills automatically"), points to the registry for exact node types, and adds the Builder-SDK branch. The Argument Sources table, with its numbered rules (isSensitive required, never pin secrets), fills a real gap. Adding create-file.md to References also fixes an orphan link.

3. deeprag/planning.md, batch-transform/planning.md, both impl-json.md field tables

Low — leftover stale load-attachments text. The PR now says "Never use" load-attachments, but two lines in the same files still list it as one of "all four toolType values":

  • deeprag/impl-json.md:112
  • batch-transform/impl-json.md:117

They describe the shared test checker, but an agent reading them will see load-attachments presented as a peer tool. Reword both, e.g. "covering analyze-attachments, deep-rag, batch-transform (plus load-attachments, which the skill forbids)", or drop the list. Fix this →

4. inline-in-flow.md lines 299, 378

Low — the node-suffix list now exists in three places. The Tool Registry is the authority, yet analyzefiles, createfile, summarize, batchtransform, httprequest is also restated in the canvas-node table (299) and in the node-type tree (378). When a sixth built-in tool is added, those copies will drift. Line 299 already links to the registry, so keep only "<tool-name> is not the toolType — see § Tool Registry" there, and remove the list from 378.

5. SKILL.md, lowcode.md, agent-definition.md, guardrails.md

OK. The routing rows are correct.

  • Low: the new SKILL.md:61 row puts a markdown link in the 4th column, where the neighbouring rows use backticked paths (e.g. lowcode/capabilities/built-in-tools/analyze-attachments.md).

6. tests/tasks/uipath-agents/lowcode/http_request_tool/

OK, with a coverage gap (see What's Missing).

  • Tags: uipath-agents first, integration, mode:build, lifecycle:generate, and feature:http from the closed vocabulary. The run_command grading is behaviour-based, and the advisory checks use pass_threshold: 0.0. The checker is defensive (skip dirs, JSON errors handled).
  • is_inline_agent means "the parent folder holds *.flow". That matches the documented <FlowProjectDir>/<projectId>/resources/… layout, and a standalone agent's solution folder never holds a .flow, so I don't see a false-pass path.

What's Missing

  • No passing coder-eval run claim for the new task. /lint-task flags this as High, and the PR body already notes it. Run it before merge.
  • No positive test of the new walkthrough (Medium). The only task is a negative guard, and the empty-workspace case passes it. Nothing checks that an agent following http-request.md produces a valid inline-agent resource: correct toolType, UUID folder equal to id, isSensitive on pinned entries, and uip agent validate --inline-in-flow passing. A smoke task that hands the agent an existing Flow project and asks for an inline agent with a pinned-GET HTTP tool would cover the documented path.
  • CODEOWNERS, skill-status.json and skills.sh.json: no change needed, since no skill folder was added or renamed.

Area Ratings

Area Status Notes
Frontmatter OK Unchanged
E2E Tests Issue Negative guard only; no passing-run claim; no positive-path task
Skill Body OK One routing row added
References & Assets OK Low: stale load-attachments lines; node-suffix list in three places
Repo Hygiene OK No secrets or personal paths; api.example.com placeholder host; scoped to uipath-agents + tests

Issues for Manual Review

  • Runtime facts. Defaults (GET, 30 s timeout, https:// auto-prefix), the 5-redirect cap, the per-hop public-host check, and the failure where an entry missing isSensitive passes validation but stops the agent loading were read from runtime source, per the PR body. A domain owner should confirm them; no live call was made.
  • Flow node availability. uipath.agent.resource.tool.builtin.httprequest is described as available "on tenants that offer it". Confirm how widely it's rolled out, so agents don't hit the stop path often.
  • The other flow node names. Confirm analyzefiles, createfile, summarize and batchtransform against uip maestro flow registry. They are newly documented here and not otherwise verified in the PR.
  • Sibling skill alignment. The Builder-SDK hand-off depends on uipath-maestro-flow (agent-resources.md httpRequest kind) agreeing with this doc that compile writes the resource and deletes a hand-written one.
  • Link checker. I couldn't run npm run skills:check-links in this sandbox. I checked every new relative link and anchor by hand, and all resolve.

Conclusion

Approve, with minor follow-ups. The content is accurate to the stated sources and fits the skill's conventions. Before merge: add the coder-eval passing-run claim and consider a positive-path task. The rest is Low polish (the stale load-attachments lines and the duplicated node-suffix lists).

@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

Claude finished @ionut-mihalache-uipath's task in 1m 10s —— View job


Coder-eval task lint (advisory)

1 task YAML changed; verdict counts: 0 Critical, 0 High, 1 Medium, 0 Low, 0 OK (plus one High finding under Evidence of passing run).

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

Evidence of passing run

  • ❌ High: the PR body says the opposite of a passing run: "Not run: coder-eval on the new task, so there is no passing-run claim yet." Please run the task and then edit the PR description to add a line like: Ran skill-agent-http-request-tool locally and it passed.

Per-task lint

tests/tasks/uipath-agents/lowcode/http_request_tool/http_request_tool.yaml: verdict: Medium

Issues:

  • [Medium] Could pass for the wrong reason: the only gating criterion (L32–38, weight 5.0) passes on an empty workspace, which the PR body confirms. An agent that does nothing, errors out on turn 1, or never loads the skill scores 100%. Both positive checks (L42–58) are pass_threshold: 0.0, so nothing separates "declined correctly" from "did nothing".
  • [Low] Could pass for the wrong reason: is_http_request_tool in check_http_request_tool.py (L35) requires $resourceType == "tool". If an agent hand-writes the forbidden resource without $resourceType, the guard misses it. uip agent validate then reports resources: 0 and the advisory validate check passes as well.
  • [Low] Meaningful coverage: is_inline_agent (L43) only treats an agent as inline when a compiled *.flow sits in its parent folder. An uncompiled Builder-SDK project that has only a .flow.ts file would be misread as standalone. The advisory validate check (L53) also hardcodes RatesSol/RatesAgent, so it fails if the agent picks a different folder layout.

Suggested fixes:

  • Add a gating conditional check: "if an Agent project exists under RatesSol/, uip agent validate must report Valid" (exit 0 when no project exists). A lazy no-op still passes, but an agent that builds a broken standalone project fails.
  • Optionally add a gating command_executed on uip\s+agent\s+(init|validate) or a .flow existence check. Either one proves the agent took one of the two acceptable routes (built the standalone agent without the tool, or built an inline agent) and didn't just stop.
  • In is_http_request_tool, match on type == "internal" and properties.toolType == "http-request" only, and drop the $resourceType requirement.
  • In is_inline_agent, also accept *.flow.ts in the parent folder: any(agent_dir.parent.glob("*.flow")) or any(agent_dir.parent.glob("*.flow.ts")).

Notes: Prompt over-specification is OK. "Use the platform's built-in HTTP request tool" is the deliberate bait, and "must stay a standalone agent project" is a ground-truth anchor. Near-duplicate is OK: builtin_tool/builtin_tool.yaml tests positive authoring of a built-in tool, and this task is a negative guard. The tags are in the closed vocabulary. CLI verb reachability: there are no command_executed patterns to check (uip agent validate only appears inside a run_command). scripts/check-cli-verbs.py could not be executed in this runner (permission denied).

Within-PR duplicates

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

Conclusion

  • ⚠ 1 task has issues, max severity Medium (High on the missing passing-run claim). Advisory only, not blocking merge.

This branch has not been deployed

No deployments
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