[Skill] Scrum pattern for pipe creation

[jpwf]

Skill

Folder / Pasta: submissions/jpwf/scrum_pattern/

What it does / O que faz: Create, read, update, delete, and troubleshoot Pipefy AI agents (conversational agents with 1–5 behaviors), covering all 7 MCP tools including pre-flight validation, plus pipe-scoped knowledge bases (plain text / document / data lookup CRUD, access probe) attached via dataSourceIds. Ships a ready-to-use, self-contained recipe for a SCRUM project-request pipe with per-transition status emails.

Contact / Contato

Email: wolfbytegames@gmail.com | jpwf01@gmail.com

Name for credit / Nome para crédito: João Pedro Weydt de Faria

The problem it solves / O problema que resolve

Building AI agents by hand means guessing event IDs, phase IDs, action types, and field IDs — and the create/update tools are all-or-nothing (RECORD_NOT_SAVED) and full-replace, so a single wrong behavior silently drops the rest or corrupts the agent. Teams also had no repeatable pattern for a request pipe that emails the requester on each status change.

What it built when you ran it / O que ela construiu quando você rodou

A SCRUM chamados pipe (Triagem → Em Análise → Em Desenvolvimento → Aguardando Informação → Validação do Cliente → Entregue, plus a lateral Arquivados/Cancelados) with a notifier agent of 4 card_moved behaviors, each firing send_email_template when a card enters a notifying phase. Because card_moved triggers on entry, it covers both forward moves and returns to Em Desenvolvimento with 4 behaviors (under the max-5 cap). Full discover → validate → create → verify flow: get_pipe for uuid/phases, get_automation_events, validate_ai_agent_behaviors, create_ai_agent, then a get_ai_agent round-trip.

What you had to fix / O que você teve que corrigir

update_ai_agent is full-replace, not patch — existing behaviors must be fetched and merged or they vanish. card_moved/field_updated behaviors need event_params (to_phase_id / triggerFieldIds) or they fire on every occurrence. Phase transition arrows (returns and the lateral Arquivados/Cancelados) are UI-only — the agent only reacts to the move and sends the email; a move_card to an unreachable phase fails with valid_destinations. Partial-failure recovery updates the same UUID (never creates a second agent).

Checklist

• I provided a contact email in this description.
• I ran this skill in a real Pipefy organization.
• SKILL.md and EVIDENCE.md are both in the folder.
• The frontmatter name matches the folder name.
• No tokens, keys, real IDs, customer names, or personal data.
• I wrote this skill, and I agree it is published under Apache 2.0 with credit to me.

[adriannoes]

Thank you for taking part in MCP Builders, João.

What we verified

We ran the recipe's discover, validate, create and verify sequence on a clean pipe.

• get_automation_events lists card_moved with to_phase_id as its accepted parameter, and send_email_template is in the action catalog.
• Two behaviors shaped as in the recipe passed validate_ai_agent_behaviors, create_ai_agent accepted them, and get_ai_agent returned them with the phase filter and action type intact.
• One finding worth adding to the skill: neither validation nor create checks that the emailTemplateId exists, so a wrong id only fails when a card moves.

The recipe itself (one card_moved behavior per destination phase, one send_email_template action each, transition rules and email templates flagged as UI-only) is correct and is the part we want to keep.

Changes requested

1. Trim the file to your recipe. About two thirds of it reproduces the toolkit's pipefy-ai-agents skill, and the frontmatter still carries that skill's name. Building on an official skill is welcome, but the entry should link it under See also and say in the PR that it builds on it, rather than include it.
2. Rename the folder to kebab-case (for example scrum-status-notifier) and set the frontmatter name to the same value.
3. Add EVIDENCE.md with proof of a real run: the agent in the pipe's AI settings or a received status email, plus what broke the first time.
4. Use the template sections (When to use, Prerequisites, Tools needed, Steps, Success criteria, Failure modes). In Steps, build the pipe and phases with create_pipe and create_phase instead of the UI, and add get_email_templates to resolve each template id.
5. Fix the phase count: the flow lists seven phases, not nine.

One smaller point: a native automation with card_moved and send_email_template sends the same notification without AI credits. Either say why the AI agent is the better fit here or switch to the automation.

Items 2 and 5 are mechanical. Item 1 is the one that matters most.

Next steps

The submission is accepted into the program. The changes above are about the format; push them to this branch and we merge. After that, we will bring the skill into the official Pipefy AI Toolkit repository, adapted to its format and CI, referencing this pull request and crediting you as a contributor there. We will reach out by email about the details.

[adriannoes]

CI update: the validate check is red with two errors, 'scrum_pattern' must be lowercase kebab-case and EVIDENCE.md is missing. The script returns early on those, so the frontmatter name mismatch and the missing template sections only show up after they are fixed. All of them are in the review above. Local run after renaming the folder:

python3 scripts/validate_submission.py submissions/jpwf/<new-folder-name>