convez.ai
Agente
parker-plextrac

pt-doots

parker-plextrac/pt-doots

Ver no GitHub
InicialPluginScore · de 0 a 100

Personal Claude Code plugin — PlexTrac orchestrated workflow with agent teams

Instalação

$npx convez-ai parker-plextrac/pt-doots

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

pt-doots

A Claude Code plugin for PlexTrac engineers. Turns a Jira ticket into a finished PR via an orchestrated team of sub-agents — research → plan → implement → test → review → commit.

Opinionated for our repos (product-core-backend, product-core-frontend, product-services-export, product-services-mcp), our standards (per-repo CLAUDE.md), and our tooling (Jira/Confluence REST, GitHub via gh CLI).

Install (from marketplace)

/plugin marketplace add parker-plextrac/pt-doots
/plugin install pt-doots@pt-doots

Then enable agent teams (one-time):

echo 'CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1' >> ~/.claude/settings.local.json

Restart Claude Code after install.

Tuning agent behavior

Per-user tweaks go in overlay files in your memory dir, not in the plugin tree. See OVERLAYS.md. Bundle vs overlay separation keeps your customizations upgrade-safe.

Prerequisites

  • Claude Code with agent teams enabled (above)
  • Workspace laid out with PlexTrac repos checked out side-by-side under one parent directory
  • PlexTrac agent-skills plugin installed and /setup completed. This is the hard requirement — pt-doots delegates verify, PR, ticket, log, and stack work to commands from agent-skills (/verify, /create-pr, /ticket, /logs, /stack-up, etc.) and its /setup provisions every credential and CLI listed below.

Agent-skills /setup installs/verifies:

  • gh CLI authenticated against PlexTrac's GitHub org — primary tool for PR work; the GitHub MCP is a fallback when gh is unavailable
  • Atlassian API token at ~/.jira-attlasian-cred — pt-doots talks to Jira and Confluence via REST directly; the Atlassian MCP is not used because its OAuth flow fails mid-workflow
  • ~/bin/jira-attachment symlink to the bundled download script (used for tickets with attached repro files)

Install agent-skills first, run /setup, then come back here.

Commands

CommandWhat it does
/pt-dootsOrchestrator — detects intent from your message
/prsPR dashboard for all open PlexTrac PRs
/prs <PR-url>Structured review of a single PR using the team
/prs self [TICKET-KEY]Run the full review swarm against your own inflight code; saves findings to notes/{TICKET}/. Auto-detects the ticket from your current branches if you omit the key. Handles multi-repo tickets (FE + BE) as parallel arms.
/bootstrap-teamOne-time setup — spawns the team-manager to create every agent locally
/team-auditRoster health check + agent performance review
/new-integrationScaffold a new EM integration across BE + FE repos
/voice-profileCustomize your personal voice overlay for the voice-stylist agent

Talking to /pt-doots

The orchestrator picks a mode from how you phrase it:

PhrasingMode
"tackle IO-2097", "do IO-2097", "implement IO-2097", "doots IO-2097"Tackle the ticket end-to-end
"check on IO-2097", "status IO-2097", "where are we on IO-2097"Status check — reads notes/IO-2097/progress.md
"save our work", "save progress", "let's save"Save current state to progress.md

The Agent Team

/pt-doots orchestrates a team of focused sub-agents. Each one has a narrow job, a locked tool surface, and a prompt tuned to PlexTrac standards.

AgentRole
scrum-masterPicks the right workflow (standard / lightweight / docs-only / custom) per ticket
researcherTraces the codebase + Confluence, writes notes/{TICKET}/research.md
developerImplements plan steps; writes production code following the repo's CLAUDE.md
implementerPlan-fidelity variant of developer — audits its own diff against the plan
test-writerWrites co-located tests in the repo's framework (mocha, jest, pytest)
code-reviewerReviews changed files against per-repo CLAUDE.md standards
acceptance-qaVerifies acceptance criteria with code evidence
edge-case-qaHunts boundary conditions, null/empty/race scenarios
code-smells-reviewerCatches Fowler-catalog smells (long methods, feature envy, primitive obsession)
test-reviewerCatches hollow assertions, over-mocking, and AI-test smells
re-reviewerVerifies prior review findings on subsequent commits (used by /prs)
documentarianUpdates READMEs and Confluence after merge (when scrum-master sets Documentation: yes, or workflow is docs-only)
voice-stylistFinal voice pass on every human-facing draft (PR comments, Slack, Jira). Used by /prs between rough draft and approval gate
team-managerCreates and tunes the team itself; used by /bootstrap-team and /team-audit

Developer modes

Pt-doots ships two implementation agents:

  • implementer (default, strict). Locks the file surface from the plan, audits its own diff against forbidden patterns, and reports every plan deviation. Slower, higher signal.
  • developer (opt-in, loose). Faster, less ceremony, more latitude. Default for lightweight workflows.

The orchestrator picks based on:

  1. Env var PT_DOOTS_DEV_MODE=loose → developer
  2. Scrum-master classifies workflow as lightweight → developer
  3. Otherwise → implementer

Other engineers who want the looser flow as their default can export PT_DOOTS_DEV_MODE=loose in their shell rc.

Review Log Discipline

Every change to an agent definition (model tier, prompt, tools, maxTurns, role) MUST append a dated entry to agents/reviews/{agent}.md with:

  • Date of the change
  • What changed (current → proposed value)
  • Why — linked ticket, observation, or learned pattern

Silent drift (changes without log entries) makes audits blind to regressions. The team-manager enforces this discipline at every /team-audit run; unexplained drift is flagged as a finding.

Workflow

Step 0    Load context + git state
Step 0.5  Scrum-master picks workflow type
Step 1    Researcher → notes/{TICKET}/research.md
Step 2    Plan with the user → notes/{TICKET}/plan.md
Step 3    Branch ({TICKET-KEY}-{short-description})
Step 4a   Developer implements → /verify
Step 4b   Test-writer writes tests → /verify
Step 4c   Quality gate (5 reviewers in parallel)
Step 4d   Developer fixes findings → /verify
Step 4e   Documentarian updates docs (when Documentation: yes or workflow is docs-only)
Step 5    Commit gate — user approves checklist
Step 6    Handoff → /create-pr

All ticket artifacts (research, plan, progress, scratch scripts) live in notes/{TICKET-KEY}/ — never committed to product repos.

Workflow types

The scrum-master picks one of four types, plus orthogonal flags (Documentation: yes/no, TDD: yes/no):

TypeWhenPipeline
standardMost tickets — features, multi-file changes, anything riskyFull pipeline; parallel quality gate (5 reviewers)
lightweightSingle-file fixes, dependency bumps, additive changesSkips acceptance-qa + edge-case-qa; smaller review surface
docs-onlyDocumentation-only tickets (READMEs, comments, reference docs)Researcher → documentarian → code-reviewer → commit
customTickets that don't fit a templateScrum-master proposes the variant with rationale

You can override the recommendation when prompted.

Customizing Your Voice

The voice-stylist agent rewrites human-facing drafts (PR comments, Slack pings, Jira replies) into your voice before they're shown for approval. It reads two layers of rules:

  1. Bundle at agents/voice-stylist/profile.md — universal good-prose rules shipped with the plugin (banned-phrases starter list, prefix scheme, em-dash rule, plain-verb mappings). Works out of the box for everyone.
  2. User overlay in your local user memory (~/.claude/projects/{workspace}/memory/voice_*.md) — personal additions that override or extend the bundle. Your audience tiers, signature emojis, phrases YOU don't use, prefix tweaks.

To set up or refresh your overlay, run /voice-profile. The command walks you through a short interview and writes the overlay files for you. You can also just ask Claude directly: "help me edit my voice profile."

Without an overlay, the agent runs on the bundle alone — still useful, just less personal.

Local development

git clone https://github.com/parker-plextrac/pt-doots ~/workspaces/plextrac/pt-doots
claude --plugin-dir ~/workspaces/plextrac/pt-doots

Edit any agent under agents/ or command under commands/ and reload Claude Code to pick up changes.

Ver o original no GitHub

Bom para

Relacionados