ALFRED PI

INSTALLpi install git:github.com/686f6c61/Alfred-Pi

then /reload · no pi? pi.dev

/// extension for the pi agent

Alfred-Pi is an extension for pi, the terminal agent: it detects the craft of each prompt and loads only that context, watches what you have spent today and relays a failing model. If you already run pi, the command above and a /reload are the whole install. If not, start at pi.dev (0.84 or later).

v0.3.0 · MIT · 0 dependencies ·GitHub

[ SECTION / 002 ]THE TURN

pi already runs, already has tools, already keeps the session. Alfred sits on top and handles the rest: keys, models, what you have spent today, and which craft this prompt belongs to.

You type «audit this repo». You do not pick a pack. Alfred reads the sentence, sometimes a file at the root, and loads security. The model sees that room's context.md. The other ten stay off. If the turn later bends toward tests, skills you enabled are still in the menu.

Before it writes models.json or auth.json it shows the diff and waits. If the budget looks ugly it warns; it does not kill the reply. If the current model has failed twice, the next link in the chain comes in on the following turn, not mid-stream.

[ SECTION / 003 ]FROM THE REPO, TODAY

Packs11

Skills52

Prompts26

Presets22

Eleven crafts covered, from security to data analysis, each with its own skills and prompts. The presets include Ollama and local servers, on equal footing with the cloud ones. The figures above are counted from the repository when the page is built.

[ SECTION / 004 ]ONE PER TURN

The radar looks for whole words in what you just typed. «express» does not fire on «expression». If nobody matches, it looks at the root: aDockerfile, an astro.config.mjs, asonar-project.properties. Tie: the name that comes first alphabetically wins. What gets injected is about fifteen lines, from that room.

001securitySecurityFires when the turn smells like risk: audit this repo, a sonar-project.properties at the root, CVE, secrets or OWASP. «audit the security of this repo» injects the defensive context and the model pairs an OWASP pass with a real reading of the code. Each finding arrives with severity, file:line, an attack story and a minimum fix.002ai-agentsAI AgentsFires when you ask to parallelise agent work: review these PRs, in parallel, worktrees, fan-out. /fanout splits the task, estimates cost from the harness price table, assigns a cheap model to each unit and a collector merges results with a merge protocol.003docsDocsFires on «document this», «adr», «changelog», «this README». /docs-audit . classifies each page by Diátaxis mode, flags mixed modes and snippets that no longer run.004escritura-esEscritura (ES)Fires on «fix this text», «review the wording», accents, English-Spanish translation. /revision-es README.md edits in place, shows the diff and explains the applied rule only when the change is not obvious, keeping the author's voice.005clean-codeClean CodeFires on «refactor», «review this PR», technical debt, SOLID, names that lie. /refactor src/foo.ts plans behaviour-preserving steps, one by one; if the code that matters has no tests, it writes characterisation tests first.006web-fullstackWeb / FullstackFires on api, endpoint, scaffold, Astro, Playwright, an astro.config.mjs or a Program.cs at the root. /scaffold-crud invoices generates a migration, a validated API, views with states and tests, copying the repo's patterns instead of imposing its own.007complianceComplianceFires on GDPR, cookies, licences, copyleft, accessibility. /privacy-check src/forms/signup maps the personal-data flow and flags consent, third parties and retention, with the exact reference for each duty.008landing-designLanding DesignFires on landing, hero, conversion, CTA, a screenshot or a mockup. /landing-review inspects the rendered page on desktop and mobile. Source alone is not evidence.009devops-infraDevOps / InfraFires on Dockerfile, compose, a 502 in production, «deploy», certificates, kubernetes. /diagnose-502 finds the failing layer with real checks. Restart is never the first diagnostic step.010data-analisisData / AnalysisFires on csv, sql, dashboard, «this query is slow», notebooks, dbt. /explore data.csv profiles the file, cleans with a reason for each step and delivers findings with numbers that trace back to a query that runs again.011qa-testingQA / TestingFires on a test plan, coverage, «this test fails sometimes», contract testing. /flaky classifies the smell, delivers a deterministic reproduction and fixes the cause. A flake is a suite bug; retries never silence it.

[ SECTION / 005 ]WHERE NEXT

Starting from zero? Install, then first steps. Already running? This is how a turn goes. A specific doubt (working locally, Chrome, the 0.2 move)? It is in the FAQ.

The full technical docs live in docs/*.md, mounted here as-is.