itachi./ 1824
← All work

Agents / Infrastructure / 2025–2026

Polli

An agent platform. Across every interface.

I shipped merged contributions across Polli’s agent interfaces, tool routing, repository retrieval, and streamed execution.

PythonMCPCloudflareGitHub Actions
Trace a request Engineering decisions ↗Sources & status ↗

Merged contributions

CONTRIBUTION & EVIDENCE
Run the workflow

Follow the caller boundary.

Follow a request through caller context, tool visibility, retrieval, and response streaming. Click a step, connection, or trace to inspect it.

Execution trace / Click to inspect

Start the walkthrough, or inspect any step before running it.

Simplified from the project architecture. Runs locally; no live models, credentials, or external services.

Polli / Inspector

Workflow overview

Receive

An OpenAI-compatible request enters the API and receives a caller context.

Why it matters

Polli connects caller context, selected models, repository evidence, and tools. Exposing a tool is distinct from validating every action it can perform.

Synthetic payloads explain the boundaries. They are illustrative examples, not project API schemas or live production traces.

The engineering

Inside
the decisions.

Connect conversational interfaces, repository intelligence, and scoped tools into one agent platform.

An active AI platform needs more than model endpoints. Maintainers need current repository context, repeatable deployment, useful issue triage, and tools that behave consistently across services.

01

The design decision

Index repository changes incrementally and expose explicit tools to the agent. Keep retrieval, routing, and execution as separate responsibilities.

02

The trade-off

Incremental indexing reduces repeated work, but it needs a deletion path as well as an update path. Container isolation also needs to keep caller credentials outside the execution environment.

03

Connecting Discord and HTTP conversations

Developed Polli across Discord, GitHub workflows, and OpenAI-compatible Chat Completions and Responses APIs, including streamed responses and session context.

04

Selecting models and exposing tools

Developed tool routing, model-complexity selection, repository and web search, visual tooling, and caller-scoped tool visibility. Extended the platform MCP server with multimodal capabilities and dynamic model discovery.

05

Updating the index when source changes

Moved code embeddings from a bot-managed local index to an event-driven Cloudflare Vectorize pipeline, processing changes incrementally and deleting stale vectors.

06

Isolating shell execution

Added OpenAI-compatible streaming and isolated shell execution in disposable Cloudflare containers, excluding caller credentials from the shell.

↳

Read the system flow

Polli · Public explanation of the implementation boundaries.

  1. RequestDiscord, GitHub, or an HTTP client enters through its own interface.
  2. Context & toolsSession context, repository retrieval, and caller-scoped tools inform execution.
  3. ExecutionThe agent routes tools; shell work uses an isolated container boundary.
  4. ResponseThe interface receives the result, including streamed HTTP output.
↳

Ownership

My contributions span agent interfaces, tool routing, retrieval, streaming, and execution. Polli is a shared Pollinations project; the linked changes identify my work.

↳

Incremental repository retrieval

A repository index must account for changed and deleted source files. Moved embeddings to Cloudflare Vectorize and made updates incremental. Changed paths are reindexed; deleted paths have their vectors removed. Merged implementation; the initial PR does not establish end-to-end production performance.

Indexing · PR #12602 ↗
↳

Isolated shell execution

Agent shell commands need a separate execution boundary. Added disposable Cloudflare containers and excluded caller credentials from the shell. Shell work runs behind an explicit container boundary. Merged PR with reported test and live-check results.

Isolation · PR #14787 ↗
↳

Streamed API responses

HTTP clients need to consume agent output as it arrives. Added OpenAI-compatible streaming to Polli’s API path. Clients can receive incremental output through a familiar API interface. Merged implementation and review history.

Streaming · PR #14438 ↗

Merged contributions

Inspect
the work.

These contributions were merged into Pollinations. The linked pull requests document implementation choices, review, and validation. The work spans the agent layer, retrieval, APIs, and deployment.

Next project: Ouroboros ↗

Source material

Polli source directory · shared projectVisual tools & routing · PR #14722Issue automation · PR #10127MCP server · PR #6232Incremental code index · PR #12602Container isolation · PR #14787Streaming API · PR #14438
Public implementation and merged contributions. Each link identifies the work it supports.

Opening the project index…

Browse all work ↗

Interactive exhibit