Podo

Inspiration

During an incident, the useful facts are split across metrics, traces, logs, deployments, commits, CI, and source code. Engineers spend too much time rebuilding that chain manually, while a chat model given a pile of logs can produce a plausible answer without proving it.

What it does

For a runtime incident, Podo builds a living evidence graph from telemetry through deployment, commit, file, and function. For a GitHub Actions incident, it separately binds evidence to the exact workflow run, head, jobs, and steps. Both flows return a structured diagnosis whose material claims cite validated evidence. After explicit approval, Codex prepares a minimal patch and regression test in an isolated worktree. Podo independently requires red-before-green behavior and package validation, seals the exact Git tree, and requires a second approval before producing the matching pull request or preview. Failed validation can enter the separately gated issue fallback, never delivery.

How we built it

Podo is a Bun and TypeScript monorepo. Core owns incidents, evidence, investigations, approvals, remediation, delivery, and audit history. The Dashboard, CLI, and OpenTUI consume the same typed client. Normalized OpenTelemetry-compatible replay, Graphify import, and GitHub are replaceable adapters. Codex App Server is the supervised execution runtime, with generated protocol types pinned to Codex CLI 0.144.5.

The deterministic judge scenario reproduces a cache-growth defect across checkout, inventory, and notification services. It proves the full path from telemetry to causal graph, evidence-backed diagnosis, approved red-green fix, sealed diff, and pull-request preview without external writes.

How we used Codex and GPT-5.6

We used Codex agents with GPT-5.6 throughout the hackathon to build parallel module slices, write regressions before fixes, inspect runtime evidence, review security boundaries, and integrate the product into one vertical flow. Inside Podo, the same Codex App Server architecture powers investigation and remediation behind Core-owned approval and validation gates. For judging, Podo performs a live App Server handshake and uses deterministic remediation turns so the result is reproducible; an optional live mode exercises the production default gpt-5.6-sol model directly.

Challenges

The hard part was not generating a patch. It was preserving evidence integrity and authority across every boundary: untrusted telemetry and model output, private Codex protocol state, explicit local approvals, isolated Git worktrees, red-before-green tests, exact-tree delivery, retries, crashes, and stale client state.

Accomplishments

one command launches the complete incident-to-PR judge experience; the same Core contracts drive Dashboard, CLI, and OpenTUI; remediation cannot run without explicit approval; a regression must fail before the fix and pass afterward; failed validation cannot create a pull request; GitHub Actions failures support exact-run retry or tested remediation; live service traffic can reproduce the canonical cache-growth incident.

What we learned

Agentic reliability comes from keeping authority outside the model. Structured output helps, but the durable guarantees are trusted evidence references, typed contracts, fail-closed state transitions, exact artifact identity, and independent tests.

What's next

Durable Core state and restart reconciliation, authenticated actor identity, complete audit persistence, broader platform testing, and production deployment hardening.

Demo video

We produced the demo video with Codex and the HyperFrames plugin, using the same agent-assisted workflow to shape and render the project story.

Built With

• bun
• codex
• github-actions
• gpt-5.6
• hyperframes
• next.js
• opentelemetry
• opentui
• playwright
• react
• typescript
• vitest