HackStack

Project Info

Premonition

Devpost

Inspiration

I often found that when debugging, a loop starts which immediately becomes a chain of context switches - especially when not coding in a native AI IDE: copy the error, find the right repo, open an AI tool, explain context, wait for suggested fix, and then work out if the patch is safe to try. The idea of Premonition is to explore a smaller interface at that moment: what if copying an error were enough to prepare a reviewable fix, whilst the developer kept the final decision?

What it does

Premonition is a native macOS menu-bar app and ambient layer for Codex. It watches the clipboard locally, but ordinary clipboard content stops before any model call. A deterministic heuristic first checks for a supported error shape, then resolves extracted paths only against Git repositories the developer explicitly allowlists. Only after those checks pass does Premonition send the copied error verbatim to Codex, pinned to GPT-5.6 Sol. Sol drafts one bounded unified diff inside an ephemeral, read-only, no-approval sandbox. Premonition parses and bounds the diff, rejects out-of-repository or malformed changes, and runs git apply --check without modifying the working tree. If the candidate is mechanically applicable, the menu-bar eye changes state and the developer can review a focused semantic diff and a short Sol rationale. The only outcomes are explicit: Apply, Copy Patch, or Dismiss. Apply rechecks that the worktree is literally clean and then uses git apply; the edit remains unstaged. Premonition never auto-applies, stages, commits, runs tests, invokes repository hooks, or executes post-apply commands. How I built it with Codex and GPT-5.6 GPT-5.6 Sol has two inspectable roles in Premonition. At runtime, every real patch and rationale call is made through the Codex CLI with the model pinned to gpt-5.6-sol. The runtime contract is deliberately narrow: read-only speculation, one candidate, canonical diff validation and a human decision before any repository mutation. During Build Week, Codex with GPT-5.6 Sol was also the build partner for the specification-led implementation: architecture, privacy and threat-model reviews, SwiftUI/AppKit implementation, tests, native UX refinement, fixture replay, signing and release evidence, and submission preparation. I retained the product-scope, privacy, acceptance and publication decisions. The durable record is visible in PREMONITION_SPEC.md, BUILDLOG.md, DEVLOG.md and docs/build-week/sol-ledger.md, with phase and session trailers in material commits. Outside this, I leveraged a purpose-build set of Skills in a Codex Toolkit plugin that aided the progress of the build, and kept it on track. These included $build-premonition which was called nearly 100 times over the 5 days I worked on this. Safety and privacy boundary Premonition's safety spine is local gate → allowlisted repository → admission cap → read-only Sol speculation → one bounded candidate → mechanical validation → explicit human decision. A ready patch expires after ten minutes, and tracked or untracked worktree changes block Apply. The limits are as important as the mechanism. The gate is heuristic. An admitted error is sent verbatim to Codex and may contain secrets; v0.1 has no secret-redaction pass and makes no claim about provider-side retention. git apply --check proves mechanical applicability, not correctness, security or completeness. Live model output can vary. A judge-runnable path The public v0.1.0 release is a signed, notarised and stapled Apple Silicon app for macOS 14+. Judges can download it without rebuilding. The repository also contains deterministic shallow and deep fixtures for demo purposes without needing to expose their own repos. Replay Fixture replaces only the model hop and is visibly labelled; the real gate, repository resolver, admission checks, diff parser, bounds validator, git apply --check, review UI and human action still run without network access. Prior work and what is new Premonition had specification v1.1 planning on 12 July 2026, before the Submission Period - I started working on this the moment that the challenge was announced, of course trying out the various Sol, Terra and Luna models. GPT-5.6 Sol-assisted specification and workflow refinement is recorded on 13 July through commit 2e83e84. Implementation began on 14 July with S0 commit 3422b0c. No application implementation is attributed to the pre-period plan; the dated boundary and direct commit links are documented in docs/build-week/prior-new-boundary.md. What I learned The interesting problem was not merely obtaining a plausible patch. It was designing the admission, validation, privacy and interaction boundaries around speculation so the experience could feel ambient without becoming autonomous. The deterministic fixture path also proved useful beyond demos: it makes the complete product loop inspectable even when network access or live model variability would otherwise get in the way.

What's next

Premonition v0.1 intentionally stops at one local, reviewable candidate. Future work may explore bi-directional MCP integration and a direct OpenAI API executor, but neither is present in this submission. Automatic mutation, generic error comprehension, secret redaction and IDE integration are not implied roadmap commitments.

Analysis

Compare with all teams

View

Metric

Figures cover GitHub contributors during the hackathon window. A co-authored commit counts in full for each author, so per-member totals add up to more than the whole-team figures.

Technology

Found in codeClaimed only
  • CSSIn code
  • HTMLIn code
  • JavaScriptIn code
  • PythonIn code
  • RubyIn code
  • SwiftIn code

6 of 6 appear in the indexed code.

AI coding agents

  • CodexConfig

Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.

Codebase size

Source size

627 KB

Source files

91

Counts recognized source files only; vendored directories, binaries and lockfiles are excluded, so this is smaller than the repository on disk.

0 stars