pi-devin-plus
v0.1.3
Published
Run your Devin / Cognition account models (Opus 5.5, GPT-5.6 Sol, Fable 5, SWE-2...) inside the pi coding agent - CLI login, live model catalog, full tool calling. Enhanced fork of kashyab12/pi-devin.
Maintainers
Readme
pi-devin-plus
Run your Devin / Cognition account models — Claude Opus 5.5, GPT-5.6 Sol, Fable 5, SWE-2 and 50+ more — inside the pi coding agent, with pi's own tools, sessions and UI. Unofficial, not affiliated with Cognition.
Login once with the Devin CLI; after that it behaves like any other pi provider.
Quick start
pi install npm:pi-devin-plusThen inside pi:
/login devin # instant if the Devin CLI or Desktop is already signed in
/model devin/claude-opus-5 # pick any model your account can runPick a thinking level with Shift+Tab or /thinking — medium / high / max map to the matching model variant automatically.
Requirements
- Pi Coding Agent 0.86+
- A signed-in Devin CLI (
devin auth status) — a signed-in Devin Desktop also works - Node 22.19+
What it does
- Login —
/login devinreuses the Devin CLI credential store (~/.local/share/devin/credentials.toml). If that is empty but Devin Desktop is signed in, the Desktop token is imported instead of forcing a browser round-trip. - Model catalog — the live catalog comes from
devin models list --format json, so the model list always matches your account. Reasoning models register under one family id (e.g.devin/claude-opus-5); the thinking level picks the variant, and levels the family doesn't ship are hidden from the picker. - Chat — completions stream from Devin's cloud endpoint straight into pi, with full tool calling. Your previous turns — including tool calls and the model's own reasoning — are replayed so multi-step agent work stays coherent.
Commands:
/devin-status— CLI path, version, auth state, credential source/devin-refresh— reload the model catalog without restarting
How this differs from upstream (kashyab12/pi-devin)
This package is a superset of upstream 0.2.0. Upstream's fixes (pi ≥ 0.86 transcript support, system-prompt field) are included, plus:
| | Effect you can see |
|---|---|
| Thinking + sealed-signature replay | Long agent loops keep the model's reasoning instead of restarting blind every turn |
| Devin Desktop sign-in reuse | No forced browser login if Desktop is already signed in; /devin-status shows the credential source |
| Family model ids + thinking levels | devin/claude-opus-5 instead of devin/claude-opus-5-high; thinking level switches variants |
| Socket-close crash fix | A dropped connection mid-stream no longer kills the process |
| CLI-aligned request fields | Requests match what the Devin CLI itself sends (trajectory + prompt ids) |
Which Devin package should you install?
Five pi packages talk to Devin. They are mutually exclusive (all register the devin provider):
| | pi-devin-plus (this) | pi-devin (upstream) | pi-devin-local | pi-devin-oauth | pi-devin-auth |
|---|---|---|---|---|---|
| pi ≥ 0.86 (tools + system prompt work) | ✓ | ✓ | ✗ broken — model ignores tools, agent runs one turn | ✓ | ✗ |
| Thinking replay across turns | ✓ | ✗ | ✓ | ✗ | ✗ |
| Devin Desktop sign-in reuse | ✓ | ✗ | ✓ | ✗ | ✗ |
| One id per model family + thinking levels | ✓ | ✗ (-high baked into ids) | ✓ | ✗ | ✗ |
| Requires Devin CLI | ✓ | ✓ | ✓ | ✗ (own OAuth) | ✗ (token paste) |
| Live model catalog | ✓ | ✓ | ✓ | ✓ | ✗ hardcoded |
| Maintenance | active | active | stale (last publish 2026-09-13; fix PRs unmerged) | active but self-described "reverse engineered, can break" | stale |
Short version:
- Use pi-devin-plus (this package) if you want upstream pi-devin plus the improvements above — or if you were on pi-devin-local, which is broken on pi ≥ 0.86.
- Use pi-devin if you want the minimal, upstream-only experience.
- Use pi-devin-oauth only if you refuse to install the Devin CLI and accept a reverse-engineered login that can break any time.
Don't co-install
Install only one of pi-devin-plus, pi-devin, pi-devin-local, pi-devin-auth or pi-devin-oauth — they all register the devin provider and overwrite each other.
Why a fork at all
The first Devin pi packages treated Devin as plain cloud chat, and models like Opus 5, Fable 5 and Sol answered:
This model is only in Devin Local.Those models are only reachable through the Devin CLI's model catalog. Upstream pi-devin solved that; this fork keeps that solution and adds the improvements listed above.
License
MIT. Unofficial — not affiliated with Cognition.
