Switch to Pi by Canvinus · Pull Request #1 · appx-org/agent-server

Canvinus · 2026-05-23T19:57:46Z

Summary

Adds the Pi-backed Appx agent-server contract needed by the Appx Pi integration.

Adds provider auth APIs for stored API keys, Pi subscription auth flows, and non-secret provider status.
Adds model and thinking controls for sessions, including LiteLLM/custom provider support for openai-completions, openai-responses, and anthropic-messages compatible APIs.
Adds extension UI bridge endpoints and SSE delivery for Pi extension prompts.
Adds AgentRuntimeRegistry so one agent-server process can keep shared auth/model state while isolating session runtimes by Appx project directory.
Updates OpenAPI and README docs, including provider-neutral SSE/client guidance.

Why

Appx is moving from a single OpenCode backend toward Pi under the hood. The frontend and Appx proxy need a stable HTTP/SSE contract that hides provider transport differences, supports subscriptions/custom endpoints, and keeps project sessions isolated.

Validation

npm run build passed.
npm test passed, 28 tests.
Project-scoped smoke test in the Appx VM verified sessions created in one project directory did not appear in another.

Notes

Pairs with appx-org/appx#2, which installs/proxies this service and adds the Pi UI.

Canvinus · 2026-05-25T16:21:41Z

@codex

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 341785e421

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-05-25T16:28:18Z

 const host = optional("AGENT_SERVER_HOST", "127.0.0.1");
 const port = Number(optional("AGENT_SERVER_PORT", "4001"));
-const token = process.env.AGENT_SERVER_TOKEN?.trim();
+const token = (process.env.AGENT_SERVER_TOKEN ?? process.env.APPX_AGENT_SERVER_TOKEN)?.trim();


Use APPX token when AGENT_SERVER_TOKEN is empty

The token selection uses process.env.AGENT_SERVER_TOKEN ?? process.env.APPX_AGENT_SERVER_TOKEN, so an explicitly empty AGENT_SERVER_TOKEN (e.g. set to "" by environment templating) suppresses APPX_AGENT_SERVER_TOKEN and leaves token empty. In that case the auth middleware is skipped and all /v1/* routes become unauthenticated even though a valid APPX token is present, which is a security regression in deployments that rely on the fallback variable.

Useful? React with 👍 / 👎.

neuromaxer · 2026-05-25T19:36:16Z

+	extensionPaths: optionalList("PI_EXTENSION_PATHS"),
+	skillPaths: optionalList("PI_SKILL_PATHS"),
+	promptTemplatePaths: optionalList("PI_PROMPT_PATHS"),
+	themePaths: optionalList("PI_THEME_PATHS"),
+	noExtensions: truthy("PI_NO_EXTENSIONS"),
+	noSkills: truthy("PI_NO_SKILLS"),
+	noPromptTemplates: truthy("PI_NO_PROMPTS"),
+	noThemes: truthy("PI_NO_THEMES"),


Do we really need all of those? Can't we just point to project/.pi/ dir and resolve to skills and extensions there? What are the cases when we have skills from multiple paths?

neuromaxer · 2026-05-25T19:42:40Z

+const gpt55ThinkingLevelMap: Partial<Record<ThinkingLevel, string | null>> = {
+	off: "none",
+	minimal: "minimal",
+	low: "low",
+	medium: "medium",
+	high: "high",
+	xhigh: "xhigh",
+};
+
+const deepSeekV4ThinkingLevelMap: Partial<Record<ThinkingLevel, string | null>> = {
+	minimal: null,
+	low: null,
+	medium: null,
+	high: "high",
+	xhigh: "max",
+};


That's quite brittle hardcoding for a specific model. Are there libs that abstract that away?

Can pull well-known LLMs data from https://github.com/earendil-works/pi/blob/main/packages/ai/src/models.generated.ts

neuromaxer · 2026-05-25T19:48:23Z

we'll prly need to split the routes into meaningful subdirs. It's getting harder to navigate them as they're lumped together

neuromaxer · 2026-05-25T19:51:39Z

That's the biggest change. Need to review carefully

neuromaxer · 2026-05-25T19:54:31Z

So this one apparently creates and manages runtimes for multi- setting

neuromaxer · 2026-05-25T19:56:38Z

+
+function truthy(name: string): boolean {
+	const v = process.env[name]?.trim().toLowerCase();
+	return v === "1" || v === "true" || v === "yes" || v === "on";


That's a weird function. Why do we need support for all of those cases?

neuromaxer · 2026-05-25T20:23:33Z

+| `GET`  | `/v1/custom/providers`     | List custom `models.json` providers without secrets   |
+| `PUT`  | `/v1/custom/providers`     | Create or update a custom provider                    |
+| `DELETE` | `/v1/custom/providers/{provider}` | Remove a custom provider                       |


Is openorange considered a custom provider?

neuromaxer · 2026-05-25T20:37:03Z

+`thinkingLevel` is one of `off`, `minimal`, `low`, `medium`, `high`, `xhigh`.
+The runtime rejects changes while a session is streaming with HTTP `409`.
+Pi clamps valid but unsupported thinking levels to the selected model's
+supported set and returns the effective level in the response.


Thinking level is model-specific. Can we infer what thinking levels model expects using pi-ai? See earendil-works/pi#3208

Well-known models are defined in https://github.com/earendil-works/pi/blob/main/packages/ai/src/models.generated.ts

neuromaxer · 2026-05-26T21:04:20Z

 generated from it (see "Consuming from another app" below).

-Mounted under `/v1`:
+In `single` mode, all routes are mounted under `/v1`:


Isn't single mode just a trivial case of multi mode? Do we really need to hard-separate them?

neuromaxer · 2026-05-27T11:32:07Z

+export type AgentAuthProviderRow = {
+  provider: string;
+  name: string;
+  configured: boolean;
+  credentialType?: "api_key" | "oauth";
+  source?: "stored" | "runtime" | "environment" | "fallback" | "models_json_key" | "models_json_command";
+  label?: string;
+  supportsApiKey: boolean;
+  supportsSubscription: boolean;
+  modelCount: number;
+  availableModelCount: number;
+};
+
+export type AgentAuthPrompt = {
+  message: string;
+  placeholder?: string;
+  allowEmpty?: boolean;
+};
+
+export type AgentOAuthFlowState = {
+  id: string;
+  provider: string;
+  providerName: string;
+  status: "starting" | "prompt" | "auth" | "waiting" | "complete" | "error" | "cancelled";
+  authUrl?: string;
+  instructions?: string;
+  prompt?: AgentAuthPrompt;
+  progress: string[];
+  error?: string;
+  expiresAt: string;
+};


I'd say auth and model config belongs at the project registry level (one level above runtime)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…nto runtimes Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Split the HTTP layer so credential routes are served by a new createCredentialsApp(credentials) factory that takes an AgentCredentialsService, while session routes stay under createSessionsApp(runtime). URL paths are unchanged — only the internal mounting changes. Changes: - Add createCredentialsApp(credentials, options) in src/routes.ts that mounts auth/*, custom/*, sessions/models, and healthz routes - Remove credential routes from createSessionsApp, keeping only session routes (sessions, sessions/{id}, prompt, abort, settings, extension-ui, events SSE) - Simplify CreateSessionsAppOptions to Record<string, never> (no more credentialRoutes/sessionRoutes/healthRoute flags) - Update server.ts to mount both apps: credentials at /v1 (always), sessions at /v1 (single mode) or /v1/projects/:projectId (multi mode) - Update openapi.ts to use registry-based approach and mirror the new mounting structure - Update test/server.test.ts: rebuild startServer helper to use AgentRuntimeRegistry and mount both apps separately; update project-scoped tests to use the new structure All 41 tests pass. OpenAPI contract preserved: all 18 documented paths still present. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…alsService Removes ~700 lines of duplicated credential code from AgentRuntime after Tasks 4–10 moved it to AgentCredentialsService. HTTP routes now reach credentials via createCredentialsApp; runtime.ts retains only session lifecycle, extension UI bridge, and resource loader construction. Deleted methods: - modelKey, defaultThinkingForModel, modelRow (now credentials.modelRow) - listModels, listAuthProviders - setProviderApiKey, removeProviderCredential, assertProviderId - customProviderApi - OAuth flow state machine: oauthFlowState, updateOAuthFlow, scheduleOAuthFlowCleanup, activeOAuthFlowForProvider, oauthLoginErrorMessage, waitForOAuthFlowUpdate, startProviderSubscriptionLogin, continueProviderSubscriptionLogin, getProviderSubscriptionLogin, cancelProviderSubscriptionLogin - readModelsJson, writeModelsJson - listCustomProviders, upsertCustomProvider, removeCustomProvider Deleted fields: pendingOAuthFlows, modelsJsonPath Deleted types: PendingOAuthFlow, CUSTOM_PROVIDER_APIS Updated: - sessionModelSettings calls credentials.modelRow - sessionModelDefaults calls credentials.defaultThinkingForModel - setSessionModelInternal calls credentials.defaultThinkingForModel Re-exported types from credentialsService: AgentModelRow, AgentAuthProviderRow, AgentAuthPrompt, AgentOAuthFlowState, AgentCustomProviderApi, AgentCustomProviderModel, AgentCustomProviderRow, UpsertCustomProviderRequest Test fix: LiteLLM test now instantiates AgentCredentialsService with thinking defaults from litellmRuntimeConfig before calling listModels. runtime.ts: 1241 → 749 lines (-492, ~40% reduction) All 41 tests pass. TypeScript compiles cleanly. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…onsibilities

Pi switch refactor

Canvinus added 9 commits May 22, 2026 10:45

feat: add pi model controls and extension bridge

6839e4e

feat: expose pi resource overlays

1464f91

chore(deps): pin pi sdk to 0.75.4

6112c2b

fix: make agent-server cli executable

5e93fae

feat: expose pi provider auth

4ef86e4

feat: support pi subscription auth and custom providers

8ebfe58

fix: reuse active pi subscription login

edd6d6f

feat: add project-scoped pi runtimes

957477b

docs: clarify provider-neutral streaming

e1a9252

Canvinus mentioned this pull request May 23, 2026

Switch to agent-server appx-org/appx#2

Merged

Canvinus marked this pull request as ready for review May 23, 2026 20:00

Canvinus added 3 commits May 23, 2026 22:11

fix(agent-server): return 404 for missing prompt sessions

aa3851e

feat(agent-server): support single and multi project modes

21d7e08

docs(agent-server): remove opencode migration wording

341785e

chatgpt-codex-connector Bot reviewed May 25, 2026

View reviewed changes

neuromaxer reviewed May 25, 2026

View reviewed changes

Comment thread src/runtime.ts Outdated

neuromaxer May 25, 2026

Copy link
Copy Markdown

Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That's the biggest change. Need to review carefully

neuromaxer reviewed May 25, 2026

View reviewed changes

add feature doc, pass .env during dev

5edce5c

neuromaxer reviewed May 26, 2026

View reviewed changes

neuromaxer reviewed May 27, 2026

View reviewed changes

neuromaxer and others added 3 commits May 27, 2026 14:18

docs(plan): add credentials extraction + thinking-level dedup plan

d5751e1

chore(deps): add @earendil-works/pi-ai for shared thinking-level helpers

9c29e19

feat(thinking): add src/thinking.ts wrapping Pi's clamp helpers

1753d3b

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

neuromaxer and others added 27 commits May 27, 2026 15:13

feat(credentials): move custom-provider models.json CRUD

c7e5949

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

feat(registry): construct shared AgentCredentialsService and inject i…

b6b5496

…nto runtimes Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

chore(exports): re-export credentials service and thinking helpers

d720ce9

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

docs(readme): update library-mode example to use createCredentialsApp

8f170af

add builder isolation doc

397b220

add docs on project runtime split and agent services

d05a4c1

split runtime into ProjectRuntime and ProjectSession to separate resp…

7229659

…onsibilities

use AgentSessionServices shared across sessions

029c811

organise files into sub-dirs

21598db

rename AgentRuntimeRegistry to ProjectRegistry

c9c1a65

factor out config from server

6d70f41

stick to pi convention in agent files

e5494a3

collapse multi and single modes, simplify project structure convention

f9caf76

update docs

35f664b

update readme

c0a48e2

align types initial attempt

1305e62

pass pi types from typia to zod

6f73da4

add operationId to each route

239951a

add gh action to check openapi freshness and pi breaking changes

0cb017e

factor out openapi contract generation logic

cd17a24

add pre-commit hook with schema generation and biom code formatter

49f4d39

allow deleting sessions

9aa23aa

allow deleting sessions p2

d3014ec

surface pi llm errors in events

e22a433

Merge pull request #2 from appx-org/pi-switch-refactor

8faec1a

Pi switch refactor

neuromaxer changed the title ~~[codex] Add Pi model controls and project-scoped runtimes~~ Switch to Pi Jun 9, 2026

disable oasdiff for now

0c3cd56

neuromaxer merged commit d7f10df into main Jun 9, 2026
1 check passed

Conversation

Canvinus commented May 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Why

Validation

Notes

Uh oh!

Canvinus commented May 25, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot May 25, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

neuromaxer May 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Canvinus commented May 23, 2026 •

edited

Loading

neuromaxer May 25, 2026 •

edited

Loading