@hienlh/ppm 0.20.5 → 0.20.6

package/docs/architecture/ai-chat-and-providers.md

@@ -2,7 +2,87 @@
 
 > Part of the [PPM system architecture](../system-architecture.md).
 
+## Codex context settings
+
+**AI Settings → Codex** exposes **Context window (tokens)** and **Auto-compact
+threshold (tokens)**. These map to Codex's documented `model_context_window` and
+`model_auto_compact_token_limit` keys under `ai.providers.codex`.
+
+Both controls offer presets, Codex default and Custom entry. Presets are numeric
+shortcuts, not advertised model capabilities. Existing non-preset values appear
+as Custom. Both values are optional positive safe integers. The compact threshold cannot
+exceed an explicitly configured window. The form saves the pair together;
+`PUT /api/settings/ai` also validates merged partial updates before persisting.
+Clearing a field sends `null`, which removes the PPM override. An omitted field
+in a partial update preserves the existing setting.
+
+PPM passes configured values in the `config` map of app-server `thread/start`
+and `thread/resume`, including account rotation. It does not edit account
+`config.toml` files. Unset values inherit native Codex configuration and model
+defaults. Changes apply on the next provider connection/resumption; an already
+running session keeps its current limits. Restart PPM and resume an existing
+session to apply changed limits to it.
+
+Increasing these values does not increase the model/account's supported limit.
+Codex can reserve part of the configured window, so runtime token usage may
+report a smaller effective window. Local compacted rollouts inspected on
+2026-09-15 reported 258,400 tokens (272,000 × 95%); this is runtime evidence for
+this installation, not a hardcoded application default.
+
+References: [Codex configuration](https://learn.chatgpt.com/docs/config-file/config-reference)
+and [App Server](https://learn.chatgpt.com/docs/app-server).
+
 ## Provider Layer (AI Adapters)
+
+### Shared instructions and memory
+
+**Settings → AI → Share rules and memory between providers** controls
+`ai.share_provider_context`. It defaults to `true`, including existing configs.
+PPM checks sources for each project-chat message and sends a snapshot only when
+it differs from the last delivered snapshot for that provider/session. Live
+follow-ups use the same check. Compaction, failed delivery, explicit resume,
+server restart, or eviction from the bounded 512-session cache causes a fresh
+snapshot on the next message. Slash commands retain their native parser behavior.
+Disabling stops future injections, but does not
+remove context already present in an existing conversation. Start a new session
+for a clean context. The isolated HTTP API proxy is outside this project-chat flow.
+
+Sources include project `AGENTS.md`, `CLAUDE.md`, `CLAUDE.local.md`, provider
+`.{providerId}/rules/` and `.{providerId}/memory/` directories,
+user Claude rules, Codex account `AGENTS.md`, and Claude auto-memory for the exact
+project path. PPM does not import account credentials, runtime settings/hooks,
+conversation transcripts, or another project's auto-memory.
+
+PPM reads existing native memory directly, including knowledge created before
+PPM was installed. It does not create or maintain an intermediate memory store.
+Claude's native `MEMORY.md` index is prioritized and its topic files remain
+readable at their original paths. `CLAUDE_CONFIG_DIR` is respected. Codex native
+memory is read from its memory database with exact project attribution from its
+thread database; databases are opened read-only. New memory stays in the native
+provider's storage. Conversation transcripts are not treated as memory.
+
+Snapshots omit rules already loaded by the receiving provider (Codex AGENTS and
+rules, Claude instructions/rules/auto-memory, Cursor rules). Codex's native memory
+is supplied only to other providers. Unknown providers retain conventional sources.
+Memory directories contribute their index and a directory reference, with topic
+files read on demand. A compact revision tracks file metadata changes, including
+omitted topics, within the bounded scan.
+
+Snapshots are bounded to 48 content files, 4 KB excerpts per file and 12,000 content
+characters, with limited directory traversal and a 1,500-character directory list.
+Large files retain their original paths and a truncation notice so agents can
+read the remainder on demand. Unreadable files are skipped. Rule scopes/frontmatter
+remain part of the supplied instructions.
+
+New registered providers automatically receive the common context and contribute
+conventional `.providerId/rules` and `.providerId/memory` directories. For custom
+paths, implement `AIProvider.getSharedContextSources(projectPath)`. Implement
+`supportsSharedContext = true` to consume `SendMessageOpts.sharedContext` separately
+from raw user text; `CliProvider` already does this. Other implementations receive
+a prefixed prompt automatically. Keep titles based on raw user text and use
+`stripSharedContext` when importing native transcripts. Claude, Codex and Cursor
+already implement this separation.
+
 **Component:** Provider interface + implementations
 
 **Responsibilities:**
@@ -263,6 +343,14 @@ if (entry.phase !== "idle" && entry.abort) {
 
 ### WebSocket Reconnection Sync
 
+Chat sockets receive a server heartbeat every 15 seconds. The client reconnects
+after 45 seconds without any incoming frame, including sockets that still report
+OPEN, and checks for a stale connection when the tab becomes visible. Replay
+finishes before queued live frames are applied. If a turn becomes idle without a
+usable finalized answer, the client reloads history; healthy completions retain
+their richer live metadata. Recovery responses are discarded after new activity
+or a session switch, and older initial loads cannot overwrite applied recovery.
+
 ```
 FE WebSocket closes (network issue, tab closes)
     ↓

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hienlh/ppm",
-  "version": "0.20.5",
+  "version": "0.20.6",
   "description": "Personal Project Manager — mobile-first web IDE with AI assistance",
   "author": "hienlh",
   "license": "SEE LICENSE IN LICENSE",

package/schemas/ppm-config.schema.json

@@ -36,6 +36,11 @@
       "type": "object",
       "properties": {
         "default_provider": { "type": "string", "default": "claude" },
+        "share_provider_context": {
+          "type": "boolean",
+          "default": true,
+          "description": "Share project rules and memory between AI providers"
+        },
         "providers": {
           "type": "object",
           "additionalProperties": {
@@ -51,9 +56,21 @@
       "properties": {
         "type": {
           "type": "string",
-          "enum": ["agent-sdk", "mock"]
+          "enum": ["agent-sdk", "cli", "mock"]
         },
         "api_key_env": { "type": "string" },
+        "model_context_window": {
+          "type": ["integer", "null"],
+          "minimum": 1,
+          "maximum": 9007199254740991,
+          "description": "Codex context window in tokens; null or omitted inherits Codex configuration"
+        },
+        "model_auto_compact_token_limit": {
+          "type": ["integer", "null"],
+          "minimum": 1,
+          "maximum": 9007199254740991,
+          "description": "Codex automatic compaction threshold in tokens; cannot exceed an explicit context window; null or omitted inherits Codex configuration"
+        },
         "model": {
           "type": "string",
           "description": "Model ID (e.g. claude-opus-5, claude-fable-5, claude-sonnet-5)",

package/src/providers/claude-agent-sdk.ts

@@ -18,10 +18,11 @@ import type {
   ModelOption,
 } from "./provider.interface.ts";
 import { configService } from "../services/config.service.ts";
+import { withSharedContext, stripSharedContext } from "../shared/provider-context.ts";
 import { mcpConfigService } from "../services/mcp-config.service.ts";
 import { listInheritedClaudeMcpServers } from "../services/claude-code-mcp.service.ts";
 import { updateFromSdkEvent } from "../services/claude-usage.service.ts";
-import { getSessionProjectPath, setSessionMetadata, getSessionTitles, insertTurnUsage, getSessionAccount } from "../services/db.service.ts";
+import { getSessionProjectPath, setSessionMetadata, getSessionTitle, setSessionTitle, getSessionTitles, insertTurnUsage, getSessionAccount } from "../services/db.service.ts";
 import { SUBSCRIPTION_PROMPT_CACHE_TTL_MS, API_KEY_PROMPT_CACHE_TTL_MS } from "../services/subprocess-retention.ts";
 import { buildTurnUsage, formatTurnUsageLog } from "../shared/turn-usage.ts";
 import { accountSelector } from "../services/account-selector.service.ts";
@@ -213,6 +214,7 @@ interface PendingApproval {
  * Uses canUseTool callback for tool approvals and AskUserQuestion.
  */
 export class ClaudeAgentSdkProvider implements AIProvider {
+  readonly supportsSharedContext = true;
   id = "claude";
   name = "Claude";
 
@@ -571,7 +573,7 @@ export class ClaudeAgentSdkProvider implements AIProvider {
       const sessions: SessionInfo[] = sdkSessions.map((s) => ({
         id: s.sessionId,
         providerId: this.id,
-        title: dbTitles[s.sessionId] ?? s.customTitle ?? s.summary ?? s.firstPrompt ?? "Chat",
+        title: dbTitles[s.sessionId] ?? s.customTitle ?? s.summary ?? (stripSharedContext(s.firstPrompt ?? "") || "Chat"),
         createdAt: new Date(s.lastModified).toISOString(),
         updatedAt: new Date(s.lastModified).toISOString(),
       }));
@@ -619,7 +621,7 @@ export class ClaudeAgentSdkProvider implements AIProvider {
         return {
           id: info.sessionId,
           providerId: this.id,
-          title: dbTitles[info.sessionId] ?? info.customTitle ?? info.summary ?? info.firstPrompt ?? "Chat",
+          title: dbTitles[info.sessionId] ?? info.customTitle ?? info.summary ?? (stripSharedContext(info.firstPrompt ?? "") || "Chat"),
           createdAt: new Date(info.lastModified).toISOString(),
           updatedAt: new Date(info.lastModified).toISOString(),
         };
@@ -725,13 +727,14 @@ export class ClaudeAgentSdkProvider implements AIProvider {
    * Push a follow-up message into an existing streaming session's generator.
    * Called by WS handler for follow-up messages (Phase 2).
    */
-  pushMessage(sessionId: string, content: string, opts?: { priority?: 'now' | 'next' | 'later'; images?: Array<{ data: string; mediaType: string }> }): void {
+  pushMessage(sessionId: string, content: string, opts?: import("./provider.interface.ts").SendMessageOpts): void {
     const ss = this.streamingSessions.get(sessionId);
     if (!ss) {
       console.warn(`[sdk] pushMessage: no streaming session for ${sessionId}`);
       return;
     }
-    const msgContent = buildMessageParam(content, opts?.images);
+    const modelInput = withSharedContext(content, opts?.sharedContext);
+    const msgContent = buildMessageParam(modelInput, opts?.images);
     ss.controller.push({
       type: 'user',
       message: msgContent,
@@ -740,7 +743,7 @@ export class ClaudeAgentSdkProvider implements AIProvider {
       priority: opts?.priority ?? 'next',
     });
     // Track latest message for retry paths (fixes stale firstMsg bug)
-    ss.lastUserContent = content;
+    ss.lastUserContent = modelInput;
     ss.lastUserImages = opts?.images;
     console.log(`[sdk] pushMessage: session=${sessionId} priority=${opts?.priority ?? 'next'}`);
   }
@@ -794,6 +797,7 @@ export class ClaudeAgentSdkProvider implements AIProvider {
     message: string,
     opts?: import("./provider.interface.ts").SendMessageOpts & { forkSession?: boolean; priority?: 'now' | 'next' | 'later'; images?: Array<{ data: string; mediaType: string }> },
   ): AsyncIterable<ChatEvent> {
+    const modelInput = withSharedContext(message, opts?.sharedContext);
     // SDK requires valid UUID session IDs. Short/random IDs can leak from
     // tab derivation or URL parsing — migrate to a real UUID early.
     const UUID_RE = /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i;
@@ -826,7 +830,7 @@ export class ClaudeAgentSdkProvider implements AIProvider {
     // Follow-up: push into existing streaming session, yield nothing
     const existingStream = this.streamingSessions.get(sessionId);
     if (existingStream) {
-      const msgContent = buildMessageParam(message, opts?.images);
+      const msgContent = buildMessageParam(modelInput, opts?.images);
       existingStream.controller.push({
         type: 'user',
         message: msgContent,
@@ -835,7 +839,7 @@ export class ClaudeAgentSdkProvider implements AIProvider {
         priority: opts?.priority ?? 'next',
       });
       // Track latest message for retry paths (fixes stale firstMsg bug)
-      existingStream.lastUserContent = message;
+      existingStream.lastUserContent = modelInput;
       existingStream.lastUserImages = opts?.images;
       console.log(`[sdk] sendMessage follow-up: session=${sessionId} pushed to generator`);
       return; // Events flow through first-message's consumer loop
@@ -853,6 +857,11 @@ export class ClaudeAgentSdkProvider implements AIProvider {
     const count = this.messageCount.get(sessionId) ?? 0;
     const isFirstMessage = count === 0;
     this.messageCount.set(sessionId, count + 1);
+    // Native firstPrompt may truncate before the shared-context block ends.
+    // Keep the opening user title available after restart without replacing a rename.
+    if (isFirstMessage && opts?.sharedContext && getSessionTitle(sessionId) === null) {
+      setSessionTitle(sessionId, meta.title);
+    }
 
     // Check if this session should fork from another
     const forkSourceId = this.forkSources.get(sessionId);
@@ -1145,7 +1154,7 @@ export class ClaudeAgentSdkProvider implements AIProvider {
       // model a bare path instead — the round trip the caller passed them in to avoid.
       const firstMsg = {
         type: 'user' as const,
-        message: buildMessageParam(message, opts?.images),
+        message: buildMessageParam(modelInput, opts?.images),
         parent_tool_use_id: null,
         session_id: sessionId,
       };
@@ -1165,7 +1174,7 @@ export class ClaudeAgentSdkProvider implements AIProvider {
       // Also returns the raw content/images for re-populating the new streaming session.
       const buildRetryMsg = () => {
         const ss = this.streamingSessions.get(sessionId);
-        const content = ss?.lastUserContent ?? message;
+        const content = ss?.lastUserContent ?? modelInput;
         const images = ss?.lastUserImages;
         const retryContent = turnProgressed ? "Continue from where you left off." : content;
         return {
@@ -1184,7 +1193,7 @@ export class ClaudeAgentSdkProvider implements AIProvider {
       // On crash retry, use buildRetryMsg to get the latest user message (not the stale firstMsg)
       const initRetry = crashRetryCount > 0 ? buildRetryMsg() : null;
       initialCtrl.push(initRetry?.msg ?? firstMsg);
-      const initContent = initRetry?.lastUserContent ?? message;
+      const initContent = initRetry?.lastUserContent ?? modelInput;
       const initImages = initRetry?.lastUserImages ?? opts?.images;
 
       const initialQuery = query({
@@ -2219,7 +2228,7 @@ function findMissingSessions(
           if (head[end] === '"') break;
           end++;
         }
-        const raw = head.slice(start, end).replace(/\\n/g, " ").replace(/\\"/g, '"').trim();
+        const raw = stripSharedContext(head.slice(start, end).replace(/\\n/g, "\n").replace(/\\"/g, '"').trim()).replace(/\n/g, " ");
         if (raw.length > 0) {
           title = raw.length > 120 ? raw.slice(0, 120) + "…" : raw;
         }
@@ -2236,4 +2245,3 @@ function findMissingSessions(
   }
   return results;
 }
-

package/src/providers/cli-provider-base.ts

@@ -9,6 +9,7 @@ import type {
 import { spawn, type ChildProcess } from "node:child_process";
 import { parseNdjsonLines } from "../utils/ndjson-line-parser.ts";
 import { configService } from "../services/config.service.ts";
+import { withSharedContext } from "../shared/provider-context.ts";
 
 /**
  * Abstract base class for CLI-spawning AI providers.
@@ -16,6 +17,7 @@ import { configService } from "../services/config.service.ts";
  * Subclasses only implement event mapping + arg building.
  */
 export abstract class CliProvider implements AIProvider {
+  readonly supportsSharedContext = true;
   abstract readonly id: string;
   abstract readonly name: string;
   abstract readonly cliCommand: string;
@@ -118,7 +120,7 @@ export abstract class CliProvider implements AIProvider {
     const config = this.getProviderConfig();
     const args = this.buildArgs({
       sessionId: isResume ? sessionId : undefined,
-      message,
+      message: withSharedContext(message, opts?.sharedContext),
       model: config?.model,
       permissionMode: opts?.permissionMode || config?.permission_mode,
       isResume,

package/src/providers/codex-app-server/codex-history.ts

@@ -1,4 +1,5 @@
 import type { ChatMessage, ChatEvent, SessionInfo } from "../provider.interface.ts";
+import { stripSharedContext } from "../../shared/provider-context.ts";
 import { readdirSync, readFileSync, statSync } from "node:fs";
 import { join, resolve } from "node:path";
 import { homedir } from "node:os";
@@ -135,7 +136,7 @@ export function parseRolloutJsonl(
     if (rec.type === "event_msg") {
       if (p.type === "user_message" && typeof p.message === "string") {
         if (pendingEvents.length) flushAssistant("", ts); // tools with no final text
-        messages.push({ id: `rollout-${i++}`, role: "user", content: p.message, timestamp: ts });
+        messages.push({ id: `rollout-${i++}`, role: "user", content: stripSharedContext(p.message), timestamp: ts });
       } else if (p.type === "agent_message" && typeof p.message === "string") {
         flushAssistant(p.message, ts);
       } else if (p.type === "item_completed") {
@@ -145,7 +146,7 @@ export function parseRolloutJsonl(
         const mapped = mapRolloutItem(p.item);
         if (mapped.kind === "user") {
           if (pendingEvents.length) flushAssistant("", ts);
-          messages.push({ id: `rollout-${i++}`, role: "user", content: mapped.text, timestamp: ts });
+          messages.push({ id: `rollout-${i++}`, role: "user", content: stripSharedContext(mapped.text), timestamp: ts });
         } else if (mapped.kind === "assistant") {
           flushAssistant(mapped.text, ts);
         } else if (mapped.kind === "events") {
@@ -196,7 +197,8 @@ export function parseRolloutJsonl(
         if (item.type !== "message") continue;
         const role = item.role === "assistant" ? "assistant" : item.role === "user" ? "user" : null;
         if (!role) continue;
-        const content = contentToText(item.content);
+        const rawContent = contentToText(item.content);
+        const content = role === "user" ? stripSharedContext(rawContent) : rawContent;
         if (content) messages.push({ id: `rollout-${i++}`, role, content, timestamp: ts });
       }
     }

package/src/providers/codex-app-server/codex-provider.ts

@@ -10,6 +10,7 @@ import type {
   UsageInfo,
 } from "../provider.interface.ts";
 import { configService } from "../../services/config.service.ts";
+import { withSharedContext } from "../../shared/provider-context.ts";
 import { setSessionMetadata, getSessionProjectPath, setSessionProvider, setSessionCodexAccount, getSessionCodexAccount, getSessionTitles, insertTurnUsage } from "../../services/db.service.ts";
 import {
   resolveCodexAccountForSession,
@@ -291,6 +292,7 @@ function buildUserInputResponse(questions: unknown, data: unknown): ToolRequestU
  * notification stream. Token-by-token streaming is the load-bearing capability.
  */
 export class CodexAppServerProvider implements AIProvider {
+  readonly supportsSharedContext = true;
   readonly id = "codex";
   readonly name = "Codex";
 
@@ -448,7 +450,7 @@ export class CodexAppServerProvider implements AIProvider {
     // gets. Images lead: the text usually refers to them ("what is this?").
     const input: UserInput[] = [
       ...(opts?.imagePaths ?? []).map((path) => ({ type: "localImage" as const, path })),
-      { type: "text" as const, text: message, text_elements: [] },
+      { type: "text" as const, text: withSharedContext(message, opts?.sharedContext), text_elements: [] },
     ];
     const turnModel = codexModel(opts?.model);
     const turnEffort = opts?.effort ?? this.config?.effort;
@@ -651,6 +653,7 @@ export class CodexAppServerProvider implements AIProvider {
     client.notify("initialized");
 
     const resumeBase = {
+      ...this.contextConfigOverrides(),
       cwd: live.cwd,
       sandbox: live.permission.sandbox,
       approvalPolicy: live.permission.approvalPolicy,
@@ -687,6 +690,17 @@ export class CodexAppServerProvider implements AIProvider {
     return client.request("thread/resume", { threadId, path, ...resumeBase });
   }
 
+  /** Read on connect/resume only; changing settings never interrupts a live turn. */
+  private contextConfigOverrides(): { config?: Record<string, number> } {
+    const providerConfig = this.config;
+    const config: Record<string, number> = {};
+    for (const key of ["model_context_window", "model_auto_compact_token_limit"] as const) {
+      const value = providerConfig?.[key];
+      if (value != null) config[key] = value;
+    }
+    return Object.keys(config).length ? { config } : {};
+  }
+
   private async connect(sessionId: string, opts?: SendMessageOpts): Promise<LiveSession> {
     const meta = this.sessions.get(sessionId);
     const cwd = meta?.projectPath || getSessionProjectPath(sessionId) || process.cwd();
@@ -720,7 +734,7 @@ export class CodexAppServerProvider implements AIProvider {
     await client.request("initialize", { clientInfo: CLIENT_INFO, capabilities: CAPABILITIES }, CONTROL_REQUEST_TIMEOUT_MS);
     client.notify("initialized");
 
-    const resumeBase = { cwd, sandbox: permission.sandbox, approvalPolicy: permission.approvalPolicy, ...(model ? { model } : {}) };
+    const resumeBase = { cwd, sandbox: permission.sandbox, approvalPolicy: permission.approvalPolicy, ...(model ? { model } : {}), ...this.contextConfigOverrides() };
     // Only treat as a resume when a rollout for this id is attributable to THIS
     // project (fail-closed cwd guard) — never resume another project's thread.
     const result = found

package/src/providers/codex-app-server/codex-rollout-header.ts

@@ -1,4 +1,5 @@
 import { mapRolloutItem } from "./codex-rollout-items.ts";
+import { stripSharedContext } from "../../shared/provider-context.ts";
 
 /**
  * Line primitives for a codex rollout JSONL, plus the header a session list
@@ -49,7 +50,7 @@ const TITLE_MAX_CHARS = 80;
 
 /** First user prompt → a one-line label. */
 function toTitle(text: string): string | undefined {
-  const flat = text.replace(/\s+/g, " ").trim();
+  const flat = stripSharedContext(text).replace(/\s+/g, " ").trim();
   if (!flat) return undefined;
   return flat.length > TITLE_MAX_CHARS ? `${flat.slice(0, TITLE_MAX_CHARS - 1)}…` : flat;
 }

package/src/providers/cursor-cli/cursor-history.ts

@@ -3,6 +3,7 @@ import { homedir } from "node:os";
 import { join } from "node:path";
 import { existsSync, readdirSync } from "node:fs";
 import type { ChatMessage, SessionInfo } from "../provider.interface.ts";
+import { stripSharedContext } from "../../shared/provider-context.ts";
 
 const DEFAULT_CHATS_DIR = join(homedir(), ".cursor", "chats");
 
@@ -38,7 +39,7 @@ export async function listCursorSessions(providerId: string, chatsDir?: string):
               const hex = typeof row.value === "string" ? row.value : Buffer.from(row.value).toString("utf-8");
               const json = Buffer.from(hex, "hex").toString("utf-8");
               const meta = JSON.parse(json);
-              if (meta.name) title = meta.name.split("\n")[0].slice(0, 80);
+              if (meta.name) title = stripSharedContext(meta.name).split("\n")[0]?.slice(0, 80) || title;
               if (meta.createdAt) createdAt = new Date(meta.createdAt).toISOString();
             }
           } catch { /* use defaults */ }
@@ -139,6 +140,7 @@ function parseDagBlobs(
           if (parsed.role === "user") {
             const match = content.match(/<user_query>\s*([\s\S]*?)\s*<\/user_query>/);
             if (match?.[1]) content = match[1];
+            content = stripSharedContext(content);
           }
           messages.push({
             id: blob.id,
@@ -159,7 +161,7 @@ function parseDagBlobs(
             messages.push({
               id: blob.id,
               role: messages.length % 2 === 0 ? "user" : "assistant",
-              content: textParts,
+              content: messages.length % 2 === 0 ? stripSharedContext(textParts) : textParts,
               timestamp: new Date().toISOString(),
             });
             continue;

package/src/server/routes/settings.ts

@@ -3,6 +3,7 @@ import { configService, FILE_CONFIG_KEYS } from "../../services/config.service.t
 import { getConfigValue, setConfigValue, listPairedChats, getPairingByCode, approvePairing, revokePairing, getPPMBotMemories, getDb } from "../../services/db.service.ts";
 import {
   validateAIProviderConfig,
+  validateCodexContextConfig,
   validateDefaultProvider,
   VALID_PROVIDERS,
   DEFAULT_CONFIG,
@@ -178,16 +179,21 @@ settingsRoutes.get("/ai", (c) => {
   return c.json(ok(stripSensitiveFields(ai)));
 });
 
-/** PUT /settings/ai — update AI provider settings, writes to yaml */
+/** PUT /settings/ai — update AI settings, persists to SQLite */
 settingsRoutes.put("/ai", async (c) => {
   try {
     const body = await c.req.json<{
       default_provider?: string;
+      share_provider_context?: boolean;
       providers?: Record<string, Partial<AIProviderConfig>>;
     }>();
 
     const currentAi = configService.get("ai");
 
+    if ("share_provider_context" in body && typeof body.share_provider_context !== "boolean") {
+      return c.json(err("share_provider_context must be a boolean"), 400);
+    }
+
     // Validate each provider config
     if (body.providers) {
       for (const [name, providerConfig] of Object.entries(body.providers)) {
@@ -201,6 +207,7 @@ settingsRoutes.put("/ai", async (c) => {
     // Merge: body overrides current values (shallow merge per provider)
     const updated = {
       ...currentAi,
+      share_provider_context: body.share_provider_context ?? currentAi.share_provider_context ?? true,
       ...(body.default_provider && { default_provider: body.default_provider }),
     };
     if (body.providers) {
@@ -214,6 +221,13 @@ settingsRoutes.put("/ai", async (c) => {
           ...currentAi.providers[name],
           ...config,
         } as AIProviderConfig;
+        for (const key of ["model_context_window", "model_auto_compact_token_limit"] as const) {
+          if (updated.providers[name]![key] === null) delete updated.providers[name]![key];
+        }
+        const contextErrors = validateCodexContextConfig(updated.providers[name]!);
+        if (contextErrors.length) {
+          return c.json(err(`Provider "${name}": ${contextErrors.join(", ")}`), 400);
+        }
       }
     }
 

package/src/server/ws/chat.ts

@@ -1288,7 +1288,7 @@ export const chatWebSocket = {
         if (provider && "pushMessage" in provider && parsed.type === "message") {
           const effort = getSessionEffort(sessionId) ?? undefined;
           const thinkingBudget = getSessionThinking(sessionId);
-          (provider as any).pushMessage(sessionId, parsed.content, {
+          await chatService.pushMessage(providerId, sessionId, parsed.content, {
             priority: parsed.priority ?? 'next',
             images: parsed.images,
             imagePaths: parsed.imagePaths,
@@ -1413,7 +1413,7 @@ export const chatWebSocket = {
           }, 0);
         });
       } else if (provider && "pushMessage" in provider) {
-        (provider as any).pushMessage(sessionId, instruction, { priority: "next" });
+        await chatService.pushMessage(providerId, sessionId, instruction, { priority: "next" });
         entry.turnEvents = [];
         entry.pendingApprovalEvent = undefined;
         setPhase(sessionId, "thinking");