fix token usage average

Refactor/0.1.0-beta

.agents/skills/cursor-prefix-stability/SKILL.md

@@ -0,0 +1,82 @@
+---
+name: cursor-prefix-stability
+description: Implement and review Cursor BYOK conversation projection with append-only provider history and prefix-cache stability. Use when changing runtime prompts, request context, canonical messages, checkpoint hydration, compaction, message identity, or provider history serialization under server.
+---
+
+# Cursor prefix stability
+
+Preserve the longest valid provider prefix across turns. Treat provider-visible history as an append-only log unless an explicit compaction operation replaces it.
+
+## Architecture
+
+Keep the invariant in the provider-independent conversation layers:
+
+```text
+server/
+├── prompt/cursor/*/runtime.md       Per-turn runtime content
+├── src/cursor/request/              Request context and runtime compilation
+├── src/cursor/projection/           Canonical ↔ Cursor checkpoint projection
+├── src/cursor/checkpoint/           Stable roots, turns, and hydration
+├── src/run/                         Provider-independent model history
+└── src/provider/                    Provider-specific serialization only
+```
+
+Do not solve prefix instability independently in each provider adapter. Produce one stable canonical history before dispatching to OpenAI Responses, OpenAI Chat, Anthropic, or another provider.
+
+## Required invariants
+
+- When no compaction occurs, the complete provider-visible history from turn N must be an exact structural prefix of turn N+1. Never edit, remove, merge, reorder, renormalize, or regenerate an earlier message.
+- Keep `PromptSpec.instructions` and the stable tool prefix byte-stable when their inputs have not changed. Deterministic ordering is required; do not use unordered iteration in provider-visible output.
+- Separate conversation/request context from the per-turn runtime message. The runtime message contains current-turn material such as the user query, selected context, open files, action context, mode reminders, and timestamp.
+- Project rules, skills, subagents, environment/Git context, and MCP metadata as a stable `request-context:*` message:
+  - append it on the first applicable turn;
+  - do not append it again when its compiled content is identical to the latest projected request context;
+  - when its content changes, append a new request-context message immediately before the current runtime message;
+  - never represent a context change by rewriting the system prompt, replacing an earlier context message, or mutating a checkpoint root.
+- Give every appended context update a unique event identity. Compare the latest context by content, not only by identifier, so `A → B → A` appends the final `A` again while retries of the same event remain idempotent.
+- Preserve `request-context:*` wire identity through checkpoint encoding and hydration. Deduplication must still work after process restart or conversation resume.
+- Automatic compaction is an explicit prefix reset. Compact obsolete history, retain exactly the latest request-context message, then place the summary and current initial messages in deterministic order. Manual compaction may reproject current context on the next user turn.
+- Background completions and injected runtime events must not manufacture duplicate request context unless they actually start a user turn whose context changed.
+
+## Change workflow
+
+Before editing, trace the whole path that applies:
+
+```text
+AgentRunRequest
+→ request context hydration/compilation
+→ CanonicalMessage identity and persistence
+→ checkpoint encode/decode
+→ projected ModelRequest history
+→ provider serialization
+```
+
+Determine which data is conversation-level and which is turn-level. If a proposed change moves or rewrites an earlier provider-visible value, redesign it as a new append-only event unless the operation is explicitly compaction.
+
+Use TDD for changes in this path. Start with a failing behavioral test, then implement the smallest provider-independent change.
+
+## Verification
+
+Cover the affected behavior with structural assertions, not token-count estimates alone:
+
+- Two turns with identical request context: the first request history is an exact prefix of the second, the system instructions are identical, and only one `request-context:*` message exists.
+- Changed context: one new context message appears at the tail before the new runtime query; all earlier messages remain unchanged.
+- Context reversion `A → B → A`: three distinct context events are retained in order.
+- Retry of one runtime event: no duplicate or conflicting context message is persisted.
+- Checkpoint round-trip: request-context identity and content survive encode/hydrate.
+- Automatic compaction: only the latest context is retained outside the summary.
+- Runtime templates render without embedding conversation-level rules or MCP metadata in every user query.
+
+Run focused tests first, then the relevant server suites:
+
+```bash
+cargo test --lib
+cargo test --test runtime_modes
+cargo test --test prefix_stability
+cargo test --test checkpoint_recovery
+cargo test --test compaction
+cargo clippy --lib -- -D warnings
+cargo fmt --all -- --check
+```
+
+Do not repair unrelated dirty-worktree failures while validating. Report them separately.

.agents/skills/release/SKILL.md

@@ -0,0 +1,77 @@
+---
+name: release
+description: Prepare, authorize, publish, troubleshoot, and verify Cursor BYOK desktop GitHub Releases. Use for version bumps, release tags, GitHub Actions release runs, updater manifests, signing, or release-readiness checks.
+---
+
+# Desktop release
+
+Release through `.github/workflows/release.yml`. Preserve both updater formats: Tauri uses `latest.json`; legacy `v0.0.49` clients use `update.json`.
+
+## Publication authority
+
+- Only the repository author, GitHub user `leookun`, may authorize a live release.
+- Before any live mutation, require an explicit release instruction from the author in the current task and verify `gh api user --jq .login` returns `leookun`.
+- Treat all of these as publication actions: pushing a `v*` tag, rerunning the release workflow, and publishing or editing a GitHub Release. Pushing a release commit to `main` only prepares the release and must never trigger publication by itself.
+- Without that authorization, restrict work to inspection, local edits, validation, and a release-ready commit or branch. Do not infer publication permission from requests such as “prepare”, “check”, or “ready to release”.
+- Never print, commit, or upload `.tauri/cursor-byok.key` anywhere except the repository's `TAURI_SIGNING_PRIVATE_KEY` Actions Secret when the author explicitly requests that secret configuration.
+- Never delete, replace, or move an existing tag or published Release without separate explicit authorization.
+
+## Version and GitHub Release policy
+
+- Do not use GitHub prereleases. Keep `prerelease: false` for every release and publish the completed release as Latest.
+- Use `vMAJOR.MINOR.PATCH` for a stable tag, for example `v0.1.0`.
+- Use standard SemVer `vMAJOR.MINOR.PATCH-beta.N` for a test tag, for example `v0.1.0-beta.1`. A beta is still a normal GitHub Release, not a GitHub prerelease. Make its title or body visibly say Beta.
+- This normal-Release rule is required because both installed update clients resolve assets through GitHub's `/releases/latest/download/` path, which excludes GitHub prereleases.
+- Windows beta builds must use the NSIS bundle. WiX/MSI rejects nonnumeric prerelease identifiers such as `beta.1`; do not weaken the SemVer tag to accommodate MSI.
+- Release only from a `v*` tag whose commit is contained in `origin/main`. The tag must equal `v<version>` from the desktop manifests.
+- Keep ordinary `main` pushes and manual workflow dispatch disabled as release triggers. The author pushes the matching tag only after the release commit is present on `origin/main`.
+- Never republish an already published version. Select a new version instead.
+
+## Release sources
+
+Keep the desktop version identical in the manifests and their locks:
+
+```text
+cursor-byok/
+├── Cargo.lock
+├── apps/desktop/
+│   ├── package.json
+│   ├── package-lock.json
+│   └── src-tauri/
+│       ├── Cargo.toml
+│       └── tauri.conf.json
+├── scripts/cursor-proto/proto/
+│   ├── agent_v1.proto
+│   └── aiserver_v1.proto
+└── .github/workflows/release.yml
+```
+
+The two listed Proto files are required build inputs and must be committed. Keep the other locally extracted Proto files ignored unless the build starts depending on them.
+
+## Prepare and validate
+
+1. Inspect `git status`, fetch `origin/main`, and preserve unrelated user changes. Confirm the release commit is based on the current remote head.
+2. Choose stable or beta numbering explicitly. Update the desktop version in both manifests and lockfiles; do not change the independent `cursor-server` version merely to release the desktop app.
+3. Confirm the updater public key in `tauri.conf.json` matches `.tauri/cursor-byok.key.pub` without exposing the private key.
+4. Confirm `TAURI_SIGNING_PRIVATE_KEY` exists in GitHub Actions. `TAURI_SIGNING_PRIVATE_KEY_PASSWORD` must be absent when the local key has no password.
+5. Confirm neither the intended tag nor Release already exists.
+6. From `apps/desktop`, run:
+
+   ```bash
+   npm run check
+   npm run tauri:build -- --debug --no-bundle
+   ```
+
+7. Validate the workflow YAML and inspect the staged diff. Ensure `.tauri/`, unrelated local files, and unrelated user changes are not staged.
+8. Use the `tauri-action@v1` input `uploadUpdaterJson: true`; `includeUpdaterJson` is not a valid v1 input.
+
+## Publish and verify
+
+After the author explicitly authorizes publication:
+
+1. Commit only the reviewed release set and push it to `main`. Confirm the release commit is present in `origin/main`; this push must not start the release workflow.
+2. Create the matching tag on that commit, for example `v0.1.0-beta.1`, and push only that tag. This tag push is the publication trigger.
+3. Follow the triggered `Release desktop app` run through completion. Report the run URL and stop on failure; diagnose locally before asking the author to authorize another live attempt.
+4. Verify `v<version>` exists, is published rather than draft, has `prerelease: false`, and is the repository's Latest release.
+5. Verify the Release contains signed Tauri updater artifacts plus `latest.json`, and the legacy platform archives plus `update.json`.
+6. For a beta, report clearly that it is a test version even though GitHub represents it as a normal Latest Release.

.agents/skills/release/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Desktop Release"
+  short_description: "Prepare, authorize, publish, and verify desktop releases"
+  default_prompt: "Use $release to prepare and verify a Cursor BYOK desktop release."

.github/workflows/release.yml

@@ -2,11 +2,8 @@ name: Release desktop app
 
 on:
   push:
-    branches:
-      - main
     tags:
       - "v*"
-  workflow_dispatch:
 
 permissions:
   contents: write
@@ -24,6 +21,28 @@ jobs:
       version: ${{ steps.version.outputs.version }}
     steps:
       - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Verify release author
+        env:
+          REPOSITORY_OWNER: ${{ github.repository_owner }}
+        shell: bash
+        run: |
+          if [[ "${GITHUB_ACTOR}" != "${REPOSITORY_OWNER}" ]]; then
+            echo "Only ${REPOSITORY_OWNER} may publish a release" >&2
+            exit 1
+          fi
+
+      - name: Verify updater signing key
+        env:
+          TAURI_SIGNING_PRIVATE_KEY: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY }}
+        shell: bash
+        run: |
+          if [[ -z "${TAURI_SIGNING_PRIVATE_KEY}" ]]; then
+            echo "TAURI_SIGNING_PRIVATE_KEY is not configured" >&2
+            exit 1
+          fi
 
       - name: Read and verify app version
         id: version
@@ -34,8 +53,12 @@ jobs:
           cargo_version=$(sed -n '/^version = / { s/version = "\([^"]*\)"/\1/p; q; }' apps/desktop/src-tauri/Cargo.toml)
           test "${version}" = "${package_version}"
           test "${version}" = "${cargo_version}"
-          if [[ "${GITHUB_REF_TYPE}" = "tag" ]]; then
-            test "${GITHUB_REF_NAME}" = "v${version}"
+          test "${GITHUB_REF_TYPE}" = "tag"
+          test "${GITHUB_REF_NAME}" = "v${version}"
+          git fetch origin main:refs/remotes/origin/main
+          if ! git merge-base --is-ancestor "${GITHUB_SHA}" origin/main; then
+            echo "Release tag must point to a commit contained in origin/main" >&2
+            exit 1
           fi
           echo "version=${version}" >> "${GITHUB_OUTPUT}"
 
@@ -68,7 +91,7 @@ jobs:
             target: ""
           - platform: windows-x86_64
             os: windows-latest
-            args: ""
+            args: "--bundles nsis"
             target: ""
           - platform: macos-aarch64
             os: macos-15
@@ -120,10 +143,10 @@ jobs:
           projectPath: apps/desktop
           tagName: v__VERSION__
           releaseName: Cursor BYOK v__VERSION__
-          releaseBody: Download the installer for your platform from the assets below.
+          releaseBody: ${{ contains(needs.prepare.outputs.version, '-') && 'Beta release. Download the installer for your platform from the assets below.' || 'Download the installer for your platform from the assets below.' }}
           releaseDraft: true
           prerelease: false
-          includeUpdaterJson: true
+          uploadUpdaterJson: true
           args: ${{ matrix.args }}
 
       - name: Package legacy Linux updater asset

.gitignore

@@ -30,4 +30,6 @@ server-go/log/
 .cursor-app-formatted/
 proto/extensions-cursor-app/
 cursor-tab-server/cursor-tab-server-linux-amd64.tar
-/scripts/cursor-proto/proto/
+/scripts/cursor-proto/proto/*
+!/scripts/cursor-proto/proto/agent_v1.proto
+!/scripts/cursor-proto/proto/aiserver_v1.proto

Cargo.lock

@@ -1128,7 +1128,7 @@ checksum = "52560adf09603e58c9a7ee1fe1dcb95a16927b17c127f0ac02d6e768a0e25bc1"
 
 [[package]]
 name = "cursor-byok-desktop"
-version = "0.1.0"
+version = "0.1.0-beta.1"
 dependencies = [
  "axum",
  "cursor-server",

apps/desktop/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "cursor-byok-desktop",
-  "version": "0.1.0",
+  "version": "0.1.0-beta.1",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "cursor-byok-desktop",
-      "version": "0.1.0",
+      "version": "0.1.0-beta.1",
       "license": "MIT",
       "dependencies": {
         "@floating-ui/dom": "^1.8.0",

apps/desktop/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cursor-byok-desktop",
-  "version": "0.1.0",
+  "version": "0.1.0-beta.1",
   "description": "Cursor BYOK desktop management application",
   "type": "module",
   "scripts": {

apps/desktop/src-tauri/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "cursor-byok-desktop"
-version = "0.1.0"
+version = "0.1.0-beta.1"
 edition = "2021"
 publish = false
 

apps/desktop/src-tauri/tauri.conf.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://schema.tauri.app/config/2",
   "productName": "Cursor BYOK",
-  "version": "0.1.0",
+  "version": "0.1.0-beta.1",
   "identifier": "dev.cursorbyok.desktop",
   "build": {
     "beforeDevCommand": "npm run dev",

apps/desktop/src/components/charts/DailyTokenUsageChart.tsx

@@ -77,9 +77,10 @@ export function DailyTokenUsageChart({
   const maximumTotal = data.reduce((maximum, day) => Math.max(maximum, totalTokens(day)), 0);
   const axisMaximum = Math.max(1, maximumTotal / DATA_HEIGHT_RATIO);
   const emptyBarHeight = axisMaximum * EMPTY_BAR_RATIO;
-  const averageLevel = data.length === 0
+  const nonZeroTotals = data.map(totalTokens).filter((total) => total !== 0);
+  const averageLevel = nonZeroTotals.length === 0
     ? 0
-    : data.reduce((sum, day) => sum + totalTokens(day), 0) / data.length;
+    : nonZeroTotals.reduce((sum, total) => sum + total, 0) / nonZeroTotals.length;
 
   const option = useMemo<EChartsCoreOption>(() => ({
     animationDuration: 450,

server/prompt/cursor/agent/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
 You are now in Agent mode. You have EXITED your previous mode. Continue with the task in the new mode.
 </system_reminder>
 <system_reminder>

server/prompt/cursor/ask/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
 You are now in Ask mode. You have EXITED your previous mode. Continue with the task in the new mode.
 </system_reminder>
 

server/prompt/cursor/compaction/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<timestamp>{{TIMESTAMP}}</timestamp>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<timestamp>{{TIMESTAMP}}</timestamp>
 <user_query>
 {{USER_QUERY}}
 </user_query>

server/prompt/cursor/debug/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
 You are now in Debug mode. You have EXITED your previous mode. Continue with the task in the new mode.
 </system_reminder>
 

server/prompt/cursor/multitask/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
 You are now in Multitask mode. You have EXITED your previous mode. Continue with the task in the new mode.
 </system_reminder>
 

server/prompt/cursor/plan/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
 You are now in Plan mode. You have EXITED your previous mode. Continue with the task in the new mode.
 </system_reminder>
 

server/prompt/cursor/subagent/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
 You are currently working inside a Task subagent. Your parent agent has delegated a clearly bounded assignment to you. Complete that assignment directly with the tools available in this session. The Task tool is unavailable inside subagents, so delegation cannot be nested.
 </system_reminder>
 <timestamp>{{TIMESTAMP}}</timestamp>

server/src/cursor/projection/decode.rs

@@ -30,17 +30,15 @@ pub fn decode(data: &[u8], internal_id: String) -> Result<CanonicalMessage> {
         .and_then(Value::as_str)
         .unwrap_or_default()
         .to_string();
+    let is_request_context = role == Role::User && wire_id.starts_with("request-context:");
+    let is_prompt_context =
+        is_request_context || role == Role::User && wire_id.starts_with("selected-context:");
     let origin = match role {
         Role::System => Origin::Prompt,
         Role::Assistant => Origin::Assistant,
         Role::Tool => Origin::Tool,
         Role::User if wire_id.starts_with("runtime:") => Origin::Runtime,
-        Role::User
-            if wire_id.starts_with("request-context:")
-                || wire_id.starts_with("selected-context:") =>
-        {
-            Origin::Prompt
-        }
+        Role::User if is_prompt_context => Origin::Prompt,
         Role::User => Origin::User,
     };
     let runtime_event_id = wire_id.strip_prefix("runtime:").map(str::to_string);
@@ -49,7 +47,7 @@ pub fn decode(data: &[u8], internal_id: String) -> Result<CanonicalMessage> {
         Role::Tool => MessageContent::ToolResult(decode_tool_result(&value)?),
         _ => decode_text(&value)?,
     };
-    let message_id = if runtime_event_id.is_some() {
+    let message_id = if runtime_event_id.is_some() || is_request_context {
         wire_id
     } else {
         internal_id

server/src/cursor/projection/tests.rs

@@ -131,6 +131,23 @@ fn runtime_wire_identity_survives_checkpoint_hydration() {
     );
 }
 
+#[test]
+fn request_context_identity_survives_checkpoint_hydration() {
+    let wire = json!({
+        "role": "user",
+        "id": "request-context:digest",
+        "content": "<rules>current rules</rules>",
+    });
+    let message = decode(
+        serde_json::to_vec(&wire).unwrap().as_slice(),
+        "cursor-root:blob-id:20".into(),
+    )
+    .unwrap();
+
+    assert_eq!(message.message_id, "request-context:digest");
+    assert_eq!(message.origin, crate::model::Origin::Prompt);
+}
+
 #[test]
 fn cursor_user_image_uses_image_field() {
     let wire = json!({

server/src/cursor/prompting/assets.rs

@@ -134,7 +134,6 @@ impl PromptAssets {
 }
 
 const RUNTIME_VARIABLES: &[&str] = &[
-    "REQUEST_CONTEXT",
     "OPEN_FILES",
     "SELECTED_CONTEXT",
     "ACTION_CONTEXT",

server/src/cursor/request/prepare.rs

@@ -192,6 +192,20 @@ pub(crate) async fn prepare(
         }
         _ => None,
     };
+    let request_context_message = match event_id.as_deref() {
+        Some(event_id) if !compacting && !background_completion => {
+            let message_id = format!("request-context:{event_id}");
+            match store.message(&conversation_id, &message_id).await? {
+                Some(message) => Some(message),
+                None => runtime::compile_request_context(
+                    event_id,
+                    &request_context,
+                    base_messages.as_deref().unwrap_or_default(),
+                )?,
+            }
+        }
+        _ => None,
+    };
     let initial_messages = if compacting {
         Vec::new()
     } else {
@@ -209,21 +223,27 @@ pub(crate) async fn prepare(
                 turn_user = Some(user);
                 vec![message]
             }
-            (Some(user), Some(event_id)) => match existing_runtime {
-                Some(message) => vec![message],
-                None => vec![
-                    runtime::compile(
-                        event_id,
-                        checkpoint_mode,
-                        &user,
-                        &request_context,
-                        &action_context,
-                        compiler,
-                        blob_sync,
-                    )
-                    .await?,
-                ],
-            },
+            (Some(user), Some(event_id)) => {
+                let runtime = match existing_runtime {
+                    Some(message) => message,
+                    None => {
+                        runtime::compile(
+                            event_id,
+                            checkpoint_mode,
+                            &user,
+                            &request_context,
+                            &action_context,
+                            compiler,
+                            blob_sync,
+                        )
+                        .await?
+                    }
+                };
+                request_context_message
+                    .into_iter()
+                    .chain(std::iter::once(runtime))
+                    .collect()
+            }
             (None, None) => Vec::new(),
             _ => {
                 return Err(Error::Protocol(

server/src/cursor/request/runtime.rs

@@ -87,10 +87,6 @@ pub async fn compile(
             .map(|env| env.time_zone.as_str()),
     )?;
     let mut values = BTreeMap::from([
-        (
-            "REQUEST_CONTEXT",
-            section(context::compile_context(request_context, &time.today)),
-        ),
         ("OPEN_FILES", section(open_files(user))),
         (
             "SELECTED_CONTEXT",
@@ -122,6 +118,41 @@ pub async fn compile(
     .await
 }
 
+pub(super) fn compile_request_context(
+    event_id: &str,
+    request_context: &pb::RequestContext,
+    history: &[CanonicalMessage],
+) -> Result<Option<CanonicalMessage>> {
+    let time = Time::now(
+        request_context
+            .env
+            .as_ref()
+            .map(|env| env.time_zone.as_str()),
+    )?;
+    let text = context::compile_context(request_context, &time.today);
+    if text.is_empty() {
+        return Ok(None);
+    }
+    let message = CanonicalMessage::text(
+        format!("request-context:{event_id}"),
+        Role::User,
+        Origin::Prompt,
+        text,
+    );
+    Ok(should_project_request_context(history, &message).then_some(message))
+}
+
+fn should_project_request_context(
+    history: &[CanonicalMessage],
+    current: &CanonicalMessage,
+) -> bool {
+    history
+        .iter()
+        .rev()
+        .find(|message| message.message_id.starts_with("request-context:"))
+        .is_none_or(|previous| previous.content != current.content)
+}
+
 pub async fn compile_background(
     event_id: String,
     user: &pb::UserMessage,
@@ -252,3 +283,55 @@ impl Time {
         })
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn request_context_is_only_projected_when_its_content_changes() {
+        let first = CanonicalMessage::text(
+            "request-context:first",
+            Role::User,
+            Origin::Prompt,
+            "<rules>same</rules>",
+        );
+        let duplicate = CanonicalMessage::text(
+            "request-context:second",
+            Role::User,
+            Origin::Prompt,
+            "<rules>same</rules>",
+        );
+        let changed = CanonicalMessage::text(
+            "request-context:third",
+            Role::User,
+            Origin::Prompt,
+            "<rules>changed</rules>",
+        );
+        let runtime = CanonicalMessage::text(
+            "runtime:turn",
+            Role::User,
+            Origin::Runtime,
+            "<user_query>next</user_query>",
+        );
+
+        assert!(should_project_request_context(&[], &first));
+        assert!(!should_project_request_context(
+            &[first.clone(), runtime.clone()],
+            &duplicate
+        ));
+        assert!(should_project_request_context(
+            &[first.clone(), runtime.clone()],
+            &changed
+        ));
+        assert!(should_project_request_context(
+            &[first.clone(), changed, runtime],
+            &CanonicalMessage::text(
+                "request-context:fourth",
+                Role::User,
+                Origin::Prompt,
+                "<rules>same</rules>",
+            )
+        ));
+    }
+}

server/src/run/engine.rs

@@ -460,11 +460,8 @@ impl RunEngine {
             .iter()
             .map(|message| message.message_id.as_str())
             .collect::<HashSet<_>>();
-        let compactable = messages
-            .iter()
-            .filter(|message| !current_ids.contains(message.message_id.as_str()))
-            .cloned()
-            .collect::<Vec<_>>();
+        let (compactable, retained_request_context) =
+            auto_compaction_partition(messages, &current_ids);
         if compactable.is_empty() {
             return Ok((revision, None));
         }
@@ -533,7 +530,8 @@ impl RunEngine {
             },
             runtime_event_id: Some(event_id),
         };
-        let mut replacement = vec![summary_message];
+        let mut replacement = retained_request_context.into_iter().collect::<Vec<_>>();
+        replacement.push(summary_message);
         replacement.extend(prepared.initial_messages.iter().cloned());
         let revision = self
             .store
@@ -565,6 +563,29 @@ impl RunEngine {
     }
 }
 
+fn auto_compaction_partition(
+    messages: &[CanonicalMessage],
+    current_ids: &HashSet<&str>,
+) -> (Vec<CanonicalMessage>, Option<CanonicalMessage>) {
+    let latest_request_context = messages
+        .iter()
+        .rposition(|message| message.message_id.starts_with("request-context:"));
+    let compactable = messages
+        .iter()
+        .enumerate()
+        .filter(|(index, message)| {
+            Some(*index) != latest_request_context
+                && !current_ids.contains(message.message_id.as_str())
+        })
+        .map(|(_, message)| message.clone())
+        .collect();
+    let retained = latest_request_context
+        .and_then(|index| messages.get(index))
+        .filter(|message| !current_ids.contains(message.message_id.as_str()))
+        .cloned();
+    (compactable, retained)
+}
+
 fn should_auto_compact(prepared: &PreparedRun, messages: &[CanonicalMessage]) -> bool {
     if prepared.action != RunAction::Start {
         return false;
@@ -724,7 +745,7 @@ fn failure_message(failure: &RunFailure) -> String {
 
 #[cfg(test)]
 mod tests {
-    use super::{estimate_context_tokens, hydrate_tool_images};
+    use super::{auto_compaction_partition, estimate_context_tokens, hydrate_tool_images};
     use crate::{
         model::{
             CanonicalMessage, ContentPart, Origin, ProjectedContent, ProjectedMessage, PromptSpec,
@@ -732,6 +753,7 @@ mod tests {
         },
         store::Store,
     };
+    use std::collections::HashSet;
 
     #[test]
     fn context_estimate_grows_with_prompt_history() {
@@ -756,6 +778,42 @@ mod tests {
         assert!(estimate_context_tokens(&prompt, &long) > estimate_context_tokens(&prompt, &short));
     }
 
+    #[test]
+    fn auto_compaction_preserves_only_the_latest_request_context() {
+        let first_context = CanonicalMessage::text(
+            "request-context:first",
+            Role::User,
+            Origin::Prompt,
+            "old rules",
+        );
+        let old_runtime =
+            CanonicalMessage::text("runtime:first", Role::User, Origin::Runtime, "old query");
+        let latest_context = CanonicalMessage::text(
+            "request-context:second",
+            Role::User,
+            Origin::Prompt,
+            "new rules",
+        );
+        let current_runtime = CanonicalMessage::text(
+            "runtime:current",
+            Role::User,
+            Origin::Runtime,
+            "current query",
+        );
+        let messages = vec![
+            first_context.clone(),
+            old_runtime.clone(),
+            latest_context.clone(),
+            current_runtime,
+        ];
+        let current_ids = HashSet::from(["runtime:current"]);
+
+        let (compactable, retained) = auto_compaction_partition(&messages, &current_ids);
+
+        assert_eq!(compactable, vec![first_context, old_runtime]);
+        assert_eq!(retained, Some(latest_context));
+    }
+
     #[tokio::test]
     async fn read_image_is_loaded_only_for_the_provider_projection() {
         let directory = tempfile::tempdir().unwrap();

server/tests/prefix_stability.rs

@@ -323,7 +323,6 @@ fn every_captured_mode_owns_and_renders_its_runtime_template() {
         .unwrap(),
     );
     let values = BTreeMap::from([
-        ("REQUEST_CONTEXT", String::new()),
         ("OPEN_FILES", String::new()),
         ("SELECTED_CONTEXT", String::new()),
         ("ACTION_CONTEXT", String::new()),

server/tests/runtime_modes.rs

@@ -19,9 +19,9 @@ use cursor_server::{
 use prost::Message;
 
 #[tokio::test]
-async fn current_mode_and_referenced_context_are_consumed_by_one_runtime_message() {
+async fn unchanged_request_context_is_not_repeated_and_preserves_the_provider_prefix() {
     let (_directory, store) = fixtures::temp_store().await;
-    let references = references(&store).await;
+    let first_references = references(&store).await;
     let provider = fake_provider::FakeProvider::default();
     provider.push(vec![
         ModelEvent::Start {
@@ -32,6 +32,15 @@ async fn current_mode_and_referenced_context_are_consumed_by_one_runtime_message
         ModelEvent::TextEnd,
         ModelEvent::Done(FinishReason::Stop),
     ]);
+    provider.push(vec![
+        ModelEvent::Start {
+            model_call_id: "model-2".into(),
+        },
+        ModelEvent::TextStart,
+        ModelEvent::TextDelta("answer again".into()),
+        ModelEvent::TextEnd,
+        ModelEvent::Done(FinishReason::Stop),
+    ]);
     let assets = PromptAssets::load(
         std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
             .join("prompt/cursor")
@@ -39,7 +48,7 @@ async fn current_mode_and_referenced_context_are_consumed_by_one_runtime_message
     )
     .unwrap();
     let registry = CursorSessionRegistry::new(
-        store,
+        store.clone(),
         Arc::new(provider.clone()),
         PromptCompiler::new(assets),
         Default::default(),
@@ -49,12 +58,13 @@ async fn current_mode_and_referenced_context_are_consumed_by_one_runtime_message
     handle
         .command(CursorCommand::Append {
             seqno: 0,
-            message: Box::new(run_request(references)),
+            message: Box::new(run_request(first_references)),
         })
         .await
         .unwrap();
 
     let mut seqno = 1;
+    let mut checkpoint = None;
     loop {
         let frame = tokio::time::timeout(std::time::Duration::from_secs(5), output.recv())
             .await
@@ -65,15 +75,21 @@ async fn current_mode_and_referenced_context_are_consumed_by_one_runtime_message
             break;
         }
         let message = pb::AgentServerMessage::decode(payload).unwrap();
-        if let Some(pb::agent_server_message::Message::KvServerMessage(kv)) = message.message {
-            handle
-                .command(CursorCommand::Append {
-                    seqno,
-                    message: Box::new(kv_ack(kv.id)),
-                })
-                .await
-                .unwrap();
-            seqno += 1;
+        match message.message {
+            Some(pb::agent_server_message::Message::KvServerMessage(kv)) => {
+                handle
+                    .command(CursorCommand::Append {
+                        seqno,
+                        message: Box::new(kv_ack(kv.id)),
+                    })
+                    .await
+                    .unwrap();
+                seqno += 1;
+            }
+            Some(pb::agent_server_message::Message::ConversationCheckpointUpdate(state)) => {
+                checkpoint = Some(state);
+            }
+            _ => {}
         }
     }
 
@@ -89,12 +105,23 @@ async fn current_mode_and_referenced_context_are_consumed_by_one_runtime_message
         .tools
         .iter()
         .any(|tool| tool.name == "GenerateImage"));
-    assert_eq!(request.history.len(), 1);
+    assert_eq!(request.history.len(), 2);
+    assert!(request.history[0]
+        .message_id
+        .starts_with("request-context:"));
+    let ProjectedContent::Parts(context_parts) = &request.history[0].content else {
+        panic!("request context message must use typed parts")
+    };
+    let [ContentPart::Text { text: context_text }] = context_parts.as_slice() else {
+        panic!("request context message must contain one text part")
+    };
     assert_eq!(
-        request.history[0].message_id,
+        request.history[1].message_id,
         "runtime:run-request:ask-request"
     );
-    let ProjectedContent::Parts(parts) = &request.history[0].content else {
+    assert!(!request.prompt.instructions.contains("workspace rule"));
+    assert!(!request.prompt.instructions.contains("<mcp_meta_tools>"));
+    let ProjectedContent::Parts(parts) = &request.history[1].content else {
         panic!("runtime message must use typed parts")
     };
     let [ContentPart::Text { text }] = parts.as_slice() else {
@@ -109,6 +136,15 @@ async fn current_mode_and_referenced_context_are_consumed_by_one_runtime_message
         "<definition_path>/tmp/mcp-test/lookup.json</definition_path>",
         "<input_schema>{&quot;properties&quot;:{&quot;query&quot;:{&quot;type&quot;:&quot;string&quot;}},&quot;type&quot;:&quot;object&quot;}</input_schema>",
         "Call a listed tool directly with CallMcpTool without calling GetMcpTools first.",
+    ] {
+        assert!(
+            context_text.contains(expected),
+            "missing request context section: {expected}"
+        );
+    }
+    assert!(!context_text.contains("complete skill body"));
+    assert!(!context_text.contains("complete MCP server instructions"));
+    for expected in [
         "Ask mode is active.",
         "<user_query>\nexplain this\n</user_query>",
     ] {
@@ -117,9 +153,65 @@ async fn current_mode_and_referenced_context_are_consumed_by_one_runtime_message
             "missing runtime section: {expected}"
         );
     }
-    assert!(!text.contains("complete skill body"));
-    assert!(!text.contains("complete MCP server instructions"));
+    assert!(!text.contains("<rules>"));
+    assert!(!text.contains("<mcp_meta_tools>"));
     assert!(text.contains("/workspace/src/main.rs"));
+
+    let second = registry.get_or_create("ask-request-2").await.unwrap();
+    let mut second_output = second.subscribe();
+    second
+        .command(CursorCommand::Append {
+            seqno: 0,
+            message: Box::new(run_request_with_state(
+                references(&store).await,
+                checkpoint.expect("first Run must publish a checkpoint"),
+            )),
+        })
+        .await
+        .unwrap();
+    let mut second_seqno = 1;
+    loop {
+        let frame = tokio::time::timeout(std::time::Duration::from_secs(5), second_output.recv())
+            .await
+            .unwrap()
+            .unwrap();
+        let (flags, payload) = connect::decode_frames(&frame).unwrap().pop().unwrap();
+        if flags & connect::END_STREAM_FLAG != 0 {
+            break;
+        }
+        let message = pb::AgentServerMessage::decode(payload).unwrap();
+        if let Some(pb::agent_server_message::Message::KvServerMessage(kv)) = message.message {
+            second
+                .command(CursorCommand::Append {
+                    seqno: second_seqno,
+                    message: Box::new(kv_ack(kv.id)),
+                })
+                .await
+                .unwrap();
+            second_seqno += 1;
+        }
+    }
+
+    let requests = provider.requests();
+    assert_eq!(requests.len(), 2);
+    assert_eq!(
+        requests[1].prompt.instructions, requests[0].prompt.instructions,
+        "unchanged request context must not rewrite the system prompt"
+    );
+    assert_eq!(
+        requests[1].history[..requests[0].history.len()],
+        requests[0].history,
+        "the previous provider history must remain an exact prefix"
+    );
+    assert_eq!(
+        requests[1]
+            .history
+            .iter()
+            .filter(|message| message.message_id.starts_with("request-context:"))
+            .count(),
+        1,
+        "identical request context must not be appended again"
+    );
 }
 
 #[tokio::test]
@@ -252,10 +344,10 @@ async fn missing_context_parts_use_current_cursor_response_and_cache_its_content
     let requests = provider.requests();
     assert_eq!(requests.len(), 1);
     let ProjectedContent::Parts(parts) = &requests[0].history[0].content else {
-        panic!("runtime message must use typed parts")
+        panic!("request context message must use typed parts")
     };
     let [ContentPart::Text { text }] = parts.as_slice() else {
-        panic!("this fixture has no images")
+        panic!("request context message must contain one text part")
     };
     assert!(text.contains("<mcp_meta_tool_server name=\"live-mcp\" identifier=\"live-mcp\">"));
     assert!(text.contains("<mcp_tool name=\"current-tool\">"));
@@ -449,6 +541,31 @@ fn run_request(references: References) -> pb::AgentClientMessage {
     }
 }
 
+fn run_request_with_state(
+    references: References,
+    state: pb::ConversationStateStructure,
+) -> pb::AgentClientMessage {
+    let mut message = run_request(references);
+    let Some(pb::agent_client_message::Message::RunRequest(request)) = message.message.as_mut()
+    else {
+        unreachable!("run_request always returns a RunRequest")
+    };
+    request.conversation_state = Some(state);
+    let Some(pb::conversation_action::Action::UserMessageAction(action)) = request
+        .action
+        .as_mut()
+        .and_then(|action| action.action.as_mut())
+    else {
+        unreachable!("run_request always contains a UserMessageAction")
+    };
+    action
+        .user_message
+        .as_mut()
+        .expect("run_request always contains a UserMessage")
+        .message_id = "wire-user-2".into();
+    message
+}
+
 fn kv_ack(id: u32) -> pb::AgentClientMessage {
     pb::AgentClientMessage {
         message: Some(pb::agent_client_message::Message::KvClientMessage(

server/tests/text_turn.rs

@@ -259,18 +259,27 @@ async fn text_turn_runs_from_bidi_request_through_checkpoint_and_end_stream() {
         .contains("powered by Test Model"));
     let projected = &requests[0].history;
     assert_eq!(projected[0].role, Role::User);
-    let ProjectedContent::Parts(runtime) = &projected[0].content else {
-        panic!("runtime context must be text")
+    assert!(projected[0].message_id.starts_with("request-context:"));
+    let ProjectedContent::Parts(context) = &projected[0].content else {
+        panic!("request context must be text")
+    };
+    assert!(matches!(
+        context.as_slice(),
+        [cursor_server::model::ContentPart::Text { text }]
+            if text.contains("<user_info>")
+    ));
+    let ProjectedContent::Parts(runtime) = &projected[1].content else {
+        panic!("runtime user message must be text")
     };
     assert!(matches!(
         runtime.as_slice(),
         [cursor_server::model::ContentPart::Text { text }]
-            if text.contains("<user_info>")
-                && text.contains("<user_query>\nhello\n</user_query>")
+            if text.contains("<user_query>\nhello\n</user_query>")
+                && !text.contains("<user_info>")
     ));
     assert_eq!(
         projected.len(),
-        1,
+        2,
         "the raw UserMessage is not projected twice"
     );
 
@@ -278,9 +287,15 @@ async fn text_turn_runs_from_bidi_request_through_checkpoint_and_end_stream() {
         .load_current_messages(&cursor_server::model::ConversationId::new("conversation"))
         .await
         .unwrap();
-    assert_eq!(messages[0].message_id, "runtime:run-request:request");
+    assert!(messages[0].message_id.starts_with("request-context:"));
     assert_eq!(messages[0].role, Role::User);
-    assert_eq!(messages.len(), 2, "runtime user plus final assistant");
+    assert_eq!(messages[1].message_id, "runtime:run-request:request");
+    assert_eq!(messages[1].role, Role::User);
+    assert_eq!(
+        messages.len(),
+        3,
+        "request context plus runtime user and final assistant"
+    );
     let stored_runs: Vec<String> = sqlx::query_scalar("SELECT run_id FROM runs ORDER BY run_id")
         .fetch_all(store.pool())
         .await