fix: resolve open Linear regressions

README.md

@@ -116,6 +116,8 @@ from their close control, with Delete/Backspace while focused, or by middle-clic
   without changing local refs or `FETCH_HEAD`. Active PRs can be closed from a
   confirmed overflow action. Closed PRs can be reopened;
   merged PRs and every terminal discussion/thread surface remain read-only.
+  Completed Azure PR Code views remain available after the provider deletes
+  the source branch by using its recorded target and merge commits.
   GitHub and Azure DevOps Services authentication stays in the signed-in `gh` /
   `az` CLI. Azure DevOps Server 2020+ is available through an optional,
   independently updated, signed `strand-azdo` REST helper configured in Settings → Hosting;
@@ -218,8 +220,9 @@ from their close control, with Delete/Backspace while focused, or by middle-clic
   Generation is cancellable, scans conservative sensitive-file
   signals before provider launch, keeps the generated description content-sized,
   and can retry explicitly with the other provider without changing your
-  default. Repository-family writing profiles keep terminology and style
-  consistent across worktrees.
+  default. Provider failures are reduced to concise, actionable hints; raw CLI
+  session, prompt, and patch transcripts are never displayed. Repository-family
+  writing profiles keep terminology and style consistent across worktrees.
 - **Fast by design** — reads go through [gix](https://github.com/GitoxideLabs/gitoxide),
   writes through git2 and your system `git`. Performance targets live in
   [`PRD.md`](./PRD.md) §8 and are measured in

ROADMAP.md

@@ -2295,6 +2295,14 @@ only `v1.1.0`. Universal macOS notarization returned `Accepted` (request
 Authenticode `NotSigned`, and the Git tag has no cryptographic signature, so
 the external publisher and remaining release-checklist gates stay open.
 
+**Linear bug sweep completed (2026-07-24):** DAN-24/25/26 are resolved.
+Completed Azure PR Code views now use durable target/merge history, so
+source-branch cleanup cannot blank the patch. AI vendor failures are reduced to
+bounded, actionable messages and transcript-shaped output never exposes echoed
+prompt or patch content. A real deleted-branch Git integration test covers the
+Azure path, and a live WebView2 pass confirmed Pierre uses its light palette
+when Strand resolves the system theme to light.
+
 ---
 
 ## Cross-cutting tracks (run in parallel with all milestones)

TASKS.md

@@ -1552,7 +1552,11 @@ tree: watch the agent work, review fast, accept or reject safely.
     diff; line totals stay in the selected-file header rather than crowding
     every tree row. The file header exposes the same
     persisted stacked/split controls in context. Azure comparisons
-    fetch source/target objects without updating repository refs or FETCH_HEAD.
+    fetch source/target objects without updating repository refs or FETCH_HEAD;
+    completed Azure PRs reconstruct the historical patch from
+    `lastMergeTargetCommit` → `lastMergeCommit`, so provider source-branch
+    cleanup cannot break Code (`completed_azure_diff_survives_a_deleted_source_branch`,
+    DAN-24).
   - ☑ 1.0 discussion threads and comment creation: Timeline reads GitHub
     issue comments, GitHub review-thread comments, and Azure thread comments
     (including inline file context) as safe
@@ -2105,6 +2109,11 @@ extraction above as prerequisite. **Do not start before 1.0 ships**
 - ☑ Broken vendor-CLI installs stay distinct from signed-out sessions
   (`AiProviderStatus.error`, auth-failure classification, and `--version`
   login preflight prevent false “browser opened” messages)
+- ☑ Vendor CLI failures become bounded, actionable messages instead of raw
+  session transcripts (`user_facing_cli_failure`, DAN-26): known rate-limit,
+  quota/billing, model, context, timeout, and network failures are classified;
+  prompt/patch text after vendor transcript boundaries is never used for
+  classification or shown to the user.
 - ☑ AI provider execution hardened and cancellable (`AiProviderAdapter`,
   canonical CLI resolution, isolated read-only Codex cwd/argv, bounded
   stdout/stderr, process-group/Windows Job Object teardown, and shared

crates/strand-tauri/src/ai/bin.rs

@@ -660,49 +660,45 @@ mod tests {
         assert!(err.contains("timed out"), "unexpected error: {err}");
     }
 
-    #[cfg(not(windows))]
-    #[test]
-    fn run_capture_cancels_process_group() {
+    fn assert_process_group_cancelled_promptly(
+        program: &'static str,
+        args: &'static [&'static str],
+    ) {
         let cancel = AiCancelHandle::new();
-        let trigger = cancel.clone();
-        std::thread::spawn(move || {
-            std::thread::sleep(Duration::from_millis(100));
-            trigger.cancel();
+        let worker_cancel = cancel.clone();
+        let (started_tx, started_rx) = std::sync::mpsc::sync_channel(0);
+        let worker = std::thread::spawn(move || {
+            started_tx.send(()).unwrap();
+            run_capture_cancellable(
+                Path::new(program),
+                args,
+                None,
+                None,
+                SUGGEST_TIMEOUT,
+                Some(&worker_cancel),
+            )
         });
-        let started = Instant::now();
-        let err = run_capture_cancellable(
-            Path::new("/bin/sh"),
-            &["-c", "sleep 30 & wait"],
-            None,
-            None,
-            SUGGEST_TIMEOUT,
-            Some(&cancel),
-        )
-        .unwrap_err();
+        started_rx.recv().unwrap();
+        std::thread::sleep(Duration::from_millis(100));
+        let cancelled_at = Instant::now();
+        cancel.cancel();
+        let err = worker.join().unwrap().unwrap_err();
         assert_eq!(err, "cancelled");
-        assert!(started.elapsed() < Duration::from_secs(3));
+        assert!(cancelled_at.elapsed() < Duration::from_secs(3));
+    }
+
+    #[cfg(not(windows))]
+    #[test]
+    fn run_capture_cancels_process_group() {
+        assert_process_group_cancelled_promptly("/bin/sh", &["-c", "sleep 30 & wait"]);
     }
 
     #[cfg(windows)]
     #[test]
     fn run_capture_cancels_process_group() {
-        let cancel = AiCancelHandle::new();
-        let trigger = cancel.clone();
-        std::thread::spawn(move || {
-            std::thread::sleep(Duration::from_millis(100));
-            trigger.cancel();
-        });
-        let started = Instant::now();
-        let err = run_capture_cancellable(
-            Path::new("cmd.exe"),
+        assert_process_group_cancelled_promptly(
+            "cmd.exe",
             &["/C", "ping", "-n", "30", "127.0.0.1"],
-            None,
-            None,
-            SUGGEST_TIMEOUT,
-            Some(&cancel),
-        )
-        .unwrap_err();
-        assert_eq!(err, "cancelled");
-        assert!(started.elapsed() < Duration::from_secs(3));
+        );
     }
 }

crates/strand-tauri/src/ai/mod.rs

@@ -72,15 +72,84 @@ pub(crate) fn map_cli_failure(
     provider_label: &str,
     err: String,
 ) -> String {
+    let diagnostic = cli_diagnostic_prefix(&err);
     if let Some(health_error) = &status.error {
         health_error.clone()
-    } else if !status.logged_in || is_auth_failure(&err) {
+    } else if !status.logged_in || is_auth_failure(diagnostic) {
         format!("{AI_AUTH_REQUIRED} {provider_label} is not signed in.")
     } else {
-        err
+        user_facing_cli_failure(provider_label, diagnostic, diagnostic.len() == err.len())
     }
 }
 
+fn cli_diagnostic_prefix(err: &str) -> &str {
+    ["<untrusted-", "\nuser\n", "\nUser\n"]
+        .iter()
+        .filter_map(|marker| err.find(marker))
+        .min()
+        .map_or(err, |end| &err[..end])
+}
+
+fn user_facing_cli_failure(provider_label: &str, err: &str, allow_detail: bool) -> String {
+    let lower = err.to_ascii_lowercase();
+    if lower.contains("rate limit") || lower.contains("too many requests") || lower.contains("429") {
+        return format!("{provider_label} is rate-limited right now. Wait a moment, then try again.");
+    }
+    if lower.contains("quota")
+        || lower.contains("billing")
+        || lower.contains("insufficient credits")
+    {
+        return format!(
+            "{provider_label} could not generate a suggestion because the provider quota or billing limit was reached."
+        );
+    }
+    if (lower.contains("model") && lower.contains("not found"))
+        || lower.contains("unsupported model")
+        || lower.contains("invalid model")
+    {
+        return format!(
+            "{provider_label} does not support the selected model. Choose another model in Settings → AI and try again."
+        );
+    }
+    if lower.contains("context length")
+        || lower.contains("context window")
+        || lower.contains("maximum context")
+        || lower.contains("too many tokens")
+    {
+        return format!(
+            "{provider_label} could not fit these changes in its context window. Stage or select fewer changes, then try again."
+        );
+    }
+    if lower.contains("timed out") {
+        return format!("{provider_label} took too long to respond. Check your connection and try again.");
+    }
+    if lower.contains("network")
+        || lower.contains("connection reset")
+        || lower.contains("connection refused")
+        || lower.contains("could not resolve host")
+        || lower.contains("dns")
+    {
+        return format!("{provider_label} could not be reached. Check your connection and try again.");
+    }
+
+    // Vendor CLIs can emit their full session preamble and echoed stdin on a
+    // failed run. That stdin contains repository paths and patches, so never
+    // surface or persist an unclassified transcript. A short single-line
+    // diagnostic is safe and usually already actionable.
+    let detail = err.trim();
+    if allow_detail
+        && !detail.is_empty()
+        && detail.len() <= 240
+        && !detail.contains(['\n', '\r'])
+        && !detail.contains("<untrusted-")
+    {
+        return format!("{provider_label} could not generate a suggestion: {detail}");
+    }
+    format!(
+        "{provider_label} could not generate a suggestion. Check the selected model, provider status, and connection, then try again."
+    )
+}
+
 #[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
 #[serde(rename_all = "lowercase")]
 pub enum AiProvider {
@@ -277,7 +346,7 @@ mod tests {
     }
 
     #[test]
-    fn map_cli_failure_preserves_error_when_logged_in() {
+    fn map_cli_failure_explains_rate_limits_when_logged_in() {
         let status = AiProviderStatus {
             provider: AiProvider::Anthropic,
             installed: true,
@@ -286,7 +355,10 @@ mod tests {
             error: None,
         };
         let err = map_cli_failure(&status, "Claude Code", "rate limited".into());
-        assert_eq!(err, "rate limited");
+        assert_eq!(
+            err,
+            "Claude Code is rate-limited right now. Wait a moment, then try again."
+        );
     }
 
     #[test]
@@ -316,6 +388,67 @@ mod tests {
         assert!(err.contains("Settings → AI"));
     }
 
+    #[test]
+    fn map_cli_failure_never_exposes_vendor_transcripts_or_prompts() {
+        let status = AiProviderStatus {
+            provider: AiProvider::Openai,
+            installed: true,
+            logged_in: true,
+            account_hint: None,
+            error: None,
+        };
+        let raw = "OpenAI Codex v0.144.3\nworkdir: C:\\repo\nsession id: secret\n\
+                   <untrusted-patches>private source patch</untrusted-patches>";
+        let err = map_cli_failure(&status, "Codex", raw.into());
+        assert_eq!(
+            err,
+            "Codex could not generate a suggestion. Check the selected model, provider status, and connection, then try again."
+        );
+        assert!(!err.contains("C:\\repo"));
+        assert!(!err.contains("private source patch"));
+    }
+
+    #[test]
+    fn map_cli_failure_explains_an_invalid_model_without_echoing_the_transcript() {
+        let status = AiProviderStatus {
+            provider: AiProvider::Openai,
+            installed: true,
+            logged_in: true,
+            account_hint: None,
+            error: None,
+        };
+        let err = map_cli_failure(
+            &status,
+            "Codex",
+            "session preamble\nError: selected model not found\n<untrusted-patches>secret</untrusted-patches>".into(),
+        );
+        assert_eq!(
+            err,
+            "Codex does not support the selected model. Choose another model in Settings → AI and try again."
+        );
+        assert!(!err.contains("secret"));
+    }
+
+    #[test]
+    fn map_cli_failure_does_not_classify_words_echoed_from_the_prompt() {
+        let status = AiProviderStatus {
+            provider: AiProvider::Openai,
+            installed: true,
+            logged_in: true,
+            account_hint: None,
+            error: None,
+        };
+        let err = map_cli_failure(
+            &status,
+            "Codex",
+            "session preamble\n<untrusted-patches>not logged in; invalid model; rate limit</untrusted-patches>".into(),
+        );
+        assert_eq!(
+            err,
+            "Codex could not generate a suggestion. Check the selected model, provider status, and connection, then try again."
+        );
+    }
+
     #[test]
     fn truncates_recent_subjects_on_utf8_boundaries() {
         let subject = "é".repeat(100);