Public

conjfrnk/blackbird

Updated: 8/13/2026

Languages

Swift74.2%Rust20%Shell4.5%C0.7%Metal0.2%Python0.1%Other0.3%
3 Models30 Tasks

A minimal, macOS-native terminal emulator.

Harness

1

Mini-SWE-agent
17 / 30

$3.55

11m29s

2

Mini-SWE-agent
16 / 30

$0.70

5m52s

3

Mini-SWE-agent
15 / 30

$0.07

4m03s

Key Takeaways

  • Muse Spark 1.2 with Mini-SWE-agent resolves 16 of 30 tasks, one behind Claude Opus 5 and one ahead of GPT-5.6 Luna.
  • GPT-5.6 Luna with Mini-SWE-agent resolves 15 of 30 tasks at $0.07 per test, the lowest supplied cost.
  • Claude Opus 5 with Mini-SWE-agent reaches 17 of 30 tasks at $3.55 per test, while Muse Spark 1.2 reaches 16 at $0.70.

Model Comparison

Accuracy

56.67%

Claude Opus 5

53.33%

Muse Spark 1.2

Task outcomes

30 tasks

Both
Claude Opus 5 only
Muse Spark 1.2 only
Neither
Not attempted

Cost / test

$3.55

Claude Opus 5

$0.70

Muse Spark 1.2

Cost distribution

$0.00$4.70$9.39

Latency

11m 29s

Claude Opus 5

5m 52s

Muse Spark 1.2

Latency distribution

0s13m 53s27m 45s

Cost Analysis

Cost / Test vs. Accuracy
ACCURACYCOST

Average Token Use / Test

Token Usage
InputOutputReasoningCache readCache write
Claude Opus 5
3.6M
Muse Spark 1.2
3.4M

Cost is the clearest tradeoff in this comparison. Claude Opus 5 leads at 56.67% for $3.55 per test. Muse Spark 1.2 is the lower-cost option at 53.33% for $0.70 per test.

Latency Analysis

Latency vs. Accuracy
ACCURACYLATENCY

Average Response Time / Test

Response Time
Claude Opus 5
11m 29s
Muse Spark 1.2
5m 52s
GPT-5.6 Luna
4m 3s

Latency separates several models with similarly strong scores. Claude Opus 5 leads at 56.67%, while GPT-5.6 Luna is fastest at 4m 3s with 50.00% accuracy.

Tasks with failures

Models
Claude Opus 5
Muse Spark 1.2
GPT-5.6 Luna

Task detail

076d3a4

Issue statement

When a terminal session's persistent OSC 8 hyperlink URI cache reaches its 1 MiB intern budget, new hyperlink URIs lose attribution without leaving any diagnostic breadcrumb. Record this budget-exhaustion condition when it first occurs so support logs explain why links stopped working. The diagnostic must be emitted at most once per terminal session even if later snapshots continue to encounter new over-budget URIs; normal hyperlink interning and the existing budget behavior must remain unchanged.

View Hidden Tests
diff --git a/core/src/tests.rs b/core/src/tests.rsindex 30660d3..9fb0dc6 100644--- a/core/src/tests.rs+++ b/core/src/tests.rs@@ -10,6 +10,42 @@ fn alacritty_terminal_is_linked() {     let _ = std::mem::size_of::<alacritty_terminal::term::Config>(); } +/// The global OSC 8 URI budget is a recoverable support condition, not an+/// invisible failure: the first over-budget snapshot must latch its diagnostic,+/// and later snapshots must preserve the one-shot state.+#[test]+fn osc8_intern_budget_exhaustion_is_logged_once() {+    unsafe {+        let term = bb_term_new(40, 30, 100);+        assert!(!(*term).osc8_intern_cap_logged);++        let bulk = "a".repeat(4000);+        for i in 0..300u32 {+            let uri = format!("https://example.com/{i:03}-{bulk}");+            let seq = format!("\x1b]8;;{uri}\x1b\\X\x1b]8;;\x1b\\");+            bb_term_input(term, seq.as_ptr(), seq.len());+        }++        let first = bb_term_take_snapshot(term);+        assert!(!first.is_null());+        assert!(+            (*term).osc8_intern_cap_logged,+            "crossing the URI intern budget must record that its diagnostic fired"+        );++        let second = bb_term_take_snapshot(term);+        assert!(!second.is_null());+        assert!(+            (*term).osc8_intern_cap_logged,+            "the diagnostic latch must remain set on subsequent over-budget snapshots"+        );++        bb_snap_release(first);+        bb_snap_release(second);+        bb_term_free(term);+    }+}+ #[test] fn new_and_free_roundtrip() {     unsafe {