Benchmark run task11-gpt-6-luna-medium-mcp-matrix-retry-failed-cell-20260927

This page contains exactly the publishable benchmark observations recorded in run task11-gpt-6-luna-medium-mcp-matrix-retry-failed-cell-20260927; it does not select or aggregate them. Return to the run index or view aggregate metrics.

Interactive browser

Detailed observations

Empirical telemetry measuring wall-clock latency, token usage, tool turns, and compiler correctness when AI coding agents perform code modifications.

Every benchmark pairs Vanilla LLM (standard file editing tools) against Semedit MCP (deterministic AST compiler operations).

Benchmark Methodology & Transparency

• Reproducibility & Provenance: Self-contained txtar archives define test scenarios. Links below resolve to commit-anchored GitHub source files for published commits, or cryptographic SHA-256 fingerprints for uncommitted local fixtures.
• Model Cost: Tables report model-specific credit consumption where a target model has a declared schedule: (uncached input × input credits + cached input × cached credits + (reasoning + visible output) × output credits) / 1,000,000. Thinking tokens use the output rate. ADR-0043 defines the per-million-token credit rates.
• Multi-Level Correctness Oracle: Four validation levels evaluate each trial:
  1. Level 1 (Mutation Policy): Restricts file modifications strictly to authorized paths.
  2. Level 2 (AST Invariants): Compiler AST verification of required symbols, imports, and relative declaration ordering.
  3. Level 3 (Clean Build): Strict compilation validation with zero compiler or typecheck errors.
  4. Level 4 (Test Suite): Unit and integration test suite execution.
• Isolated Sandboxing: Each trial executes in an ephemeral directory with pristine git and module state.

Test case: task-11-mixed-sink-api-migration

Target: codex/gpt-6-luna/medium

default prompt · prescriptive MCP instructions · write semedit restriction

Vanilla LLM Prompt:

Do not edit tests. You are forbidden to modify protected files: “go.mod”, “go.sum”.

Audit event kinds may contain inconsistent whitespace and casing. Add the public audit.NormalizeKind(kind string) string function in audit/normalize.go to remove surrounding whitespace and lowercase the kind. Apply it in Dispatcher.Record before persistence, and classify normalized probe events as control traffic, like health. Do not change the metrics or legacy export protocols. Preserve the module metadata and tests, and leave the workspace verified. Do not use semantic editing MCP tools; use standard file editing. When done, output DONE.

Semedit MCP Prompt:

Do not edit tests. You are forbidden to modify protected files: “go.mod”, “go.sum”.

Audit event kinds may contain inconsistent whitespace and casing. Add the public audit.NormalizeKind(kind string) string function in audit/normalize.go to remove surrounding whitespace and lowercase the kind. Apply it in Dispatcher.Record before persistence, and classify normalized probe events as control traffic, like health. Do not change the metrics or legacy export protocols. Preserve the module metadata and tests, and leave the workspace verified. Use semedit semantic tools for supported source code modifications. Shell commands for builds and tests are allowed. When done, output DONE.

// Package audit leaves room for normalization rules owned by the delivery boundary.
package audit

Baseline vs Semedit Agentic in Small Context

• Vanilla (Baseline vs Semedit Agentic in Small Context) Edit: File audit/normalize.go modified (2 lines -> 10 lines)
• Semedit MCP (Baseline vs Semedit Agentic in Small Context) Edit: File audit/normalize.go modified (2 lines -> 8 lines)

/bin/zsh -lc "pwd && \
rg --files -g 'AGENTS*' -g 'PROJECT_BRIEF.md' -g 'go.mod' -g 'go.sum' -g 'Makefile' -g 'audit/**' -g 'docs/**' | head -80 && \
git status --short"

/bin/zsh -lc "pwd && \
rg --files -g '"'!*.sum'"' -g '"'!go.mod'"' | head -80"

/bin/zsh -lc "cat audit/normalize.go && \
sed -n '1,260p' audit/event.go && \
rg -n \"Dispatcher|Record\\(|health|probe|control\" audit"

{
  "include_unexported": true,
  "path": "audit"
}

/bin/zsh -lc "rg -n \"type Dispatcher|func \\(.*\\) Record|persist|Sink|Classify\" . -g '*.go' -g '"'!**/*_test.go'"' && \
sed -n '1,240p' audit/sink.go && \
cat audit/summary.go && \
cat go.mod"

/bin/zsh -lc 'cat audit/normalize.go && \
cat Makefile'

/bin/zsh -lc 'cat service/dispatch.go && \
rg --files'

{
  "symbol": "Classify"
}

/bin/zsh -lc 'gofmt -w audit/normalize.go audit/summary.go service/dispatch.go && \
go test ./... && \
git diff -- audit/normalize.go audit/summary.go service/dispatch.go go.mod go.sum && \
git status --short -- audit/normalize.go audit/summary.go service/dispatch.go go.mod go.sum'

{
  "symbol": "Record"
}

/bin/zsh -lc "cat audit/normalize.go && \
cat audit/summary.go && \
sed -n '9,20p' service/dispatch.go && \
git status --short --untracked-files=no"

{
  "auto_organize_imports": true,
  "edits": [
    {
      "params": {
        "access_modifier": "public",
        "file": "audit/normalize.go",
        "source": "func NormalizeKind(kind string) string {\n\treturn strings.ToLower(strings.TrimSpace(kind))\n}"
      },
      "tool": "semantic_insert_function"
    },
    {
      "params": {
        "auto_organize_imports": true,
        "body": "event.Kind = audit.NormalizeKind(event.Kind)\nreturn d.sink.Write(event)",
        "file": "service/dispatch.go",
        "symbol": "(*Dispatcher).Record"
      },
      "tool": "semantic_replace_body"
    },
    {
      "params": {
        "body": "switch NormalizeKind(kind) {\ncase \"health\", \"probe\":\n\treturn \"control\"\ndefault:\n\treturn \"data\"\n}",
        "file": "audit/summary.go",
        "symbol": "Classify"
      },
      "tool": "semantic_replace_body"
    }
  ]
}

/bin/zsh -lc 'git status --short && \
git diff -- go.mod go.sum'

/bin/zsh -lc 'go test ./...'

{
  "language": "go"
}

Baseline vs Semedit Agentic in Large Context