Default repository audits to gpt-5.3-codex high so runs are more stable than Spark while preserving override controls. Add date-prefixed output directories and update docs/workflow defaults so repeated test runs stay easier to organize. (#26)

Author: anand-testcompare

.github/workflows/daytona-e2e.yml

@@ -10,7 +10,11 @@ on:
       model:
         description: "Model to use"
         required: false
-        default: "zai-coding-plan/glm-4.7-flash"
+        default: "openai/gpt-5.3-codex"
+      variant:
+        description: "Model variant"
+        required: false
+        default: "high"
       analyze_timeout_sec:
         description: "Analyze timeout seconds"
         required: false
@@ -26,6 +30,7 @@ permissions:
 env:
   DAYTONA_API_KEY: ${{ secrets.DAYTONA_API_KEY }}
   DAYTONA_API_URL: ${{ vars.DAYTONA_API_URL }}
+  OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
   ZHIPU_API_KEY: ${{ secrets.ZHIPU_API_KEY }}
 
 jobs:
@@ -58,6 +63,7 @@ jobs:
           bun run analyze -- \
             --out-dir .memory/daytona-e2e/findings \
             --model "${{ inputs.model }}" \
+            --variant "${{ inputs.variant }}" \
             --install-timeout-sec "${{ inputs.install_timeout_sec }}" \
             --analyze-timeout-sec "${{ inputs.analyze_timeout_sec }}" \
             "${{ inputs.repo_url }}"

README.md

@@ -179,9 +179,11 @@ bun run start -- --no-open
 ### Defaults and behavior
 
 - Default model selection:
-  - Standard: `zai-coding-plan/glm-4.7-flash`
+  - Standard: `openai/gpt-5.3-codex`
+  - Standard variant: `high`
   - Vision mode (`--vision`): `zai-coding-plan/glm-4.6v`
 - Override with `--model`, `--variant`, `OPENCODE_ANALYZE_MODEL`, or `OPENCODE_ANALYZE_VARIANT`
+- If you override the model, variant is opt-in (no forced default variant for custom/env model overrides)
 - Auto-installs missing `git` and `node/npm` inside sandbox
 - Forwards provider env vars (`OPENAI_*`, `ANTHROPIC_*`, `XAI_*`, `OPENROUTER_*`, `ZHIPU_*`, `MINIMAX_*`, etc.)
 - Syncs local OpenCode config files from `~/.config/opencode` when present
@@ -196,7 +198,7 @@ bun run start -- --no-open
 ```bash
 bun run analyze -- --input example.md
 bun run analyze -- https://github.com/owner/repo-one https://github.com/owner/repo-two
-bun run analyze -- --out-dir findings --model zai-coding-plan/glm-4.7-flash --target us
+bun run analyze -- --out-dir findings --model openai/gpt-5.3-codex --variant high --target us
 bun run analyze -- --vision
 bun run analyze -- --analyze-timeout-sec 3600 --keep-sandbox
 ```
@@ -208,15 +210,15 @@ If no URLs and no `--input` are provided, the script uses `example.md` when it e
 ## Output Layout
 
 - `<out-dir>/index.md` - summary across all URLs
-- `<out-dir>/<NN-slug>/findings.md` - final report for each repository
-- `<out-dir>/<NN-slug>/README.*` - copied repository README (if found)
-- `<out-dir>/<NN-slug>/opencode-run.log` - raw OpenCode run output
-- `<out-dir>/<NN-slug>/opencode-models-anthropic.log` - Anthropic model-list preflight output (only when `anthropic/*` model is requested)
+- `<out-dir>/<YYYY-MM-DD-NN-slug>/findings.md` - final report for each repository
+- `<out-dir>/<YYYY-MM-DD-NN-slug>/README.*` - copied repository README (if found)
+- `<out-dir>/<YYYY-MM-DD-NN-slug>/opencode-run.log` - raw OpenCode run output
+- `<out-dir>/<YYYY-MM-DD-NN-slug>/opencode-models-anthropic.log` - Anthropic model-list preflight output (only when `anthropic/*` model is requested)
 
-Example retained in this repo:
+Recommended naming convention for manual runs:
 
 ```bash
-bun run analyze -- --input example.md --out-dir findings-confidence-3 --analyze-timeout-sec 3600 --keep-sandbox
+bun run analyze -- --input example.md --out-dir findings-2026-03-03 --analyze-timeout-sec 3600 --keep-sandbox
 ```
 
 ---

src/analyze-model.test.ts

@@ -20,8 +20,8 @@ describe("resolveAnalyzeModel", () => {
     delete process.env.OPENCODE_ANALYZE_VISION_MODEL;
 
     const resolved = resolveAnalyzeModel({});
-    expect(resolved.model).toBe("zai-coding-plan/glm-4.7-flash");
-    expect(resolved.variant).toBeUndefined();
+    expect(resolved.model).toBe("openai/gpt-5.3-codex");
+    expect(resolved.variant).toBe("high");
   });
 
   test("uses vision default when requested", () => {
@@ -34,6 +34,15 @@ describe("resolveAnalyzeModel", () => {
     expect(resolved.variant).toBeUndefined();
   });
 
+  test("does not force default variant when model is overridden by env", () => {
+    process.env.OPENCODE_ANALYZE_MODEL = "openai/gpt-5.3-codex";
+    delete process.env.OPENCODE_ANALYZE_VARIANT;
+
+    const resolved = resolveAnalyzeModel({});
+    expect(resolved.model).toBe("openai/gpt-5.3-codex");
+    expect(resolved.variant).toBeUndefined();
+  });
+
   test("respects env overrides", () => {
     process.env.OPENCODE_ANALYZE_MODEL = "zai-coding-plan/glm-5";
     process.env.OPENCODE_ANALYZE_VARIANT = "high";
@@ -44,15 +53,15 @@ describe("resolveAnalyzeModel", () => {
   });
 
   test("cli args override env", () => {
-    process.env.OPENCODE_ANALYZE_MODEL = "zai-coding-plan/glm-4.7-flash";
+    process.env.OPENCODE_ANALYZE_MODEL = "openai/gpt-5.3-codex";
     process.env.OPENCODE_ANALYZE_VARIANT = "low";
 
     const resolved = resolveAnalyzeModel({
-      model: "zai-coding-plan/glm-4.6v",
+      model: "anthropic/claude-sonnet-4-6",
       variant: "high",
     });
 
-    expect(resolved.model).toBe("zai-coding-plan/glm-4.6v");
+    expect(resolved.model).toBe("anthropic/claude-sonnet-4-6");
     expect(resolved.variant).toBe("high");
   });
 

src/analyze-model.ts

@@ -12,11 +12,13 @@ export type ResolvedAnalyzeModel = {
 };
 
 export function resolveAnalyzeModel(input: AnalyzeModelInput): ResolvedAnalyzeModel {
+  const modelOverride = input.model ?? process.env.OPENCODE_ANALYZE_MODEL;
   const defaultModel = input.vision
     ? (process.env.OPENCODE_ANALYZE_VISION_MODEL ?? "zai-coding-plan/glm-4.6v")
-    : "zai-coding-plan/glm-4.7-flash";
-  const model = input.model ?? process.env.OPENCODE_ANALYZE_MODEL ?? defaultModel;
-  const variant = input.variant ?? process.env.OPENCODE_ANALYZE_VARIANT;
+    : "openai/gpt-5.3-codex";
+  const model = modelOverride ?? defaultModel;
+  const defaultVariant = !input.vision && !modelOverride ? "high" : undefined;
+  const variant = input.variant ?? process.env.OPENCODE_ANALYZE_VARIANT ?? defaultVariant;
 
   return {
     model,

src/analyze-repos.ts

@@ -116,6 +116,13 @@ function slugFromRepoUrl(url: string): string {
   return sanitizeSlug(`${owner}-${repo}`);
 }
 
+function formatDatePrefix(date: Date): string {
+  const year = String(date.getFullYear());
+  const month = String(date.getMonth() + 1).padStart(2, "0");
+  const day = String(date.getDate()).padStart(2, "0");
+  return `${year}-${month}-${day}`;
+}
+
 function normalizeUrlCandidate(raw: string): string | undefined {
   const cleaned = raw.trim().replace(/[),.;]+$/g, "");
   if (!cleaned.startsWith("http://") && !cleaned.startsWith("https://")) {
@@ -250,7 +257,7 @@ function parseCliOptions(): CliOptions {
 Examples:
   bun run analyze -- --input example.md
   bun run analyze -- https://github.com/agenticnotetaking/arscontexta
-  bun run analyze -- --input links.md --out-dir findings --model zai-coding-plan/glm-4.7-flash
+  bun run analyze -- --input links.md --out-dir findings --model openai/gpt-5.3-codex --variant high
   bun run analyze -- --vision
 
 Options:
@@ -260,8 +267,8 @@ Options:
       --install-timeout-sec <n>  OpenCode install timeout (default: 900)
       --analyze-timeout-sec <n>  Per-repo analysis timeout (default: 2400)
       --target <name>            Daytona target override
-      --model <provider/model>   OpenCode model (default: zai-coding-plan/glm-4.7-flash)
-      --variant <name>           Model variant (example: xhigh)
+      --model <provider/model>   OpenCode model (default: openai/gpt-5.3-codex)
+      --variant <name>           Model variant (default: high, when using built-in default model)
       --vision                   Prefer vision-capable default model (zai-coding-plan/glm-4.6v)
       --keep-sandbox             Keep each sandbox instead of deleting it
   -h, --help                     Show this help
@@ -676,7 +683,8 @@ async function analyzeOneRepo(params: {
   const { daytona, options, config, url, index, total } = params;
   const slug = slugFromRepoUrl(url);
   const runPrefix = `${String(index + 1).padStart(2, "0")}-${slug}`;
-  const localDir = path.join(options.outDir, runPrefix);
+  const datedRunPrefix = `${formatDatePrefix(new Date())}-${runPrefix}`;
+  const localDir = path.join(options.outDir, datedRunPrefix);
   const findingsPath = path.join(localDir, "findings.md");
   let readmePath: string | undefined;