v0.1.1 — Rename VERSION to TEMPLATE_VERSION

v0.1.1 — Rename VERSION to TEMPLATE_VERSION

Released: 2026-05-11

Breaking Changes

VERSION → TEMPLATE_VERSION — The version tracking file was renamed to avoid colliding with project-level VERSION files that consuming projects may already have.

If your project copied VERSION from the template, rename it:

git mv VERSION TEMPLATE_VERSION

New Features

None.

Improvements

None.

Bug Fixes

None.

Dependency Updates

None.

Upgrade Instructions

1. Rename the file

git mv VERSION TEMPLATE_VERSION
git commit -m "chore: rename VERSION to TEMPLATE_VERSION"

2. Update TEMPLATE_VERSION value

echo "0.1.1" > TEMPLATE_VERSION

v0.1.0 — Initial versioned release

v0.1.0 — Initial versioned release

Released: 2026-05-11

Breaking Changes

None — this is the first versioned release.

New Features

None.

Improvements

LLM provider retry with exponential backoff

supabase/functions/_shared/langfuse/shared.ts — new exported fetchWithRetry function.

All 13 provider fetch() calls across anthropic.ts, openai.ts, and gemini.ts now use it instead of bare fetch().

Behaviour:

• Retries on HTTP 429 (rate limit), 500, 502, 503, 504
• Respects the Retry-After response header when present (Anthropic and OpenAI send this on 429)
• Falls back to exponential backoff with ±25 % jitter when no header is present
• Drains the response body before retrying to avoid connection pool leaks
• After exhausting retries (default: 3), returns the final failed response so callers can still read the body
• Capped at 30 s per wait

To use in a custom edge function:

import { fetchWithRetry } from '../_shared/langfuse/shared.ts';

const response = await fetchWithRetry(url, { method: 'POST', ... });
// or with custom limits:
const response = await fetchWithRetry(url, init, { maxRetries: 5, initialDelayMs: 500 });

The provider-specific tracing functions (traceLLMAnthropic, traceLLMOpenAI, traceLLMGemini, etc.) already use it — no changes needed if you call those.

Maestro MCP — clarified two-track testing in AGENTS.md

AGENTS.md now clearly separates:

• Regression track — maestro test YAML flows, run before production or in CI
• Interactive track — Maestro MCP tools (list_devices, take_screenshot, inspect_view_hierarchy, taps) for PR-scoped checks during development

Copy the updated AGENTS.md or apply the relevant sections to your own agent context file.

Streaming Gemini — missing response.ok guard added

supabase/functions/_shared/langfuse/gemini.ts — streamLLMGemini previously had no HTTP error check before starting the stream. It now throws a typed error on non-2xx responses before the TransformStream is attached.

No API change — callers already wrapped this in try/catch.

Bug Fixes

None.

Dependency Updates

None.

Upgrade Instructions

This is the first versioned release. For projects already based on the template:

1. Pull the retry improvement (recommended for all projects)

Copy the updated supabase/functions/_shared/langfuse/shared.ts into your project.
Then update the import in each provider file to add:

import { fetchWithRetry } from "./shared.ts";

Replace every await fetch(url, init) call inside anthropic.ts, openai.ts, and gemini.ts with:

await fetchWithRetry(url, init)

Also add the missing response.ok guard in streamLLMGemini (see the diff in this release).

2. Update .gitignore (optional)

Add these entries if you use agent worktrees or LLM flow runs:

.llmflows/
.worktrees/

3. Update VERSION

echo "0.1.0" > VERSION