Prompt-Pinned Model Resolution

Highlights

This release adds verified, prompt-pinned Ollama model-family resolution for the supported OpenCode launch workflow.

Model resolution

• Resolves the highest stable numeric Ollama Cloud model within configured families.
• Defaults to MiniMax M for the builder and GLM for the reviewer.
• Correctly compares numeric versions such as 5.10 > 5.2.
• Supports strict exact-model overrides.
• Never crosses providers or model families automatically.
• Never downloads heavyweight local model variants.

Verification

Every previously unseen exact model must pass:

• Ollama Cloud manifest readiness.
• A tools-disabled inference smoke test.
• A restricted fixture-read tool-call smoke test.
• OpenCode runtime and agent model-mapping checks.

Prompt reproducibility

• Pins the verified builder/reviewer pair for the lifetime of each prompt.
• Reuses the same pair for implementation, review, fixes, and signoff.
• Changes models only through an explicit --refresh-models.
• Keeps prompt locks and resolver state local and gitignored.

Concurrency and recovery

• Adds per-prompt resolution locks.
• Adds a globally serialized, schema-3 shared model cache.
• Isolates last-known-good entries by provider, families, and exact overrides.
• Merges concurrent cache updates without losing unrelated contexts.
• Preserves existing prompt locks when refresh or persistence fails.
• Uses one immutable resolver result for both builder and reviewer selection.

Current verified pair

• Builder: ollama/minimax-m3:cloud
• Reviewer: ollama/glm-5.2:cloud

Validation

• 54 deterministic resolver and launcher tests.
• 53 passed.
• 1 opt-in live test skipped in standard CI.
• Live Ollama Cloud compatibility test passed separately.
• Public-pack validation passed.
• Python compilation passed.
• Zsh syntax checks passed.
• Tiny example tests passed.
• OpenCode and Ollama doctor checks passed.

Compatibility

• Prompt-lock schema: 2
• Shared-cache schema: 3
• Resolver verification policy: 2.1
• Supported platforms: macOS and Linux
• OpenCode upgrades remain manual

v0.1.1: Polish release

Polish release for the public starter kit.

Adds:

• GitHub Actions sanity workflow
• model override environment variables
• clearer README command examples
• explicit redaction placeholders
• complete screenshot-free demo session example
• release checklist for future versions

v0.1.0: Public starter kit

Initial public starter kit for a receipt-first OpenCode + Ollama multi-agent workflow.

Includes:

• builder/reviewer OpenCode agents
• .opencode/skills guardrails
• headless run scripts
• local session folder templates
• redaction scanner
• public-pack validator
• lean developer docs
• tiny public-safe demo app

This release intentionally excludes private prompts, raw sessions, screenshots, local paths, credentials, and private project artifacts.

The goal of this repo is simple: stop vibe-coding blind. Every agent run should leave enough evidence to answer what was planned, what changed, what tests ran, what failed, and whether a reviewer signed off.