Repository navigation
Releases
Releases · QianJinGuo/prompt-contract
v0.3.1 — contract watch resident mode + model adaptation
First release under the PromptContract name, shipping two major additions plus reliability fixes.
New
prompt-contract watch(macOS resident mode) — select text anywhere, press the global hotkey (default ⌥B, configurable via--hotkeyor config), and the enhanced prompt replaces your selection: clipboard backed up and restored, focus re-validated before pasting. Startup is gated byprompt-contract spike-0evidence (decision D7). Swift hotkey helper is compiled on demand — registering the hotkey needs no accessibility permission.- Model adaptation P1 — reasoning-block stripping (
think/thinking/reasoning, incl. unclosed streaming chunks), vendor presets (deepseek / qwen / glm / moonshot / groq / openrouter / lmstudio — zero-config base URLs, explicit config always wins), native Anthropic Messages SSE (max_tokensenforced, thinking deltas dropped at the transport layer).
Fixed
- Piped-stdin reads no longer use sync fd reads — an intermittent
EAGAINfailure on macOS runners. - Collapsed duplicated
prompt-prompt-contract-style command prefixes that 0.3.0 shipped in docs and CLI strings. - Providers tests: fixed-sleep assertions → predicate polling (deflakes CI); restored anthropic provider test coverage.
Added
- Task-level eval harness (
npm run eval:tasks) with curated fixtures — deterministic format gate, then hands original/enhanced pairs to an external runner for real downstream measurement. - Playground hosted at https://qianjinguo.github.io/prompt-contract/
npm: npm i -g prompt-contract · npx prompt-contract "your vague idea"
v0.3.0
v0.2.1
v0.2.0
v0.1.1
v0.1.0 — M0 minimal set
First public release: the M0 minimal set.
- Engine (
@prompt-boost/core): profile-driven single-shot enhancement, deterministic output cleaning, structured error codes, six hard-constraint rule assertions (template and eval share one spec) - CLI (
pb): one-key enhance,profiles,check(rule gate),doctor - MCP server:
enhance_prompttool (agent-invoked, BYOK) +boost-*prompts (user-invoked, zero-key — uses the client's own model) - Playground: single-file BYOK UI with streaming, rule badges, revert
- Providers: OpenAI-compatible SSE with connection warmup; Ollama with
keep_alivemodel pinning
Verified: 43/43 tests, 9/9 deterministic eval cases, engine overhead P50 0.004ms.
pb watch (resident hotkey mode) is intentionally gated by decision D7 — see docs/ACCEPTANCE.md.
Full Changelog: https://github.com/QianJinGuo/prompt-boost/commits/v0.1.0