Releases: phdgil/pi-solar-workflow
Release list
v0.3.0 - Lightweight workflows and verified handoffs
pi-lite-workflow
Lightweight research -> interview -> plan -> execute workflows for tool-capable models in pi. MIT licensed; no additional controller dependency.
Changes
- Rename the project and four skills to pi-lite-workflow and lite-research/interview/plan/execute. Retain old solar-* commands and saved-session identifiers for migration.
- Preserve original intention and research context through each stage. Interview ambiguity is informational; the user may finish at any score.
- Hand a reviewed, conflict-free plan to execution automatically within the requested local scope, without another confirmation. The alignment report is single-model self-review, not independent semantic proof.
- Fix rejection of valid plans written with bold Step N labels. Numbered lists/headings and checkboxes are supported; fenced examples do not count as steps.
- Fix helper reload caching. Keep provider-specific Solar retry/tool-choice handling separate from the generic workflow context.
Install
pi install git:github.com/phdgil/pi-lite-workflow@v0.3.0
piExisting installations: follow docs/INSTALL.md to remove the exact old package registration before installing the renamed source. Preserve keys, model settings, and sessions. Old command aliases remain available.
Verification and limitations
69 unit/package tests passed. Fresh GitHub-tag installations passed the full real-pi reload/alias/workflow/alignment tests with both Upstage Solar Pro4 and a non-Solar mock-medium fixture. Model traffic used loopback mocks only. Solar Pro4 Max remains the only live-tested model; this is not a cross-model quality benchmark or a guarantee for every provider.
Experimental prerelease, tested with pi 0.85.0 on Windows.
v0.1.0 - First experimental release
First experimental public release for pi 0.85.0 on Windows.
Install
pi install git:github.com/phdgil/pi-solar-lite@v0.1.0Then launch pi. The package includes all four skills and the runtime extension. See README and docs/INSTALL.md for model setup without shared credentials.
Workflow
Research -> interview -> plan -> execute: establish context, sharpen user intention, prepare executable steps and checks, then verify explicitly authorized work.
User-controlled interview
Ambiguity is informational only: no 5% cutoff or blocker floor. Use /solar-interview finish at any score, or voluntarily continue. Missing next questions are valid. Unresolved/deferred issues and original answers remain in the handoff; stopping does not claim every issue is solved. Finishing also works with an interrupted or unassessed request.
Licensing and validation
MIT for original work, with upstream notices and a reference/license audit. No API keys, private sessions, experimental controllers, or vendored community packages are distributed.
58 unit/package tests pass. Clean-profile installation from GitHub and real-pi scripted tests pass, including optional questions, finish at 25%, restart, unassessed cancellation, and max reasoning requests. Model traffic in release tests uses a local fixture, not a paid API. Other OSes/models and semantic interview quality remain unverified.