SHINE Claude Code Framework — a comprehensive system for managing Claude Code sessions, profiles, hooks, and project workflows with git integration, memory management, and context monitoring.
SHINE tools lack critical definition quality. Tool descriptions are present but generic (avg ~80 chars, well below 194-char baseline). Input schemas are visible in code but lack formal JSON Schema structure, parameters are documented in comments, not in a machine-readable schema object. No output schemas documented. Parameters lack type constraints (enums, ranges). Error handling is minimal (fail() calls with generic messages). No tool annotations (readOnlyHint, destructiveHint). Tools like 'commit' and 'commit-to-subrepo' perform destructive operations but lack confirmation patterns. Parameter descriptions are sparse (e.g., 'phase' is just 'Phase number (default: 1)', no format guidance). No pagination support despite tools returning lists (roadmap, verify). Composition is weak: 'init' and 'state' should be separate from 'commit' operations, but state mutation is conflated with read operations.
Stage all changes and commit to git with a formatted SHINE commit message (shine(scope): message).
Commit changes to a git subrepo at a specified path with a SHINE-prefixed message.
Print a JSON health report of the SHINE installation and environment.
Regenerate skills/INDEX.md from available skill files.
Initialize SHINE project structure for a given phase, returning project root, phase directory, plan files, and configuration paths.
First-run memory wizard to set up typed memory structure.
Parse and return phases from ROADMAP.md, optionally filtered by phase number.
Read or update phase state variables stored in STATE.md. Supports --set key=value to update state entries.
No formal JSON Schema for tool inputs. Parameters documented in comments, not in machine-readable schema objects. LLMs cannot parse constraints or types programmatically.
Destructive operations (commit, commit-to-subrepo, state --set, onboard, index-skills) lack confirmation or dry-run patterns. Agents can irreversibly modify git history or state without safeguards.
Output schemas not documented. Tools return JSON objects (roadmap, verify, doctor, state) but LLMs have no specification of returned fields, types, or structure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 28 | 2026-07-28+ | v2 |
Run verification checks on a phase (goal-backward, task-completion, etc.) and return results.
Error messages are generic ('git add failed', 'cwd does not exist'). No recovery guidance. LLMs cannot determine if errors are retryable or require user intervention.
Parameter descriptions are minimal (10 - 30 chars). 'phase' is 'Phase number (default: 1)' with no format guidance. 'check' in verify lacks enum values. Descriptions fall far below 194-char baseline.