Release v1.5.0 — Prompt Intelligence System · anotherben/claude-harness

What's New

classify-prompt.sh — 12-category intent classifier replacing the binary refine-prompt.sh hook
- Categories: reference, bug, ops, feature, refactor, investigate, design, business, marketing, content, meta, question
- Two-pass classification: explicit investigation verbs override bug symptoms
- Confidence scoring: ambiguous prompts (3+ category hits) pass through instead of misrouting
- Compound prompts: "fix X then deploy Y" detects and routes both intents sequentially
- Correction detection: user corrections like "test it yourself" logged to .cortex/knowledge.jsonl
- Positive reinforcement: "perfect" / "nailed it" signals logged alongside corrections
- Hot-correction escalation: 3+ same-category corrections become auto-enforced for the session
log-classification.sh — PostToolUse hook for misclassification tracking
- Detects when the user invokes a different skill than the classifier suggested
- Writes misclass annotations for ongoing classifier improvement
prompt-intelligence skill — Session-start learned behavior loader
- Reads accumulated cortex feedback annotations
- Builds LEARNED BEHAVIORS context block for the session
- Bootstrap seeds from existing memory files on first run

All 12 enterprise skills now load domain-specific cortex lessons at entry
Each skill reads its own tag (e.g., feedback:BUILD, feedback:DEBUG)
Corrections during skill execution write back domain-tagged annotations — creating a feedback loop

One-command Codex agent setup for any repo: ~/claude-harness/codex-init.sh --auto
Creates .codex/ structure, AGENTS.md, project-prefixed repo-skills, manifest
Auto-detects project name, type, and Obsidian vault path
--auto flag for non-interactive use inside Codex sessions

4 new files (classify-prompt.sh, log-classification.sh, prompt-intelligence/SKILL.md, codex-init.sh)
13 modified files (12 enterprise skills + ensure-environment.sh)