What It Does For You¶
A complete map of capabilities — what fires, when, and what it prevents.
The Full Picture¶
Lifecycle: When Each Capability Fires¶
Capability Reference¶
| Capability | What It Does | Hook / Component | Level |
|---|---|---|---|
| Infrastructure Checks | Validates DB, git, venv, custom commands at startup | on_session_start.py |
1 |
| Tiered Loading | Loads core rules always, task-specific rules on demand | on_session_start.py |
1 |
| Manifest + Sentinel | Tracks session state, file reads, stage progression | on_session_start.py |
1 |
| Gate Enforcement | Blocks non-Read tools until Tier 1 is loaded | gate_check.py |
2 |
| Prompt Health | Warns at configurable prompt count thresholds | on_prompt_submit.py |
2 |
| Context Reset Detection | Detects /clear and re-triggers startup |
on_prompt_submit.py |
2 |
| Tier 2 Keyword Triggers | Scans tool inputs for task keywords, loads matching files | gate_check.py |
3 |
| Drift Detection | Compares expected counts against live state | cross_check.py |
3 |
| Auto-Heal | Fixes safe drift items automatically (bounded, 2 passes) | cross_check.py |
3 |
| Write-Back Suggestions | Proposes manifest updates for persistent drift | cross_check.py |
3 |
| Rule Zero | Scans edited files for scattered content, warns to consolidate | on_edit.py |
4 |
| Edit Tracking | Counts edits, periodic save reminders | on_edit.py |
4 |
| Self-Verification | Blocks exit if infrastructure was edited but not re-checked | on_stop.py |
4 |
| Clean Shutdown | Requires clean repos, transcript, audit pass before exit | on_stop.py |
4 |
| Session Continuity | Persistent backlog (JSON or DB) loaded at startup | backlog.json / DB |
4 |
| No-Truncation | Verifies DB stores weren't silently truncated | on_stop.py |
4 |
| Audit Runner | On-demand infrastructure checks, same validators as startup | audit.py |
4 |
| Sensitive Data Scan | Pre-commit hook + consistency checker block personal data | pre-commit + consistency_check.py |
4 |
| Anti-Hallucination Rules | 14 research-backed rules in 5 phases for faithful LLM outputs | rules/ |
Any |
| Feedback Loop | Failure → Learning → Rule → Check → Hook evolution | Pattern | Any |
What Problem Does Each Capability Solve?¶
Read this column when deciding what to enable
| Capability | Without It | With It |
|---|---|---|
| Gate Enforcement | Agent starts working before loading rules — applies defaults, not your conventions | Tools blocked until rules are in context — every session starts with full knowledge |
| Rule Zero | Information scatters across files and conversations, lost at session end | Every edit is categorized and routed — nothing gets lost |
| Self-Verification | Agent says "done" but the fix wasn't tested | Exit blocked until verification re-runs — "done" means verified |
| Drift Detection | Config says 3 rules but you have 5 — silent mismatch grows | Expected vs actual compared every session — drift caught early |
| Anti-Hallucination | LLM adds plausible but unsourced details to summaries | 14 rules enforce source-faithful extraction — retract if no quote |
| Feedback Loop | Same mistakes repeat across sessions | Each failure becomes a rule, each rule becomes a check — system learns |
| Session Continuity | User re-explains context every session | Agent picks up where last session left off — backlog persists |
| Sensitive Data Scan | Personal data accidentally committed to public repos | Pre-commit hook + scanner block matches before they reach git |
Adoption Levels¶
Start with Level 1 and grow as needed. Each level adds capabilities without breaking previous ones.
Ready to set it up? Run the Setup Wizard to generate your configuration at your chosen level.