Features
Capabilities beyond the basic loop - quality gates, free-tier routing, memory, rules, streaming UI, and slash utilities. Use this page as a map; deep dives live in the linked guides.
Agent harness
Plain English starts a multi-step tool loop. The model can read, search, write, edit, run shell, verify, touch git, and track todos until it calls finish_task.
- Tool registry documented in Tools
- Loop semantics in Agent loop
- Default max iterations: 32 (
VRITRA_AGENT_MAX_ITER)
Streaming UI
- Live narration: Calling model… Create… Edit… $ command…
- Divider-style code / diff previews (not noisy boxes)
- Approve · always this session · deny
- Disable streaming with
VRITRA_AGENT_STREAM=0
Stub / hollow content rejection
Theatrical empty HTML/CSS/JS scaffolds are rejected by content-quality checks. The agent should not mark a hollow page as done. When prompting, ask for real sections, copy, and behavior.
If a finish looks empty, restate the structure you need and run again - see Troubleshooting.
Free-tier aware failover
- Hard faults (credits, unavailable, empty free replies) switch models
- Paid OpenRouter without credits does not burn every free id
- A single OpenRouter 429 leaves other free models available
- Routing prefers free / Groq over expensive dead-ends
- Free paths may use lower max tokens to avoid afford errors
Details: Providers.
Approvals & trust
Per-workspace modes (ask / auto / safe), session always-allow, dangerous shell classification, trust file at .vritraai/trust.json. Full guide: Approvals.
Memory & rules
.vritraai/agent_memory.json- workspace-scoped memory- Optional rules:
.vritraai/rules.md,.vritraai/rules,AGENTS.md - Keys stay in
~/.config-vritrasecz/vritraai/config.json
Slash utilities
Same CLI without leaving the product: /review, /security, /optimize, /refactor, /doc, /project, /explain, /learn, /cheatsheet, /summarize. Index: Commands.
Multi-file delivery
write_filesfor batch creates- Exact-match
replace_in_file/apply_edits - Verify hooks before finish when the task needs proof