Merge pull request #31 from huggingface/agent-improvements

Fix premature agent loop exit in long-running autonomous tasks

Only show plan_tool output in CLI, hide all other tool outputs

Use Haiku 4.5 for research sub-agent

Use Haiku for research sub-agent when main model is Anthropic

Research stats tick every second via background timer

Live research stats: tool count, tokens, elapsed time in header

Show full sub-agent tool call lines, no truncation

Revert "Sub-agent log: tool_name → output_preview on one line"

Sub-agent log: tool_name → output_preview on one line

Log 20-char output preview for sub-agent tool calls

Sub-agent display: 2 lines, gray, 20 char truncation

Indent ASCII banner to match all other output

Indent all agent output, remove turn separator line

Render streamed responses as markdown, indent agent output for alignment

Buffer streamed chunks and render full markdown on completion

Switch banner to doom font with connected letterforms

Hide research sub-agent tool_call and tool_output from CLI display

Tool output: last 10 lines for main agent, rolling 3-line display for sub-agents

Fix Ctrl+C at idle prompt: no false interrupt, no stuck state

Replace compact panel with medium ASCII art banner

Fix streaming: line-buffered output instead of per-token Live re-render

Overhaul CLI visual design with rich-powered formatting

feat: add headless CLI mode with local filesystem tools and rich terminal rendering

refactor: append dynamic session context after static system prompt for cache stability

Deduplicate file re-reads via content hashing for local and sandbox tools

feat: LLM-based pruning for complex tool outputs (hf_jobs)

fix: also skip plan_tool from deterministic pruning

fix: skip research tool outputs in deterministic pruning

Add deterministic pruning of old tool outputs before LLM compaction

Add per-tool error limits and per-turn request cap

feat: add doom-loop detection for repeated tool call patterns

fix: move import/signature validation to sandbox, keep host-side lightweight

fix: update system_prompt_v3.yaml (the actual active prompt) to use research tool

feat: emit tool_log events from research sub-agent for UI liveness

feat: add research sub-agent tool, slim down main agent system prompt

Improve tool output quality: ANSI stripping, tail-biased truncation, noise suppression

feat: render cancelled tool state in UI after user interrupt

feat: cancel sandbox creation on user interrupt

feat: kill sandbox processes and cancel HF jobs on user interrupt

fix: auto-retry on truncated tool calls + correct activity status on refresh

fix: retry sandbox API calls on transient failures

fix: live event streaming after reconnection + LiteLLM timeout

chore: double sandbox command timeouts

chore: increase sandbox sleep_time to 45 min

fix: clear pending_approval before tool execution starts

fix: gracefully handle MCP and OpenAPI failures during startup

refactor: simplify tool call validation and make interrupts cancel tool execution

feat: prompt for HF token interactively when not found

fix: correct context_length init and emit tool_call events for malformed calls