- Detect empty LLM responses (no content, no tool calls) instead of
silently treating them as task completion. Retries once without tools
before warning the user.
- Gate /no_think system message and chat_template_kwargs to Qwen/QwQ
models only — sending /no_think to llama3.x caused empty responses.
- Add model_profiles config section for per-model overrides (token
budget, thinking, temperature, max_tokens) matched by name prefix.
Applied at startup and on /model switch.
- Update SessionManager on /model switch so session files record the
correct model.
- Add NDJSON fallback in SSE stream parser for Ollama compatibility.
- Improve read_file error to suggest find_files on FileNotFoundError.
- Add diagnostic logging for empty streams and empty results.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace built-in Header with a custom HeaderPanel showing model name,
mode badge, and live token usage. Add AgentMode enum (normal/plan/auto)
with mode-aware permission gating — Plan mode restricts to read-only
tools, Auto mode auto-approves everything. Includes /mode slash command
and Ctrl+P keybinding to cycle modes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Bug 1: Assistant text content was silently dropped when the LLM response
included both content and tool calls (e.g. finish with a summary). Now
content is displayed before tool call execution regardless.
Bug 2: Shell redirect operators (>, >>, <<) allowed bypassing file-write
permissions when the base command (e.g. cat) was in the allowed list.
Redirects now require explicit user approval in permissions, and the
shell tool itself blocks them as defense-in-depth.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The /no_think directive was appended directly to the user's message text,
causing the LLM to interpret it as user input (e.g., referencing it as a
directory path). Now injected as a separate system message after the last
user message. Also adds which, jq, type, and file to shell allowed_commands
(were mistakenly placed in permissions.auto_approve).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds `llm.thinking` config option (default: true) that when disabled:
- Injects /no_think into the last user message for Qwen 3.x compatibility
- Sends chat_template_kwargs in API payload for backends that support it
- Silently and immediately nudges on reasoning-only responses instead of
showing warnings and wasting retry iterations
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add a skill package system where each skill is a directory with a skill.yaml
manifest and prompt markdown files. Skills support /command triggers, scoped
config overrides (temperature, max_tokens, tool filtering), chain dependencies
with cycle-safe resolution, and a finish_skill completion signal.
Includes four built-in skills: explore, brainstorm, write-document, and plan.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Assistant message panels now use Markdown() instead of raw strings, so
bold/italic/lists render properly. Also nudge the model immediately after
tool errors instead of wasting 2 retry iterations in reasoning-only mode.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The streaming widget shows content during streaming then hides on
completion, but nobody was writing the final response to the RichLog.
The assistant message flashed briefly then disappeared.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Implement 6 new agent tools — write_file, make_dir, delete_file,
str_replace, patch_apply, run_command — bringing the agent from
read-only observer to active code modifier. All write/shell operations
are gated through the existing permissions service.
Also fix a bug where qwen3.5 thinking mode produces reasoning tokens
but no content after tool results, causing the agent to silently exit.
The loop now detects reasoning-only responses, retries twice, then
injects a nudge message to break the model out of its thinking loop.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
resolve_safe_path now detects when the LLM passes the workspace root as
a relative path (e.g. "home/user/project/file.py") and strips the prefix
before joining. Also updated the system prompt to explicitly instruct the
model to use relative paths for all tool arguments.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Implement the core autonomy layer — AgentLoop streams LLM responses,
parses tool calls, executes them with permission checks, feeds results
back, and repeats until the task completes or finish is called.
- Add FinishTool for explicit loop termination
- Add tools parameter to LLMClient.stream_chat() for function calling
- Add compact tool result display (status line, not full output)
- Refactor REPL to delegate to AgentLoop.run_turn()
- Fix Ollama null content rejection (always send content as string)
- Add finish to auto_approve permissions
- 9 unit tests for agent loop (34 total, zero regressions)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
