- /compact: summarizes evicted context messages via LLM call, replaces
them with a single summary system message
- /clear: resets session and display
- ContextBuilder.EvictedMessages(): returns messages outside bounded window
- Session.Compact(): drives the summarization and history replacement
- file_read and file_write require explicit y/n confirmation before executing
- New agentConfirming UI state with confirm [y/n] status bar indicator
- y/yes approves, n/no denies, anything else denies and falls through to
normal prompt handling
- file_read output includes line numbers for precise file_write targeting
- Fix tool output display: remove per-line squashWhitespace (preserves indentation)
- file_write description hints to preserve formatting
- Pass caller context into Executor.Execute so cancellation reaches
the subprocess immediately
- Use Setpgid + SIGKILL on process group to kill grandchild processes
- Add context to ToolExecutor signature and thread it through dispatch
- Reset agent state to idle after drainAgent
- Rollback incomplete session turn on interrupt
When the model emits a text-only turn containing narration phrases
("let me", "i'll", "i will", etc.) with no tool calls, the loop now
injects an ephemeral "Continue. Act now." user message and runs another
step rather than treating the turn as completion. Capped at 2 nudges
per run to prevent infinite loops. UI shows "[nudge: continuing…]" so
the user can see what happened.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Return RateLimitError from openai backend on HTTP 429, parsing the
Retry-After header (integer seconds or HTTP-date). The agent loop retries
up to 3 times with exponential backoff (5s/10s/20s) when no header is
given, emitting per-second countdown ticks. The status bar renders
"retry {N}s" during the wait.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Five improvements for cost/focus efficiency:
1. Budget now accounts for fixed per-request overhead (system prompt +
tool schemas), which were previously invisible to ContextBuilder and
could silently exceed intended limits by 30-50%.
2. Tail counting now only counts user and plain-assistant (no tool_calls)
messages toward TailMessages quota. Processed tool exchanges between
conversational turns are included but do not crowd out genuine context.
Trailing in-progress exchanges are always preserved and free.
3. BoundedHistory / BoundedHistoryWithNotice deduplicated into a single
buildBounded(injectNotice bool) private helper.
4. O(n²) prepend in the greedy inclusion loop replaced with
append+slices.Reverse.
5. exec.Executor.MaxOutputChars replaces the hardcoded 8000-char const.
Initialised from OLLIE_TOOL_OUTPUT_CHARS env var, defaults to 8000.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Loop struct and New constructor removed; Run is now a package-level
function taking (ctx, Config, State)
- Stop condition de-nested: natural stop (no tool calls → MarkComplete
+ break) and step-limit stop (step >= maxSteps-1 → break) are now
separate, sequential checks instead of a redundant outer/inner pair
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Both backends already implement streaming; the non-streaming fallback
was dead code that added complexity.
- Backend interface now requires only ChatStream; the separate
StreamingBackend interface and Chat method are removed
- Both OpenAIBackend and OllamaBackend lose their Chat methods and the
stream=bool parameter on their internal doChat helpers
- Loop.Run is simplified: one streaming path, no streamed flag, no
skippedCalls map, no shouldStop helper, no runStreamStep indirection
- "call" events are now emitted solely in the act phase, not in the
stream phase, so dedup tracking is no longer needed
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two bugs that combined to produce unrecoverable 400 errors from backends:
1. loop.go: runStreamStep returned nil on stream-closed-without-done,
causing Run to commit a partial assistant message (with no ToolCalls)
to state even when the backend had already sent tool call frames.
Now returns a real error so state.Update is never called and the
session history stays clean for a retry.
2. context.go: BoundedHistory and BoundedHistoryWithNotice could evict
an assistant[tool_calls] message while its paired tool[result] messages
remained in the tail window, producing a tool message with no preceding
assistant — rejected by strict backends (OpenAI, DeepInfra) with a 400.
Fixed by dropping assistant+tool pairs atomically in the hard-limit loop
and adding a sanitizeHistory pass that strips any remaining orphaned
tool messages before the history is returned.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
In the streaming path all tool calls were being added to skippedCalls,
causing the post-stream Act loop to suppress every 'call' emit. The
'-> tool(args)' lines therefore never appeared in the UI.
Fix: emit the 'call' OutputMsg inside runStreamStep once the stream
reports Done and all tool-call arguments are fully accumulated. The
Act loop still skips re-emitting them (correct), while the UI now
sees one 'call' event per tool invocation.
OutputMsg.Usage carries the raw backend.Usage value when Role=="usage",
rather than formatting it into Content. main.go can now read em.Usage
directly instead of parsing the formatted string.
Add agent/context.go with a ContextBuilder that enforces a character budget
(soft + hard) over the assembled history, evicting oldest non-system messages
first and replacing large tool outputs with truncated summaries.
Update agent/session.go to use ContextBuilder in History(), so the loop
always sees a bounded context window regardless of how many steps have run.
The budget defaults are:
- SoftLimit: 24000 chars (~6k tokens at 4 chars/token)
- HardLimit: 96000 chars (~24k tokens)
- MaxToolOutputChars: 2000
- TailMessages: 6 (always keep last N messages verbatim)
All limits are configurable via ContextConfig.
Add Usage{InputTokens, OutputTokens} to backend.Response. Both
OllamaBackend (prompt_eval_count/eval_count) and OpenAIBackend
(prompt_tokens/completion_tokens) populate it. The loop emits a
'usage' OutputMsg after each Chat call; the UI displays it as
[↑N ↓N tokens].
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Emit a 'call' message before each tool execution so the UI shows
execute_code(...) with full args before the result. Result lines
are now indented with '= ' to visually pair with the call.
System prompt now explicitly says to call execute_code immediately
after planning, without pausing for confirmation, to prevent smaller
models stalling after the planning step.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
BeadState duplicated what execute_code already provides via tool scripts.
Removed it entirely. Added SystemPrompt field to agent.Config, prepended
as a system message each loop iteration. Built-in prompt covers
execute_code usage, sandbox behavior, and skill discovery.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
main.go now uses backend.New() for LLM selection (OLLIE_BACKEND env),
agent.Loop for observe/decide/act/update/terminate, and ollie/exec for
built-in execute_code. Multi-turn sessions persist across user inputs.
MCP tools route through mcpExecutor; unknown tools fall back to built-in.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Implements the core agent loop with two state backends:
- Session: ephemeral in-memory, lives for one process lifetime
- BeadState: bead-backed via 9beads tool scripts (claim/read/complete)
Loop is backend-agnostic; takes a ToolExecutor and emits OutputMsgs.
MaxSteps defaults to 1 for simple prompts; callers set higher for tasks.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Levi Neely
Levi Neely