Run/RunStream took the new turn as a bare string, which had nowhere
to carry ContentPart attachments. Both now take an llm.Message
(Role is forced to RoleUser regardless of what the caller sets), so a
caller building a multimodal turn just fills in Content/Parts on it
instead of the loop needing a second, parallel parameter.
subagent.go and every test call site are updated to wrap their string
prompt as llm.Message{Role: llm.RoleUser, Content: ...} — SubAgent.Run
itself is untouched, it still takes a plain task string.
Message gains an optional Parts []ContentPart alongside the existing
plain-text Content, so a turn can carry text plus image/video
attachments. Content stays the single source of truth for every
existing text-only caller (sidebar.go, memory_tools.go, etc. are
untouched); Parts only matters to a provider client when non-empty.
openai and llamacpp (both OpenAI-compatible) serialize Parts into the
standard text/image_url content-array shape; llamacpp additionally
passes video through as a best-effort video_url part, since llama.cpp
itself has no video support but the whole point of this client is the
user's own OpenAI-compatible server sitting in front of a
video-capable model — the server decides whether it understands it,
not this client. anthropic converts image parts to its base64 image
content block, and rejects a video part outright with a clear error:
the Messages API has no video block type at all, so sending one would
just produce a confusing 400 instead.
ProviderCapabilities gains SupportsVideo, true only for llamacpp.
Two failure modes seen live with Qwen3.6 on llama.cpp ended turns silently
mid-task:
- The model writes its tool call as plain text inside its reasoning, the
server never parses it, and the round ends with nothing executed. The
loop now detects the markers and nudges the model to re-issue the call
for real (max 2 per turn).
- llama.cpp silently ignores the max_thinking_tokens field, so a model in
a reasoning spiral ran until max_tokens (seen live: 25k+ tokens of
nonstop thinking, ~20 min). The llamacpp client now enforces the budget
client-side during Stream: once exceeded while the round is still pure
reasoning, it cuts with FinishThinkingBudget and aborts the request
(freeing the server slot); the loop answers with its own corrective
nudge, on a separate counter.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
NetworkPolicy validates scheme/host and re-validates resolved IPs at dial
time and on redirects (DNS-rebinding defense), with cloud metadata
endpoints always blocked. Redact masks known credential shapes (OpenAI/
Anthropic/GitHub/AWS/Slack/Google keys, PEM blocks, JWTs) in tool output.
WrapUntrusted fences fetched web content against prompt injection, paired
with UntrustedContentInstruction for the system prompt.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Fragments now carry a memory type in metadata (legacy fragments count as
procedural) with SearchByType filtering, and EpisodeCapture summarizes a
finished turn with the local LLM and stores it as episodic memory, so the
agent can answer "what did we do yesterday?". Includes an E2E test against
a live llama.cpp server (gated) and taxonomy unit tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Formatting only (struct field alignment, import ordering) across the
files that didn't comply — no semantic changes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Run() reported "max iterations reached" even when a valid final answer
arrived exactly on the last allowed iteration, throwing the response
away; a completed flag now distinguishes success from budget exhaustion.
- Tool schemas are marshaled once per Run/RunStream instead of once per
loop iteration — they never change between iterations.
- RunStream's content gate (which hides raw deltas during a tool-call
round) also swallowed that round's token usage, so callers only ever saw
the final round's count and context tracking lagged exactly when the
context grew fastest. Usage is now forwarded in its own chunk.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- llamacpp: 4MB SSE scanner buffer (the 64KB bufio.Scanner default killed
streams whose single line exceeded it, e.g. a write tool call carrying a
whole file) and an empty-choices guard in toResponse instead of a panic;
request payload now uses bytes.NewReader (drops a full string copy).
- anthropic: Capabilities() reported a 1M-token context window for any
non-haiku model. Callers use that number to decide when to compact, so
compaction would have fired far too late and requests overflowed the
real window. Default is now the standard 200k, configurable via
Config.ContextWindow for extended-window models/plans.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The Stream() path was broken end to end:
- requests always went out with "stream": false, so the SSE parser found
no data lines and every stream ended empty
- Config.Model was discarded at construction, and the agent loop never
sets req.Model, so requests carried an empty model (hard API error)
- tool-call deltas were ignored entirely: the agent never executed tools
over a stream with this provider (which also backs the ollama type)
- usage was neither requested nor parsed, so token tracking stayed at 0
Now mirrors the proven llamacpp client: stream flag + stream_options
.include_usage, per-index tool-call fragment accumulation flushed on
finish_reason, usage passthrough, a 4MB SSE scanner buffer (64KB default
kills the stream on large tool arguments), and an empty-choices guard in
toResponse instead of a panic.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Bug found while building rony-harness's "Edited Files" info panel: it
scanned the transcript for tool-call messages carrying write/edit
arguments, but those messages never appeared — not even for a plain,
successful top-level write with no delegation involved.
Root cause: RunStream's content-streaming gate (`if !hasToolCalls &&
(hasContent || hasUsage) { yield(chunk, nil) }`) suppresses yielding
*any* chunk once a tool call is seen in that iteration, including the
chunk carrying the tool call itself. So chunk.ToolCalls was executed
internally (hence approvals and results worked) but never yielded to
the caller. Every caller-side "which tool got called" hook depending on
the stream (not the Approver callback) was therefore dead code.
Fix: yield a dedicated chunk carrying just the executed ToolCalls right
after running them, independent of the content-streaming gate below.
Updated TestRun_Stream_WithToolCalls, which asserted the old (buggy)
1-chunk behavior.
Implements docs/phase2.md §5 (Sub-agents), pulled forward from the
harness's item 2 work: SubAgent/SubAgentRegistry let a caller run a
nested agent.Loop with its own persona/tools/iteration cap and get its
final response back. Run doesn't set Approver/Sandbox, so a single Ask
approval on the caller's own delegating tool covers the whole nested
run (Ask-permission tools execute unprompted when Config.Approver is
nil). rony-harness consumes this for its delegate tool (builder/planner).
- pkg/llm/providers/anthropic: new client implementation (was previously
imported by rony-harness but never committed here, so a fresh clone
wouldn't build)
- pkg/llm/providers/llamacpp: Config/Client gain the full local-model
sampling surface (max_tokens, context_window, top_k/top_p/min_p,
presence/repetition penalty, max_thinking_tokens) to match the
llamacpp-local* entries added to configs/ai_providers.yaml
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Export persona.DiscoverAgentsMD and add agent.Config.AgentsMD so project
and global AGENTS.md rules actually reach the model. Previously
buildInitialMessages always called AssembleSystemPrompt with an empty
string, so no AGENTS.md content was ever injected despite the discovery
logic already existing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- agent/loop.go: Record assistant message with ToolCalls before tool results,
and set ToolCallID on tool-result messages so follow-up requests have complete
context (prevents models from losing track of already-attempted tools).
- llm/providers/llamacpp/client.go: Buffer fragmented tool call deltas during
streaming, assemble them into complete calls when finish_reason arrives.
Add ToolCalls, ToolCallID, Name fields to request building.
- llm/providers/openai/client.go: Send ToolCalls, ToolCallID, Name when
building chat requests so messages are wire-format correct.
- llm/types.go: Add ToolCalls field to Message struct for serialization
back into conversation history.
- agent/integration_test.go: Move integration test skip from TestMain to a
per-test skipUnlessIntegration() so it doesn't hide other package tests.
- sandbox & tools: Add edge-case tests (relative traversal, array paths,
non-path strings, zero-value guards, sentinel errors).
Backend.Upsert never received the fragment's Content, so ChromaDB (and
any backend) stored the vector but silently dropped the actual text —
saved memories had nothing to retrieve later. Backend.Search now also
takes the raw query text, and a failed/missing embedding no longer
hard-fails Add/Search: it degrades to a nil vector so a lexical-capable
backend can still index/find the content (Chroma has no such fallback
and now says so explicitly instead of misbehaving).
Adds pkg/rag/backends/sqlitevec: a zero-dependency backend (pure-Go
SQLite, no external service) that does cosine similarity when a real
embedding vector is available and falls back to FTS5/BM25 full-text
search otherwise. Adds pkg/rag/embeddings.OpenAICompatible, covering
both a local llama.cpp server (`--embeddings` enabled) and real OpenAI
(or any OpenAI-shaped /embeddings endpoint) through the same client.
Also fixes token usage tracking for llama.cpp streaming: the client
never requested `stream_options.include_usage` nor parsed a usage-only
SSE event, and even when present, the agent loop's RunStream dropped
any chunk with no Delta/ReasoningDelta — silently discarding the only
chunk that carries usage.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- Add optional history parameter to Run/RunStream for conversation context
- Add Reasoning/ReasoningDelta fields to completion and stream types
- Update llama.cpp adapter to propagate reasoning content from responses
- Default persona language now adapts to the user's language dynamically
- Change status from 'design/spec only' to actual implementation
- Add pkg/*/README.md as a reference in the docs table
- Update build/test section for existing codebase
- Add test coverage summary table
- Update testing conventions with current mock types
- Add ErrPersonaNotFound and ErrConfigNotFound to sentinel errors
- Add ChatTemplateKwargs field to llm.CompletionRequest
- Propagate kwargs through agent loop in both Run() and RunStream()
- Pass kwargs to llama.cpp client chat request
- Fix tool schema marshaling to include type/function wrapper
- Fix stream indentation logic in RunStream with responseBuilder
- Remove indirect marker from uuid dependency
