Batch 3c: when an agent has llama_extra_args in AGENTS.md, provider.ts
routes inference through LLAMA_SIDECAR_URL instead of LLAMA_SWAP_URL.
X-Agent-Flags header built from the agent's flags. Boot-time guard
refuses to start if any agent has llama_extra_args but LLAMA_SIDECAR_URL
is unset. PrefixFingerprint gains a route field (swap/sidecar) for
per-turn visibility. 9 provider tests.
AGENTS.md tool gap: all agents (except Prompt Builder) were missing 8
tools that were added after the original tool lists were written:
request_read_access, view_truncated_output, ask_user_input, git_status,
get_blast_radius, get_hot_files, get_middleware, get_routes. The missing
request_read_access caused silent "permission denied" when reading files
outside the project root.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Five fixes for latent regressions surfaced during the v1.13.x.cosmetic
revert investigation. None alter schema or compaction; all cleanup
against the v1.13.1-A AI SDK migration's hidden surface.
(1) provider.ts — includeUsage: true on createOpenAICompatible.
@ai-sdk/openai-compatible defaults this false, omitting
stream_options.include_usage from the request body; llama-swap never
emitted the usage block, so result.usage.inputTokens/outputTokens
resolved undefined and tokens_used / ctx_used landed NULL in every
assistant row since v1.13.1-A. No historical backfill.
(2) MessageList.tsx — hasText = m.content.trim().length > 0.
AI SDK v6 streaming occasionally emits a leading "\n" text-delta on
tool-call-only turns; the literal newline passed length > 0 and
rendered an empty bubble + ActionRow between every tool call. Trim
catches it without changing semantics for genuine content.
(3) MessageBubble.tsx — same trim on hasContent for the no-tool-calls
path. Defensive symmetry with MessageList.flatten.
(4) payload.ts — buildMessagesPayload skips assistant rows with
status='failed' AND assistant rows with status='complete' + empty
content + no tool_calls. Without this, a trailing empty/failed
assistant + the next attempt's placeholder produced "Cannot have 2
or more assistant messages at the end of the list" rejections from
the OpenAI-compatible upstream after cap-hit + Continue.
(5) budget.ts — BUDGET_NO_AGENT 15 → 30. Every tool in ALL_TOOLS is
read-only today; the 15-cap was forward-looking for write tools that
haven't landed. No-agent mode now matches BUDGET_READ_ONLY.
47 LoC across 5 files. 190/190 server tests pass.
Verified live: new assistant turns populate StatsLine token data;
single-tool-call turns no longer render the stray empty-bubble +
ActionRow between tool calls; Continue after cap-hit no longer hits
the trailing-assistant API rejection.
- Add ai@^6 and @ai-sdk/openai-compatible@^2 to apps/server.
- New services/inference/provider.ts: createOpenAICompatible against
llama-swap (baseURL threaded from config.LLAMA_SWAP_URL, cached per
baseURL). No apiKey — Authelia + Tailscale gate llama-swap, not keys.
- streamCompletion rewritten as an adapter over streamText. AI SDK
fullStream parts (text-delta, tool-call, finish, error) map back to
the legacy {content?, tool_calls?, finishReason} StreamResult shape
that executeStreamPhase already consumes. No layer above
streamCompletion changes.
- toModelMessages converts BooCode's OpenAI-shaped history to AI SDK
ModelMessage[]; tool messages need toolName which we look up by
scanning earlier assistant tool_calls for the matching id.
- buildAiTools wraps BooCode's JSON-schema tool defs via
tool({ inputSchema: jsonSchema(parameters) }) with NO execute —
BooCode dispatches tools in tool-phase.ts, not the AI SDK loop.
- XML fallback parser preserved as-is — qwen3.6 still emits XML tool
calls in text content that the structured tool-call layer misses.
- reasoning-delta parts dropped with a debug-level counter — captured
properly in v1.13.1-C.
- Abort path: streamText({ abortSignal }) wires ctx.signal through, but
AI SDK v6 swallows the abort (fullStream iterator exits cleanly
rather than throwing). Post-iteration `if (signal?.aborted) throw` so
handleAbortOrError owns the row and writes status='cancelled'. Caught
by smoke D; would have shipped as status='complete' on stop otherwise.
- Usage frame reads result.usage (inputTokens / outputTokens v6 names)
AFTER stream drain. Single trailing publish through the existing 500ms
throttle. Known regression: ChatThroughput's live mid-stream tick
(v1.12.2) is gone — it now shows a single value at stream end.
TODO(v1.13.1-followup): interpolate outputTokens during streaming
via a delta-cadence counter (e.g. part.text.length/4 token proxy)
and publish every 500ms; reconcile against result.usage at finish.
- Write-path dual-write from v1.13.0 unaffected.
Read path stays on JSON columns. v1.13.1-B flips reads to message_parts.
Smoke verified end-to-end against running container:
- A. Plain text: status='complete', 1 text part.
- B. Single tool prompt → multi-tool chain (4 calls): every assistant
with tool_calls has 2 parts (text+tool_call), every tool row has
1 part (tool_result).
- C. Multi-step covered by B's chain.
- D. Stop mid-stream: status='cancelled' written via handleAbortOrError
after the post-iteration abort throw.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
