From 8126d78b346d05eba9ac6b21217dc427a339a2a1 Mon Sep 17 00:00:00 2001 From: indifferentketchup Date: Fri, 22 May 2026 14:07:11 +0000 Subject: [PATCH] docs: capture v1.13.7-v1.13.9 invariants in CLAUDE.md MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Five additions surfacing session-discovered constraints future Claude sessions need: - AI SDK v6 includeUsage:true requirement (avoids re-introducing the v1.13.1-A→v1.13.7 NULL-tokens regression) - \n text-delta trim guards in MessageList/MessageBubble + payload.ts failed/empty-assistant skip rules (avoid undoing v1.13.7) - 0.85 × ctx_max overflow formula (v1.13.9) replacing the stale ctx_max - 20k line - New services/system-prompt.ts bullet documenting the v1.13.8 fingerprint instrumentation surface - New services/inference/budget.ts bullet with current BUDGET_NO_AGENT=30 and read-only-tools rationale --- CLAUDE.md | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/CLAUDE.md b/CLAUDE.md index 545a1be..918b6e1 100644 --- a/CLAUDE.md +++ b/CLAUDE.md @@ -47,10 +47,12 @@ Tests: `pnpm -C apps/server test` runs the vitest suite. No test harness on `app Key services: - **`services/inference/`** — Public surface re-exported via `inference/index.ts`; callers import from `./services/inference/index.js` explicitly (NodeNext doesn't honor directory-index resolution). Layout: `turn.ts` (runAssistantTurn / runInference / createInferenceRunner; exports `InferenceFrame`, `InferenceContext`, `TurnArgs`, `StreamResult`), `stream-phase.ts` (streamCompletion as a v1.13.1-A AI SDK adapter + executeStreamPhase), `provider.ts` (`upstreamModel(baseURL, modelId)` wrapping `createOpenAICompatible` against llama-swap), `tool-phase.ts` (executeToolPhase; value back-edges into turn.ts for the runAssistantTurn recursion — cycle safe because deref at call time, not module top-level), `sentinel-summaries.ts` (runCapHitSummary + runDoomLoopSummary + their sentinel inserters), `error-handler.ts` (handleAbortOrError, finalizeCompletion), `payload.ts` (buildMessagesPayload, loadContext, maybeFlagForCompaction, `OpenAiMessage`), `sentinels.ts` (`detectDoomLoop`, `DOOM_LOOP_THRESHOLD`, sentinel predicates), `budget.ts` (resolveToolBudget), `xml-parser.ts` (qwen3.6 XML tool-call fallback — KEEP, AI SDK doesn't handle inline-XML tool calls), `parts.ts` (v1.13.0 dual-write helpers: `partsFromAssistantMessage`, `partsFromToolMessage`, `insertParts`), `prune.ts` (v1.13.4 two-tier compaction; `selectPruneTargets` is the pure decision helper), `types.ts` (`StreamPhaseState`, `DB_FLUSH_INTERVAL_MS`). **`TurnArgs`** is the per-turn state envelope threaded through the `executeToolPhase → runAssistantTurn` recursion; reset in `runInference` at user-message boundary. Add new per-turn state to `TurnArgs`, not module-level closures. -- **AI SDK v6 streamCompletion adapter** (v1.13.1-A; `services/inference/stream-phase.ts`). `streamText` is the underlying call; the BooCode layer above (executeStreamPhase, finalize, dual-write) is shape-preserved via an adapter. Three gotchas the LSP/test suite won't catch: +- **AI SDK v6 streamCompletion adapter** (v1.13.1-A; `services/inference/stream-phase.ts`). `streamText` is the underlying call; the BooCode layer above (executeStreamPhase, finalize, dual-write) is shape-preserved via an adapter. Five gotchas the LSP/test suite won't catch: - **Abort signals are swallowed.** `streamText`'s `fullStream` iterator exits cleanly when `abortSignal` fires — no throw. Post-iteration `if (signal?.aborted) throw ` is required; without it the row finalizes as `complete` instead of `cancelled`. Comment in stream-phase.ts pins this; don't refactor it away. - **Usage lands only at stream end** via `await result.usage` (`inputTokens` / `outputTokens` v6 names → mapped to `promptTokens` / `completionTokens` for the existing onUsage callback). Mid-stream live tok/s is gone vs v1.12.2; ChatThroughput shows a single value at stream end. - **Tools have NO `execute` field.** BooCode dispatches tools in tool-phase.ts, not the AI SDK loop. Only `description` + `inputSchema: jsonSchema(parameters)` — surfacing tool-call parts via `fullStream` and stopping is what we want. + - **`includeUsage: true` MUST be set on `createOpenAICompatible`** in `services/inference/provider.ts`. The adapter defaults it false, omitting `stream_options.include_usage` from the request body; llama-swap then never emits the usage block and `result.usage.inputTokens/outputTokens` resolve to `undefined`. Latent regression from v1.13.1-A through v1.13.7 — every assistant row in that window has `tokens_used`/`ctx_used` NULL. Don't remove this flag during refactor. + - **Tool-call-only turns may emit a leading `\n` text-delta** as the assistant content. `MessageList.flatten`'s `hasText` and `MessageBubble`'s `hasContent` both `.trim()` before the length check — otherwise whitespace-only content renders an empty bubble + ActionRow between every tool call (v1.13.7 fix). `payload.ts:buildMessagesPayload` also skips `status='failed'` AND complete-but-empty (no content, no tool_calls) assistant rows to avoid "Cannot have 2 or more assistant messages at the end of the list" upstream rejections after cap-hit + Continue. - **AI SDK ModelMessage conversion** (`toModelMessages` in stream-phase.ts). Tool messages need a `toolName` for `ToolResultPart` — BooCode's OpenAI-shape history doesn't carry it, so a forward-scan builds a `tool_call_id → toolName` map from prior assistant `tool_calls`. Tool outputs wrapped as `{ type: 'json' | 'text', value }` matching the v6 `ToolResultOutput` union. Assistant messages with reasoning emit a `ReasoningPart` first in the content array (v1.13.1-C). - **`experimental_repairToolCall`** (v1.13.3) wired into `streamText` to keep the stream alive when qwen3.6 emits malformed tool args. Pass-through implementation — logs the bad call and returns it unmodified; `executeToolPhase`'s existing zod-reject error path routes it to the model on the next turn. - **`chat_status` frame shape** (published via `broker.publishUser`) — `status: 'streaming' | 'tool_running' | 'waiting_for_input' | 'idle' | 'error'` (widened from `working|idle|error` in v1.12.1). Frontend `useChatStatus` derives `idle_warm` (<30s since idle) vs `idle_cold`. `ChatThroughput` renders inline beside `StatusDot` only when streaming or tool_running, fed by 500ms-throttled `'usage'` WS frames (`completion_tokens` + `ctx_used` + `ctx_max`). The `POST /api/chats/:id/discard_stale` endpoint exists to mark a stuck-streaming row as `failed` when the frontend's 60s no-token-activity timer (`ChatPane` content-length watcher) gives up. @@ -58,7 +60,9 @@ Key services: - **Periodic 60s sweeper** in `apps/server/src/index.ts` (v1.13.3 + v1.13.5). Same `setInterval` runs `sweepStaleStreaming` (marks `messages.status='streaming'` older than 5 min as `failed`, publishes `chat_status='idle'` so the UI dot drops) and `cleanupTruncations` (TTL + orphan reap of tmpfs truncation files). `app.addHook('onClose')` clears the timer. No-op when nothing to reap. - **`services/broker.ts`** — In-memory pub/sub with two channel types: per-session (message streaming) and per-user (sidebar updates). No persistence; clients reconnect on restart. - **`services/tools.ts`** — Tool registry (`ALL_TOOLS`, `READ_ONLY_TOOL_NAMES`, `TOOLS_BY_NAME`). Filesystem tools (view_file/list_dir/grep/find_files) go through three guard layers: `path_guard.ts` (workspace scope), `secret_guard.ts` (filename deny list), `url_guard.ts` (SSRF/private-IP block for web_fetch). v1.11.8+ web tools (`web_search`, `web_fetch`) are opt-in per chat via `session.web_search_enabled` (resolved with `project.default_web_search_enabled` fallback) and filtered out of the LLM's tool schema when false. v1.13.5 truncation: when a tool slice cuts content, `services/truncate.ts` stashes the full text on tmpfs at `BOOCODE_TRUNCATION_DIR` (default `/tmp/boocode-truncations`, 0o700) keyed by an opaque `tr_<12 base32 chars>` id, and the `view_truncated_output(id)` tool retrieves it. 5MB cap (matches `view_file`'s `MAX_FILE_BYTES`), 7-day TTL, reaped by the periodic sweeper. Tmpfs path means container restart loses retrieval — acceptable, the model usually has moved on. -- **`services/compaction.ts`** + **`services/model-context.ts`** — v1.11.0 anchored rolling summary (single `summary=true` assistant row per chat, supersedes itself on each compaction). Triggered when `chats.needs_compaction` is set after an inference turn exceeds `usable(ctx_max) = ctx_max - 20k`. **`ctx_max` comes from `model-context.getModelContext()` which fetches `${LLAMA_SWAP_URL}/upstream//props`** — NOT from `parsed.timings.n_ctx` (the stream completion's `timings` doesn't carry n_ctx; that read was dead code until v1.11.3 ripped it out). v1.13.6: `buildHeadPayload` embeds `reasoning_parts` as a `...` prose prefix on the assistant `content` (OpenAI wire shape has no structured reasoning field; the summarizer reads text). Standalone tag when content is empty (tool-call-only turn). `buildHeadPayload` + `OpenAiMessage` exported for test access — keep them exported. +- **`services/compaction.ts`** + **`services/model-context.ts`** — v1.11.0 anchored rolling summary (single `summary=true` assistant row per chat, supersedes itself on each compaction). Triggered when `chats.needs_compaction` is set after an inference turn exceeds `usable(ctx_max) = floor(0.85 × ctx_max)` (v1.13.9 opencode-pattern early trigger; was `ctx_max - 20k` pre-v1.13.9, which gave only 7.6% headroom at 262k and 0 budget for ≤20k contexts). **`ctx_max` comes from `model-context.getModelContext()` which fetches `${LLAMA_SWAP_URL}/upstream//props`** — NOT from `parsed.timings.n_ctx` (the stream completion's `timings` doesn't carry n_ctx; that read was dead code until v1.11.3 ripped it out). First inferences after a boocode boot may have `ctx_max=NULL` if llama-swap hasn't loaded the model yet; negative cache TTL is 60s, recovers on next turn. v1.13.6: `buildHeadPayload` embeds `reasoning_parts` as a `...` prose prefix on the assistant `content` (OpenAI wire shape has no structured reasoning field; the summarizer reads text). Standalone tag when content is empty (tool-call-only turn). `buildHeadPayload` + `OpenAiMessage` exported for test access — keep them exported. +- **`services/system-prompt.ts`** — `buildSystemPrompt` is the string-returning shim; `buildSystemPromptWithFingerprint` is the canonical impl returning `{prompt, fingerprint, drift}`. v1.13.8 instrumentation: SHA-256 of the assembled prefix is logged per `buildMessagesPayload` call (msg `prefix-fingerprint`, level=info); a `Map` observer fires `prefix-drift` (level=warn) on hash change with a field-level `changed_inputs` diff. Smoke proved the prefix is byte-stable across turns in steady-state — the originally-planned `system_prompt_cache` DB table was dropped as redundant against the v1.12.0 input-layer mtime caches (BOOCHAT.md here + AGENTS.md global+per-project in `agents.ts:safeStat`). +- **`services/inference/budget.ts`** — tool-call budgets: `BUDGET_READ_ONLY = 30`, `BUDGET_NON_READ_ONLY = 10` (forward-looking; no write tools yet), `BUDGET_NO_AGENT = 30` (v1.13.7; was 15 — every tool in `ALL_TOOLS` is read-only today, so no-agent mode shares the read-only-agent cap). Per-agent `max_tool_calls` from AGENTS.md frontmatter overrides. - **`messages_with_parts` view** (v1.13.1-B; `schema.sql`). Read sites that need `tool_calls` / `tool_results` / `reasoning_parts` SELECT from this view, NOT `messages` directly. `COALESCE`s parts-table rows over the legacy JSON columns, so pre-v1.13.0 history still resolves. Writes still target `messages`; the v1.13.0 dual-write into `message_parts` keeps both halves in sync. New payload-assembly code must use the view — calling `messages.tool_calls` directly will miss anything written post-v1.13.1-B if the JSON column ever drifts (and dual-write makes that easy to miss). Shapes: `tool_calls jsonb[]`, `tool_results jsonb` single object, `reasoning_parts jsonb[]` of `{text}`. - **`services/file_ops.ts`** — Shared file operation implementations used by both inference tools and HTTP routes. - **`services/auto_name.ts`** — Non-streaming LLM call to generate 4-word session titles after first assistant reply.