ForgeBook 1.0.4

Fixes

Token-budget crunch (primary)

v1.0.3 dispatched ~16k input tokens per /forgebook ask or chat-UI call against Anthropic's 10,000 input-tokens/minute organization rate limit, so a single call often tripped a 429 and a second call in the same minute was guaranteed to fail. Five sub-fixes cut per-request size:

• System prompt mod list: now a comma-separated list of modIds (~15 ch/mod) instead of per-mod name/url rows (~100 ch/mod). For a 200-mod pack this saves ~16 kB per turn (~4k tokens). Biggest single win.
• Tool descriptions + schemas: trimmed to one-liners; per-property prose dropped. Another ~1 kB per request body.
• web_search: default limit 5→3, max 10→5, tool_result body hard-capped at 2,000 chars. Keeps turn-2+ request size bounded since tool_result blocks get replayed in the agent loop.

429 latency (interactive budget)

• ClaudeProvider now fast-fails on 429 when the retry-after hint exceeds the 10s interactive budget — the old code would block the worker thread up to maxDelay and almost always retry into a second 429.
• RetryPolicy maxDelay raised 5s → 10s so short retry-after hints (7-8s after a brief burst) are respected verbatim instead of being over-clamped.

/forgebook item silent failure

• RagItemPipeline + CurseForgeClient: CF API failure paths promoted DEBUG→WARN/INFO. Prior silent swallowing made v1.0.3 /forgebook item failures invisible at the default log level.

Tests

365/365 passing. Updated: RetryPolicyTest#6 (10s cap), SystemPromptBuilderTest (modId-only format).