{"content":"IDBots dev journal — perf/dsh-agent-speed d80f0ad5: aligned the DeepSeek V4 family output-token ceiling with upstream deepseek-harness, 32,768 -> 256,000.\n\nWhy: thinking-mode reasoning shares the output budget, so the old 32K ceiling truncated effort-high/effort-max steps mid-thought and triggered the paid truncated-turn auto-continue (max-tokens turn/end reasons appear in IDBots session logs, none upstream). Upstream's DEFAULT_MAX_TOKENS is 256K and its flash catalog declares no per-model cap, so 256K restores parity. The API itself allows up to 384K; billing is by actual tokens, so the higher declared ceiling costs nothing for short replies.\n\nImplementation: coworkModelLimits constants raised for deepseek-flash + v4-pro; the runtime-config generator's native-route fallback aligned; a one-shot idempotent startup migration (services/deepSeekOutputCeilingMigration) rewrites provider rows still pinning the legacy 32,768 default — user-customized values and non-DeepSeek rows are untouched (provider-model rows take precedence over the catalog, so without the migration upgraded installs would keep 32K forever). The MetaApp bridge's own 32K validation limit is a separate API contract and stays unchanged. Tests updated (coworkModelLimits) + new migration suite; 42/42 pass.","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}