v0.7.1

Review follow-ups:- Fix agent/disposed listener leak (single global listener via disposer map)- Legacy deepseek-v4-flash / deepseek-v4-pro catalog ids declare inputModalities [text, image]- Skill content templated from configured routes (buildSkillContent)- 47 tests passing (router lifecycle + omit-to-inherit regression guards)

v0.7.0

• V4.1 Flash (deepseek-flash, 2026-09-10): planner, executor, and vision defaults all move to deepseek-flash (native multimodal, 1M context, 384K max output). deepseek-v4-pro / deepseek-v4-flash / deepseek-v4-flash-vision-exp are retired server-side (all route to V4.1 Flash; deepseek-v4-pro follows on Sep 14, 2026) — the role split stays in the config so the planner flips back with a one-line change when V4.1-Pro launches. Vision routing is now a no-op by default (same model) and kept for explicitness and the future split. - Bundle patch: llm-deepseek catalog lists deepseek-flash with inputModalities [text, image] and corrected maxTokens 384000 (was 256000 from the v4 preview docs); retired v4 ids kept as compat aliases. - Prompt section, pro-flash-routing skill, GUI card copy, README, and tests updated for the unified default.

v0.6.3

Fix vision session-shape gaps from review:

• Fix sessionHasImage missing tool/result events: tool screenshots now route to vision (shapes verified against harness SessionEventMap; robust data.message.content ?? data.content extraction).
• Add testable shouldUseVision(events, vision) wiring helper; agent/request uses it; re-export sessionHasImage + shouldUseVision from the public entry.
• Clarify sticky vision in README and code comments (entire-log scan until compaction/pruning) and mark hasImageContent deprecated (payload never carries messages).

v0.6.2

Fix vision routing never firing: the agent/request waterfall payload does not carry messages, so image detection now reads the session event log (user/message and assistant/message content, including images nested in tool-result blocks).

v0.6.1

Fix the vision-support bundle patch: the llm-deepseek catalog row and attachment-local limits now use direct id-targeted entries, so installing the plugin actually adds deepseek-v4-flash-vision-exp to the catalog and raises image admission limits.

v0.6.0

Vision routing (opt-in, default off): image-bearing requests are stamped with deepseek-v4-flash-vision-exp on deepseek-official, from every role. Includes the vision catalog entry + raised attachment-local image limits, a Vision section on the Model router card, and a conditional prompt line.

v0.5.0

Browser half: the package now ships a client bundle the harness serves automatically. Settings -> Plugins renders a Model router card with a live enabled switch (plus overridden badge, reset button, and read-only route lines). Changes apply live and persist in settings.yaml under model-router:.

v0.4.0

GUI on/off toggle: new enabled config (default true), surfaced as a live settings section. Turning it off stops all request rewrites and unregisters the prompt section and skill. Applies immediately, no restart; persists in settings.yaml under model-router:.

v0.3.0

Add error-driven effort escalation: escalateOnError/escalateTo/recoverySteps bump reasoning effort after a failed execution step. Also releases the per-role reasoningEffort + maxTokens and mode: plan features from 0.2.0.

v0.2.0

Add per-role reasoningEffort and maxTokens, plus a plan routing mode that reserves pro for plan mode. Updated skill and prompt section to keep the planner context lean.