All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

• Client dependency-surface migration. The browser half no longer type-imports the removed @deepseek-ai/dsh-client-runtime package: it rides the cordis Context plus the published client packages, and the frozen tool-call block is read through a local structural contract (the published 0.1.1-rc.2 line keeps the block union in the removed runtime package, and the unreleased 0.1.2-alpha.1 host owns it in the unpublished dsh-client-ui-chat). The tsdown external list and the tsconfig paths drop the runtime entries; peers, devDeps, and the client inject manifest drop dsh-client-runtime.
• Envelope-less host calibration. The session-event gate docs now cover 0.1.2-alpha.1, which removed the ignorable envelope and fails closed on unknown event types at read: the probe reports no support there and every draw/generated commit degrades to the in-memory fallback ledger (already the fail-safe behavior).
• Tests derive the call-id brand from the dsh-tools execution contract (tests/call-id.ts) instead of importing CallId from dsh-llm, which host master renamed to ToolCallId.

• Declare the web-client inject packages (@deepseek-ai/dsh-client-connection, @deepseek-ai/dsh-client-locale, @deepseek-ai/dsh-client-runtime, @deepseek-ai/dsh-client-ui-settings, @deepseek-ai/dsh-client-ui-slots, @deepseek-ai/dsh-client-ui-tool) as optional peerDependencies so the bundle composition is explicit and standalone installs stay clean.

• Pluggable engine provider seam (Replicate / fal.ai).

• Presenter pure-function suite (tests/present.spec.ts, 11 tests): presentDrawResult and presentDrawPanel are pinned directly on their arguments — settled/error/foreign/null-call-head blocks, meta fallbacks, regenerate-args parsing, engine-row projection, and the cooldown flag.

• presentDrawPanel now takes an optional now clock (default Date.now) so the cooldown flag is a pure function of its inputs and testable without fake timers; the image_generate tool surface and the settings tab are unchanged.

• The five-language README "Development" section reported a stale 77 tests, 11 suites; it now matches the actual 107 tests, 16 test files.
• src/version.ts referenced a non-existent tests/version.spec.ts; the comment now names the real tripwire (tests/session-events.spec.ts).

• Upgraded every @deepseek-ai/dsh-* devDependency to 0.1.1-rc.2 and the dshWorkshop.compatibility.dshVersions declaration to 0.1.1-rc.2 (rc.2 compatibility release); peers stay >=0.1.0-rc.8 <0.2.0 and @deepseek-ai/cordis stays ^4.0.1.
• Re-verified the rc.2 credential read/write surface (resolve/describe/set/unset) and the attachment saveImage/ImageAttachmentRef/ImageAttachmentLimits face; both remain source-compatible, so no plugin code changes were required.
• The monthly compat workflow now installs and smokes against DeepSeek Harness 0.1.1-rc.2.

• Upgraded every @deepseek-ai/dsh-* peer to >=0.1.0-rc.8 <0.2.0 and every devDependency to 0.1.0-rc.8 (rc.8 compatibility release); @deepseek-ai/cordis stays at ^4.0.1.
• tests/harness.ts pins the rc.8 ImageAttachmentLimits face (adds maxImageDimension).

• Sessions that used image generation no longer refuse to reopen on rc.6/rc.7 hosts (issue #2). draw/generated is declared only by this package, so it sits outside the host's static KNOWN_SESSION_EVENT_TYPES whitelist, and those hosts' Session.append cannot stamp the ignorable envelope — every generated image left the log unloadable after restart with SessionFormatUnsupportedError. The append now goes through an adaptive gate (src/event-gate.ts): the event is logged only when the host knows the type or a mount-time probe on a detached SessionStore proves envelope support. On gated hosts the accounting payload rides a new in-memory fallback ledger (WeakMap-keyed by session), so quota stays exact for the live session and the log stays reloadable; durable, log-folded accounting resumes on hosts with a plugin event surface.
• Regression coverage: tests/event-gate.spec.ts pins the gate decisions, the rc.6 probe reading, fallback-ledger quota accounting, and the append-failure degradation; the assembly regenerate test now asserts the rc.6-safe behavior (clean log, exact quota).

• Initial release: unified image_generate tool with config-driven engine routing, health-aware fallback, durable attachments, per-session quota accounting, and credential-reference key storage.

• Unified image_generate tool: standard parameters (prompt/size/count/quality/style/engine) translated per engine; canonical value carries durable attachment references plus quota and routing facts; render embeds the images as attachment content blocks.
• Config-driven engine chain with two shipped presets (OpenAI Images gpt-image-1, Zhipu CogView cogview-3-flash) and any OpenAI-compatible endpoint; per-engine size maps, quality/style support, response format, and media type.
• Health-aware fallback router: chain order with engine-override promotion, consecutive-failure cooldown, disabled-engine skipping, and per-attempt audit views.
• Per-session quota accounting folded from the durable draw/generated session event (generation calls + image bytes), enforced before engine spend and before attachment storage.
• draw Typert Remote service (draw/status, draw/probe, draw/setCredential, draw/unsetCredential, draw/regenerate) with a hand-written host ./typert manifest shared with the client Remote contribution.
• Browser half: keyed tool.call.toolview result card (engine/quota facts + regenerate action) and a Plugins settings tab (engine chain, credential references, probes, quota limits).
• API keys stay credential references (environment-variable names) resolved per call through the official ctx.credentials seam — values are never logged or snapshotted.

• HttpRequest.body carries Uint8Array<ArrayBuffer>; the fetch transport spreads optional headers/body so exactOptionalPropertyTypes holds.
• Session.append uses the two-argument form (the pinned 0.1.0-rc.6 peers have no append-envelope option).

• translateRequest defaults size/count for engines reached with a normalized request missing them.