Agent Smith 1.0.3

Release 1.0.3 of Agent Smith.

Install

1. Download the DMG asset below.
2. Open it and drag Agent Smith to Applications.
3. Launch the app from Applications.

Requires macOS 26.2 or later. The app is Developer ID signed and notarized by Apple.

Changes

• Release 1.0.3 (00826f7)
• Derive the inspector's display state in the model (94ce398)
• docs: onChange warning investigation — exact sites, A/B, skipped actions (8c2b5ce)
• docs: onChange per-frame warning investigation (open) (8a782e0)
• Attach a file named twice in one tool call once (6b937fe)
• Fix two gaps from the final review (a3e7df5)
• task_complete: don't attach a deliverable's evidence file twice (280b420)
• Close three withdrawal races found in review (8001872)
• Close the remaining withdrawal and provider-merge gaps (6323191)
• Fix defects from reviewing the withdrawal and model-refresh work (c69b147)
• Security Agent and summarizer pick up model changes live (issue #9 gap 2) (5e3aa05)
• Cancelling a watch withdraws its handed-off deliveries (bce8da1)
• Fix defects from the final task-watch reviews (ec3020c)
• Failure-injection tests for task events; fix held effects at launch (06b3bfe)
• Task watch authoring: Smith tools, Task Detail section, Timers tab (15c5f94)
• macOS notifications for task watches, with click-through to Task Detail (294c431)
• Start holds and typed start origins, enforced at one gate (a93f0fe)
• Task watches: model, firing in the status write, typed broker settlement (a7cdad6)
• Durable Smith briefing: transition effects, held release, acknowledged delivery (2acfec1)
• Task status funnel: one writer, typed causes, ordered events (ee8fb43)
• Truthful persistence: durable watermarks, store-owned tasks.json (35cab93)
• docs: revise task-watch plan after code review (6e4e089)
• Fix cold-boot recovery of submitted work (9056501)
• README badges; social preview image (8e402bb)
• Roadmap: link good-first issues #24-#34 and scheduled-time bug #23 (378c0cf)
• Roadmap: link good-first issues #19-#22 (1b3856a)
• docs: align Smith briefing phase with the durability decision (3d5490c)
• docs: resolve task-watch design decisions; record architecture in CLAUDE.md (d34f445)
• Roadmap: link P0/P1 items to issues; drop items not pursued (fa8d735)
• docs: plan task state events and task watches (56baba0)
• Move internal plans and audits out of the repo root (884c1c2)
• Audit the roadmap against the code; add contributor docs (12ae7ba)
• Apply worker-capacity changes to the live run (f363cbb)
• Merge pull request #7 from drewster99/copilot/fix-invalidating-cached-login-path (f333886)
• Merge pull request #8 from drewster99/copilot/fix-mcpservereditor-sheet-save-issue (b6130cf)
• Fix findings from the 24-hour recheck of inspector and task-action work (2a47066)
• docs: close remaining inspector acceptance items with evidence (4d691f6)
• Validator model line offers no empty stats popover; fix MLX test invocation doc (341af23)
• docs: inspector data-source rules; plan acceptance status (62f807e)
• Inspector phase 7: fixes from review and live smoke test (8fb2080)
• Inspector phase 6: Memory mutations and consolidation in the activity feed (5a0c82d)
• Inspector phase 5: typed Memory query activity with exact results (de5aff4)
• Inspector phase 4: Summarizer call capture and pop-out-only card (770aa39)
• Inspector phase 3: Validator pop-out backed by task verdict ledgers (aaff80e)
• Inspector phase 2: Security Agent request capture and pop-out window (7ebdabc)
• Inspector phase 1: honest LLM-call retention and failed-call records (94d2e75)
• Surface server-declared LLM failures at once; shrink context after 3 memory refusals (06b2f97)
• docs: plan inspector improvements (1a53806)
• Post Retry/Run Again as system notices; add inline Resume/Undelete (3c22389)
• Post user task-action notices to Smith as system, not as the user (9957345)
• Pre-clear Smith stop controls (927336f)
• Stabilize security evaluator prompt prefixes (19a7e93)
• Suppress malformed empty-response markers (9eff9d2)
• Fix regressions found in 72-hour review (a728172)
• Clarify completed-task follow-up handling (cbaaa5a)
• Show effective output token limits (18b6a9d)
• Recover truncated Qwen tool calls (8ecfc12)
• Address Brown termination by task (c94c29c)
• Declare local network access (e0479e5)
• Add batch deep model probing (785e617)
• Update SwiftLLMKit to 0.0.207 (3f56930)
• Keep deep probes on probe-safe requests (aa9acc7)
• Surface capability probe failures (fa1deb3)
• Show model probe depth (b9736b0)
• Fix local metadata mapping selection (29fd4e9)
• Add deep model capability probing (729c30b)
• Add local metadata mapping and flag help (42c1df7)
• Initial plan (9b49241)
• Initial plan (b831f6d)