v1.1.0

1.1.0 (2026-10-08)

Features

• --escalate-model — resume the 2nd+ gate repair on a stronger model (#407) (5f4dcd9)
• current-generation models — 5.5/Fable 5.1 prices, aliases repointed, alias resolution not a swap (#417) (8c37e59)
• decision model as fork scorer and eval judge (#413) (f12f011)
• desktop: collapsible icon sidebar + shell alignment + single @/ui seam (#312) (4cb02fa)
• desktop: Kanban board glow-up + project-color hover + rail collapse fix (#315) (7591793)
• desktop: live cost/budget strip on the run viewer (#317) (b9bc59e)
• desktop: show in-flight runs as rows in the Runs table, not a card (#320) (5ec052a)
• Editor chrome: indent guides, bracket matching, status bar (Editor UX 2/7) (SebaBoler/vanguard#357) (#365) (1c04130)
• Editor interaction: multi-cursor and VSCode keybindings (Editor UX 3/7) (SebaBoler/vanguard#358) (#367) (619ea42)
• Editor UX upgrade (SebaBoler/vanguard#352) (#363) (0dd0c8c)
• Linear tasks open GitLab MRs; review-mr skips reviewed heads (#398) (8cd4eea)
• log-only difficulty probe via a decision model (Clef / System One API) (#410) (d95a353)
• review: let review-mr and review-pr take --max-turns (#402) (39e503c)
• s10: task drafts — the Docs page dies, authoring becomes task-first (#349) (f682b78)
• s4.3: create a task from a doc — the app's first irreversible write (#335) (9c8050a)
• s5.1: flows go live — .vanguard/flows/*.hcl discoverable, runnable, writable (#336) (d4272a2)
• s5.2: visual flow editor — Workflow screen reads and writes real flow HCL (#338) (6ab753e)
• s6.1: custom providers go live — repo-configured Anthropic-compatible endpoints (#340) (89320fe)
• s6.2: custom providers in the app — NewRunForm merge, Settings editor, config data-safety (#341) (b93b195)
• s7: shared-types seam — one generated wire contract, mirror class deleted (#342) (828fe3c)
• s8.1: hygiene bundle — double-scroll, drag cancel, input feedback, strip bleed, nav guard (#339) (#343) (62c4fbe)
• s8.2: in-app flow rename + delete (additive deleteFlow method) (#345) (fac70ff)
• s9.1: board read path in core — listTasks/fetchSpec sidecar methods (#346) (45ea16a)
• s9.2: the Rust board dies — one brain (#348) (5ca0227)
• sidecar: give short calls their own pipe; keep cancel pointed at the run child (#334) (69162fb)
• spec: --base flag + auto-fetch origin baseline for the spec pass (#311) (5936a56)
• task-page: full-page editor with a tabbed chat drawer (#350) (24d0122)
• vanguard metrics push + stats --branch — durable run metrics on an orphan branch (#419) (e8c76ef)
• vanguard:model=<m> issue label pins the implementer model per task (#409) (16f7481)
• watch: add --spec-only to run the loop-v1 spec pass alone (#400) (0155ebc)

Bug Fixes

• codex/cursor-only review needs no Anthropic credential; log why a watched item failed (#422) (abe3256)
• desktop: default runs to claude, drop dead zai/--llm-proxy hardcode (#321) (57d0669)
• desktop: Fleet watch-loop toggle falsely reads 'stopped' after navigating away (SebaBoler/vanguard#318) (#319) (4d14e2b)
• dogfood: proxy EPIPE crash-loop killed run #352 again; live-run strip trapped the Runs page (#354) (e016ec0)
• fail fast when every assistant turn is a synthetic API error (#392) (eed565a)
• follow-ups from the last #398 review (#399) (2315cfa)
• gate runs on the pinned sandbox claude CLI, add doctor --fix (#393) (70729db)
• keep truncations visible after repair, annotate gateway swaps in stats, map Claude aliases on openrouter (#406) (6d37155)
• linear: list over GraphQL — vanguard watch --linear was broken on every poll (#333) (4203660)
• log a self-contradicting review verdict (clean verdict, high finding) (#412) (48a485c)
• loop-v1 spec pass researches against --base (#401) (e50d60b)
• pipeline: an incomplete implementer fails the quality gate — no more residue PRs (#356) (d83c3d7)
• pr-review: failed reviews retry instead of posing as success; pin event PR past stale label scan (#344) (727a00a)
• review verdict first — accept a stated verdict without the completion signal, stop calling small diffs too large (#411) (a23f663)
• route repair loops on the implementer model, keep all-attempt cost, stats by model (#404) (45975c1)
• sandbox: coerce undefined execa streams in ExecResult (#382) (194bb6f)
• sandbox: supervise the egress proxy — proxy death bricked run #352 (#353) (6ba7c86)
• spec: resume incomplete tech-spec, keep text on empty result (#380) (39bd9d5)
• spec: tech-spec turn cap 15→30, wire --max-turns, planner 10→15 (#347) (a61940d)
• task-page: dogfood r3 — unbreak doc chat, split conversation/doc naming into InlineEdits (#351) (e68db34)
• watch: carry the stack into failure comments (#381) (5ea9006)