Release Notes
Changed
- Fixed remaining Milestone 3 headless startup-resume parity for primary-agent runs: completed headless runs now persist sessions for later
--continue / --session usage, headless resume now reuses persisted conversation history (not just session metadata), and legacy resumed metadata with primaryAgentId but no primaryAgentPrompt now rehydrates prompt behavior from current config with focused regressions for resume context + metadata restore behavior.
- Implemented and tightened Milestone 3 primary-agent startup/session behavior:
config.yaml now supports first-class primary_agents plus optional default primary_agent, CLI accepts --agent <id> with explicit provider/model precedence compatibility, and mixed --agent + --provider/--model startup behavior is now covered by focused precedence tests. Session persistence/restore now stores and re-applies both primaryAgentId and primaryAgentPrompt metadata, resume/session-switch flows now restore the active primary-agent prompt behavior (not just the profile ID), startup resume still skips session metadata when an explicit --agent override is present, and runtime prompt layering now composes startup primary-agent behavior with Plan/Code mode suffix behavior so mode switching no longer overwrites primary-agent behavior.
- Closed the Milestone 2 backend-foundation subagent follow-up with a narrow backend-first pass: the effective catalog now includes the default
subagent alias, subagent config path/read/write ownership is further centralized in ava-config (including canonical compatibility path helpers plus backend-owned project/global write seams), backend/runtime consumers now use those helpers instead of local filename joins where practical, backend-gate now mandates focused ava-config + delegated runtime signoff checks, delegated smoke now asserts material SubAgentComplete payload fields, and the remaining user-facing legacy config hint now points to .ava/subagents.toml with explicit legacy .ava/agents.toml compatibility wording.
- Implemented the backend foundation slice of the subagent rework milestone: added a dedicated
ava-agent::subagents module for backend-owned subagent runtime/catalog helpers, introduced subagents.toml as the preferred config path with explicit legacy agents.toml compatibility input, switched backend/benchmark config writes to subagents.toml, added a built-in general subagent profile plus config schema support for per-subagent description, and added focused config/runtime tests for precedence and trusted-vs-untrusted project loading behavior.
- Extended the backend subagent foundation seam with an explicit effective-catalog API (
AgentStack::effective_subagents), centralized project/global subagent config path helpers in ava-config so backend-owned writes/read-target selection stays subagents.toml-first, and added focused compatibility regressions for trusted/untrusted legacy .ava/agents.toml fallback behavior.
- Added cwd overrides for CLI, TUI, and headless runs:
ava --cwd <path> now takes priority, AVA_WORKING_DIRECTORY provides the same fallback for automation, and startup switches the real process working directory early so trust loading, file scanning, workspace display, and agent runtime/project discovery all stay aligned on that cwd without needing a separate repo-root flag.
- Added a scoped Milestone 1 subagent rework blueprint in
docs/architecture/subagent-rework-blueprint-m1.md, linked it from the architecture index/manifest plus the active backlog, and locked the next delegation milestones around a backend-owned catalog/config seam, headless-first validation, and later TUI/desktop/web adoption instead of mixing the UI rework into the first implementation slice.
- Finished the default-path XDG migration for fresh installs so AVA now uses lowercase XDG app directories (
$XDG_CONFIG_HOME/ava, $XDG_DATA_HOME/ava, $XDG_STATE_HOME/ava, $XDG_CACHE_HOME/ava) for global config, credentials/data, logs/state, and caches, while existing ~/.ava installs still resolve there for compatibility until migrated; global instruction discovery now reads $XDG_CONFIG_HOME/ava/AGENTS.md on the XDG path.
- Added an
AVA_PURE startup env var for automation/isolation flows: when set to a truthy value (1, true, yes, on), AVA now skips all power-plugin auto-loading from both $XDG_CONFIG_HOME/ava/plugins and project-local .ava/plugins; added focused stack coverage and documented the new env var in the runtime reference.
- Fixed project instruction discovery so startup prompt assembly now cascades trusted
AGENTS.md files from the repository or trusted boundary down to the active working directory instead of loading only the current directory's file, and runtime file-context activation now concatenates the matching AGENTS.md chain within the current workspace subtree instead of choosing only the nearest file.
- Fixed first-message session auto-titling across the desktop/web shell path: AVA now treats both current and legacy placeholder names (
New Chat and New Session) as eligible for automatic renaming, web session creation no longer persists the placeholder as a real backend title, and successful web-run persistence now backfills a title from the first user message when none was explicitly set so conversation naming behaves consistently with the TUI heuristic.
- Tightened the TUI subagent inspection UX: completed subagent cards now show the specialist label more clearly, open child conversations on click instead of hijacking bare
Enter, and child-session views now reconstruct persisted tool-call/tool-result history so delegated chats render much closer to the main conversation instead of collapsing into broken result-only output.
- Simplified the public docs surface so user-facing entrypoints stay focused on install, first run, configuration, local usage, troubleshooting, and reference:
README.md now points users to docs/index.md instead of the internal docs map, docs/index.md and docs/README.md now clearly separate public vs internal paths, docs/how-to/install.md was trimmed into a shorter install chooser, tutorial/how-to workflow duplication was reduced by making docs/tutorials/your-first-workflow.md the primary public walkthrough, and contributor-heavy verification/build pages are now labeled and linked from contributor docs instead of the main public path.
- Trimmed the remaining public docs pages to sound more like product docs and less like source-code audits: desktop download, CI/headless, Ollama, troubleshooting, security/trust, and install/release pages now use shorter user-facing guidance, fewer implementation caveats, and less maintainer- or code-grounding prose while keeping the deeper detail in the reference and contributor sections.
- Made the docs structure line up more cleanly with Diataxis and current repo reality: added section landing pages for
tutorials/, how-to/, and explanation/, updated root and section _meta.json files to reflect the public/internal split, removed contributor-only pages from the public how-to nav, fixed the stale docs/project/_meta.json entry for the old eval-plan filename, removed the broken README screenshot reference, and corrected the user-facing post-install hint in install-from-source.sh from ava auth add to the real ava auth login command.
- Tightened the reference section boundary: moved
scripts-and-build.md into docs/contributing/ so repo workflow commands no longer appear as public reference, marked web-api.md as feature-gated implementation reference for ava serve, and updated the reference navigation/manifest to distinguish user-facing reference from internal integration detail more clearly.
install-from-source.sh now defaults CARGO_BUILD_JOBS to the detected CPU count and auto-enables sccache when available if the caller has not already set CARGO_BUILD_JOBS or RUSTC_WRAPPER, so guided source CLI/desktop builds use explicit full-machine Cargo parallelism plus compiler caching by default while still allowing power users to override or disable either behavior per invocation.
- Tightened the public docs entry flow so normal users, power users, and AI tooling can all choose the right path faster:
README.md, docs/index.md, docs/README.md, docs/how-to/install.md, docs/how-to/download-desktop.md, docs/how-to/run-locally.md, docs/how-to/configure.md, docs/tutorials/first-run.md, docs/reference/README.md, docs/reference/install-and-release-paths.md, and docs/troubleshooting/common-errors.md now use clearer decision-first guidance, earlier Windows/desktop visibility, explicit CLI-vs-desktop distinctions, and more consistent wording around configuration, local runs, source builds, release binaries, web mode, and troubleshooting entrypoints.
- Normalized the repo's public release versioning back to
0.6.0 across the Rust workspace and desktop manifest, and hardened install.sh so the public installer falls back to the newest published release that actually contains the current platform archive instead of failing on a partially published latest tag.
- Hardened the CLI self-updater for the restored
0.6.0 release line: update checks now invalidate stale cached results when the installed AVA version changes, and binary downloads now recognize the current ava-tui-*.tar.xz release assets instead of only the older ava-*.tar.gz layout.
- Tightened the Milestone 9 desktop session recovery/materialization seam with a minimal replay-fidelity fix: active-session sync snapshots now preserve tool-role entries plus replay-critical per-message metadata (
toolCalls, toolCallId, visibility flags, and related fields), the desktop recovery mapper no longer drops common tool-heavy transcript content, and focused frontend/Rust regressions now cover recovered tool-context round-tripping into the replay path.
- Tightened the final Milestone 8 broad-review blockers with a minimal safety pass: desktop session-open/switch flows now omit
workingDirectory unless it can be resolved from the target session's own project binding (no '.' or ambient-project fallback), and ava serve now also places CLI-agent discovery plus high-risk plugin mount/route GET surfaces behind the existing control-token middleware, with focused TS/Rust regressions and web-doc note updates.
- Tightened the remaining Milestone 7 browser session fallback correctness seam with one minimal pass: web fallback session writes now fail closed instead of reporting fake success on create/rename/archive/unarchive/delete errors, deleting an archived session now also clears archived frontend state, and the web session create/list seam now round-trips
project_id through session metadata so project-scoped browser session lists no longer silently drift.
- Added a compact platform download matrix to
README.md so power users can scan the expected CLI archive names and desktop bundle types for Linux, macOS, and Windows without digging through deeper install docs.
- Tightened the desktop maintainer release doc with a fuller cross-platform artifact collection/upload example and a concrete checklist for Linux, macOS, Windows, and updater payloads; refreshed the root
README.md install block into a clearer CLI-vs-desktop chooser so users can scan the two product surfaces faster and jump to the dedicated install pages.
- Added a dedicated public desktop-download guide in
docs/how-to/download-desktop.md, linked it into the docs navigation/index/manifest, and documented the current maintainer-facing release gaps that still block guaranteed desktop downloads on every version in docs/contributing/releasing.md; the public install docs now point desktop users to that dedicated page instead of folding everything into one generic install section.
- Split the public install story more explicitly into two product surfaces,
ava CLI/TUI and AVA Desktop, modeled after clearer dual-surface docs like OpenCode's: README.md now has separate CLI and desktop install sections, docs/how-to/install.md now separates CLI install paths from desktop download/build paths and adds exact CLI asset-name examples by platform, and docs/reference/install-and-release-paths.md now documents the same CLI-vs-desktop split plus the current desktop bundle types produced by the repo's Tauri release flow.
- Hardened the public installation docs for both normal users and power users: the root
README.md now routes users more clearly between fast binary install, manual release downloads, source installs, and the guided source installer; docs/how-to/install.md now explicitly splits quick/manual/source/desktop paths, documents AVA_INSTALL_DIR, PATH reload behavior, Windows/manual release guidance, and the repo's install-from-source.sh; and docs/reference/install-and-release-paths.md now reflects the PowerShell/manual release surface and the current guided source-build path more accurately.
- Hardened
ava serve with a minimal local-first security slice: the default bind host is now 127.0.0.1, browser origins default to loopback-only instead of open CORS, privileged HTTP control-plane routes plus /ws now require a simple control token, the browser frontend can source that token from VITE_AVA_SERVER_TOKEN or a one-time ?ava_token=... bootstrap, focused backend/frontend regressions now cover the new auth/origin behavior, and the web-mode docs now describe token usage plus the explicit --insecure-open-cors override.
- Tightened the Milestone 5 web-hardening follow-up without broadening into full auth: sensitive read routes that expose session history/transcripts or live agent status/queue state now also require the control token, auto-generated control tokens are no longer written to normal tracing logs (only shown on the live terminal, otherwise redacted), the web docs now mention the
access_token query alias for WebSocket/control-token compatibility, and focused Rust/frontend/browser regressions were updated for the stricter route protection.
- Closed the Milestone 5 security blocker on persisted plan routes:
GET /api/plans and GET /api/plans/{filename} now sit behind the existing control-token middleware, and plan filename resolution now rejects traversal-style inputs (including encoded parent-path attempts) with focused backend regressions for auth gating and traversal rejection.
- Fixed the desktop session-open cwd binding regression:
notifySessionOpened now preserves the target workingDirectory through the frontend/Tauri bridge, switchSession() resolves cwd from the target session's project metadata instead of the ambient project, and focused store/bridge/backend regressions now cover override-project creation plus non-current-project session switching so replay-style backend binding uses the correct project root.
- Fixed the stale overlapping session-switch activation race in the frontend session lifecycle: async persisted-session finalization is now gated inside
activatePersistedSession() with an explicit currentness predicate from switchSession(), so an older load that resolves after a newer switch no longer re-selects the old session, overwrites visible artifacts, or re-persist the old last-session marker; focused activation/lifecycle regressions now cover both helper-level stale-finalization skipping and the overlapping-switch old-load-last case.
- Finished the tight Milestone 4 per-run context/session alias parity follow-up: web submit + replay flows now resolve frontend→backend session aliases before session-scoped backend calls, browser status/rehydrate correlation now does the same so active runs and pending interactive requests reattach after alias divergence, focused web regressions cover alias-aware submit/replay/rehydrate behavior, and the remaining bounded exception is still the manual TUI/headless
/compact path not consuming persisted run-context metadata (EX-003).
- Tightened the V1 benchmark-backed signoff safety slice in
scripts/testing/signoff-v1-headless.sh: report enforcement now fails closed on metadata mismatches (provider, model, task_filter, run_count, suite_name) in addition to per-task success, and new focused coverage in scripts/testing/signoff-v1-regression.sh exercises provider/model auth gating plus report-enforcement success/failure branches; docs now consistently describe verify-v1 as preflight and signoff-v1-headless as the authoritative V1-evals proof path (explicitly separated from the architecture M4–M7 milestone chain).
- Implemented the Milestone 3 repo-owned benchmark-backed headless V1 signoff path via
scripts/testing/signoff-v1-headless.sh, wired just v1-signoff + pnpm signoff:v1, and added strict benchmark report pass/fail enforcement for an explicit required-now task slice (small_coding_http_status_class, normal_coding_retry_backoff, selected tool_reliability, one stress_coding, one product smoke); verify-v1 remains a local preflight-only aggregate.
- Completed the Milestone 2 headless approval follow-up on the real runtime path: headless/benchmark-style stacks no longer reuse the interactive yolo/permissive short-circuit for unattended approvals, so dangerous approval-worthy actions now reach the existing approval bridge and fail closed while safe work still proceeds automatically; added backend runtime regressions and upgraded
ava-smoke so the local backend gate exercises this path instead of relying on slash-only coverage.
- Tightened the remaining Milestone 1 composer image-follow-up seam with a minimal safe UX guard: when a run is already active, image-bearing submits in
use-input-state are now explicitly blocked with a clear error instead of silently queuing text without images, focused input-state regression coverage now locks that path, and message-input image MIME typing is narrowed to the accepted image set.
- Wired desktop initial-submit image attachments end-to-end: pending composer images now flow into the first
agent.run(...) submit path, frontend/Tauri submit_goal contracts now accept an optional images payload, and the desktop backend now maps/forwards those images into the existing run path (with focused TS/Rust regression coverage for IPC invocation and submit image mapping, while preserving compatibility when images is omitted).
- Tightened the Milestone 1 image-submit follow-up with one small cross-surface pass: the optimistic local user message now retains submitted images, image-bearing submits queued behind an active same-session run now preserve image payloads for later dispatch instead of dropping them, and the web
/api/agent/submit contract now accepts/forwards the same optional image payload as desktop (with focused frontend + backend regression tests for local-state retention, queue preservation, and web submit image mapping).
- Re-scoped the Milestone 5
scripts/testing/verify-v1.sh entrypoint and matching just v1-gate / pnpm verify:v1 surface as a local V1 preflight aggregate (not a full final signoff). The preflight still runs just check, pnpm verify:mvp, just backend-gate, and deterministic Playwright parity coverage for e2e/app.spec.ts + e2e/web-mode.spec.ts, while explicitly excluding the benchmark-backed headless proof required for full signoff in docs/project/v1-evals.md (this preflight does not run that proof path, and fuller live-provider headless + broader desktop/TUI parity coverage remains outside it).
- Fixed Milestone 4 QuestionDock accessibility review findings: replaced misleading
role="dialog" with accurate role="region" (component is an inline widget, not a modal), added aria-labelledby to the free-text textarea so it has a real accessible label, split the active freeform multiple-choice row so its textbox is no longer nested inside a clickable radio button, updated the Playwright selector to use data-testid="question-dock", replaced the custom button-based radio controls with native input type="radio" plus label pattern for proper keyboard and screen-reader semantics (arrow-key navigation now works within the radiogroup), and added clear focus-visible styling for the textarea, freeform input, option controls, and action buttons.
- Added a debug-only web interactive seam for deterministic browser question coverage (
POST /api/debug/inject-question) mirroring the existing approval injector flow, and added a focused Playwright web-mode test that verifies QuestionDock visibility, answer submission, and clear-on-resolve behavior end-to-end.
- Clarified Milestone 1
0.6 -> V1 proof/docs posture by splitting authoritative headless-first backend proof from lighter parity/smoke checks in docs/project/backlog.md, docs/project/roadmap.md, docs/project/v1-evals.md, and docs/benchmark/deep-dive.md; required Milestone 1 proof now explicitly includes selected tool_reliability coverage and a minimal product smoke, and the roadmap workstream numbering around the HQ cleanup sections was cleaned up.
- Added a new Milestone 2 implementation backend automation gate (
scripts/testing/backend-automation-gate.sh) with a just backend-gate entrypoint plus docs updates (docs/testing/README.md, docs/how-to/test.md, scripts/README.md) to codify mandatory no-secrets backend checks and optional live-provider checks in the local validation flow.
- Refined the Milestone 2 backend automation gate: required no-secrets smoke now runs a deterministic headless
"/help" command (instead of --help), and optional live-provider smoke now pairs provider to explicit models while asserting BACKEND_GATE_OK in output.
- Aligned the milestone documentation wording for final polish by removing the stale
ava-3.3.1-evals related-doc reference and tightening V1 checklist language so docs/project/roadmap.md and docs/benchmark/deep-dive.md explicitly describe headless authoritative proof with lighter parity/smoke checks on TUI/desktop/web.
- Added the Milestone 1 implementation-phase planning update to lock backend V1 proof posture in
docs/project/backlog.md, docs/project/roadmap.md, and docs/project/v1-evals.md: backend validation is now explicitly headless-first, proof quality is grounded in real coding outcomes with build/test validation (not contract checks alone), dangerous actions only require explicit approval, and OpenCode + Goose are recorded as primary/secondary comparison references for backend/runtime and automation-shape parity.
- Reset the active product-planning language from
3.3 / 3.3.1 to 0.6 -> V1 in the key project docs, collapsed docs/project/backlog.md into a simple V1 checklist plus archived prior backlog themes, renamed the active eval-plan doc to docs/project/v1-evals.md, and updated the frontend package version to 0.6.0 so the current work is described consistently.
- Tightened the browser session-mutation seam without broadening the backend API: web duplicate/fork now flow through a dedicated alias-aware
web-session-mutations adapter, fork lineage is now backend-owned and survives reload/list/detail refetches, web clone/fork now preserve hidden persisted messages instead of dropping non-user_visible entries, duplicate no longer incorrectly records a parent session, cloned-session previews come back from the backend contract instead of staying blank, and web branchAtMessage now fails explicitly instead of creating a frontend-only branch that cannot persist.
- Completed the Milestone 8 hook-policy docs ownership cleanup: retired orphaned
scripts/dev/check.sh, designated scripts/dev/git-hooks.sh as the canonical hook/check seam in contributor and script docs, and tightened docs/how-to/test.md, docs/reference/scripts-and-build.md, and docs/contributing/development-workflow.md wording so staged pre-commit behavior is now documented as file-scoped rustfmt --check via the canonical hook entrypoint.
- Tightened the final Milestone 7 hook-safety gap in
scripts/dev/git-hooks.sh: pre-commit now validates materialized staged snapshots instead of reading working-tree file contents, so partially staged same-file unstaged edits no longer cause false hook failures, and scripts/testing/git-hooks-regression.sh now covers staged-snapshot Rust/frontend partial-commit safety on top of the existing hook-routing regressions.
- Implemented Milestone 9 public Ollama docs: added
docs/how-to/ollama-local-models.md for Ollama-only local-model setup grounded in current provider defaults/endpoint precedence/env overrides and practical verification paths, added focused docs/troubleshooting/ollama-local-models.md, updated navigation/index/manifest metadata, and cross-linked the new guidance from provider/env/troubleshooting/config docs without overclaiming broader local-model support.
- Tightened the remaining Milestone 7 hook policy correctness gaps in
scripts/dev/git-hooks.sh: staged Rust pre-commit checks now use direct file-scoped rustfmt --check so unrelated unstaged Rust formatting cannot block a staged-only commit, docs-only pre-push classification is now narrow enough that runtime-affecting Markdown under code trees still triggers the Rust gate, plugin workspace paths now count as frontend-sensitive for pre-push validation, and scripts/testing/git-hooks-regression.sh covers the new runtime-markdown/plugin-TypeScript/file-scoped-Rust cases.
- Tightened the remaining Milestone 7 pre-commit safety gap in
scripts/dev/git-hooks.sh: staged-file checks now stay non-mutating (cargo fmt --check and biome check instead of write-and-restage behavior) so partial commits keep their index/worktree separation intact, and scripts/testing/git-hooks-regression.sh now covers partial-stage Rust/frontend safety alongside the existing pre-push routing cases.
- Implemented Milestone 8 public CI/headless automation docs: added
docs/how-to/ci-headless-automation.md grounded in current headless CLI/JSON queue behavior and .github/workflows/ci.yml, updated public nav/index/manifest (docs/how-to/_meta.json, docs/index.md, docs/manifest.md), and linked the new page from related docs (docs/README.md, docs/how-to/run-locally.md, docs/how-to/test.md, docs/reference/README.md, docs/reference/commands.md) with conservative non-interactive guidance.
- Tightened the remaining Milestone 7 hook reliability gaps in the repo-owned
scripts/dev/git-hooks.sh layer: pre-push file classification now includes deletions so delete-only frontend/Rust pushes still trigger the right local gates, stdin-driven pre-push ref updates are now consumed when Git supplies them (with the prior inferred-range logic preserved as fallback), and a new scripts/testing/git-hooks-regression.sh script covers delete-only docs/frontend/Rust routing plus the no-stdin fallback path.
- Expanded
pre-push docs-only routing to include .github/CONTRIBUTING.md and .github/pull_request_template.md, with regression coverage added for both paths in scripts/testing/git-hooks-regression.sh so these docs edits stay on the lightweight docs path.
- Tightened the Milestone 7 hook-policy follow-up without broadening scope: hooks now route through one repo-owned
scripts/dev/git-hooks.sh entrypoint, pre-push is path-aware enough to run pnpm typecheck + pnpm lint for frontend-sensitive pushes while skipping the Rust gate for docs-only/frontend-only pushes, the staged pre-commit path no longer depends on mapfile so it stays compatible with stock macOS Bash 3.2, and docs now describe just ci as a broader local pass rather than a literal CI-equivalent clone.
- Added the Milestone 7 public troubleshooting common-errors page (
docs/troubleshooting/common-errors.md) grounded in live provider/auth/CLI error paths (ava-tui/ava-config/ava-llm), updated troubleshooting section navigation (docs/troubleshooting/README.md, docs/troubleshooting/_meta.json), and linked it from the public docs index (docs/index.md) while keeping Linux/Tauri-specific pages separate.
- Reworked the local git-hook verification policy for Milestone 7:
pre-commit now routes through a repo-owned staged-file script (cargo fmt for staged Rust files, biome --write plus oxlint for staged frontend files), pre-push now avoids package-wide ava-agent tests that pulled in unrelated provider/prompt-family coverage, and the new scripts/dev/run-rust-throttled.sh wrapper keeps heavy local Rust checks friendlier on interactive machines while CI remains the authoritative full gate.
- Added a public-facing install/release docs flow for Milestone 6: README quick start now includes installer script + GitHub Releases + source install paths,
docs/how-to/install.md now documents binary/manual/source routes with conservative platform notes, and a new docs/reference/install-and-release-paths.md page maps CLI distribution and desktop release ownership to install.sh, dist-workspace.toml, and .github/workflows/release.yml.
- Fixed the remaining web session-archive persistence gap: browser archive/unarchive flows now round-trip through the web backend instead of staying frontend-local, archived status is persisted in the web session metadata and reloaded through
/api/sessions filtering, and the browser DB fallback keeps using the frontend→backend session alias seam for archive/unarchive operations after IDs diverge.
- Tightened the remaining Milestone 5 live stress validation gaps:
e2e/agent-stress.spec.ts now requires the explicit assistant success marker for the simple-goal UI smoke, rejects clear assistant error output, verifies the API submit smoke by reading back the session transcript and asserting the latest assistant message, fails loudly when a run never settles, and skips live agent-submission smokes when no providers are configured.
- Tightened the remaining Milestone 5 live stress assertion:
e2e/agent-stress.spec.ts now requires a newly-added assistant artifact after submission (using stable message-role hooks) instead of falling back to generic message-area text, and its default backend URL now matches the Playwright web backend port.
- Hardened the Playwright shell smoke helpers for Milestone 5:
e2e/app.spec.ts and e2e/agent-stress.spec.ts now use the same robust changelog-dismiss flow as e2e/web-mode.spec.ts, so the What's New modal no longer silently intercepts composer clicks during shell readiness/input smoke checks.
- Tightened stale Playwright smoke harness coverage for Milestone 5:
e2e/agent-stress.spec.ts now waits on the current frontend app shell instead of navigating to the backend serve URL, uses the semantic "Message composer" textbox locator instead of a bare textarea selector, and drops stale --auto-approve guidance; nearby e2e/app.spec.ts composer smoke checks now use the same semantic locator pattern.
- Milestone 5 docs/DX cleanup: refreshed architecture entrypoints (
docs/architecture/entrypoints.md, docs/architecture/crate-map.md, docs/architecture/README.md) and roadmap guidance (docs/project/roadmap.md) to make current canonical owner seams explicit while preserving milestone artifacts as historical references; added concise module-level ownership headers across ava-agent::control_plane (mod.rs, commands.rs, events.rs, interactive.rs, orchestration.rs, queue.rs, sessions.rs) so contributors can identify command/event/interactive/session/queue boundaries directly from code.
- Completed Milestone 4 final verification: fixed syntax error in
PlanFullScreen.test.tsx (removed orphaned }) that was closing describe block prematurely); added closePlan to MainArea.plan-integration.test.tsx mock to support MainArea.tsx calling planOverlay.closePlan() on fullscreen close; fixed PlanOverlay resolve failure test by adding mutable mockAnnotations array so rejection path can be tested with feedback present (prevents immediate closePlan() when no feedback). All plan-viewer tests now passing (16 PlanFullScreen tests, 3 PlanOverlay tests, 10 MainArea integration tests with 2 skipped). [N minutes ago]
- Fixed remaining Milestone 4 concrete issues: wired fullscreen plan path in
AppShell.tsx so PlanFullScreen takes precedence over PlanOverlay when viewingPlanId is set (condition changed from !planOverlayOpen() to !planOverlayOpen() || viewingPlanId()); made missing-requestId path in resolvePlan() reject with error instead of resolving silently so await/catch hygiene behaves consistently; updated resolveApproval() and resolveQuestion() for same consistency; added tests for fullscreen routing precedence and missing-requestId rejection behavior. ChatView.tsx approval/question resolve handlers (handleApprovalResolve, handleQuestionResolve) now properly await the async agent.resolveApproval()/agent.resolveQuestion() calls and catch rejections; optimistic side effect (addAutoApprovedTool) now only runs after backend success to avoid persisting invalid state. PlanDock.tsx inline plan actions (handleApprove, handleReject, handleEdit) now await/catch agent.resolvePlan() to prevent unhandled rejections and keep dock consistent on failure. Bridged diff state from plan overlay/store into PlanFullScreen wrapper via new optional props (previousPlan, showDiff, hasDiff, onToggleDiff) instead of silently hardcoding false.
- Cleaned up unhandled rejection in resolve-failure test path: updated
useAgentActions.resolveApproval(), resolveQuestion(), and resolvePlan() to all return Promise<void> (was void); fixed parameterized test "keeps pending $label UI visible when resolve IPC rejects" in useAgentActions.test.ts to properly catch rejection via await scenario.actPromise.catch() instead of leaving promise rejection unhandled; removed unused flushMicrotasks helper. MainArea approve/revise handlers are now async and properly await agent.resolvePlan() — executePlan()/refinePlan()/closePlanViewer() are only called on success, keeping the plan viewer open on resolve failure so users can retry; added focused integration tests for resolve failure scenarios on both approve and revise paths.
- Finished the remaining browser approval-validation seam: the debug-only web E2E helper now creates a synthetic correlated run plus pending approval that flows through the real status/rehydrate/resolve path, the ApprovalDock Playwright case is deterministic instead of opportunistically skipped, and the debug backend now exposes an explicit run-cleanup endpoint so approval E2E teardown does not leave synthetic runs behind.
- Fixed Milestone 4 final concrete issues:
PlanFullScreen global shortcuts (approve/copy/download) now check for focused input/textarea/contenteditable elements and skip when an input is focused (especially the comment textarea); plan identity logic uses [...s.files].sort() and [...s.dependsOn].sort() to avoid mutating plan props; MainArea.plan-integration.test.tsx now includes direct assertions for agent.resolvePlan() on approve (called with 'approved' response) and revise-with-annotations (called with 'rejected' and serialized feedback), plus negative case for revise-without-feedback (resolvePlan not called).
- Fixed Milestone 4 plan resolution flow:
PlanFullScreen/MainArea now properly resolves plans back to the agent using agent.resolvePlan() matching the store-backed overlay pattern (approve sends 'approved' response, revise sends serialized annotations as feedback); strengthened plan identity reset to use full step content (action, files, dependsOn) rather than only prefixes; updated focused integration tests to verify plan resolution and feedback forwarding.
- Fixed Milestone 4 concrete blockers: ensured annotation feedback flows through
MainArea to planOverlay.refinePlan(annotations) so revise actions preserve wrapper annotations; fully disabled the history/version loading path in PlanFullScreen (commented out the /api/plans fetch); verified the previously failing tests in useAgent.test.ts and useAgentActions.test.ts are now passing (desktop session sync happy-path, web retry sync, and tool-call preservation tests all green).
- Tightened the remaining web recovery/backend-sync seam: browser recovery now reconstructs durable tool-call payloads from persisted message data, replay/direct web sync keeps frontend message ownership and sync bookkeeping keyed to the origin session even when backend session IDs diverge, and enriched assistant metadata/tool-call detail is patched back into the web backend so cache-eviction reloads can recover full tool cards.
- Tightened authoritative recovery tool-call preservation: when desktop/web recovery payloads return thinner assistant tool-call data than the already-cached message state, recovery now merges by message/tool-call identity so final content can update without dropping richer output, status, timestamps, or inline content offsets.
- Fixed Milestone 4 concrete review blockers for plan viewer: repaired
PlanFullScreen.test.tsx with accurate mocks matching component imports and added direct coverage for Escape layering and Ctrl/Cmd+Enter approve shortcuts; made version history path clearly non-interactive (shows "Version history not available" instead of clickable no-op UI); added per-plan state reset effect to prevent annotation/floating UI leakage across plans; aligned top-level viewer ARIA semantics by removing misleading aria-modal/role="dialog" and using accurate role="region" since the viewer is not a true modal with focus trap/restore.
- Implemented Milestone 4 plan viewer UI decomposition: transformed
PlanFullScreen.tsx from a monolithic 1445-line component into a thin orchestration wrapper (~350 lines) that delegates to already-extracted plan-viewer components (PlanHeader, TOCSidebar, AnnotationToolstrip, PlanDocument, AnnotationsPanel, SelectionToolbar, CommentPopover, QuickLabelPicker), eliminating duplication between the legacy props-based API and the modern store-based PlanOverlay architecture while preserving behavior, visual design, keyboard shortcuts, Escape layering, and the existing open/close flow.
- Tightened the current web chat/session detach seam: off-screen web completions now mark their originating session for authoritative recovery, retain any backend session-id mapping needed to reload the right session, and switch-back recovery now replaces the hidden partial placeholder with the backend-authored final message/tool state.
- Fixed Milestone 3 review issues in the frontend agent bridge: replay-style flows (retry/edit/regenerate) now handle detached-session results with the same authoritative recovery semantics as the main run path instead of treating them as ordinary success; settlement helpers in
agent-settlement.ts now await placeholder deletion where deletion is part of settlement semantics (clean async API); added focused regression test proving the successful web replay sync path through the settlement service including backend message replacement and active-session sync marking.
- Tightened the remaining Milestone 4 queue/runtime polish: queued-row edit/remove/reorder actions now resolve against the intended visible-session section instead of drifting across hidden-session or post-complete rows, regular queue reorders no longer cross the post-complete boundary, and fully idle session switches skip caching empty runtime snapshots.
- Landed the Milestone 3 frontend agent bridge and action flow extraction: created
src/services/agent-settlement.ts as a transport-agnostic settlement coordination layer, extracted duplicated backend-sync and message-settlement logic from useAgentActions.ts (reducing ~300 lines of duplication across retry/edit/regenerate paths), clarified rust-agent-ipc.ts as the canonical runtime/session correlation owner with cross-references to the new settlement service, preserved all public UI/backend contracts, and maintained incremental migration paths with deprecated-but-functional wrappers in useAgentActions.ts for review safety.
- Tightened the remaining Milestone 4 chat/session performance hardening pass: bounded agent event history now appends in O(1) after saturation and is consumed incrementally by the shared agent store, warm session/runtime cache restores avoid redundant cache touches and snapshot reclones on switch-back, and same-session reselection no longer needlessly reloads storage or rebind the frontend session surface.
- Tightened the current session/new-chat cache seam:
createNewSession() now snapshots the session being left just like switchSession(), so opening a fresh chat no longer drops hidden off-screen completions/tool-state artifacts before a later switch-back, including the web path where inactive-session DB persistence is skipped.
- Tightened the Milestone 2
ava-agent runtime-boundary follow-up: concurrent MCP lazy-init callers now await the same in-flight initialization, MCP init/reload snapshot disabled-server state before taking the shared registry lock to avoid lock-order inversion with enable/disable paths, and run-scoped compaction settings again fall back to the stack-wide override when the per-run context leaves compaction model unset.
- Landed the Milestone 2 backend runtime/service extraction at the safest
ava-agent seam: MCP runtime state, lazy init/reload, session-scoped enable/disable handling, and config-backed server introspection now live in crates/ava-agent/src/stack/stack_mcp.rs, shrinking AgentStack ownership sprawl without changing desktop/web/frontend contracts or rewriting the agent loop.
- Tightened the Milestone 3 browser IPC regression follow-up: the cancel + late-websocket runtime test now waits for real listener registration and correlation teardown explicitly instead of depending on incidental microtask timing, eliminating the combined-suite flake without changing browser run-correlation behavior.
- Tightened the last Milestone 1 touched-path follow-up: browser cancel now clears active run/session correlation before awaiting backend cancel so late WebSocket events from the cancelled run are dropped, and desktop resolve-send failure paths now promote the next same-kind actionable prompt after clearing the failed request just like successful resolves.
- Tightened the Milestone 3 tool-surface follow-up: tool-only assistant completions now preserve their placeholder rows instead of being deleted, desktop reopen recovery keeps rich backend tool-call details intact, and focused hook/session/component regressions now cover tool-only completion settlement plus recovered tool-surface rendering.
- Tightened the remaining Milestone 1 acceptance follow-up on touched paths: initial desktop/web same-kind interactive prompts now surface only when that run owns the global actionable head, and browser IPC now accepts matching-session uncorrelated completion events in the single attached-run case without weakening stale-run filtering for mismatched sessions or run-scoped events.
- Tightened the Milestone 1 shared-orchestration follow-up again: scoped desktop/web status now exposes a same-kind approval/question/plan prompt only when that run also owns the global actionable front request, and cancellation-time promotion no longer re-emits an already-visible prompt from another run when a background run's queued request is removed.
- Tightened the Milestone 1 shared-orchestration follow-up: same-kind interactive promotion now advances to the next globally visible queued request even when it belongs to a different run, and the shared deferred-queue helpers now mutate deferred/in-flight state through one consistent lock-order path to avoid cross-surface deadlock risk.
- Tightened the Milestone 2 frontend session-cache follow-up: inactive-session authoritative replacements now refresh the bounded artifact cache, hidden run/detach settlement evicts stale per-session runtime snapshots, and authoritative rehydrate clears/replaces stale cached runtime UI instead of replaying obsolete running or pending-interactive state.
- Landed the Milestone 2 frontend session-cache slice: recent hydrated session artifacts now stay in a bounded in-memory cache for instant switch-back, per-session live runtime state (streaming/thinking/tool activity, pending approval-question-plan state, and recent event timeline) is restored optimistically on return, and backend rehydrate still revalidates/corrects stale cached state before settling the visible session.
- Started Milestone 1 shared run-orchestration deduplication in the backend seam: the duplicated desktop/web helpers for deferred follow-up/post-complete queue promotion, deferred queue restore/clear, post-complete progress parsing, and inactive scoped-status classification now live in
ava-agent::control_plane::orchestration with focused backend tests, while desktop/web frontend-facing APIs and events remain unchanged.
- Closed the remaining Milestone 1 desktop reattach race: detached-session authoritative recovery markers now stay live until a non-running backend reload succeeds, reopening a Tauri session before its off-screen run finishes no longer loses the final repair signal, and the visible session now triggers the same backend
load_session recovery when that rehydrated run later reaches terminal state.
- Tightened the remaining web validation and local-state seams: browser-mode undo history now scopes to the requested session like desktop instead of sharing one global web edit stack, and Playwright web-mode now starts both the real AVA web backend and the Vite frontend so app-load/session smoke tests exercise a reproducible browser path instead of failing on missing backend proxy assumptions.
- Closed the remaining Milestone 1 desktop off-screen completion blocker: detached Tauri runs now mark their session for authoritative recovery, reopening that session reloads the finished backend
load_session snapshot and writes it back into frontend memory + local DB, and focused regressions now cover switch-away / backend-completes / switch-back final-message recovery without introducing the deferred per-session live cache.
- Closed the remaining hidden interactive-state reattach gap in the desktop/web runtime: pending approval/question/plan payloads are now cached alongside run ownership in the backend adapters,
get_agent_status rehydrates the current pending interactive request for the scoped run, and frontend status reattach now restores those actionable prompts instead of returning to a silently stalled hidden run.
- Finished the remaining Milestone 1 detached-run follow-up in the desktop/web frontend runtime: off-screen detach now prefers run-scoped result metadata while retaining a detached-session signal fallback for compatibility, stale newer runs can no longer cause an older detached run to settle against the wrong placeholder/session, rehydrate reconcile rollback now fully clears tracked run ownership, and focused IPC/run regressions cover both the detach race and rollback cleanup.
- Tightened the Milestone 1 off-screen session follow-up in the desktop/web frontend runtime: visible-session start/retry/replay controls now key off the current session's active run instead of process-global running state, hidden-run settlement now clears stale run/session bindings plus preserves detached-session ownership for placeholder cleanup, and focused regressions now cover hidden-run progression/completion plus starting/retrying from another visible session without leaking wrong-session runtime state.
- Added
t3code to the local docs/reference-code/ research set, updated the reference index/clone instructions, and added a short audit note capturing why its thin web/desktop shell over external coding-agent runtimes is worth studying alongside the other agent repos.
- Tightened the remaining Milestone 1 chat-session continuity gap in the desktop/web frontend runtime: switching chats no longer best-effort cancels the active run, the shared agent store now keeps that run alive off-screen while hiding its state from the newly visible chat, and focused regressions lock the no-leak / return-to-session behavior without introducing the deferred per-session runtime cache yet.
- Fixed chat tool-usage continuity across desktop/web session restore and post-run backend sync: persisted tool calls are now normalized back into the frontend
ToolCall shape (including legacy arguments payloads), and web backend refresh no longer overwrites richer live tool-call details with thinner persisted snapshots right after a run completes.
- Landed the desktop/Tauri slice of Milestone 3 multi-run backend ownership:
DesktopBridge now mirrors the web run registry with per-session run exclusivity, desktop cancel/status/queue routing can target a specific run_id/session_id without tearing down other active sessions, shared desktop interactive forwarders stay attached across concurrent runs, and the minimal frontend IPC/rehydration plumbing now keeps each desktop session correlated to its own active backend run with focused Rust/frontend regressions.
- Tightened the remaining actionable Settings broad-review fixes in one minimal frontend pass: settings search now auto-selects only on query changes so manual tab switches stick, the Add MCP Server dialog resets to its default presets state on reopen, Skills now expose a restore path for migrated hidden built-ins, and shared settings toggles regained explicit accessible names through localized toggle wiring plus focused regression coverage.
- Cleared the remaining in-scope Milestone 2 web multi-run blockers with one localized pass: browser status/session-transition rehydration now scopes to the relevant session/run so other active web runs no longer hijack the current chat, same-kind approval/question/plan requests now queue independently per run while still enforcing per-run
request_id ordering, and the web replay routes now use a clearer replay-specific request shape without breaking current browser callers.
- Finished the remaining Milestone 3 settings follow-up with one tight accessibility/restore pass: provider test + refresh now persist refreshed model lists back into settings state, default-model selection and the Ollama local-model browser are reachable again in provider settings, migrated LLM/Plugins/Base URL controls regained accessible names, and focused frontend regressions now lock the restored settings flows.
- Tightened Milestone 2 web multi-run review follow-ups with a narrow backend/browser pass: web interactive forwarding now defers same-kind queued approval/question/plan events until the request is truly front-of-queue/actionable (including promotion after resolve/timeout/cancel), browser IPC control calls now thread run/session correlation through status/cancel/steer/queue helpers where available, and web control-route correlation inputs now accept consistent snake_case + camelCase aliases.
- Fixed nested Settings child-dialog Escape ownership for the real Add MCP Server flow and the shared settings-dialog layer: topmost child dialogs now register themselves as the active Escape owner, pull focus into the modal when needed, and swallow Escape before the parent Settings surface can react; added a focused regression that reproduces the real "Add Server + Escape" path where focus is still on the opener.
- Landed the Milestone 2 web multi-run backend slice in the web adapter only: web runs now register per-run ownership by
run_id + session_id, cancel/queue/status paths can target a specific run without tearing down other active sessions, the web adapter keeps one shared interactive forwarder layer instead of stealing the bridges per run, and focused ava-tui web regressions now lock cancel/status/queue isolation plus the existing timeout path against the new multi-run registry.
- Milestone 3 review fixes: restored provider functionality in
ProviderCardExpanded (models list with refresh, test connection button, base URL input), added useSettingsDialogEscape hook and improved DeviceCodeDialog focus management for reliable nested Escape handling, fixed shared component accessibility with explicit labeling support (label, id, ariaLabel props) and visible focus indication (focus rings), added 7 focused tests for accessibility improvements; total 33 tests passing.
- Extended Milestone 3 with entrance animation removal pass: removed the
animate-slide-up entrance animation from DeviceCodeDialog.tsx (the only settings entrance animation found); all loading spinners and hover interaction feedback remain intact.
- Landed Milestone 3 Settings UI consistency pass: introduced shared settings building blocks (
SettingsCard, SettingsCardSimple, SettingsSectionHeader, SettingsActionHeader, SettingsPageTitle, SettingsButton, SettingsInput, SettingsTextarea, SettingsSelect, SettingsStatusBadge, SettingsTabContainer, SettingsRow, SettingsLabelValue) in shared-settings-components.tsx that use theme-aware CSS variables instead of hardcoded hex colors; migrated LLMTab and SkillsTab fully to the shared system; partially migrated PluginsTab headers and search to shared components; added 26 focused unit tests for the shared components; kept Providers and MCPServers tabs largely as-is per scope.
- Tightened the remaining Milestone 2 Settings cleanup slice with one low-risk frontend pass: nested Settings child dialogs now own Escape without collapsing the whole Settings surface, stale settings search entries were pruned back to current controls, Advanced was rebuilt into cleaner card-based About/Agents/Extension Surfaces/Developer sections, and OAuth-backed provider test/refresh flows now reuse stored credentials instead of requiring a stale API-key copy.
- Landed Milestone 1 steering-state recovery with a narrow cross-surface slice: desktop/web agent status now expose active
runId, frontend IPC rehydrates isRunning/run correlation from backend status on startup so Enter routes into steering/queue behavior instead of trying a fresh run during an already-active backend run, and session create/switch now performs best-effort cancel confirmation before transition to reduce 409/race windows.
- Removed the Settings > Usage subscription page and its dedicated desktop/web plumbing: the settings tab/search entries, frontend bridge/types, Tauri command, web
/api/usage route, and the now-dead ava-llm::usage module are gone so the remaining settings/backend surface stays aligned.
- Landed Milestone 1 desktop diagnostics/logging polish: Settings > Developer now exposes practical run diagnostics (run ID/state, progress, last error, recent agent events, backend log location, copyable payload, and backend log tail), and the Tauri desktop backend now initializes a persistent file-backed tracing subscriber so Rust
tracing output is retained on disk instead of only appearing in the console.
- Reworked the active docs/testing surface after the recent refactors:
docs/project/backlog.md now tracks pending work only, a new docs/testing/desktop-testing.md page defines the practical desktop regression workflow, and the root/docs indexes now surface the desktop validation path directly.
- Closed the remaining Milestone 7 WS5/WS6 control-plane parity gaps in one backend-first pass: desktop/web queue cancel paths now revoke queue ownership and interactive prompts through shared lifecycle locks, steering/deferred queue semantics are enforced consistently across cancel/clear boundaries, and delegation visibility events (
plan_step_complete, subagent_complete, streaming_edit_progress) now require canonical run_id/call_id correlation end-to-end with malformed-event drops plus focused Rust/frontend regressions.
- Closed the last broad-review normalization follow-up with one narrow fix set: frontend queue-row edit/remove/reorder actions now map visible rows back to the correct current-session entries even when hidden rows from other sessions exist, and desktop/web same-kind interactive watchdogs now wait for promotion before starting a queued request's timeout window; focused regressions cover both seams.
- Fixed the last queued same-kind interactive timeout blocker for desktop/web: the shared interactive request store now refuses to auto-time-out non-front approval/question/plan request IDs, so hidden queued requests stay pending until they are promoted/visible, with focused regressions covering two-request FIFO timeout order across all three kinds.
- Closed the final broad-review interactive lifecycle blockers in one narrow pass: stale
AgentStack::run(...) test call sites now pass the new interactive_run_id argument again, the desktop/frontend agent store now queues same-kind approval/question/plan requests by request_id instead of overwriting older hidden requests, and the shared interactive request store now rejects out-of-order resolve(Some(request_id)) calls that target a queued non-front request.
- Tightened the Milestone 7 shared control-plane proof/adoption slice in one narrow pass:
ava-agent now locks the full current command+event wire contract in one backend-owned fixture matrix, and the remaining frontend control-plane event consumers now read canonical snake_case correlation fields directly for approval/interactive-clear/delegation/edit-progress handling instead of carrying duplicate camelCase fallbacks.
- Closed the last Milestone 3 cross-kind TUI modal-ownership review bug in one small pass: approvals, questions, and plans now share one deterministic promotion policy so cross-kind arrivals queue behind the currently visible interactive modal and the next pending request is promoted regardless of kind when the active owner clears; focused
ava-tui regressions now cover approval↔question/plan arbitration, hidden queued timeout arming for question/plan, and the missing no-foreground-start survival paths for background question/plan requests.
- Tightened the last bounded Milestone 3 TUI ownership gaps in one small pass: starting a fresh foreground goal no longer treats
foreground_run_id = None as "cancel every pending interactive request", and cross-run question/plan arrivals now preserve the currently visible modal while queueing the newcomer deterministically until the active owner clears; focused ava-tui regressions cover both paths.
- Tightened the last bounded Milestone 3 TUI run-ownership correctness gap: approval/question/plan requests now carry their true originating TUI
run_id from the run-scoped agent registry into the shared interactive store, TUI supersession/cancel cleanup now targets only that owning run instead of whichever run is currently foreground, and focused ava-tui regressions cover background-run approval survival across a foreground switch plus question/plan same-run supersession without cross-run cancellation.
- Tightened the last in-scope Milestone 3 TUI lifecycle gaps with one small cohesive pass: queued approval timeouts now arm only for the visible front request and re-arm when the next approval becomes active, approval modal stage/rejection text now reset on timeout/cancel cleanup, and TUI question/plan intake now supersedes any older pending sender before registering a replacement so earlier replies cannot be stranded.
- Tightened the remaining Milestone 3 TUI run/event-normalization follow-up with one narrow approval-lifecycle fix: the shared interactive request store now preserves multiple pending requests per kind by request ID instead of overwriting older senders, TUI cancellation drains the owning run's pending interactive requests instead of only the latest reply handle, and a focused
ava-tui regression now proves two queued approvals resolve in order without losing either sender.
- Tightened the current TUI Milestone 3 run/event-normalization slice with one narrow follow-up: fixed the
plan_approval.rs borrow checker regression plus the new AgentActivity equality compile blocker, added focused ava-tui regressions that prove plan resolve + timeout lifecycle cleanup through the shared interactive store, and covered the stale plan-modal clear path when a new foreground TUI run supersedes the previous one.
- Landed the first bounded interactive-TUI Milestone 7 follow-up slice on run/event normalization: TUI approval/question/plan modals now resolve through the shared
ava-agent::control_plane::interactive lifecycle store with canonical request IDs, shared timeout/cancel cleanup, and focused regressions for timeout/resolve/cancel behavior, while the TUI run routers now surface the remaining obvious required control-plane events (plan_step_complete, streaming_edit_progress, and background complete) instead of silently dropping them.
- Finished the remaining Milestone 2 queue/session normalization pass with one small cohesive fix set: deferred follow-up/post-complete enqueue paths now require the active backend run session to match any caller-provided session ownership instead of appending into another session's preserved deferred queue, frontend session-switch queue clearing now removes only the session being left while preserving queued work owned by other sessions, and focused Rust/frontend regressions now cover backend deferred-queue ownership plus real session-switch and backend-managed queue-row boundaries.
- Tightened the remaining Milestone 2 web session/history ownership gaps with one small cohesive pass: web submit no longer regresses toward process-global
last_session_id when callers omit session_id, replay/edit/regenerate routes now require explicit web session_id ownership instead of silently borrowing shared global state, the web test fixture state builder was updated for the current WebStateInner shape, and focused Rust/frontend regressions now lock both the server-side ownership rules and browser-side session_id forwarding.
- Extended the shared session contract with canonical prompt-context and replay-payload helpers in
crates/ava-agent/src/control_plane/sessions.rs, moved desktop/web retry-edit-regenerate plus tool-introspection session-history loading onto that backend-owned seam, and added focused regressions that lock shared image/history semantics plus desktop's newly normalized "user-only edit target" rule.
- Tightened the current Milestone 7 shared-backend contract slice on the web run-correlation seam: web adapters now keep submit/retry/regenerate/edit runs on an adapter-owned
run_id even when callers omit one, frontend web listeners now ignore terminal and interactive events that arrive without a matching active run_id, and focused adapter/frontend regressions cover submit/compaction precedence, queue-clear conformance, and the remaining interactive-clear smoke path.
- Landed the first backend-owned shared-contract normalization slice for Milestone 7 beyond WS1/interactive bootstrap:
ava-agent now ships canonical control_plane/events.rs, sessions.rs, and queue.rs modules, desktop/web run-start + compaction paths reuse the shared requested > last > new session-precedence helper, queue-clear semantics now come from one backend contract helper, and web event projections now include the missing approval/delegation/edit-progress correlation fields required by the new contract tests.
- Finished the last WS2 review cleanup on the interactive lifecycle seam: desktop approval/question/plan request handles now retain their originating
run_id through resolve/cancel/timeout cleanup so stale interactive clear events cannot be mis-tagged onto a newer run, web websocket projections for the same interactive slice now carry matching run correlation where implemented, and focused Rust/frontend regressions now cover stale interactive filtering plus correlated timeout/cancel cleanup. TUI interactive adoption remains the bounded follow-up surface.
- Closed the last desktop-facing blocker pair with one small cohesive pass: onboarding's single-auth provider cards now establish a positioned containing block so the full-card connect button cannot bleed across neighboring tiles, and the Tauri agent-event bridge now carries
run_id on late-arriving per-run streaming/progress events so the frontend drops stale run-A updates before they can contaminate active run B state.
- Closed the last desktop broad-review selection/proof gap in one tight pass: composer auto-selection and
QuickModelPicker now keep provider+model paired in session state so desktop submit_goal reliably receives both values, and focused frontend regressions now cover provider/model forwarding through the Tauri submit + terminal-event seam.
- Closed the last desktop-facing broad-review follow-up with one tight pass: restored desktop session snapshots/materialization now carry user image attachments through the Tauri active-session sync seam so retry/regenerate/edit can preserve multimodal context after backend rebind, and device-code OAuth success now commits its connected-state transition immediately after credential storage so Settings/onboarding cannot lose the success update if the dialog closes.
- Closed the remaining desktop-facing broad-review correctness follow-up with one narrow pass: desktop session open/switch/create now forwards a frontend session snapshot into the Tauri active-session sync path so missing backend sessions can be lazily materialized with matching IDs/message history instead of leaving retry/regenerate/edit permanently blocked, and Settings-provider OAuth success now drives a real
updateProvider(...) patch that marks the provider connected while clearing stale API-key precedence in the desktop settings state.
- Landed the desktop+web slice of Milestone 7 WS2 on the shared interactive lifecycle seam:
ava-agent now owns canonical pending approval/question/plan request state, request IDs, and timeout policy under control_plane/interactive.rs; desktop and web now require request_id on interactive resolve calls, reject missing/stale IDs without consuming newer pending state, and emit the same interactive_request_cleared event on successful resolve/cancel/timeout paths so subscribed frontends clear stale approval/question/plan UI consistently. TUI adoption remains a bounded follow-up adapter slice rather than implied complete milestone closure.
- Closed the final Milestone 4 desktop timing/correctness nits: Tauri terminal settlement now tracks a unique per-run correlation ID plus run-local completion ownership so a delayed terminal/invoke outcome from an earlier run cannot contaminate the next run, and desktop approval/question/plan docks stay visible until their resolve IPC actually succeeds (with focused regressions for both seams).
- Finished the last Milestone 4 desktop stress/polish correctness follow-up: run settlement now still persists the originating session's placeholder/final output/cancel note after a fast session switch while keeping active-session UI cleanup guarded, and Tauri now emits explicit interactive-request cleared/timeout events so stale approval/question/plan docks are removed from the desktop UI when the backend auto-times-out or auto-clears them.
- Finished the last two Milestone 3 desktop run-reliability correctness fixes: the Tauri
create_session path now persists new backend sessions before any active-session bind checks run, and core-bridge now retries same-session backend sync after an earlier missing/error result instead of permanently reusing the failed cache; focused Rust/frontend regressions cover both seams.
- Closed the last Milestone 3 desktop review follow-up: syncing a real API key into the desktop secure store now clears stale OAuth metadata so API-key auth can actually take over, onboarding/API-key completion now wins over stale OAuth markers, archive/delete fallback session switches reuse the normal backend rebind path, and targeted regressions now cover the App onboarding seam, fallback session rebinding, and the symmetric stale-sync late-error repair branch.
- Tightened Milestone 7 WS1 contract correctness from review feedback: canonical command specs now match implemented cancellation/interactive-resolve terminal semantics and correlation metadata, headless JSON queue parsing accepts canonical queue command names while rejecting unknown tiers instead of silently downgrading, web
resolve_plan route coverage now includes success and bad-request paths, and the remaining desktop completion-mode divergence is explicitly tracked as a versioned backend-contract exception (EX-002).
- Landed Milestone 7 WS1 command/control-plane bootstrap in the runtime:
ava-agent now owns a canonical command inventory/completion contract under control_plane/commands.rs, desktop/web/TUI/headless queue-command semantics now reuse that shared map, the web API client now maps resolve_plan correctly, and focused Rust/frontend regressions cover the new contract plus the browser resolve_plan route.
- Refreshed dependency security baselines across the repo: the frontend/tooling lockfile now resolves the open
dompurify, markdown-it, picomatch, rollup, vite, flatted, minimatch, brace-expansion, ajv, smol-toml, and yaml advisories via direct version bumps plus targeted pnpm overrides, while the Rust manifests/lockfiles now move rand, bytes, time, quinn-proto, rustls-webpki, lru, and lz4_flex forward to non-alerted versions where the current dependency graph allows it.
- Raised the workspace Rust baseline to 1.86 so AVA can take the current secure
ratatui 0.30 and tantivy 0.26 lines that clear the remaining lru/lz4_flex dependency alerts in the core workspace and desktop app.
- Added Milestone 5 parity runtime artifact
docs/architecture/cross-surface-runtime-audit-m5.md with prioritized cross-surface findings (commands, approvals/questions/plans, event projection & delivery, tool introspection, session lifecycle, queue/cancel, and subagent visibility), and updated architecture/doc discoverability in docs/architecture/README.md, docs/architecture/_meta.json, and docs/README.md.
- Tightened Milestone 3 desktop agent-run reliability in a backend-first slice: OAuth-backed providers now clear stale API-key state and prefer OAuth routing in the live backend credential contract, Tauri auth writes/deletes only fall back to local storage when the native bridge is unavailable (native failures now surface instead of falsely claiming success), restored/switched desktop sessions now use an acknowledged active-session sync contract before retry/regenerate/edit flows proceed, and focused regressions cover desktop auth wiring plus terminal
error settlement alongside complete settlement.
- Finished the remaining in-scope Milestone 3 desktop reliability follow-up: session open/switch/restore now treat active-session sync as best-effort while retry/regenerate/edit preflights stay strict, Tauri agent IPC keeps listening through a short late-terminal grace path so trailing updates are not dropped when
invoke resolves early, onboarding/provider OAuth now distinguishes pending device-code flows from immediately connected PKCE flows without persisting "(oauth)" as an API key, and direct regressions cover the desktop auth/session command seams plus happy-path action ordering.
- Closed the last in-scope Milestone 3 desktop run-reliability gaps from review: stale desktop session-sync completions now trigger a rebind to the newest winning session instead of letting an older switch steal the backend again, onboarding OAuth clears conflicting provider API-key state so routing keeps preferring OAuth, and Tauri terminal-settlement now returns one consistent
SubmitGoalResult shape regardless of whether the terminal event or the invoke path finishes first.
- Fixed the last narrow Milestone 2 MCP session-visibility blocker in the backend runtime: session-disabled MCP servers are now filtered back out when rebuilding both run-scoped agent registries and interactive/introspection tool registries, so disabling a server in Settings hides its tools immediately for the rest of that session instead of letting them reappear on the next tool-surface refresh.
- Added Milestone 7 implementation-planning artifact
docs/architecture/backend-correction-roadmap-m7.md, turning Milestone 5 drift findings and Milestone 6 contract decisions into dependency-ordered backend correction workstreams with canonical owner seams, adopter rollout order, first code slices, and required conformance gates; discoverability links were updated in docs/architecture/README.md, docs/architecture/_meta.json, docs/README.md, and docs/project/backlog.md.
- Fixed the Milestone 2 MCP visibility responsiveness regression: read-only surfaces like Settings > MCP and the Tool List now return cached/config-backed MCP state immediately instead of synchronously waiting on lazy MCP startup, while explicit reloads and real agent runs still populate live statuses/tools.
- Added Milestone 6 contract planning artifact
docs/architecture/shared-backend-contract-m6.md as the canonical shared-backend definition based on Milestones 4 and 5, covering contract scope/philosophy, owner seams, required adopters, command/completion semantics, approval-question-plan lifecycle, event schema coverage rules, session continuity, queue/cancel semantics, delegation visibility requirements, headless non-interactive rules, explicit open decisions, and conformance test requirements; discoverability was updated in docs/architecture/README.md, docs/architecture/_meta.json, and docs/README.md.
- Added/refined the Milestone 4 architecture artifact
docs/architecture/cross-surface-runtime-map-m4.md, which traces CLI/TUI, desktop/Tauri, and web from entrypoint through transport/bridge into the shared AgentStack seam, summarizes event/approval/command/session/tool-introspection flow per surface, and highlights parity-relevant divergence points without starting the cross-surface behavior audit; discoverability is updated via docs/architecture/README.md.
- Finished the last narrow Milestone 2 Settings MCP follow-up: backend MCP server status payloads now mark whether the current enable/disable action is actually supported, the Settings tab disables non-reenableable config-disabled toggles instead of advertising a broken enable path, and live MCP fetches now use generation-based cancellation so stale responses cannot repopulate invalidated state after local MCP edits; focused frontend regressions cover both the disabled-toggle contract and the stale-fetch race.
- Tightened the remaining Milestone 2 desktop/web Settings/tooling follow-up: the Tool List dialog now sends explicit fresh goal/history/image context on same-session refetches (with desktop/web backends preferring that live frontend context over session-only fallback when present), and the Settings MCP tab now clears stale live-cache state after local MCP mutations so add/remove changes render immediately; focused frontend tests cover both the MCP mutation refresh path and the tool-dialog context refresh path.
- Fixed the remaining Milestone 2 desktop/web MCP/tool visibility parity gap with a narrow backend-first pass: web tool introspection now has a real
/api/tools/agent backend route and preserves session context, the Tool List dialog refetches when the active session changes, desktop MCP IPC now exposes live status plus enable/disable commands like web, and Settings refreshes from backend live MCP state instead of leaning on stale saved-form data.
- Fixed the last narrow Milestone 1 onboarding guide focus regression: dismissing the Settings-launched guide now restores focus to the Settings affordance instead of falling back to a generic shell target, and the onboarding regressions/browser smoke assertions now match the current guide dismiss label plus Escape-close behavior.
- Added a Milestone 3 backend-correction planning artifact section in
docs/project/backlog.md (## Milestone 3 Backend Correction Plan) that turns Milestones 1 and 2 into priority-ordered execution tracks with canonical owner seams, dependent adapters, success criteria, and explicit routing between backlog implementation work and architecture-doc planning work; also clarified that docs/architecture/plugin-boundary.md remains HQ-specific rather than the general backend-correction intake surface.
- Added
docs/architecture/agent-backend-capability-comparison-m2.md as a concise Milestone 2 artifact: a side-by-side AVA vs OpenCode/Pi/Claude Code/Codex backend-capability matrix (tools, commands, skills, delegation, write/edit, planning/todo, permissions, review, runtime traits), explicitly grounded on the Milestone 1 AVA audit and including prioritized correction-reference systems.
- Fixed the Milestone 1 onboarding shell review findings with a narrow desktop/web follow-up: onboarding now honors the selected light/dark/system mode, reopens cleanly from Settings by closing the settings surface first and rendering above it, replaces misleading import/workspace copy with truthful guidance, and uses a real modal dialog/focus path so keyboard users land inside the blocking overlay.
- Tightened Milestone 1 desktop/web onboarding dismissal behavior: the app now clears
onboardingOpen before running follow-up completion/skip persistence work, and new frontend regression tests cover both dismiss and full-complete flows to ensure the onboarding overlay fully unmounts and Settings remains usable afterward.
- Finished the remaining Milestone 1 onboarding/shell usability follow-up: each onboarding step now moves focus to its current heading, dismissing onboarding restores focus to a stable shell target (including the Settings -> Open Guide reopen path), and the browser smoke coverage now matches the current five-step onboarding UX instead of the removed import-link flow.
- Fixed the last Milestone 1 onboarding guide blocker: reopening onboarding from Settings now uses an explicit guide mode with a visible dismiss control on every step, Escape-to-close support for keyboard users, and regression coverage for closing the guide after advancing beyond step 1 while restoring focus back to the shell.
- Fixed desktop tool-list introspection context threading: the Tool List dialog now sends a frontend-serializable chat context payload through the Rust bridge into
list_agent_tools, and the Tauri command now maps that narrow context into effective_tools_for_interactive_run(...) inputs so interactive tool visibility/delegation analysis reflects the active chat state.
- Fixed web backend startup for existing local installs after the 3.3 baseline comment-only edits to historical HQ SQL migrations:
ava-db now repairs the known legacy SQLx checksum drift for migrations 003_hq.sql and 004_hq_agent_costs.sql before strict migration validation, so ava serve can boot against older ~/.ava/ava.db files instead of failing with migration X was previously applied but has been modified.
- Added a new architecture doc,
docs/architecture/agent-backend-capability-audit-m1.md, that inventories the current coding-agent backend capability surface (skills, command surfaces, subagents, tool execution, write/edit, planning/todo, permissions, review flow, and key registries) with concrete code references and initial gap notes.
- Milestone 2 of TUI/Desktop tool-surface grounding is now in place: explicit delegation detection can inherit a recent explicit subagent request for short follow-up turns, and interactive tool introspection now uses the same run-time visibility/delegation analysis path as
run() for tool filtering/delegation while staying on cached MCP state for read-only previews instead of a global boolean shortcut.
- Added a narrow interactive-tool introspection path so TUI/Desktop can query the effective run-time tool surface (including run-scoped
subagent metadata when delegation is enabled) without registering subagent in the long-lived AgentStack::new() registry.
- Milestone 3 prompt/tool grounding now includes a narrow fake-claim hardening pass: the shared system prompt and GPT-family prompt text now explicitly forbid claiming file edits, writes, or todo updates without successful tool evidence, and
ava-agent now rejects obviously ungrounded completion claims for those categories unless matching successful tools actually ran in-session.
- Milestone 5 Alibaba+Qwen prompt tuning now includes a minimal recovery-discipline overlay refinement from a fresh
prompt_regression --repeat 3 baseline on qwen3-coder-plus: once the target function is known, hosted Qwen is now explicitly steered away from repeated shell listing loops and toward direct file-tool edit+verify cycles (especially on wrong_first_edit_recovery).
- Added a narrow Alibaba+Kimi Milestone 4 overlay refinement from
prompt_regression --repeat 3 failure shape: hosted Kimi now gets explicit direct-edit (no diagnosis-only stop), first-try-correct recovery discipline, and one-shot tuple guidance for tool_choice_discipline.
- Finalized the Alibaba+MiniMax Milestone 3 overlay with stronger but more generic execution guidance: hosted MiniMax models are now steered toward assertion-led direct edits, immediate reruns of the same verification command, minimal normalization changes without brittle fallback paths, and atomic direct-tool updates for default-value fixes.
- Added a narrow Alibaba+MiniMax provider-family overlay for prompt-regression behavior: hosted MiniMax models are now steered to convert failing assertions into direct target-file edits, avoid diagnosis-only turns when the fix is already implied, and rerun verification immediately after edits.
- Refined the Alibaba+MiniMax overlay for the remaining Milestone 3 instability while keeping it behavior-focused: minimal string-normalization fixes now prefer standard-library edits over extra dependencies, and default-value fixes now prefer one atomic direct-file update over staged edits or shell-heavy rediscovery.
- Tightened the Alibaba+Qwen overlay after the first tuned run still stopped at diagnosis on the recovery fixture: hosted Qwen models are now told not to end a turn with summary-only output once the failing assertions already imply a concrete edit.
- Added a narrow Alibaba+Qwen provider-family overlay after prompt-regression showed hosted Qwen diagnosing the recovery fixture correctly but stopping before the direct edit; the overlay now pushes it to ignore unrelated files, patch the identified target file, and rerun verification immediately.
- Fixed the
prompt_regression_wrong_first_edit_recovery fixture so it now materializes as a standalone Cargo package with its own Cargo.toml, which removes the parent benchmark-workspace trap from cargo test and makes recovery benchmarking reflect model behavior instead of workspace leakage.
- Clarified the Alibaba+GLM recovery overlay after a deeper benchmark read: recovery tasks do not ask the model to intentionally make a bad first edit, so hosted GLM models are now told to attempt the most likely correct fix first and only enter recovery mode when verification disproves that fix.
- Refined the Alibaba+GLM recovery overlay again for Rust benchmark fixtures: hosted GLM models are now told to treat the fixture-root project test command as the source of truth and to avoid trusting stale
test_runner/test_bin style binaries during recovery.
- Tightened the Alibaba+GLM provider-family overlay again based on repeat-run failures: after a bad first edit, hosted GLM models are now steered to use the failing assertion as the source of truth, avoid guessing a second implementation, and rerun the same verification command instead of inventing alternate test binaries or wrapper scripts.
- Added a narrow Alibaba+GLM prompt overlay based on prompt-regression failures so hosted GLM models stop over-exploring after locating the target file, recover more directly after a failed verification, and stay focused on edit-then-verify loops; also hardened the
read_before_edit prompt-regression task to score real verification tool use instead of requiring literal cargo test transcript text.
- Tightened benchmark TPS reporting so wall-clock throughput is labeled explicitly, solo benchmark paths now include sub-agent token usage in their token totals, HQ harness runs forward real worker token/cost usage instead of relying on fabricated worker token splits, and solo benchmark reports now also emit TTFT-normalized
generation_tps plus a GenTok/s table column.
- Tightened project trust boundaries again: untrusted projects no longer auto-discover/start
.ava/plugins, and trusted project instruction loading no longer inherits ancestor AGENTS.md files from outside the explicitly trusted project root.
- Added a first-class runtime skills listing path:
/skills now shows the live filesystem-discovered SKILL.md set (global plus trust-gated project-local skills) using the same discovery model as prompt assembly, with matching TUI/headless help and command-surface updates.
- Followed up the runtime-skills milestone with tighter project-local
@include boundaries, a flat-layout .../skills/SKILL.md naming fix, and a lightweight headless /skills path that avoids unnecessary full app startup.
- Continued modularizing the benchmark runtime for better DX: shared Tier 2 code extraction/compile-test validation now lives in
benchmark_validation.rs, shared LLM-as-judge logic now lives in benchmark_judge.rs, shared benchmark/harness table rendering now lives in benchmark_format.rs, both benchmark runners reuse the shared validation path, and benchmark/judge output truncation plus timestamp rendering are now hardened against UTF-8 and malformed input panics.
- Removed the dead bundled-skill registry/types module from
ava-agent so the live skill surface is unambiguous again: slash commands remain their own wired command system, while skills are documented and implemented strictly as filesystem SKILL.md discovery in instructions.rs.
- Prompt-note resolution now supports a separate provider-name dimension (in addition to
ProviderKind) so provider-family overlays can be layered on top of model-family notes; initial support adds a lean Alibaba+Kimi overlay while preserving existing behavior for providers without overlays.
- Hardened loop-prone provider coverage by adding Kimi-specific stuck-detector regression tests (
for_provider_model("kimi", "k2p5") plus a nudge, nudge, stop similarity-loop check), and fixed benchmark score-input handling so code-task task_pass no longer treats compile success alone as a resolved pass when tests fail.
- Extended doom-loop threshold coverage for Alibaba-hosted models without provider-only blanket matching: missing Alibaba Qwen variants now resolve to aggressive loop policy explicitly, existing Alibaba GLM/Kimi/MiniMax behavior is preserved, and new stuck-detector regressions lock that policy in.
- Added
docs/testing/ and docs/operations/ sections plus a benchmark deep-dive page, and expanded project/docs navigation so benchmark, validation, testing, and maintainer runbook material now have first-class homes in the docs tree.
- Added a dedicated
docs/benchmark/ section with overview, suite/workflow, report/comparison, and prompt-benchmarking entry pages so benchmark behavior and usage now live in a first-class docs section instead of only project planning pages.
- Added the first provider/system-prompt benchmark runtime slice: benchmark mode now accepts prompt-family/variant/file/version/hash metadata, repeat + seed controls, optional benchmark output paths, raw-per-run plus aggregate report artifacts, benchmark-to-agent prompt override plumbing, and generic left/right report comparison flags (with legacy AVA/OpenCode aliases preserved).
- Added 3.3.1 eval Round 7 first
prompt_regression benchmark lane: six prompt-sensitive Tier 3 tasks (verify-before-finish, targeted edit only, minimal patch, read-before-edit, wrong-first-edit recovery, and tool-choice/subagent discipline), plus suite/category wiring and deterministic workspace setup + validation integration.
- Added
docs/project/provider-prompt-benchmarking.md, then expanded it into current implementation/usage documentation for the shipped 3.3.1 prompt-benchmark workflow, including live CLI flags, prompt-regression coverage, repeat-run behavior, comparison flow, and remaining gaps.
- Fixed the 3.3.1 benchmark runner workspace scoping so benchmark and harness tasks run against
~/.ava/benchmarks/workspace as their actual tool workspace instead of falling back to the repo root, which unblocked frontier Tier 3 tasks that read and edit benchmark fixture files.
- Fixed background sub-agent spawning so optional background helpers (such as a review pass) can run once when allowed without immediately tripping the internal delegation budget check.
- Tightened benchmark result semantics so code-task quality now requires Tier 2/3 validation success instead of allowing regex-only matches to report
PASS when tests fail, and relaxed overly implementation-specific recovery/maintenance regex expectations to accept equivalent correct fixes.
- Added 3.3.1 eval Round 6 first AVA-vs-OpenCode comparison runner in
ava-tui: it loads two existing benchmark JSON reports, aligns tasks by name, computes per-side aggregate stats plus time/cost savings and win counts, prints a readable CLI summary, and can optionally emit a structured comparison JSON artifact.
- Added 3.3.1 eval Round 5 first real
lsp_smoke and product_smoke benchmark lanes with deterministic Tier 3 fixtures and validation (LSP config/project/toolchain smoke plus session/config/tool-discovery/permission flows), without implying unsupported full LSP client behavior.
- Added 3.3.1 eval Round 4 first real
mcp_integration benchmark lane with deterministic local stdio mock MCP servers for filesystem, git, and multi-server workflows, including project-local .ava/mcp.json setup and audit-log-backed validation.
- Added 3.3.1 eval Round 3
tool_recovery coverage with real Tier 3 tasks (missing-file recovery, targeted edit recovery, and verification-discipline), plus workspace setup and validation wiring for the new lane.
- Expanded the 3.3.1 benchmark corpus with the first real project-scale coding tasks for
small_coding, stress_coding, large_project, test_heavy, and maintenance; added corresponding Tier 3 workspace setup/validation wiring (including stricter test-focused validation for the new test_heavy task).
- Reorganized the active docs set into clearer purpose-based sections: roadmap and backlog now live under
docs/project/, extension and credential references now live under docs/reference/, docs/README.md now explains the docs taxonomy by audience, and the root README.md plus contributor compatibility entrypoints were updated to match the new structure.
- Expanded the docs so the codebase is represented as a usable retrieval layer: added dedicated references for providers/auth and command surfaces, split extension docs into page-sized sections for plugins, MCP, commands/hooks, tools, and instructions, added architecture entrypoint and contributor workflow docs, and corrected stale path and trust-model details to match the current runtime.
- Made the active docs website-ready: added consistent frontmatter across active pages, added section
_meta.json navigation manifests, created a troubleshooting index page, and aligned sidebar ordering so the Markdown can be imported into a future docs site with minimal extra work.
- Tightened release-hardening coverage around current 3.3 routing and TUI behavior: cheap-route tests now assert against the computed cheapest configured candidate, stale TUI interaction assertions were updated to the current UI contract, and the release verification path is green again across Rust and frontend checks.
- Cleaned the standard verification path by fixing the current frontend lint/reactivity warnings and clearing the remaining workspace clippy warnings in
ava-plugin, ava-acp, and ava-tui.
- Made secure credential handling the default runtime path: desktop sync now writes to the Rust secure store, normal settings persistence no longer serializes raw provider API keys, and startup prefers secure storage while still reading an existing plaintext
~/.ava/credentials.json for compatibility.
- Reworked onboarding into an optional in-app guide instead of a startup gate, with a reopen entry in Settings > General.
- Moved plugin and MCP management into
Advanced by default so the main settings surface stays focused on core product configuration.
- Started the HQ plugin reintroduction groundwork from the plugin side:
ava-hq now ships a small ava-hq-plugin binary and plugins/examples/ava-hq/ local-link artifact, but that track is now deferred behind the future roadmap while current work stays focused on core AVA.
- Normalized provider route/region aliases back into the canonical core provider surface: the repo-owned fallback model catalog now collapses legacy coding-plan/region variants into their main providers, and the frontend model/docs helpers now resolve those old IDs through the same canonical paths instead of treating them like separate providers.
ava-tui 0.6.1
Install ava-tui 0.6.1
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/Artificial-Source/AVA/releases/download/v0.6.1/ava-tui-installer.sh | sh
Install prebuilt binaries via powershell script
powershell -ExecutionPolicy Bypass -c "irm https://github.com/Artificial-Source/AVA/releases/download/v0.6.1/ava-tui-installer.ps1 | iex"
Download ava-tui 0.6.1
ava-hq 0.6.1
Install ava-hq 0.6.1
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/Artificial-Source/AVA/releases/download/v0.6.1/ava-hq-installer.sh | sh
Install prebuilt binaries via powershell script
powershell -ExecutionPolicy Bypass -c "irm https://github.com/Artificial-Source/AVA/releases/download/v0.6.1/ava-hq-installer.ps1 | iex"
Download ava-hq 0.6.1