Changelog

All notable public changes to DynosAI are documented here. The project uses semantic-version-style releases and a Keep-a-Changelog-inspired structure.

1.0.0-rc.3 - 2026-09-01

Added

• Host-owned acceptance ZIP importer writes failed children into bounded Eval Registry cases. Studio can import a local ZIP. Improvement stays in inbox and does not spawn a provider.
• Completed-work aggregates of governed_change_cost() on stats, scorecard, overview, Studio and diagnostic bundles. MCP duration and result bytes are recorded when measured.
• Deterministic PROMPT_PREFIX_1.0 authority prefix with a stable hash stored in runtime/audit. Phase context remains a suffix. The hash is not a cache-hit claim.

Changed

• Package version is 1.0.0rc3. Display surfaces may show 1.0.0-rc.3. Public status is Public RC · Eval maturity, not stable 1.0.

Security

• Schema remains v6. Unique MCP names remain 31. Unused advertised MCP tools are not scored as waste.

Validation

• Added tests/test_230.py for ZIP import bounds, inbox-only improvement, governed-change scorecard surfaces, prefix hashing and the no-waste rule.

1.0.0-rc.2 - 2026-09-01

Added

• Introduced the versioned MATRIX_1.0 live-certification schema (docs/validation/matrix-1.0.json) with Codex/Cursor × greenfield/brownfield cells.
• Studio Overview shows those four cells. dynosai_stats, project overview and diagnostic bundles expose the same provider-aware placeholder. No MCP tool was added.
• engine.claim_lease records an optional host session identity so two governed sessions can claim file-disjoint leases without spawning providers.

Changed

• Package version is 1.0.0rc2. Display surfaces may show 1.0.0-rc.2. Public status is Public RC · Certification evidence, not stable 1.0.

Security

• Schema remains v6. Historical 0.13 Quality 100 evidence is not copied into 1.0 live cells. all_passed stays false until every cell is a real provider pass.

Validation

• Added tests/test_220.py for MATRIX_1.0 placeholders, two-session disjoint claims, overlap serialization, fan-in conflicts and scheduler-off serial fallback.

1.0.0-rc.1 - 2026-09-01

Added

• Documented the 1.0 compatibility contract in docs/COMPATIBILITY.md: frozen MCP names, public CLI, schema v6, Studio ↔ App Server scope, human gates, and additive-field policy.
• Exposed optional harness features as host-owned project settings: context handles, governed team scheduling, and eval intelligence. Precedence is environment override → project setting → default. Studio Overview and Project settings show EN/ES status chips.
• Added tests/test_210.py to freeze the unique MCP tool-name set (31) and the optional-harness degraded paths.

Changed

• Package version is 1.0.0rc1 (PEP 440). Display surfaces may show 1.0.0-rc.1. Public status is Public RC · Contract freeze, not stable 1.0.
• Release gate, CI and scripts/build_release.py now share one policy: python -m pytest is the stable suite. Historical aggregate tests remain in the repository and stay excluded unless DYNOSAI_HISTORICAL_TESTS=1.
• dynosai_stats reports mcp_surface_frozen and mcp_tool_count. No MCP tool was added, removed or renamed.

Security

• Human gates, capability certification, path security and execution-profile enforcement are not kill-switches. There is no DYNOSAI_DISABLE_* family. Schema remains v6. Agents cannot mutate harness settings through MCP.
• Turning context-handle creation off keeps historical dynosai_retrieve_handle reads. Scheduler off uses serial fallback and a structured feature_disabled refusal. Eval intelligence off hides propose/mining while keeping historical cases readable.
• Website Next.js and eslint-config-next are 16.3.3 (Dependabot #14/#15).

Validation

• tests/conftest.py labels and excludes the four historical aggregate suites from the default collection.
• RC1 does not prove a 1.0 live provider matrix, installer, OS sandbox, additional certified clients, or autonomous predictive routing.

0.19.0 - 2026-08-31

Added

• Introduced host-owned provider capability manifests for the already-shipped Cursor ACP and Codex app-server transports.
• Added capability_manifests.py. Unknown clients and project extension packs raise instead of silently inheriting Codex/Cursor behavior.
• Studio Overview shows the certified provider inventory. dynosai_stats and diagnostic bundles include a capability summary. GET /api/provider-capabilities is host-owned; no MCP tool was added.

Changed

• Provider setup refuses uncertified names through the same manifest contract.
• Predictive routing remains shadow-only. Human gates remain required.

Security

• Additional adapters cannot self-certify. Schema remains v6.

Validation

• Added tests/test_200.py for shipped manifests, refused clients/packs, and secret-free inventory evidence. Studio tests require certified-provider copy in EN/ES.

0.18.0 - 2026-08-31

Added

• Introduced Strict, Balanced and Autonomous execution profiles as host-owned policy outside the model. Balanced is the project default. Autonomous uses an allowlist, not unrestricted network.
• Added execution_profiles.py. Docker, VM and remote runtimes raise a policy error rather than silently falling back.
• Optional local secret vault materializes named secrets to the local runtime only. materialize_for_model still refuses.
• Studio Project settings can select the profile. Code/merge reviews, overview, dynosai_stats and diagnostic bundles include policy evidence without secret values.

Changed

• Engine-constructed LocalExecutionRuntime now follows the selected profile (governed dependencies, profile network). Direct LocalExecutionRuntime() construction remains compatibility/unrestricted so 0.15 tests stay valid.
• Predictive routing remains shadow-only. Human gates remain required in every profile.

Security

• OS-level child-process network interception is not shipped. Evidence records enforcement: decision_only and os_network_enforcement: false.
• Schema remains v6. Vault files live under .dynosai/runtime/. Agents cannot change the execution profile through MCP.

Validation

• Added tests/test_190.py for profiles, vault materialization, refused container runtimes, and secret-free evidence. Studio tests require execution-profile settings and EN/ES OS-enforcement copy.

0.17.0 - 2026-08-31

Added

• Introduced eval intelligence on top of Eval Registry v0: failure attribution by layer (model, harness, provider, project, infrastructure, governance), mining of local validations/audit/eval/model-control traces, and bounded regression cases under .dynosai/runtime/.
• Added eval_intelligence.py. Improvement work is created in inbox only; DynosAI does not auto-start a provider. Regression evidence does not complete or approve the work.
• Studio Overview shows mined eval cases with an explicit note that predictive routing stays in shadow mode.
• dynosai_stats includes an eval-intelligence summary. Offline fixtures cover team_overlap and mined_regression without live providers.

Changed

• Predictive routing remains shadow-only. Non-model failures remain ineligible for capability learning.

Security

• Eval mining writes runtime files only. Schema remains v6. Agents cannot spawn improvement sessions through MCP.

Validation

• Added tests/test_180.py for attribution, mining, inbox-only propose, and regression evidence. Studio tests require eval-case rendering and EN/ES shadow-mode copy.

0.16.0 - 2026-08-31

Added

• Introduced governed agent-team scheduling: approved task DAGs become serial or parallel waves with file-disjoint leases, scope ceilings, token/time budgets, and follow-on reviewer/tester contracts.
• Added team_scheduler.py as a Core scheduling module. Leases are derived from existing tasks.claimed_run / runs; schema remains v6.
• Exposed team waves on engine.schedule, engine.claim_lease and engine.fan_in_report. Fan-in is wave-scoped so sequential overlap of the same file is not a false conflict.
• Studio Work cards show current-wave team slots (role, files, lease status) with an explicit note that DynosAI does not spawn extra provider processes.
• dynosai_schedule now returns the team plan plus a fan_in report. A second worker exists only if the host opens another governed session against an open lease.

Changed

• Parallel batches remain the compatibility view of a team wave. Overlapping planned files are never leased in the same parallel wave.
• Reviewer is the existing human code_review gate, not a spawned agent. Tester is the governed validation contract on the same lease.
• Predictive routing remains shadow-only.

Security

• Team scheduling does not expand agent authority, spawn extra Codex/Cursor processes, or bypass human gates.

Validation

• Added tests/test_170.py for DAG waves, claim denial, fan-in conflicts, schema-v6 integration and reviewer-as-human-gate.
• Studio asset tests require team-slot rendering and EN/ES copy that DynosAI does not start extra agents.

0.15.0 - 2026-08-31

Added

• Introduced the Verified Agent Harness: typed execution/recovery contracts, persistent context handles, dynosai_retrieve_handle, and pass-by-reference compaction of large diffs/search/validation outputs.
• Formalized ExecutionRuntime / LocalExecutionRuntime as the provider-neutral hands boundary (filesystem, processes, network decisions, dependency classification). Local remains the default; Docker/VMs are not required.
• Added minimum execution policy outside the model: path-root/escape/symlink denial, process timeouts, network profiles (unrestricted default, allowlist, default_deny, offline), dependency use-vs-install classification, credential redaction, and a refuse-closed secret broker.
• Added Validation Integrity reports (requirement → acceptance → evidence → validation) and exposed them on Studio code/merge reviews together with bounded diffs and validation results.
• Added Eval Registry v0 (nine representative offline scenarios) and a skill on/off context-size comparison. No live-provider quality claim is made from these fixtures.
• Added cost-per-successful-governed-change raw observability (tokens, tools, retries, human gates; catalog price only when configured).
• Added mechanical architecture-import checks in scripts/check_repository.py.

Changed

• Reframed 0.15 in the public roadmap around a verified harness rather than a vague loop-engine slogan. Cursor ACP remains documented as a 0.14.1 capability, not future 0.19 greenfield work.
• Studio completion reviews can expand the governed chain (requirements, acceptance, tasks, files, diff, validation, integrity) without making technical internals the default UX.
• Predictive routing remains shadow-only.

Security

• Secrets matching known credential patterns are redacted from MCP tool results before they reach prompts, evidence-shaped payloads, or Studio review diffs.
• The secret broker refuses to materialize values into model context.

Validation

• Added tests/test_160.py and tests/test_161.py for handles, runtime policy, integrity, evals, cost, MCP retrieve, and Studio reviewability.
• Schema remains v6; handles and eval records are runtime files under .dynosai/runtime/, not workflow-authority tables.

0.14.1 - 2026-08-26

Added

• Added a guided Studio setup model that explains project detection, Git readiness, DynosAI initialization and project-check approval in user-facing terms.
• Added a Review Center backed by pending human_interactions, including bounded specification, plan and quality/evidence context for human gates.
• Added /api/setup and /api/reviews to the loopback App Server.
• Added System/Light/Dark Studio themes with browser-local persistence and the same DynosAI SVG mark used by the public website.
• Added in-product Help, plain-language workflow timelines, project-check selection, support summary and user-friendly task creation guidance.
• Added in-app project management with recent projects, local path opening and a contextual folder browser, without restarting Studio.
• Added an English/Spanish localization layer with English as the default and browser-local language preference.
• Added a contextual greenfield Fibonacci walkthrough that guides the user through the normal Studio UI without creating projects, starting agents or approving workflow steps on the user's behalf.
• Added accessible shadcn/ui-style combobox controls for provider, workspace and language choices.
• Added a neutral project hub: dynosai studio no longer selects the launch directory unless --project is explicitly supplied.
• Added explicit neutral greenfield project creation from Home; technology-specific files are introduced by governed changes rather than by a starter selector.
• Refined Home semantics so returning to Home closes the active project context and returns to a neutral project hub while keeping the project in Recents.
• Replaced OS-dependent folder-picker calls with an in-Studio directory browser for both project creation locations and existing-project opening.
• Fixed sidebar horizontal overflow and stopped rendering long project paths inside the navigation rail.
• Replaced the dedicated Fibonacci tutorial page and automatic demo creation with a contextual, state-aware walkthrough that highlights the normal UI while the user performs every action.
• Replaced browser-native confirmation dialogs with Studio-native shadcn/ui-style dialogs and styled the remaining checkbox/toggle controls consistently.
• Added folder creation directly inside the Studio folder browser.
• Added project-scoped execution settings and model routing inspection/editing for Codex and Cursor across discovery, specification, planning, implementation, code review, validation and merge.
• Bounded Recent projects with an internal vertical scroll area and removed the manual Refresh button in favor of background synchronization.
• Reduced repeated page headings/copy and corrected execution-summary spacing in the New change flow.
• Simplified New change to the request itself; coding agent, workspace and per-phase model choices now live only in Project settings.
• Locked project workflow pages until DynosAI initialization completes, removed the duplicated Next step panel/button, and added blocking progress feedback for mutating actions.
• Made model routing follow the project coding agent automatically instead of exposing a second Codex/Cursor switch inside the routing panel.
• Made Studio-created changes start the configured Codex/Cursor agent asynchronously instead of remaining in Inbox/Queued, with observable running/interrupted states and a safe Continue agent retry.
• Made project path/location fields read-only and folder-browser controlled, and fixed the Fibonacci walkthrough so repeated approval cycles do not prematurely jump to the final result step.
• Fixed Studio provider execution to use provider-native headless transports instead of terminal-oriented CLI launches, so Codex/Cursor can run from the browser-managed workflow without an interactive console; Studio now owns human gates and exposes provider diagnostic tails when a turn fails.
• Fixed Windows Codex Studio execution with explicit UTF-8 JSON-RPC/MCP stdio, preventing Spanish/non-ASCII task text from being encoded through the Windows ANSI code page; headless launches also avoid creating project-local provider config as a launch side effect.
• Fixed Cursor Studio execution by replacing terminal/print-mode automation with Cursor ACP (agent acp), including session-scoped DynosAI MCP injection, explicit provider-permission handling and actionable authentication failures.
• Fixed headless CLI exit-code propagation so Studio receives the real provider failure instead of a successful wrapper process plus a large runtime JSON dump.
• Made Work diagnostics use a Studio-native persistent disclosure panel, so background refresh no longer collapses diagnostic details while the user is reading them.
• Expanded Approvals so specification and plan gates expose the actual requirements, acceptance criteria, tasks, files and risks before a user can approve them; standard gate guidance is localized by Studio.
• Preserved request-change/clarification drafts across background refresh and prevented focused approval editors from being rebuilt while the user is typing.
• Added a bounded live Agent activity stream to Work, fed by Cursor ACP/Codex provider progress plus governed workflow events.
• Added an approval history on the Approvals page and a project-level auto-approve switch that records studio_auto decisions without answering clarifications. Ordinary product-file scope requests can be auto-approved; DynosAI-owned spec/plan overlays never become a human scope gate.
• Explained Project checks as later evidence commands for a library or web app, not extra work on that screen.
• Fixed Cursor planning turns so the provider-native cursor/create_plan transport gate is acknowledged without replacing or bypassing DynosAI's authoritative plan and human approval gate.
• Fixed the contextual tutorial to advance automatically when a real approval appears and to keep following repeated approval/work cycles until the governed task is done.
• Fixed Windows WinError 32 after Cursor ACP workflow transitions: Studio now closes ACP stdin, waits for process exit, reaps the provider process tree, and refreshes the managed runtime without failing when leftover acp-sessions files are still unlocking.
• Stopped Studio refresh from jumping the page, diagnostic log or Agent activity scrollbar back to the top, and show newest activity first with a timestamp.
• Survived Windows locale subprocess output (UnicodeDecodeError in _readerthread) by decoding captured Git, pytest, tasklist and provider stdio as UTF-8 with replacement.
• Decoded Git C-quoted pathnames (core.quotepath) so Spanish overlay paths such as specs/...pequeña.../plan.md are recognized as DynosAI-owned artifacts instead of failing implementation register_result.
• Stopped exported specs/** overlays (including slash-corrupted Git octal names like /303/261/plan.md) from remaining as required task files, so an otherwise complete implementation can register and leave Implementar.
• Fixed final merge on interactive branches so leftover verified files such as tests/test_fibonacci.py are scooped into the checkpoint instead of blocking squash as collapsed ?? tests/ dirt; DynosAI-owned overlays and caches are not merge blockers.
• Unstuck Work when Final review has already been auto-approved but merge did not finish: Continue now completes the host-owned merge instead of demanding a missing Approvals card.
• After a specification is auto-approved during a Codex/Cursor turn, Studio relaunches the agent to author the plan instead of treating the already-advanced plan_review state as a finished turn.
• Stopped classifying requests that mention ValueError as bugs; exception type names are not bug reports.

Changed

• Replaced the ambiguous 0.14.0 Advanced mode with opt-in Show technical details in Settings.
• Reworked Studio navigation around global Home/Settings/Help plus a nested selected-project workflow for Overview, New change, Work, Approvals and Project checks.
• Moved raw risk, diagnostics and activity/event views out of the default non-technical experience.
• Simplified risk presentation to a user-facing change-sensitivity level while keeping the deterministic score available in technical details.
• Aligned Studio branding, colors and interaction hierarchy more closely with dynosai.com.
• Kept Git, schema v6 workflow state, validation approval and human-gate authority unchanged.
• Removed Claude from Studio and public CLI provider choices; the supported public execution surface remains Codex and Cursor.
• Separated global Studio navigation (Home, Settings, Help) from a nested selected-project workflow (Overview, New change, Work, Approvals, Project checks).
• Removed normal-user UI wording about the underlying local server/loopback runtime; those implementation details remain architectural rather than product-navigation concepts.

Validation

• Added 0.14.1 regression coverage for setup guidance, Review Center API, shared branding/favicon, theme persistence, non-technical navigation, local asset serving and technical-mode opt-in behavior.
• Stable Python regression now contains 441 tests, including Windows locale-safe subprocess capture, Studio auto-approve history, activity-log ordering, managed spec/plan overlays, Git C-quoted Spanish overlay paths, leftover verified tests/ files at merge, Continue-to-finish on Final review, auto-continue into plan authoring after a mid-turn spec approval, Cursor ACP process-tree/acp-sessions lifecycle coverage, neutral project-hub startup, project creation/switching, contextual Fibonacci walkthrough, localization, accessible combobox and current-provider-boundary coverage.
• Repository-quality, Studio-sync and website check/typecheck/lint/build gates remain required for release.

0.14.0 - 2026-08-26

Added

• Added a loopback-only DynosAI App Server as a provider-neutral interface over DynosAIApplication.
• Added dynosai studio and the first Local Studio alpha for project onboarding, recent work, blocker visibility, validation discovery, risk and diagnostics.
• Added stack-aware project detection and suggested validation/test commands.
• Added read-only validation auto-discovery for Python, Node.js/TypeScript, Rust, .NET and Java with explicit approval before profiles become governed checks.
• Added deterministic Risk Assessment v1, separate from workflow Quality, using blast-radius and sensitive-surface signals.
• Added explainable blockers and a bounded project overview read model for graphical clients.
• Added public /studio and /roadmap website surfaces plus a substantially expanded product roadmap through 1.0.

Changed

• Provider process launching and acceptance shims are now cross-platform on Windows: Python shebang fixtures/provider overrides are routed through the active interpreter, and Windows Git Guard tests invoke the generated command wrapper explicitly.
• TOML configuration tests now validate parsed values instead of raw backslash escaping, making Codex configuration assertions portable across POSIX and Windows paths.
• Project onboarding now uses detected language/test hints instead of defaulting every repository to Python/pytest when stronger project evidence is available.
• Website version checks now compare against the authoritative pyproject.toml version instead of hardcoding a release number.
• The 0.13.0 governed SDD core, schema v6 authority model, Git boundaries and predictive-router shadow policy remain intact.

Validation

• Added the 0.14 regression suite covering Studio assets/API, project/stack detection, validation approval, risk assessment, blocker explanations and CLI contracts.
• Existing stable regression suites remain release gates; historical pre-stable aggregate suites remain available for manual diagnostics.

0.13.0 - 2026-08-25

Changed

• Promoted the validated 0.12.9 RC4 behavior to the first internally stabilized public-beta baseline with no functional architecture change.
• Kept database schema v6 and existing public MCP contracts unchanged.
• Kept Predictive Router runtime authority disabled (shadow mode) despite passing quantitative offline authority gates.

Validation

• Final real-provider matrix passed 4/4: Codex/Cursor × Greenfield/Brownfield.
• Every child finished done with Quality 100 and complete Oracle checks.
• Matrix recorded 0 MCP failures, 0 MCP rejections, 0 scope requests, 0 reconnects, and 0 retries.

0.12.9

Fixed

• Acceptance retry logic learned to distinguish recovered historical MCP rejections from unresolved rejections, without retrying functional failures.

Validation

• Final RC4 source/wheel selected regression reached 280/280 PASS before beta baseline promotion.

0.12.8

Added

• Transparent one-time retry for side-effect-free provider/worker idle-timeout failures.
• Preserved failed attempts and actual usage/cost including retries in acceptance bundles.

0.12.7

Fixed

• Added provider-specific MCP result transport: Codex keeps structured-primary transport; Cursor uses full-text compatibility because its CLI does not reliably expose structured content to the model.

Added

• Consolidated Codex/Cursor Green/Brown acceptance into one matrix bundle.

0.12.6

Validation

• Release-candidate hardening for schema migration, restart/resume, corrupted state/checkpoints/telemetry, validation-learning hygiene, scope lifecycle, and human-gate polling.

0.12.5

Added

• Offline chronological Predictive Router replay and authority gates with zero provider/model calls.
• Explicit exclusion of legacy/orchestration-contaminated outcomes from learning.

0.12.4

Changed

• Reduced duplicate MCP text transport for compatible providers.
• Collapsed deterministic READY -> IMPLEMENTING after an approved plan when execution was already requested.
• Completed real certification of a two-task execution wave.

0.12.3

Added

• Execution waves for bounded dependent task chains.
• Final-wave deterministic auto-advance to code review.

0.12.2

Changed

• Reduced MCP payload/replayed context and added structural efficiency telemetry while keeping workflow semantics stable.

0.12.1

Fixed

• Corrected scope lifecycle and predictive-learning hygiene after the 0.12.0 real acceptance exposed orchestration failures that could be misclassified as model failures.

0.12.0

Added

• Predictive routing memory in shadow mode, context checkpoint reuse, targeted checkpoint loading, calibrated estimator updates, and phase-aware context strategies.

0.11.x

Added / Changed

• Introduced and hardened the adaptive Model Control Plane: phase-aware budgets, complexity, safe routing boundaries, exact transition accounting, context-pressure separation, and initial-route verification.

0.10.x

Added / Changed

• Improved execution/token efficiency, telemetry, observability, and model-control planning.

0.9.x

Added / Changed

• Unified agent configuration, managed provider runtime, token telemetry, provider evidence, and acceptance baselines.

0.8.x

Added / Changed

• Automated real-provider acceptance, machine diagnostics, strict provider evidence, and provider/activity model routing.

0.7.x

Added / Changed

• Combined greenfield/brownfield certification, MCP contract hardening, portable read surface, and migration evidence improvements.

0.6.x

Added / Changed

• Brownfield context, task-aware execution, compact MCP contracts, and Cursor brownfield hardening.

0.5.x and earlier

Added / Changed

• Established the persistent SQLite-backed SDD workflow, provider configuration/bootstrap, schema migration, Git governance, and early local/provider E2E foundations.

For the curated pre-public evolution record, see docs/EVOLUTION.md.