DynosAI
Documentation navigation

Project Evolution

DynosAI has been developed through repeated local and real-provider acceptance loops. Each iteration has tried to preserve previously verified behavior while narrowing a concrete failure or efficiency problem.

Major stages

VersionsMain themeHow it was evaluated
0.4.x-0.5.xPersistent workflow foundation, schema/bootstrap, provider configurationLocal flow tests, schema upgrade tests, early Cursor runs
0.6.xBrownfield context, task-aware execution, compact provider contractsCursor brownfield E2E and migration scenarios
0.7.xCombined greenfield/brownfield certification, MCP contract hardening, portable readsLocal/provider E2E and authority/evidence tests
0.8.xProvider-native acceptance, machine doctor, strict Codex evidence, model routingReal-provider acceptance bundles and runtime audits
0.9.xUnified agent configuration, managed runtime, telemetry and stable provider baselineCursor/Codex runtime tests and measured acceptance
0.10.xExecution optimization, token telemetry, observability, roadmap for model controlAcceptance cost/token comparisons
0.11.xAdaptive Model Control Plane and boundary hardeningPhase-accurate telemetry, routing invariants, real Codex acceptance
0.12.0-0.12.2Predictive/context efficiency, checkpoint reuse and data-hygiene correctionsHistorical replay plus repeated Fibonacci acceptance
0.12.3-0.12.4Execution waves, deterministic auto-advance, transport de-duplicationReal Codex acceptance; implementation tokens reduced substantially
0.12.5Offline Predictive Router validationReplay with zero model calls; router intentionally kept shadow-only
0.12.6RC hardening: migrations, restart/resume, corrupted state/telemetry, scopes/gates257 selected source/wheel tests plus fault injection
0.12.7Cursor transport compatibility and consolidated 4-case matrixReal Codex/Cursor Green/Brown matrix; exposed provider-specific transport behavior
0.12.8Transparent infrastructure retry policySimulated retry certification and real full matrix
0.12.9Recovered-rejection-aware retry hardening280 selected source/wheel tests and final 4/4 real-provider matrix
0.13.0Stable promotion with no functional change from accepted RC4Final matrix certification, promotion invariants, offline predictive replay
0.14.0Product/control-plane layer on stable governed coreLocal App Server, Studio alpha, validation discovery, risk, explainable blockers, roadmap/web refresh
0.14.1Guided Studio UX refreshSetup guidance, themes/branding, task flow, Review Center, project checks, help, opt-in technical details
0.15.0Verified Agent HarnessContext handles, ExecutionRuntime, Validation Integrity, Eval Registry v0, execution policy; 0.14.1 Studio baseline preserved
0.16.0Governed Agent TeamsDAG waves, file-disjoint leases, claim/fan-in, reviewer/tester as contracts; no extra provider spawn; schema v6
0.17.0Eval IntelligenceFailure attribution, local-trace mining, inbox-only improvement loop, offline regression evidence; predictive routing still shadow
0.18.0Secure Autonomous RuntimeStrict/Balanced/Autonomous profiles, runtime-only vault, policy evidence; no Docker/VM or OS network interception
0.19.0Ecosystem & InteroperabilityCapability manifests for Cursor ACP and Codex app-server; uncertified clients and packs refused
1.0.0rc1Contract freezeFrozen MCP names, optional harness switches, shared release-gate policy; not production-ready 1.0
1.0.0rc2Certification evidenceMATRIX_1.0 not_run cells; two-session file-disjoint leases; 0.13 evidence kept historical
1.0.0rc3Eval maturityAcceptance ZIP importer, governed_change_cost aggregates, PROMPT_PREFIX_1.0 hash without cache-hit claims

How changes are recorded now

Public development uses three complementary records:

  1. CHANGELOG.md — user-visible release changes, grouped as Added / Changed / Fixed / Validation / Security when relevant.
  2. Git commits and pull requests — implementation-level history. Conventional Commit prefixes are recommended (feat:, fix:, docs:, refactor:, test:, chore:).
  3. docs/validation/ — selected machine-readable evidence for stable/release-candidate claims.

Historical raw release notes from the pre-public development cycle are retained under docs/history/, but new releases should avoid accumulating generated logs in the repository root.

Release philosophy

A version is not promoted because a feature appears to work once. The project has repeatedly used the pattern:

change
-> deterministic tests
-> installed-wheel tests
-> provider acceptance when required
-> inspect failures and attribution
-> fix only the demonstrated cause
-> re-run
-> freeze
-> stable promotion

This is the standard expected for future DynosAI releases.