Metaharness changelog: what's new each month
Every stable Metaharness release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 4 months covered; the current month updates daily.
October 2026
3 releases: 0.11.0 → 0.4.19
October brought refinements to Metaharness with two major API releases for contrastive testing and a maintenance update addressing validation, policy, and toolchain improvements.
Highlights
- New opt-in contrastive API with invariant edits, measured composition, deduplicated probes, and admission budgets.
- Independent confirmation path for contrastive testing on unseen task pairs before adoption.
- Security improvements including default-deny MCP policy emission and validation of receipt canonicalization.
- Enhanced validation with stricter checks on command-like project names, SDK names, and cost limits.
- Updated test toolchain to Vitest 4.1.12 with improved shell-injection and SDK alignment testing.
New
- Opt-in API for invariant and domain edits with scoped experimental memory.
- Admission budgets and measured composition features for contrastive testing.
- Deduplicated probes capability in contrastive API.
- Independent confirmation mechanism with paired statistical comparison and regression gates.
- Default-deny MCP policy for Claude Code and native Codex, Hermes, OpenClaw, Copilot, and OpenCode.
- Explicit validation for SDK names and duplicate registrations.
Improved
- Contrastive candidate performance increased from baseline 25/100 to 100/100 in synthetic benchmarks.
- Parser executions reduced from 64 calls through optimization.
- Receipt canonicalization with bounded depth to prevent invalid scores in evidence.
- Development confirmation improved from 23/24 to 24/24 in measured pilot testing.
- Policy preservation during upgrades while maintaining user edits.
- AGNTCY Directory adapter aligned with SDK 1.8 constructor.
Fixed
- Fixed root workspace lockfile dependency on Darwin 0.10.x for existing CLI.
- Reject non-finite signed numbers in evaluator to prevent invalid scores.
- Clear positive proof classifications when evidence cannot be digested.
- Reject ambiguous command-like project names while preserving doctor and overwrite protection.
- Resolved invalid SWE-bench cost limit validation across all five solver entrypoints.
September 2026
1 release: 0.4.17
September brought security hardening, correctness fixes for core workflows, and updates across the Metaharness ecosystem. Multiple packages received patches to address path traversal risks, evidence tracking, and adapter stability.
Highlights
- Fixed path traversal vulnerability in writeAtomic and strengthened host-adapter injection guards.
- Corrected sequential evidence wiring in run and replay operations.
- Improved MCP risk detection to no longer mask high-risk warnings under Grade A classification.
- Resolved promotion-rule lockout scenarios that blocked workflow progression.
Improved
- .env file guard anchoring enhanced for better protection.
- MCP detection now validated in settings.json configuration.
- Host adapters now properly preserve HarnessSpec.autonomous setting.
- Flywheel graph range handling improved to prevent RangeError exceptions.
Fixed
- Path traversal vulnerability in writeAtomic operations.
- Host-adapter injection attack surface.
- Sequential evidence not being wired into run and replay workflows.
- Flywheel graph RangeError crashes.
- Promotion-rule lockout conditions.
August 2026
1 release: 0.4.4
August brought major advances in defensive security, adaptive computation, and verifiable results for Metaharness. The month focused on hardening the framework against adversarial conditions while shipping production-ready evolution tools.
Action needed
- Solver-trajectory forward-contract enforcement: previously advisory checks (nogoldinloop, localizationnogold, bestofnselectorconformant) are now hard-enforced and will fail non-compliant configurations.
Highlights
- Darwin Shield adds zero-day defensive capabilities by applying Darwin Mode to security response itself.
- New metaharness learn command enables $0 dry-run evolution with the cand-6 genome included.
- Provably-clean SOTA feature cryptographically signs results via Ed25519 for attestable integrity.
- Non-stationary CFR/CFR+ poker solver now evolves against genuinely adversarial, shifting-strategy opponents.
- GEPA engine shipped as pluggable evaluator with full candidate lineage tracking.
New
- Darwin Shield defensive zero-day harness for security response.
- Self-learning WASM-memory cost-optimal escalator with tiered compute escalation across E2/E3/E4 cores.
- Non-stationary CFR/CFR+ poker solver with Chebyshev schedules for adversarial strategy evolution.
- GEPA Distilled Executor Genome with Agenticow branch-memory for full candidate lineage tracking.
- metaharness learn command for offline evolution with dry-run capability.
- Ed25519-signed integrity gate for cryptographically attestable results.
- @metaharness/darwin 0.8.0 as pluggable evaluator with cand-6 genome promotion.
Improved
- Self-learning escalator uses peer-reviewed mitigations for compute tier selection.
- Candidate lineage now tracked with version-control-like transparency in GEPA evolution.
- Solver trajectory checks (nogoldinloop, localizationnogold, bestofnselectorconformant) now hard-enforced instead of advisory.
June 2026
1 release: 0.1.3
June brought major expansion of Metaharness templates across 20 verticals with standardized plugin configuration, and significant hardening of the test and deployment infrastructure.
Highlights
- 20 vertical templates now ship with standardized plugin.json files for Claude Code marketplace integration.
- New per-vertical categorization and tagging system enables filtering by industry (devops, legal, research, support, trading, education, sales, gaming, crm, coding, ai, business, advertising, agentics, and more).
- 42 regression tests added to validate template propagation, substitution, and manifest consistency.
- CI validation now runs across 8 hosts to catch deployment issues before release.
New
- 19 new vertical templates (devops, legal, research, support, trading, education, sales, gaming, crm, coding, ai, business, advertising, agentics, and more).
- Plugin.json configuration file support for all 20 templates.
- Per-vertical category and tagging system for marketplace filtering.
- Static regression tests validating manifest consistency across templates.
- End-to-end regression tests verifying template substitution and placeholder cleanup.
- Multi-host CI smoke validation pipeline.
Improved
- Template consistency validation now runs on all 8 supported deployment hosts.
- Test coverage extended with 42 regression tests covering propagation and substitution paths.
Fixed
- Leftover template placeholders ({{}}) now caught by regression tests.
- Template substitution correctness validated across all verticals.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Metaharness: pricing, features and alternatives · All changelogs