Skip to content

Releases: SUNRNEHUI/agent-harness

Agent Harness v10.0.0

Choose a tag to compare

@SUNRNEHUI SUNRNEHUI released this 27 Aug 10:18

Agent Harness v10.0.0 is the first release under the shorter public name and runtime Skill ID.

Highlights

  • Breaking rename: Agent Reliability Harness / agent-reliability-harness is now Agent Harness / agent-harness.
  • Intent-to-contract entry gate: an explicit “写一个 harness” / “write a harness” request now produces a reviewable alignment packet and dispatch-readiness contract before mode selection or worker dispatch.
  • Sharper delegation: worker packets explicitly bind responsibility, scope, inputs/outputs, edge cases, verification evidence, PASS, Stop, and parent acceptance.
  • Proportional execution preserved: Native remains artifact-free; Portable materializes only the three-file contract; Audited adds only justified protected controls.
  • Compatibility preserved: legacy handoff-v1 / Full artifacts and existing arh-* payload discriminator values remain readable. Public schema URNs move to urn:agent-harness:* without changing payload structures.
  • Direct runtime download: the Release includes agent-harness-v10.0.0.zip plus a SHA-256 checksum asset.

Install

npx skills add https://github.com/SUNRNEHUI/agent-harness --skill "agent-harness"

Explicit invocation is now $agent-harness.

Migration from v9

  • GitHub repository: SUNRNEHUI/agent-reliability-harnessSUNRNEHUI/agent-harness.
  • Skill directory and ID: agent-reliability-harnessagent-harness.
  • Install the new Skill before removing the old directory. Existing v9 tags and releases remain available through GitHub's repository redirect.

Verification

  • 108 unit tests passed locally; the installed-package comparison passed separately.
  • 14 runtime behavior checks passed.
  • 16 adversarial protocol regression checks passed.
  • Canonical package build/check, Skill validation, version synchronization, Python compilation, JSON validation, and diff checks passed.
  • GitHub Actions passed discovery plus Ubuntu Python 3.10–3.14, macOS 3.14, and Windows 3.14.

Agent Reliability Harness v9.3.0

Choose a tag to compare

@SUNRNEHUI SUNRNEHUI released this 27 Aug 09:16
dcf9c01

Agent Reliability Harness v9.3.0 is the cumulative public release after v9.1.0. It combines the open Skill layout and proportional execution policy from v9.2 with versioned public contracts, cross-platform hardening, and a redesigned bilingual GitHub landing page.

Highlights

  • One canonical Skill package: installable source now lives under skills/agent-reliability-harness/, with repository-only tests and legacy material kept outside the runtime package.
  • Proportional execution: confirmed micro work may stay parent-direct, ordinary implementation uses one bounded worker, and complex work is parallel only across independent ownership boundaries.
  • Stable routing history: new runs use progress-bounded-v3; sealed progress-bounded-v2 and cost-aware-v1 maps remain readable without reinterpretation.
  • Public machine contracts: five Draft 2020-12 schemas cover Portable v2, Audited run state, acceptance registry, worker results, and arh-status-v1.
  • Automation output: status.py --json provides a versioned projection, while --require-high-confidence restores the strict exit gate without changing default human output.
  • Cross-platform integrity: canonical JSON and digest anchors use deterministic UTF-8 plus LF bytes; CI now covers Python 3.10 through 3.14, macOS 3.14, and Windows 3.14.
  • GitHub presentation: English and Chinese README pages now include truthful badges, quick start, mode comparison, Mermaid flow, documentation navigation, and aligned release history.

Install

npx skills add https://github.com/SUNRNEHUI/agent-reliability-harness --skill "agent-reliability-harness"

Compatibility

  • Portable Contract v2 and legacy handoff-v1 artifacts remain supported.
  • Human status output remains the default; JSON output is opt-in.
  • No license is asserted because the repository does not currently include a license file.

Verification

  • 108 unit tests passed; 1 opt-in installed-package test was verified separately.
  • 14 runtime behavior checks passed.
  • 16 adversarial protocol regression checks passed.
  • Five schemas, canonical package build/check, Skill discovery, README parity, version sync, and diff checks passed.
  • PR checks: 16/16 successful.
  • Merged main checks: 8/8 successful, including Windows and macOS Python 3.14.

Agent Reliability Harness v9.1.0

Choose a tag to compare

@SUNRNEHUI SUNRNEHUI released this 02 Aug 07:17
272d50f

Highlights

  • Replaced the prompt-level Direct/Lite/Full workflow with Plan-native Native, Portable, and Audited execution.
  • Added Portable Contract v2, bounded resume capsules, fail-closed continuation, owner fencing, and legacy handoff-v1 compatibility.
  • Added an observable Progress Circuit Breaker and expanded adversarial protocol regression coverage.
  • Aligned Codex routing around a Sol max parent for planning and acceptance, with Luna max for justified long or mechanical execution.
  • Added optional public Codex configuration guidance while keeping private model-cache overrides outside the distributed skill.

Validation

  • Model routing: 16/16 passed
  • Plan-native and Portable: 11/11 passed
  • Handoff/resume: 21/21 passed
  • Protocol regression: 17/17 passed
  • State Witness: 11/11 passed
  • Runtime behavior, artifact binding, lessons, Python compilation, version sync, package verification, and diff checks: passed
  • Generated runtime package matches the local installed skill

The skill records requested and resolved models separately. Luna availability is runtime-dependent; the distributed package does not modify Codex model caches or install private catalog overrides.

Agent Reliability Harness v7.4.0

Choose a tag to compare

@SUNRNEHUI SUNRNEHUI released this 18 Jul 09:22

Highlights

  • Added project-root discovery and automatic continuation for the unique active Full run across Codex and Grok runtimes.
  • Added transaction recovery, atomic owner transfer, actor-plus-epoch fencing, complete resume packets, checkpoint/handoff commands, and legacy artifact upgrade.
  • Added fail-closed ambiguity and corruption handling, repository content drift detection, and continuation status output.
  • Aligned Grok routing, Codex/Grok/universal adapters, evidence refresh commands, runtime packaging, regression tests, and bilingual README documentation.

Validation

  • Runtime behavior: passed
  • Handoff/resume regression: 21/21 passed
  • Model routing regression: 14/14 passed
  • State witness regression: 11/11 passed
  • Protocol score: 100/100 (Grade A)
  • Runtime source and installed package checks: passed
  • Python compilation, JSON validation, version sync, and git diff check: passed

Automatic takeover starts when the replacement runtime is launched and runs resume from the project root. The harness transfers durable observable state only; it does not transfer hidden reasoning, provider session internals, credentials, quota callbacks, or in-flight external side effects.

Agent Reliability Harness v7.2.0

Choose a tag to compare

@SUNRNEHUI SUNRNEHUI released this 15 Jul 07:33

Summary

Agent Reliability Harness v7.2.0 makes Production State Witness a runtime-enforced control loop for state/UI/async/concurrency work.

Changes

  • Added state_witness.md schema and semantic checker with real call-chain/source-locator requirements.
  • Added reachable failing/fixed/preserved state rows with observed-before/expected-after values and executable evidence locators.
  • Added controller gates for seal, run-set dispatched, dispatch-create, validate, and protected acceptance.
  • Added witness-set for independent review evidence and policy/flow/user_visible verification tiers.
  • Added sealed witness digests and post-seal mutation detection.
  • Updated Full/Lite/worker/evaluator templates, protocol docs, evaluation cases, scoring, packaging, and bilingual release history.
  • Added regression coverage for hollow witnesses, mismatched production states, missing seal, insufficient verification tier, post-seal mutation, and the complete stateful acceptance path.

Verification

  • State Witness regression: 11 tests passed.
  • Runtime regression: all checks passed.
  • Runtime package/source verification: passed.
  • Installed skill package drift check: passed.
  • Protocol score: 100/A.
  • Python compilation and git diff --check: passed.

Boundary

The harness validates the evidence contract and controller transitions. Actual Xcode/device user-visible evidence remains required for app-specific acceptance.

Agent Reliability Harness v7.0.0

Choose a tag to compare

@SUNRNEHUI SUNRNEHUI released this 14 Jul 13:56
1e967a7

Agent Reliability Harness v7.0.0

This release completes the public rebrand from Agent Dispatch Harness to Agent Reliability Harness.

Highlights

  • Rebrand the project and Codex Skill as Agent Reliability Harness.
  • Add the new runtime identity: $agent-reliability-harness.
  • Move the local Codex installation path to ~/.codex/skills/agent-reliability-harness.
  • Restructure the bilingual README around product positioning, guarantees, execution modes, model routing, installation, runtime support, verification, and migration.
  • Keep Direct/Lite/Full proportionality, durable Full Harness state, evidence-based acceptance, and cost-aware Codex routing.
  • Preserve migration guidance for Agent Dispatch Harness and Multi-Agent Dispatcher installations.

Codex routing defaults

  • fast: GPT-5.6 Luna, medium
  • main: GPT-5.6 Luna, xhigh
  • planner: GPT-5.6 Sol, high
  • critical_reviewer: GPT-5.6 Sol, xhigh

Terra remains excluded from the configured Codex policy.

Validation

  • 57 behavior tests passed
  • Runtime package and installed Skill verified
  • Skill protocol score: 100/100, Grade A
  • English and Chinese README heading parity: 13 sections
  • Version synchronization and diff checks passed

Agent Dispatch Harness v6.0.0

Choose a tag to compare

@SUNRNEHUI SUNRNEHUI released this 14 Jul 01:12
7502e7b

Agent Dispatch Harness v6.0.0

Universal task execution OS for coding agents (Codex / Claude / Grok / others).

Highlights

  • Density first: Direct → Spec Synthesis → Lite → Full (lightest process that still controls false completion)
  • Spec Synthesis for fuzzy / improvement-shaped goals (init_run.py --with-synthesis)
  • Progressive reference loading and shorter entry prompts
  • Optional plan quality scoring (score_harness.py) — never product acceptance
  • Universal runtime adapter + bilingual README / VERSION management
  • Preserved runtime safety from main: manager state/trace API, atomic JSON, timeout budgets, workspace binding

Install

git clone https://github.com/SUNRNEHUI/agent-dispatch-harness.git
cd agent-dispatch-harness
git checkout v6.0.0
python3 scripts/package_skill.py --verify-source
python3 scripts/package_skill.py --output /tmp/agent-dispatch-harness-runtime --force
mkdir -p ~/.codex/skills/agent-dispatch-harness
rsync -a --delete /tmp/agent-dispatch-harness-runtime/ ~/.codex/skills/agent-dispatch-harness/

Verification

  • test_runtime_behavior.py PASS
  • score_skill_protocol.py 100 / A

Agent Dispatch Harness v5.11.0

Choose a tag to compare

@SUNRNEHUI SUNRNEHUI released this 13 Jul 12:47

v5.11.0

Changed

  • Compared agent-dispatch-harness with Cjbuilds/Codex-Orchestration and documented the adopted and rejected patterns in docs/codex-orchestration-review.md.
  • Kept the active parent model as the only root orchestrator and made explicit no subagents instructions authoritative.
  • Added truthful model-route states: requested, accepted, used and confirmed, and inherited root.
  • Added scripts/status.py --require-high-confidence for CI and release gates. Default human-readable status output remains unchanged.
  • Replaced the default question loop with decision alignment that asks only for scope, ownership, irreversible-action, verification, or user-visible decisions.

Deliberate non-goals

  • No second orchestrator, default advisor, cross-provider bridge, persistent global routing mutator, full transcript fork, unbounded follow-up, or small-task ceremony bundle.
  • No fixed token or cost savings claim. Real GPT-5.6 Luna/Terra/Sol benchmarks are still required for that conclusion.

Verification

  • Runtime behavior tests: 26/26 PASS
  • Python compilation, JSON validation, version sync, runtime source verification: PASS
  • Runtime package build/install check: PASS
  • Full/Lite initialization and report validation smoke checks: PASS
  • git diff --check: PASS

v5.10.0 - GPT-5.6-aware dispatch routing

Choose a tag to compare

@SUNRNEHUI SUNRNEHUI released this 13 Jul 09:14

Highlights

  • Add GPT-5.6-aware Codex routing: use gpt-5.6-luna with low reasoning for simple sub-agents, escalate to Terra/Sol by task complexity, risk, and context.
  • Add bounded fan-out, depth-one nesting, task-local worker context, compact reports, and truthful fallback when model controls are unavailable.
  • Add machine-readable token budget state with actual/estimated/unknown accounting, exhaustion blockers, and accepted-state validation.
  • Make Superpowers-style methods optional and risk-triggered, avoiding default TDD/review/worktree ceremony on low-risk work.
  • Update bilingual docs, runtime packaging, eval cases, and release metadata to v5.10.0.

Validation

  • python3 scripts/test_runtime_behavior.py - 23/23 PASS
  • python3 scripts/tdd_gate_check.py workspace/5-6-strategy-optimization/tdd_trace.jsonl - PASS
  • python3 scripts/package_skill.py --verify-source - PASS
  • python3 scripts/sync_version.py - PASS
  • Python compilation, JSON validation, and git diff --check - PASS

Notes

This release does not claim numeric token or cost savings; representative end-to-end GPT-5.6 A/B benchmarking remains a follow-up. The release tag points to the published branch commit while PR #3 remains draft for maintainer review.

Agent Dispatch Harness v5.9.0

Choose a tag to compare

@SUNRNEHUI SUNRNEHUI released this 09 Jul 03:39

Summary

Agent Dispatch Harness v5.9.0 adds a proportional Completion Confidence Loop so multi-agent runs can answer whether many delegated tasks are actually complete, what evidence supports that claim, and what verification remains.

Changes

  • Added completion rollup output to scripts/status.py: task completion, Full acceptance status, evidence gaps, state conflicts, confidence bands, and next verification guidance.
  • Added regression coverage for high-confidence happy paths, pending/fail/blocked acceptance, missing or malformed registries, passed work without evidence, and accepted run-state conflicts.
  • Expanded verification and evaluator guidance to compare final claims against fresh evidence.
  • Added lightweight confidence/evidence-gap prompts to Lite and progress templates without adding new artifact types.
  • Synced VERSION, README, README.zh-CN, SKILL footer, and version sync default date to v5.9.0 / 2026-07-09.

Validation

  • python3 scripts/sync_version.py
  • python3 -m json.tool templates/run_state.json >/dev/null
  • python3 -m py_compile scripts/init_run.py scripts/package_skill.py scripts/status.py scripts/sync_version.py scripts/validate_report.py scripts/harness_test_run.py scripts/tdd_gate_check.py scripts/test_runtime_behavior.py
  • git diff --check
  • Full and Lite init_run.py smoke checks
  • validate_report.py for spec, progress, evaluator, and lite plan
  • scripts/status.py Full and Lite smoke checks
  • python3 scripts/test_runtime_behavior.py
  • python3 scripts/package_skill.py --verify-source
  • python3 scripts/package_skill.py --output /tmp/agent-dispatch-harness-runtime --force
  • python3 scripts/package_skill.py --check /tmp/agent-dispatch-harness-runtime