Agent Harness v10.0.0

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-harness → SUNRNEHUI/agent-harness.
• Skill directory and ID: agent-reliability-harness → agent-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

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

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

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

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

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

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

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

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

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