In short, here's what we offer:
Latest model compatibility
Ability to specify a separate model depending on the agent's role
More refined harness configuration through interviews during harness setup
https://claude.ai/artifact/NEgL1vj4Dmn1UJYJpuXRWG
Release Notes
Model Cognition Harness
1.9.02026-09-25Previous release 1.8.0
When creating a harness, it asks six questions. The answers determine the model selection, approval point, and whether to use external reviewers.
New Features
What's New in 1.9.0
Added "External Export" Question to Configuration Interview
It now asks if your code and documentation can be sent to external engines outside the tools you are currently using, and selects external reviewers accordingly.
No Need to Specify Model Names Directly
Just provide a one-line description of the role, and it automatically determines the appropriate grade and inference intensity. It also cross-references this with your agent definition file.
Predefined Alternate Models
You can now specify a sequence of alternate models to use in case the primary model is unavailable, directly within your harness settings.
Future-Proofing with Model Generations
Instead of using generation names, it refers to models by lineage. This means you don't need to update your harnesses when a model receives a new version.
Creating a New Harness
Six Key Questions Asked
When you request a harness, it first analyzes your project and asks only the relevant questions for that specific project. For example, if a project doesn't involve deployment or tag releases, questions about "releases and tag releases" won't be asked.
Each question includes a recommended answer with a brief explanation. You can accept the recommendation or choose a different answer.
①
What signifies completion of the task?
Passing tests · CI green · Output existence · Human final review
→ This defines the completion criteria, which the harness uses to determine when a task is finished.
②
What actions are irreversible?
Release · Tag release · Package deployment · DB migration · Forced push · External sending · Unknown · None
→ Tasks reaching these points are automatically classified as critical.
③
What is more painful when something goes wrong?
Errors are more painful · Delays are more painful · Both are equally painful
→ This determines the verification intensity and the grade of the deployed model.
④
What points require mandatory human review?
For each item selected in ②, "…approval just before" is automatically created · Critical tasks require approval at the planning → plan → execution stages · Autonomous progression allowed
→ These become approval checkpoints.
⑤
How should existing agents and skills be handled?
Reuse prioritized · Reference only · Ignore
→ This prevents the accumulation of role definitions with different names.
⑥
Can content be sent to external APIs? NEW
Only the tools currently in use · Allowed list only (already installed and used review engines) · No restrictions
→ This answer is checked every time an external reviewer is selected. It cannot be arbitrarily expanded through environment variables.
You don't have to answer all six questions. Unanswered questions are filled in with the safest option, and the harness is still created.
However, a note will be added to the documentation indicating that these values were assumed by the harness, not explicitly set by you. This allows you to distinguish between your choices and the harness's assumptions when reviewing the documentation later.
- ⚠ Assumption (No Answer): Completion Criteria = Test gate passage · CI green · Output path existence
- ⚠ Assumption (No Answer): Failure Cost = Errors are more painful — better late than inaccurate
- Irreversible: Release · Tag release · Unknown → Stages reaching this list are classified as critical
If you want to change the assumed value, just answer that question again. The rest of the answers will remain the same.
Model Awareness
Models are defined by their roles
When creating an agent, you don't specify a model name. You provide a one-line role description, and the system classifies it, assigning a corresponding grade and inference intensity. These are then recorded in the definition file. If an agent spans multiple classifications, it is assigned a higher grade.
Deeply Judgement·Verification·Security · Design·Planning · Implementation Inference Intensity High
Standard Coordination·Integration · Documentation·Release Medium
Lightly Collection·Search·List Low
Ambiguous roles are classified as ③ Failure Cost Response. If "delay is worse," it's treated lightly; otherwise, standard. For irreversible one-time implementations, if "error is worse," it's elevated to deep.
New generation models don't require changes. Harness doesn't specify generational names like Fable 5.1 or GPT-6 Astra. Instead, it uses lineage names. The actual model appended to the lineage name is updated on the tool side, so you won't need to modify harness documentation or agent definitions. You can directly fix a model ID if you want to bind to a specific version.
Codex·Gemini use the model set in their CLI. Harness doesn't swap models. Instead, it translates inference intensity according to each product's method — setting names and available values vary by product, and some values are ignored. It omits sending ignored settings.
After batching, the agent definition file is reread to verify that the value was actually applied. Any discrepancies are immediately reported.
Batch: Judgement role → Deep · Inference intensity high
Confirmation: Two matches in definition files
Existing Harness
How to Upgrade a Harness Made with 1.8.0
If your repository already contains a harness, upgrade the factory to 1.9.0 and then review these five points.
Fix the execution line once. Before fixing it, all external reviews for that harness will fail. The tool will tell you what to append.
$ harness-update.sh plan <harness> <factory>
LAUNCHER: needs-update(.claude 149)
Append line: REVIEW_GRADE=… HARNESS_ORCHESTRATOR=… bash "…/run-review.sh" …
Recreate the interview result block embedded in the harness documentation. As questions are added, existing blocks are marked as outdated (stale). The order remains the same.
Run updates four times if you use both Claude and Codex. Two runtimes × two skills. Skipping a step leaves old scripts behind.
It's best to answer question ⑥ once. Daily tasks will continue as usual, but using major grade verification will result in a "assumed release policy" tag, preventing review completion.
Choosing "**only the tools currently used**" for question ⑥ eliminates external reviewers. This means you'll need to manually approve each major stage.
Things to Keep in Mind
What Doesn't Work Yet
We can't predict when a lineage name will disappear. While we're strong at generational changes, if an entire lineage vanishes, it only surfaces as an error during execution. We don't yet have a function to periodically check and notify you.
Model data staleness is only checked in the factory. In harnesses, it only manifests as errors during execution.
We haven't measured how much interviews improve results. Comparing before and after has been postponed since 1.8.0.
In Windows environments with only executables lacking extensions, the installed reviewer may be omitted from the list. This doesn't apply to tools installed via npm.
my_harness 1.9.0Release PageRepository Documentation Path docs/v1.8.3/ (Development Codename)
▶ Original Source: https://claude.ai/artifact/NEgL1vj4Dmn1UJYJpuXRWG
▶ Original Source: https://github.com/cookyman74/my_harness/blob/main/README_KO.md
