agent-harness GitHub
Menu

Overview

agent-harness Agent Harness Blueprint Blueprint plugin product brief The Harness Book Deny-floor known limitations Licensing scope and decision Replay tool product brief Agent Harness roadmap Harness Specs

Docs

Advisory agent-operations measurements Mapped authority-document drift probe Mapped authority drift: pinned local trial Benchmarks and measured evidence Corpus path boundary Advisory entropy measurements Learning evaluator evidence boundaries Outcome stream validation System comparisons and canary evidence Advisory logical-job accounting Replay lab Python check inventory Replay CLI input identity and timeout limits Replay input decoding failures Replay report output boundary Replay process policy-root aliases Replay supervisor marker scope Review evidence and acceptance oracles Offline reviewer-packet diagnostics System state Process-fixture startup is not containment evidence

Agentic

Architecture additives — perfect assistant → estate owners Perfect autonomous agentic assistant — estate synthesis

Evals

Harness extension notes Evaluation merge-boundary contract: v0 Merge-gate model Outcome ledger Review under flood Agentic development evaluation taxonomy: v0

Evals / Examples

Worked evidence packet: a private product PR

Extraction

Replay v0 extraction inventory

Freeze

Legacy deny-floor v1 freeze record

Maintenance

Managed Codex agent publication Sync input and Doctor session roots

Observability

Coding-agent loop observability v0 Offline source-record identity inspector

Observability / Adapters

Claude Code to agent-loop v0: mapping note

Review evidence

Review/oracle wave: source ledger

UX evaluation

Frozen observation verifier Evidence-bound issue seeds Advisory product UX evaluation Offline scenario binder Scenario authoring contract v0 Taskdeck-first adapter sketch Workspace handoff acceptance profile Workspace receipt lab

Evidence-bound issue seeds

T4 #285. This is a formatter contract, not a GitHub publisher. A coordinator must search the target repository before filing. Use existing labels only after checking they exist.

Required structured fields

finding_id, repository, subject_revision, scenario_id, step_ids, dimension, severity, status (observed, hypothesis, needs_reproduction), evidence_sufficiency, confidence with rationale, observation, expected, actual, reproduction, artifact_refs, limitations, dedupe_key, related_issues, proposed_fix (optional), authority: advisory, gate_eligible: false.

Compute a stable dedupe key from product + journey + affected affordance + normalized failure mechanism, not the model’s title or severity. Keep exact evidence revision separately so a new build can retest the same finding. A changed severity is not a new defect. Similarity matches are suggestions, not permission to close existing issues.

Markdown output template

## Observed behavior
[Specific action, actual UI/state outcome, expected contract. No unsupported causal claim.]

## Reproduction and scope
Product revision: [full SHA]
Scenario/steps: [IDs]
Fixture/configuration: [sanitized reference]
Status: [observed / hypothesis / needs_reproduction]
Actions: [ordered minimal sequence]

## Evidence
[Artifact IDs, step/region/time references, sanitized accessible links if approved.]
Missing or unverified: [layers and environments not observed]

## Impact
Dimension: [effectiveness / usability / simplicity / clarity / feel]
Severity: [S0-S4, with impact rationale]
Confidence: [with evidence rationale, not a fabricated probability]

## Follow-through
Existing issue matches: [links or recorded none]
Proposed regression: [observable deterministic invariant, or explicitly subjective]
Proposed fix: [optional; separate from the observation]
Retest: [new revision and evidence required]

Advisory finding; not a merge/release verdict.

Do not publish this unfilled template as a product bug. The first dry-run may retain a filled draft locally rather than file it. Evidence referenced by a seed must exist and match its manifest. An inaccessible screenshot is not made public by embedding its local path.

S0/S1 concerns receive prompt human attention without automatic destructive action. S4 taste suggestions are batched under the design owner. A clean run produces an evidence summary, not an empty issue. A blocked observer produces a harness/environment finding, not an invented product defect. Future tests should prove duplicate suppression, missing-evidence refusal, safe redaction and separation of observations from suggested fixes.