Changelog

This file records notable public changes to Better Harness. Entries describe observable behavior and compatibility, not every internal refactor.

0.6.4 - 2026-08-19

Added

• Session analysis now supports DeepSeek Harness evidence, including bounded project discovery, normalized messages and tool lifecycles, workspace matching, and Inspector/commit-session integration.

• Harness Inspector can list sessions for the selected UTC day and present session and experiment evidence in denser notebook-oriented views.

Changed

• Inspector's Capability, Date, Trace, and Replay surfaces now share the repository design contract for readable typography, semantic state and categorical colours, keyboard navigation, responsive density, and bounded overflow.

• The public homepage and README make Harness Inspector easier to discover, while the self-contained report uses a flatter workbench hierarchy and a clearer date calendar and session detail flow.

Fixed

• Tool-call details keep redaction visible in the Inspector instead of making protected content look accidentally empty.

• Cross-platform source and test paths are canonicalized so Windows short-path aliases and line endings do not create false failures.

0.6.3 - 2026-08-14

Added

• npx @qoder-ai/better-harness inspector now renders the current workspace, opens the generated report, and uses a bounded 30-day UTC evidence window by default. The public Inspector page and bilingual guide lead with the same short command.

Changed

• The zero-argument Inspector quickstart scans up to 200 commits and hydrates up to 100 sessions, while explicit render invocations preserve their existing bounds and open only when --open is provided.

Fixed

• Qoder assistant messages with structured thinking, text, and tool_use content are normalized before Session Detail renders them, preventing raw transport JSON from appearing as intermediate responses while preserving tool calls in the structured activity trace.

0.6.2 - 2026-08-14

Added

• Harness Inspector now documents its local evidence pipeline, relationship strengths, privacy boundaries, and CLI workflow in a bilingual concept guide with a source-backed architecture diagram.

• The Inspector renderer accepts --open so a generated self-contained report can be opened in the default browser after it is written.

Changed

• Session View uses denser activity and commit presentation, compresses long idle windows, and keeps expanded activity focused while preserving access to the full retained trace.

• Inspector chrome now uses one workspace identity, clearer breadcrumbs, and the visible Capability label consistently across navigation and docs.

• Qoder CLI installation guidance now distinguishes the Desktop-bundled path from standalone marketplace and Git installation.

Fixed

• The Inspector sticky header is isolated from trace content so scrolling and focused expansion do not create overlap.

0.6.1 - 2026-08-13

Added

• Harness Inspector now includes a read-only Session Replay that advances through retained prompts, intermediate responses, normalized tool calls, assistant responses, and observed commit events without rerunning tools or resuming the coding-agent session.

• The GitHub Pages site has a first-class Inspector tab with an interactive, deterministic English sample. The bilingual wrapper explains the Workbench, its three evidence lanes, usage flow, evidence labels, and the command for generating a private self-contained report from a local repository.

Changed

• Session View places the elapsed-time activity chart beside the retained Turn trace, links chart selections into the corresponding calls, and shows a continuous ribbon that distinguishes observed tool execution from unattributed time.

• Capability navigation opens the declared Delivery Tree by default and keeps scope navigation separate from evidence selection. Short sessions expose their tool calls by default while repeated call runs remain compact.

Fixed

• Story-to-session candidate matching filters generic stop words before scoring overlap, reducing incorrect associations caused by broad terms such as project, session, or harness.

• Session View filters now keep visible tool-call totals and collapsed run groups aligned with the current selection.

0.6.0 - 2026-08-13

Added

• Harness Inspector is now available from the published CLI through better-harness harness-inspector and the better-harness inspector shortcut. inspector render creates a self-contained, read-only HTML workbench that relates Feature Tree stages, Stories, prompts, sessions, tool calls, files, and commits across the supported session providers. Its synchronized Evidence Drawer explains why evidence is linked, states known limitations, and distinguishes commits created during a session from files merely present in those commits.

• better-harness commit-session-link correlates bounded Git history with coding-agent sessions and renders commit-oriented provenance evidence. Long-session reports can now retain privacy-safe tool activity and file evidence for trace inspection instead of reducing execution to aggregate counts.

• A new Harness Component Snapshot contract and direct CLI captures, compares, and resolves non-authorizing rollback references for bounded project-owned Harness component state. Standard report analysis can also surface evidence-bound native Learning Capture candidates without requiring adapters to assign pattern labels.

Changed

• The repository test suite now runs on Vitest with human-readable failures in the main GitHub Actions log, source annotations, a Job Summary, and JUnit output. Existing Node assertions remain intact, while Windows, macOS, Linux, Node 22.20.0, and Node 24.x remain release gates.

• Test ownership is organized by capability, and contributor commands now use the same Vitest discovery contract locally and in CI.

Fixed

• Claude session discovery resolves underscore-based transcript directories, component snapshot failures retain bounded diagnostics, and review-trigger stop-hook results use a structured cross-platform output contract.

• CI test module paths remain valid on Windows drive-letter workspaces, and failure output identifies the owning test instead of reporting only a failed capability group.

0.5.0 - 2026-08-04

Added

• better-harness plugin status, plugin plan, plugin verify, and better-harness doctor expose a read-only lifecycle control plane over eight host profiles and eleven host surfaces. Status reports installation, enablement, observed-version relation, and verification per surface from the public configured-asset inventory; plugin plan emits typed native argv or manual steps for install, update, and remove without executing them; doctor reports bounded runtime and host diagnostics with redacted authorized roots. Unknown, mixed, foreign, or unbound host state fails closed, plugin apply stays unregistered, and Kimi Code and Grok are rejected with UNKNOWN_HOST until their native lifecycle contracts are validated.

• Kimi Code is now a supported analysis-capable source-local host. The repository installs as a Kimi Code plugin (/plugins install <repo>) through a .kimi-plugin/plugin.json manifest, gains a Kimi configured-asset provider (user ~/.kimi-code/skills and mcp.json, project .kimi-code/skills and .kimi/skills, and managed plugins from plugins/installed.json with enabled filtering and plugin-root path confinement) plus a Kimi session-evidence adapter that reads workspace-matching wire transcripts under ~/.kimi-code/sessions/<wd_*>/ses{sion}_*/agents/*/wire.jsonl, resolving the workspace mapping through workspaces.json and session_index.jsonl with a wd_<name>_* prefix fallback that records a kimi-workspace-index-absent warning. The public npm package now ships seven host metadata roots; the Qoder runtime bundle remains Qoder-specific.

• A read-only native Learning Capture review contract can now screen ordinary Task Episodes for repeated exact repair routes, emit a bounded privacy-safe packet, validate evidence-bound match or abstain decisions, and project accepted recurring-correction opportunities through the existing Learning Loop candidate model without requiring adapter-supplied pattern labels.

Fixed

• Portable HTML finding-bound fixes now record against the HTML report contract without requiring Qoder's canvas.json, and refresh findings.json, report.md, and report.html to the same repair revision. Qoder split reports retain their Canvas-sidecar validation boundary.

• Root CLI delegation failures now keep machine mode parseable: spawn errors, signal termination, and output-buffer exhaustion each emit one stable JSON error document, while normal child stdout, stderr, and numeric exit status remain capability-owned.

• Checkup plan/apply is provider-aware: only provider=qoder can emit or execute qodercli disable mutations. Other hosts keep candidates as manual-review until a provider-native apply contract exists. provider-home source resolution and fingerprints bind to the explicit host home (for example Codex uses codexHome, never Qoder home).

• Make command describe resolve exact registered leaf paths instead of returning the parent command metadata.

• The Portable HTML report route in templates/reporting/routing.md now lists WorkBuddy, so agents on WorkBuddy are routed to the self-contained HTML + Markdown output the 0.4.0 host adapter already ships. A derived support-declaration check now requires every adapter-matrix host claiming portable HTML output to appear in that routing row.

0.4.1 - 2026-08-04

Fixed

• The published npm package identity is @qoder-ai/better-harness. The previously documented @qoderai/better-harness scope was never a valid registry name, so package metadata, the lockfile, the adapter matrices, and the documentation site now all reference the hyphenated scope.

• The repository test script runs the automated suite again, so the release workflow verifies tests before publishing instead of reporting success without running them.

0.4.0 - 2026-07-30

Added

• WorkBuddy is now a supported analysis-capable source-local host. It gains a WorkBuddy configured-asset provider (user skills, marketplace plugins with settings.json enabled state, mcp.json/.mcp.json user and plugin MCP servers, the global AGENTS.md and identity context files) plus a WorkBuddy session-evidence adapter that reads workspace-matching JSONL transcripts under ~/.workbuddy/projects/, including cwd-less 5.x transcripts from exact workspace-slug directories and sparse camelCase/snake_case usage, with a WORKBUDDY_DIR override. WorkBuddy has no install shell in this repository; the skill installs by copying it into ~/.workbuddy/skills.
• Pi (pi.dev) is now a supported analysis-capable source-local host. The repository installs as a pi package (pi install <repo>) through a pi manifest in package.json, registers a /better-harness prompt template, and gains a Pi configured-asset provider (settings-declared pi packages, skills, prompt templates, extensions, and AGENTS.md context) plus a Pi session-evidence adapter that reads workspace-matching JSONL v3 transcripts under ~/.pi/agent/sessions/ with PI_CODING_AGENT_DIR and PI_CODING_AGENT_SESSION_DIR overrides. Pi's shell is the pi manifest in the existing package.json, so the public npm package still ships six host metadata roots and the Qoder runtime bundle remains Qoder-specific.

Changed

• Cursor installed-plugin inventory now leaves unknown numeric or opaque IDs unmatched instead of assigning them to cached plugins by name/order. Direct manifest IDs and workspace project MCP hints remain supported.
• harness record-fix-output now resolves Home only for Global output, so a verified Project-only result remains recordable when Home is unavailable.
• The harness analyze platform gate now names the full supported set (qoder, codex, claude, cursor, qwen, copilot, pi) when it rejects an unsupported --platform, matching the session-analysis and asset-baseline gates. The existing error prefix and exit behavior are unchanged.
• Core Change Watch now requires framework-specific evidence before labeling Rails or FastAPI, exposes bounded root Just recipes as statically discovered unverified argv entrypoints, and keeps historical-only files out of current recommended reads and action targets.
• Evidence bundles now discover and privacy-filter one frozen Session population before either Session facts or lead analysis hydrates it. Versioned redacted bindings fail closed on population, selection, or admission contradictions while preserving bounded lead selection and explicit zero-signal filtering.
• Self-contained HTML reports now expose every fluency-dimension score track as a labeled progressbar with a zero-to-100 range and the displayed rounded score. Report validation rejects incomplete, duplicated, invalid, or score-mismatched dimension progressbar contracts.
• Chinese self-contained HTML reports now use standards-based language segmentation to keep bounded word-like phrases together while preserving normal wrapping around Latin text, paths, URLs, and longer content. Runtimes without segmentation support fall back to readable escaped text, and English reports remain unchanged.
• HTML Evidence cards now display machine-owned Task Episode coverage from a summary-facts companion, with legacy at-a-glance coverage retained only as a compatibility fallback.

0.3.0 - 2026-07-27

Changed

• The public npm package now includes the Qoder, Claude Code, Codex, and Cursor plugin metadata roots with aligned public descriptions. The generated Qoder runtime bundle remains Qoder-specific.
• CI now follows the main branch, and repo-local Agent Skills use SKILL.md directly without a mirror sidecar contract.
• Claude Code now defaults /better-harness to a validated, self-contained HTML report with paired Markdown and findings artifacts. Explicit inline or no-files requests remain write-free.

Removed

• Removed pre-public identity aliases, migration-only specifications, and local compatibility readers. Better Harness is now the only product, CLI, plugin, callback, report-root, and session-reference identity.
• Removed developer-specific paths and obsolete compatibility commands from the public terminal-demo documentation.