Harness Inspector

Harness Inspector is a local, read-only delivery provenance workbench. It brings declared product intent, coding-agent sessions, normalized tool activity, Git commits, and repository paths into one view without treating correlation as proof of authorship or improvement.

Use it to answer three bounded questions:

1. What product intent, sessions, actions, files, and commits are visible?
2. Why does Inspector relate two objects, and how strong is that evidence?
3. What remains missing, inferred, or unsuitable for a stronger conclusion?

How it works

1. Choose a bounded local window. The CLI limits the inspected providers, dates, session count, and commit count. An optional Feature Tree supplies reviewed Feature and Story intent.
2. Collect independent evidence. Session adapters read supported local coding-agent evidence; Git and Entire readers collect commit, file, and checkpoint facts. A missing provider remains an explicit gap.
3. Normalize without widening the privacy boundary. Inspector retains privacy-safe prompts, provider-neutral action labels, bounded tool details, observed timing, and repository-relative paths. Raw tool payloads, hidden reasoning, credentials, and absolute home paths do not enter the report.
4. Build evidence-bounded relationships. Reviewed references, explicit commit/session metadata, observed commit calls, exact path overlap, time, and retained text have different authority. Every relationship carries facts, a source, and a limitation.
5. Render one self-contained report. The output is an interactive HTML file that reads the projected evidence. It does not write Git metadata, rerun a tool, restore code, or resume a native coding-agent session.

Reading the workbench

• Capability scopes the report by reviewed Feature and Story intent.
• Date keeps sessions and commits visible even when product mapping is absent or incomplete.
• User prompts / Agent activity / Commits and files are simultaneous lanes; one lane does not silently establish authorship in another.
• Evidence Drawer explains the selected relationship and its limitations.
• Session Trace shows retained turns, normalized actions, observed timing, idle gaps, files, and filters.
• Replay advances through retained evidence only. Sequence-only events stay labeled when an exact timestamp was not observed.

Evidence vocabulary

Label	What it supports	What it does not prove
Explicit / direct	A retained reviewed reference or explicit metadata links the objects.	Every action or changed line contributed to the delivery.
Observed same-path	Exact repository-relative paths occur in both session and commit evidence.	The session authored the final committed contents.
Candidate	Structure, retained text, or timing suggests a useful association.	A reviewed task identity or causal relationship.
Contextual	Nearby date or path history helps explain the visible delivery window.	A direct session, Story, or commit link.

Feature Tree selection and evidence confidence are separate. A Story can be a useful navigation scope while its link to a session is still only a candidate.

Generate a private Inspector

Run the shortest path inside the repository you want to inspect. npx downloads or reuses the current package, renders the current workspace, and opens the written report:

npx @qoder-ai/better-harness inspector

By default this opens the current workspace and includes activity from the latest 30 UTC days, bounded to 200 commits and 100 hydrated sessions. Date lists days with observed evidence rather than inventing empty-day activity. After installing Better Harness, or when you need explicit bounds, the expanded form remains available:

better-harness harness-inspector render --workspace . --open

The default output is:

.qoder/better-harness-runs/harness-inspector/inspector.html

Useful bounds include --platform, --since, --until, --commits, and --max-sessions. --platform all discovers evidence across Qoder, Codex, Claude Code, Cursor, Qwen Code, GitHub Copilot, Pi, Kimi Code, WorkBuddy, and Grok, while preserving provider-specific gaps.

An optional .better-harness/feature-tree.md supplies reviewed intent:

- [ ] Checkout reliability
  - [ ] Explain payment failure telemetry {#payment-failure-telemetry}

Two-space indentation declares the hierarchy. Internal nodes become Features; leaf nodes become Stories. Stable {#id} anchors are recommended when another artifact needs to reference a node.

Boundaries

• Inspector is an observation surface, not a recovery or mutation authority.
• A tool call labeled Run tests proves that the call was observed; it does not by itself prove that a test suite passed.
• Shared paths and nearby timestamps support correlation, not authorship.
• Missing token, cost, outcome, base-commit, budget, or harness-revision evidence must remain unobserved instead of becoming zero or a clean result.

Future evolution

The intended progression is Inspect → Compare → Eval → Decision.

• Compare should be a top-level workspace mode, not a third Capability or Date scope. Its first job is to assess whether two sessions are comparable.
• The first credible slice is Story-scoped observational trajectory comparison: align provider-neutral actions, show neutral deltas and confounders, and make insufficient evidence a normal result. Sharing a Story is a discovery aid, not proof of equivalent tasks.
• Stronger improvement or regression verdicts belong to later Eval evidence: bound harness revisions, controlled treatment, objective outcomes, compatible budgets, repeated trials, or held-out/later-comparable cohorts.
• Decision support may then retain, narrow, revise, or revert an intervention; Inspector Compare alone must not manufacture that authority.