← Back to .knowledge

// Blog / inspector

Visual inspector for AI coding agent trust and freshness

A trust layer is only useful if humans can see what is trusted, stale, suspect, and waiting for repair.

// direct answer

Short answer

A visual inspector turns .knowledge artifacts into an operating surface: health, trust buckets, freshness, repair queue, graph relationships, and critical paths. It does not replace the files; it makes the files reviewable by humans before agents act.

Understand what a visual inspector adds on top of repo-local knowledge files.

// real failure mode

The failure mode

A repository can contain trust reports and stale markers, but the team may never inspect them because they are buried in JSON.

The agent might know a module is suspect while the human reviewer sees only a clean-looking PR.

Without a visual surface, trust debt stays invisible until a wrong-file edit or stale-summary failure reaches review.

// visual workflow

The inspector fix

The Inspector renders the trust layer as a dashboard: health score, trust graph, module status, repair queue, and policy warnings.

It lets a reviewer see what the agent should recheck before code changes, especially around stale modules and critical paths.

The static local Inspector ships with the open core. Inspector Pro is the paid workflow layer for interactive team review.

// concrete example

When a module looks safe but is not

A PR modifies runtime execution. The code diff is small, but the module summary is stale and linked evidence is incomplete.

The Inspector surfaces the stale module, the missing evidence, and the repair item before review handoff.

The reviewer can ask the agent to re-read source and tests instead of approving a change against stale context.

// repo proof

Repo proof to inspect

Static Inspector

.knowledge/inspector/index.html

The local dashboard is generated from repo-local artifacts and can be opened without a cloud service.

Use it for
  • Visual trust review
  • Checking repair queue state
Do not use it for
  • Editing source files
  • Replacing code review
Trust report

.knowledge/maintenance/trust_report.json

The Inspector view should reflect the current trusted, suspect, stale, and low-confidence buckets.

Use it for
  • Trust bucket display
  • Review readiness checks
Do not use it for
  • Claiming product behavior without source
Freshness map

.knowledge/freshness.json

Freshness shows which knowledge artifacts may need recheck after source changes.

Use it for
  • Stale module warnings
  • Post-commit maintenance
Do not use it for
  • Ignoring current git changes
Repair queue

.knowledge/maintenance/repair_queue.json

Repair items convert uncertainty into visible work the team can assign.

Use it for
  • Prioritizing rechecks
  • Team review conversations
Do not use it for
  • Blocking every small edit

// command transcript

Commands and expected checks

node .knowledge/tools/build-visual-inspector.js
expected
The static Inspector is regenerated from current .knowledge artifacts.
inspect next
.knowledge/inspector/index.html
node .knowledge/tools/serve-inspector.js
expected
The generated Inspector can be served locally for review.
inspect next
Browser preview of the local Inspector
node .knowledge/tools/doctor.js
expected
The health status feeding the visual layer is checked.
inspect next
.knowledge/maintenance/quality_report.json
Inspector surface repo-local proof

What the human should see

01 Health: is the knowledge layer usable, cautious, or broken?
02 Trust graph: which modules are trusted, stale, suspect, or advisory?
03 Selected module: what should the agent re-read before acting?
04 Repair queue: what knowledge debt should become assigned work?

// field guide

Inspector signals

A visual layer should make hidden agent risk reviewable.
SignalQuestion it answersReviewer action
HealthIs the knowledge layer usable today?Run doctor or refresh before handoff.
TrustCan the agent use this summary for planning?Re-read source when trust is weak.
FreshnessDid code move after the summary?Rebuild or repair stale artifacts.
Repair queueWhat should not stay implicit?Assign or close the repair item.

// guardrails

What the agent should not trust blindly

  • The Inspector is a view over files, not a new source of truth.
  • A green visual state should still not override current code and tests.
  • Use the visual layer to catch stale context before it becomes a PR review problem.

// common mistakes

Common mistakes

  • Treating the dashboard as magic instead of generated state.
  • Showing graph visuals without source-of-truth rules.
  • Hiding repair items because they make the dashboard less pretty.

// quick FAQ

FAQ

Does the Inspector require cloud sync?

The static local Inspector can be generated from repository files. Inspector Pro is the paid workflow layer for team operations.

What is the most important Inspector signal?

Trust state. If a module is stale, suspect, or low-confidence, the agent should re-read current source and tests before behavior claims.

// page-specific next step

Use this page when trust needs to become visible

Generate the static Inspector from the repo-local artifacts, then use it to decide what an agent must re-read before acting.