Skip to content

Repo Analysis

The dashboard’s Repo Analysis page (/repo, in the sidebar after Rules) gathers what Itervox knows about the repository your agents work in: what they load, what it costs, which prompt version produced which runs, whether the repository is set up for agents, how the profile evals score, and where agent work goes wrong. On a wide screen each part is a tab; on a phone they are stacked. Each has a deep link:

SectionLinkWhat it shows
Capabilities/repo#capabilitiesThe skills inventory: skills, subagents, plugins, MCP servers, hooks, instruction docs, analyzer recommendations with Fix, Re-scan
Runtime usage and cost/repo#usageRuntime analytics and their recommendations, tokens per profile
Prompt versions/repo#promptsPer profile, each prompt version with its runs, success rate and duration
Agent readiness/repo#readinessThe repository checks itervox doctor runs, pass / warn / fail with the fix
Evals/repo#evalsThe latest saved eval result per profile and scenario: pass / fail, failed assertions, prompt digest, regressions
Hotspots/repo#hotspotsFiles and directories ranked by failed or stalled runs, change requests, input requests and reverted PRs

In the Mac app, each daemon’s menu has a Repo Analysis item that opens this page in the daemon’s dashboard window. The command palette lists it too (g p).

The skills inventory, as it was in Settings before (Settings now links here): the scan status and Re-scan, the analyzer’s recommendations (expand one for what it means and how to fix it; Fix applies a one-click fix, and asks first when it edits configuration), a tile per capability kind with its count and approximate context cost (click one for the list), the skill cost split by provider and the context budget bars. The Skills Inventory guide explains the scanners, rules and fixes.

What the session logs show agents actually load: how many skills are runtime-verified, how many hooks fired, and the analytics recommendations — HIGH_COST_LOW_USAGE (a costly skill not seen loaded lately), HOOK_STORM (a hook firing very often), CONFIGURED_NOT_LOADED and LOADED_NOT_CONFIGURED (inventory and runtime disagree). Without session logs the section says how to collect them (CLAUDE_CODE_LOG_DIR).

Tokens per profile breaks each agent profile’s approximate context cost down into instructions, skills, hooks, MCP tool schemas and the prompt template, so you can see which profile carries the most before it reads the issue.

Every run records the version of the prompt it was dispatched with: a digest of the profile’s SOUL.md, its INSTRUCTIONS.md, the resolved rules and the prompt template (see run history). This section lists, per profile, each version with when it was first and last seen, how many runs used it, their success rate and their average duration, newest first. Select a version to list its runs (the dashboard reads GET /api/v1/history?prompt=<digest>); Clear filter closes the list.

Use it after you change a profile’s SOUL or INSTRUCTIONS or a rule: the new version appears at the top once a run uses it, and its success rate and duration next to the previous version’s tell you whether the change helped. Runs recorded before prompt versions existed are counted below the tables but not listed.

The repository-level checks of itervox doctor, from GET /api/v1/repo/readiness. The page and doctor build their rows from the same check functions, so they agree; doctor prints them in its repo readiness: section.

CheckPasses whenOtherwise
WORKFLOW.md schema 2 markerthe workflow loads and declares itervox_schema_version: 2fail — run itervox init --update --workflow WORKFLOW.md
.gitignore carve-outsthe root .gitignore does not ignore .itervox/, or keeps .itervox/agents/**, .itervox/handoff/**, .itervox/playbooks/** and .itervox/memory.json committablewarn — profile files, agent handoffs, playbooks or the shared memory are not committed; itervox init --update adds the lines
CLAUDE.md / AGENTS.mdeach backend in use has its repository instruction file (Claude Code reads CLAUDE.md, Codex AGENTS.md)warn — add the missing file at the repository root
Project verification skill or script (verify_skill)the repository has a .claude/skills/<…verif…>/SKILL.md, a scripts/verify* file, a verify Makefile target or a verify script in package.jsonwarn — “no project verification skill or script”; Generate verify skill (operator) creates an issue asking an agent to write one. See Verification readiness
Shared bare clone fetch refspecno shared bare clone, or its fetch works on the current refspecfail when the last fetch failed; warn on the old refspec (migrated automatically) or a stale fetch
Skill and subagent referencesevery skill and subagent a profile prompt names exists for its backendwarn — fix the name or add the skill / subagent
Tracker reachableone read of the Linear or GitHub API succeeds (local issues need none)fail when it fails or does not answer within 8 seconds; warn for a GitHub token without push access

The checks run when the section opens and on Re-run checks, on the request, never on the daemon’s dispatch loop: a slow or unreachable tracker delays only this section, by at most its timeout.

The latest result of each profile eval scenario, from GET /api/v1/repo/evals. itervox evals recorded (replays the recorded transcripts; free), ITERVOX_EVALS_LIVE=1 itervox evals live and ITERVOX_EVALS_LIVE=1 itervox evals record save each judged report to .itervox/evals/<ISO8601>.json in the project (gitignored; the newest 20 are kept), and this section reads the newest of them. It never starts an eval run: run them from the project root, where the daemon’s WORKFLOW.md is.

Per profile, each scenario shows pass or fail, every assertion that failed (a missing or forbidden action, a forbidden_phrases match, a missing marker phrase), a live run’s error, when and in which mode it ran, and the prompt version it ran against — the same digest the Prompt versions section shows. A live run computes it from the workflow it used; recorded mode reads it from the recording when a live run made it, and shows “unknown” for the hand-written fixtures. A case that fails on the current digest after passing on the previous one is marked regression, with both digests: the prompt change is the first suspect.

Files and directories ranked by how often agent work on them went wrong, from GET /api/v1/repo/hotspots. Over the run history (the newest 200 runs), a run counts when it failed or stalled, asked for input, was a reviewer that requested changes (verdict block), or opened a PR that was later reverted on the base branch. Each such run is attributed to the files its branch changed. Each row shows its count per signal, the issues involved (linked to their logs), and the runs (expand N runs), with their PR.

The daemon computes it in the background, never on its dispatch loop, and refreshes it at most every 10 minutes; the section shows “Computing…” or “Refreshing…” until a result lands. Git is read-only, through the same scrubbed git runner as the rest of the daemon, with no fetch, and each run’s files are cached by run ID, so a refresh only runs git for runs it has not attributed yet. Files and Directories switch the list; a directory counts a run once however many of its files it changed.

Some signals are approximations, and the notes under the list say when one applied:

  • Files a run changed. Run history records each run’s branch, PR and (for a reviewer) verdict from this release on. While the branch exists, the files are its diff against the merge base with the base branch — the branch’s state now, not only that run’s commits. Once it is merged or deleted, they are the files of the commit on the base branch whose message names the PR ((#N) or Merge pull request #N). A run with neither is counted as not attributed.
  • Repository. The shared bare clone (worktree mode with clone_url), the workspace root (worktree mode), or the project directory otherwise; refs are as fresh as the last fetch.
  • Change requests count only reviewer runs recorded with a verdict, so reviews from before this release are left out.
  • Reverted PRs are found in the newest 500 commits on the base branch, by Revert "… (#N)" subjects, GitHub revert-N-… branches, and the PR numbers of commits a revert names.