Inspect AI Agent Runs, Logs and Costs

Triage every task from one console.

One row per pull request with its status, agent, duration and review score. Thirty fix and review passes still take one row.

A task row: Ultrafix is running on storefront-web with Claude Opus 5.5, and the run track shows review scores of 5 and 7 — One row per pull request: Ultrafix running, and the run track holding earlier scores of 5 and 7.

Click a task and its details open beside the list: every run on a timeline, the score trend, findings and the execution log. Filter by assignee to see your own queue.

Every run of one pull request on one timeline: implement, review scored 5, fix, review scored 7, and the Ultrafix run in progress with its live steps — Every run of one pull request on one timeline, with scores and pushed commits.

The Tasks console with a task open beside the list: Show delivery windows at checkout in its second Ultrafix cycle, every run on one timeline with review scores of 5 and 7, the changed files and the live execution trace — One task open beside the list: its runs, changed files and live trace.

Every run carries its cost, and a cap.

The usage log ties every model call to its tokens, cost, duration and repository.

The LLM log: each model call with its model, tokens and cost, from a $1.31 Claude Opus 5.5 implementation to a three-cent repository summary — Every model call carries its model, tokens and cost.

Task header: repository and pull request, Implementing status, the run in progress, its model and runtime, and token consumption — The task header names the run in progress, its model and runtime, and its tokens.

Set a spend cap per task, per repository or for the whole instance. A run that reaches it stops, and the work so far still arrives as a pull request.

A task timeline: the run reaches its $5.00 spend cap from .propr/workflow.yml at an estimated $5.04, then ProPR opens the pull request with the partial work — The run hits the $5.00 cap from `.propr/workflow.yml`, and ProPR still opens the pull request with its partial work.

Open the dashboard and see what needs you.

The dashboard leads with Needs attention: failures, reviews and decisions waiting on a person.

Needs attention: a goal asking a question, a plan pull request waiting for review, and a fix run that failed, each with its next action — Needs attention: a goal question, a review request and a failed fix.

Happening now lists every running task and goal with its current step, elapsed time and last output.

Happening now: five agents at work across Northwind repositories, each row showing what the agent is doing, its step and elapsed time, with one more task queued — Happening now: each agent with its step, elapsed time and last output, and why queued work waits.

Completed groups finished runs by the work they belong to, and Historical stats counts tasks, success rate and spend over 7 or 30 days. Everything updates live.

Completed work: reviews with their scores and findings, applied fixes, and newly opened pull requests — Completed: each finished review, fix and implementation with its outcome and score.

Analytics reports runs per task, first-time pass rate, time to merge and autonomy, the share of tasks that finished without a person stepping in, over any window from 24 hours to all time.

Analytics delivery metrics: 2.1 runs per task, 64% first-time pass, 3h 16m mean time to merge, 88% autonomy — The Analytics delivery band.

Agent efficacy by model compares initial and final review scores, runs to merge and merge rate, so you see which model earns each phase.

Agent efficacy by model: evaluated PRs, initial and final review score, score delta, runs to merge and merge rate for Claude Opus 5.5, GPT-6 Astra and Claude Fable 5.1 — Agent efficacy by model: initial and final scores, score delta, runs to merge and merge rate.

Recent role changes on the Access page: who added a member, and who promoted a member to admin, and when

Settings and access

Choose who can trigger tasks, which labels and comments count, which agents run by default and how much automation you allow. Admin and member roles map to GitHub accounts, with an audit trail of every change.

Tasks on a phone: each pull request is a card with its status, run track and review scores, agent and repository

CLI and mobile access

Run plans, tasks, goals and automations from the command line, terminal to pull request. The same task history and actions work on a phone.

Agent watchdog settings: a 10-minute stall timeout and a 30-minute tool stall timeout, each an adjustable default

A watchdog for stuck agents

ProPR stops an agent that goes silent and still publishes the work so far. The timeline names the rule that fired.

A task timeline: the run fails on a provider error and ProPR starts Replacement Attempt 2 right after

Replacement runs

A provider hiccup never loses a task. ProPR starts a linked replacement run with the same agent and model.

Auto-Resolve Merge Conflicts turned on as the default for repositories, which can set Always or Never themselves

ProPR resolves base-branch conflicts

When the base branch moves, ProPR merges it in, resolves the conflict and pushes a real merge commit. Turn it on per instance or per repository.

A task timeline ending in Task Cancelled: cancelled because the issue was closed

Withdrawn work stops

Close the issue or the pull request and ProPR cancels every queued and running job behind it.

Keep CI from fighting the agent.

Two per-repository settings keep CI out of the agent's way.

Cancel CI while a follow-up implements, for storefront-web: CI and End-to-end tests selected, Preview deploy left running

Cancel superseded CI runs

Name the workflows and ProPR cancels their runs for a PR the moment a follow-up starts, so CI minutes go to the commit that matters.

Checks that never block automation

Name checks by pattern, such as lint-*, and their failures stop holding up /merge, auto-merge and Ultrafix. GitHub still shows every result.

Carry task control into chat.

Plan work, steer goals, inspect tasks and diffs, and review pull requests from any chat client through MCP. Your account and repository permissions bound every control, and every result links back to its record on your instance.

MCP server settings: enabled, the connection URL to paste into a chat client, the scopes an admin allows, and a link to connected apps — The scope ceiling: no chat client gets more than an admin allows.