Insights
Insights is a project page, in the project’s navigation on desktop and on a phone. It shows how work moves from idea to merged pull request, over the last 7, 30 or 90 days (the range control at the top right; the range stays in the page’s address, so a link keeps it).
It has four tabs, all described below: Flow, Agent setups, Accounts and People.
Every number is worked out from what SAM already records about tasks, ideas, pull requests and their checks. Nothing is estimated: when SAM has no record for something, the page says so.
The four figures
Section titled “The four figures”- Idea to merged, median: for ideas whose first pull request merged in the range, the time from the idea being filed to that merge. Ideas are matched to pull requests the same way the Ideas page matches them.
- CI passes first try: of the pull requests opened in the range, the share whose first commit passed every check with no failure and no rerun. A pull request counts as failed as soon as any check on that commit fails, and as passed once its checks have finished or it has moved on (merged, closed or pushed again). Until then it is left out, and the figure says No result yet; that includes a repository with no CI, where no check ever reports. Pull requests SAM did not see from their first commit are left out, and when there are none the figure reads Not recorded.
- Waiting on a person, median: how long questions agents filed for people (asks in the Inbox) waited for an answer. A question still open counts up to now.
- Rework within 7 days: the share of merged pull requests that were followed, within 7 days, by another pull request on the same branch or for the same idea. Administrators can change the window.
Each figure shows its change against the range before it, for example ▼ 0.4 days vs previous 30 days. Green means better.
Runs per week by outcome
Section titled “Runs per week by outcome”A run is a task an agent worked on to completion, not a conversation. Each bar is a week (a day on the 7-day range), counting runs by when they ended, split into runs that merged, ended with their pull request closed without merge, or failed. A line underneath counts the rest: runs that were stopped, runs whose pull request is still open, and runs that finished without a pull request.
Why runs stopped before finishing
Section titled “Why runs stopped before finishing”Runs in the range that ended without a merge, and runs still going that are waiting on a person, grouped by why: waiting on a person, a pull request waiting to merge, CI failing, a usage limit, no machine available, workspace setup failing, running past the time limit, the agent stopping, a session expiring while asleep, a pull request closed, a person or coordinator stopping it, or another error. The reason is worked out from the task’s own record each time the page loads.
Flaky checks
Section titled “Flaky checks”A check is flaky when it failed and then passed on the same commit: nothing in the code changed, so the failure was not caused by it. The list shows the checks that did this most often, and on how many of the commits they ran on. SAM keeps check results for each pull request’s current commit and the commit it opened with, so a flip on a commit that was later pushed over is not counted.
Ideas waiting longest
Section titled “Ideas waiting longest”Open ideas a person filed in the range, oldest first (ideas an automation or incident filed are left out, as on the Ideas page), with how many runs worked on each and where it stands (not started, pull request open with CI red, stuck, and so on). Select one to open it.
Agent setups
Section titled “Agent setups”Which setup does this kind of work best on this repository. Choose what a setup is with Group by: the agent Profile a run used, the Skill it ran, or the Model (the model the profile or skill asked for; runs that asked for none appear as the agent’s default model). The choice stays in the page’s address. The runs are the same ones Flow counts.
Each setup
Section titled “Each setup”- Runs, and how many merged a pull request.
- Merged without rework: of the runs that finished, the share whose pull request merged and was not followed, within the rework window, by another pull request on the same branch. A run that failed or was stopped counts against its setup. Flow’s rework figure also counts a follow-up linked to the same idea; this tab counts the same-branch kind only, so the two can differ.
- Median to pull request: from the run being asked for (time waiting for a machine included) to the first pull request it opened, over the runs that opened one.
- Nudges per run: messages a person sent into the run’s chat while it was going — a typed
follow-up or a comment addressed to the agent. The task’s own first prompt, messages from other
agents and SAM’s own prompts are not nudges. Not recorded when the chats can’t be read. An
installation that switched durable prompt delivery off (
DURABLE_PROMPT_DELIVERY_ENABLED=false; it is on by default) does not mark follow-ups, so they are not counted there. - Tokens per merged PR: the tokens (input, output and cache) that the agent or SAM’s AI proxy reported for the setup’s runs, divided by how many of those runs merged. A run continued after waking from sleep shares its chat, and its tokens are counted once. Runs whose agent reports no tokens are left out and the figure says how many did report; with none it reads Not reported. SAM never estimates tokens.
- Memory p90 / requested: the 90th percentile of each run’s peak memory against the memory the runs asked for, with any out-of-memory kills. Not recorded when the runs have no resource telemetry.
A figure that covers only some of a setup’s runs says partial, and a setup with fewer finished runs than the suggestion minimum says few runs. Another member’s personal profile or skill is never named; it appears as A teammate’s profile (or skill). Compared puts the five busiest setups side by side on merges, time to pull request and nudges.
Suggestions
Section titled “Suggestions”When the numbers support it, the tab suggests a change to one of this project’s skills, applied with one tap through the skill’s own settings (you need permission to edit the project):
- A default profile for a skill: offered when, among the skill’s finished runs, a project profile merged without rework clearly more often (by 20 points by default) than the skill’s current default profile, or than the next-best profile when it has none. Each profile compared needs at least 10 finished runs of that skill; with fewer, one run moves the rate by ten points and the difference is noise.
- A memory request for a skill: offered when at least 10 of the skill’s runs recorded their peak memory and the 90th percentile, plus 25% headroom, differs from what the skill asks for by a gigabyte or more. It never suggests asking for less after an out-of-memory kill.
Once applied, the suggestion goes away because the skill no longer matches it.
Accounts
Section titled “Accounts”Which account each run used, and how often an account’s usage limit paused the work. An account is a stored credential: a Claude or ChatGPT subscription login, an API key, or SAM’s own platform credential.
Whose accounts you see
Section titled “Whose accounts you see”Your own accounts are listed one by one, by the name you gave them in Settings. Other members’ accounts are never named: they appear grouped by kind, for example Teammates’ accounts (2) for two other people’s Codex subscriptions, the same way the chat header shows only the kind of account another person’s run used. Shared project credentials appear as Project account. Runs whose agent session recorded no credential appear as Account not recorded.
Each account
Section titled “Each account”- Runs and Merged: the same runs Flow counts, attributed to the account the run’s agent used, and how many of them merged a pull request.
- Limit pauses: how many times the account stalled on a usage limit in the range, and for how long in total (Paused now while one is still going).
- Limits now (your own accounts): the provider’s last-reported windows, the same chip as the chat header. Open it for every window and when it resets.
- Tokens (every account): the tokens the account’s runs used in the range (input, output, cache reads and cache writes), and that total divided by the account’s merged pull requests. A run through SAM’s AI proxy is counted once, from the proxy’s own measurement, even though the agent reports the same calls. Agents that report no token counts contribute nothing, so a total can be low rather than wrong.
- Spend (API keys only): an estimate at SAM’s model catalog prices, never a bill, with the estimate per merged pull request beside it. Subscriptions and SAM’s platform credential show tokens only. A model SAM has no price for, or a custom OpenAI-compatible endpoint, is named under the figure and makes the total partial (At least); when none of the key’s usage has a price, the figure reads No estimate. SAM never guesses a price. A key with no measured usage in the range shows Not reported. A teammate’s personal key shows its tokens but its spend is Private: only the key’s owner sees it, as with its name. Shared project keys show spend to every member. Each run’s own tokens per model are on its chat’s Receipt.
Tokens and spend cover usage SAM recorded since the range began, while Merged counts runs that ended in the range, so a run that spans the range’s start counts its later usage only.
The ring around each account’s icon is its colour in the chart below.
Runs per day by account
Section titled “Runs per day by account”The runs of the range, by the day (7 days) or week (30 and 90 days) they ended, or started if still going, stacked by account. The five busiest accounts each get a colour; the rest, and runs with no recorded account, are grey (Other).
Stalled on a limit
Section titled “Stalled on a limit”Each time an account ran out: when it started, which window ran out when SAM knows (5h, Week), how many runs it held, how long it lasted and how it ended — resumed (the limit reset, or the owner started the run anyway), the run ended while it waited (it was stopped or failed before the limit reset), still waiting, or stopped retrying, needs a person (SAM gave up waiting and the run needs someone to look at it). Runs paused by one exhausted window count as one stall, not one per run. The list shows the most recent ones.
A stall is worked out each time the page loads from two records: the latest usage-limit wait of each task (a run paused mid-way, or held before it started), and the windows the provider reported as exhausted. SAM keeps only a task’s latest wait and a window’s latest report, so an older stall that a later one replaced is not shown. A window reported without a reset time counts for at most its own length. Another member’s personal accounts contribute their runs’ waits, but not the provider windows they reported, as on the project’s other limit views. If a section can’t be read, the tab says so rather than showing it empty.
People
Section titled “People”Where work is waiting on people right now, and how quickly people pick it up. It never counts how much anyone did: there is no per-person tally of answers, reviews or merges anywhere on the tab.
The four figures
Section titled “The four figures”- Waiting on people now: open questions, review asks and human tasks agents filed in the Inbox (not GitHub pull request reviews; those are under Review pickup). System error items are left out.
- Longest wait: how long the oldest of them has been open.
- Review pickup, median: on the project’s GitHub pull requests, the time from a review being requested to the first review. Per project, never per reviewer.
- Answered by the first person asked: of the questions, review asks and tasks first given to one person and closed in the range, the share that person answered without handing it on.
Waiting on
Section titled “Waiting on”Open items grouped by who they wait on, the longest wait first: each person, then Not taken yet for team tasks nobody has taken. A team task one member took waits on that member. Someone who has left the project shows as A former member. Each group says how many are open, how old the oldest is and how many have an agent waiting on them, and lists the oldest ten (an administrator can change how many).
You see an item only if you could open it in your own Inbox: your own items, team tasks, and tasks your runs filed. Any other item shows only how long it has waited, as An item you cannot open; its title, kind and flags are not sent to your browser. The group’s counts still include it.
Reassign appears on a human task you are allowed to hand on: one assigned to you, one your run filed, or a team task nobody has taken or that you took, if your project role can change tasks. Choose a member or Back to the team (whoever has the task now is not offered), then Hand it on. It is the same action as handing a task on from the Inbox, with the same checks: the new person must be an active member of the project. Questions cannot be reassigned.
Review pickup
Section titled “Review pickup”The median time from a review request to the first review, the time 9 in 10 requests were picked up within, and how many requests made in the range are still waiting. SAM records review requests from the GitHub App’s webhooks, so pickup starts from when this was first recorded for the project (the panel says since when); earlier requests read Not recorded, never estimated. Until a request in the range has been picked up, the panel says so instead of showing a time.
- A request to a person ends at that person’s first review (a comment review counts).
- A request to a team ends at the first review by anyone except the pull request’s author.
- A request that is withdrawn, or still open when the pull request is merged or closed, is not a pickup and is left out. Requests to and reviews by bots are ignored.
- Asking the same person again after they reviewed starts a new request.
Answered by the first person asked
Section titled “Answered by the first person asked”Counts questions, review asks and human tasks that were answered or marked done in the range. An item first given to the team is left out, since no single person was asked. An item counts as answered by the first person when it was never handed on and the person it was given to answered it. Handing a task on, including the automatic hand-back to the team when its assignee leaves the project, means it was not.
On the Overview
Section titled “On the Overview”The Project Overview shows a small This week pane for the last 7 days: pull requests Merged, Idea → merged and CI first try, and the top blocker, with a link here. The top blocker is the most common reason runs stopped (other than a person or coordinator stopping them on purpose) or the flakiest check, whichever affected more work.
For administrators
Section titled “For administrators”How many ideas “idea to merged” looks at, how many flaky checks and waiting ideas are listed, the
rework window, how many limit waits the Accounts tab examines, how many stalls it lists and how
many token usage rows it reads, how many runs Agent setups examines and when it makes a suggestion,
and how many items People lists per person and how many review requests it examines are runtime
settings under Admin → Settings → Operations (insights.flow*, insights.accounts*,
insights.setups*, insights.people*).
See Admin settings.