Stats
Stats summarizes your archive for one time range: how much you worked with agents, which harnesses and models you used, what it would cost at API prices, and where the work landed. To see what changed between two periods, open the Changes page, described in Compare periods.
- App
- Stats
- CLI
hansard stats
Use Stats in the app
Section titled “Use Stats in the app”

- Select Stats in the dock. The sidebar lists the Stats pages, and the subtitle under the page title shows the current scope. The row of controls under the title applies to every chart on the page.
- In that row, choose a time range: 1M, 3M, 6M, 1Y, All, or Custom. Custom adds From and To date fields.
- Under Show by, choose Provider, Lab, Model, Project, or Device to set how charts group the data.
- Under Include, choose which kinds of sessions count: Agent, Chat, Subagents, and SDK jobs. Chat covers imported ChatGPT, Claude.ai, and Gemini conversations. The first three are on by default. SDK jobs is off.
- Under Filter, choose Provider, Lab, Device, or Person to list the groups in your data. Clear a group to leave it out of every chart. Reset includes all groups again.
- Choose a page in the sidebar list. Pages sit in groups: You, Activity, Agents and models, Work, Spend, Telemetry, and Archive. When the sidebar is folded, or the window is too narrow to show it, choose the page from the menu in the page title instead.
- To see what a card counts and what it leaves out, select the i button next to its title.
Other controls sit beside the titles of the cards and sections they change. A control that several cards follow appears on each of them, and they always show the same choice:
| Control | Options | Where | Effect |
|---|---|---|---|
| Token direction | In, Out, Cache | Usage & cost: Token usage and Provider mix | Chooses which token directions count in token charts. |
| Activity metric | Messages, Conversations | Overview: Activity and Activity by provider | Messages counts prompts a person typed each day. Conversations counts each included conversation once. |
| Chart scale | Totals, 100% | The daily and monthly stacked charts | Totals shows actual values. 100% scales each stack to compare its composition; hover still shows actual values. |
| Runtime evidence | Telemetry, Standardized | Speed: How did each agent run? | Telemetry adds the approvals, failures, and retries agents reported to this computer. Standardized counts only what transcripts record, the same way for every agent and device. Appears when this computer has telemetry counts. |
The Where does activity happen? card on Overview shows one-shot SDK jobs when Include has SDK jobs selected, the same setting the headline totals follow.
Stats opens with the most recently calculated numbers and refreshes them in the background, so values can change a moment after the page loads. For numbers scoped to one project, see Projects and coordinators.
Compare projects and agents
Section titled “Compare projects and agents”Under Projects → Compare projects, choose up to four projects. Without a saved choice, the three projects with the most activity appear. The cards compare messages, active days, spend, model and provider shares, code changes, and devices over the selected range. Judged outcomes appear once a project has five judged conversations. The ranking chart shows how project activity changes over time.
The Agents page compares coding agents, chatbots, personal agents, and background agents. Its Agent to agent section shows recorded delegation, messages, board posts, questions, and agent trees. Missing records remain unknown. These classifications refresh from archived facts without reimporting.
The Team picker in the header scopes Stats to one Team. Overview then shows Team members, with each member’s active days, messages, and spend.
Under Speed → Agent runtime, How did each agent run? compares recorded approvals, waits, failures, and retries. Where did each agent’s requests go? uses the last 14 days of activity on this computer. The Runtime evidence choice beside the card’s title selects Telemetry or Standardized counts. An agent that records no wait time does not imply a zero wait.
Look back on a year
Section titled “Look back on a year”Year in review turns one year of your archive into a deck of full-screen cards to step through and share.
The review includes your conversations with coding agents and web assistants. It excludes subagents and SDK jobs. Its year and source choices work independently of the Stats filters. Message totals follow conversation start dates. Active days and streaks count days when you started conversations.
- In Stats, select Year in review at the top of the page.
- Choose All assistants or one source, and choose a year.
- Move between cards with the Left and Right arrow keys, select the left or right side of a card, or swipe on a trackpad.
- Select Share to send the current card as an image, or Save to Downloads to save it. Where the system has no share sheet, Share reads Copy image. On the summary card, Copy as text copies the summary.
Depending on how much the year holds, the cards cover when you started, what the year came to at API prices, the days you showed up, your busiest month, your longest streak and longest break, your busiest day, the assistant you used most, the model that did most of your work, the models that joined your work that year, your top projects, the hours you work, and a summary. The cover draws every day of the year, and each card carries your weekly rhythm across its top. A year with fewer than seven active days shows only the opening, totals, days, and summary cards. A year still under way opens with “So far”.
Use Stats from the CLI
Section titled “Use Stats from the CLI”hansard stats prints an overview, harness and model tables, classification
results when available, and tool, command, skill, MCP server, slash-command,
and file-reference rankings:
hansard statshansard stats --provider codex-cli --since 30dhansard stats --repo github.com/example/app --top 25To print only the tool, skill, MCP server, and command sections:
hansard stats --toolsTo get the complete result for a script or an agent:
hansard stats --json --top 100To scope Stats to one Team you joined, as the team picker in the Stats header does, or to read totals over a range of days without folding every conversation:
hansard stats teamshansard stats --team <team-id>hansard stats totals --start 2026-10-01 --end 2026-10-24stats totals reads the day rollups the watcher keeps, and exits with status 1
while they are still being built.
| Option | Effect |
|---|---|
--provider <provider-or-alias> |
Limits Stats to one harness or source. Accepts the same names as hansard history. |
--since 30d|90d|all |
Limits Stats to a date window. Without it, Stats covers the whole archive. |
--repo <repo-key> |
Limits Stats to one project, such as github.com/example/app. |
--top N |
Sets rows per table (default 12). With --json, also caps the tool, skill, MCP server, and command lists. |
--include-sdk-jobs |
Adds one-shot SDK jobs to token, spend, and tool totals. Conversation and activity measures still leave them out. |
--tools |
Prints only the tool, skill, MCP server, and command sections. |
--json |
Prints the full Stats data, including model identities and pricing versions. |
Read the headline cards
Section titled “Read the headline cards”The twenty cards at the top of Overview summarize the selected time range under the current Include and Filter choices, the same scope the subtitle under the page title names. Two lead: Conversations and Estimated spend, or Tokens when nothing could be priced, each with how it changed against as many days just before the range; hover the change for the earlier figure. Active days, Session time, Busiest hour, and Favorite provider follow in smaller cards, and the rest open under Show 14 more numbers. Changes, Tokens & spend, Tools & harnesses, and Projects each open with their own two headline figures. With All selected, the totals and favorites also count activity without a known date, and charts over the whole archive are captioned All time. Shorter ranges and calendar cards such as Active days use dated activity only.
| Card | What it shows |
|---|---|
| Conversations | Included conversations, split into agent, chat, and subagent, plus the one-shot SDK jobs kept separate or included. |
| Messages | Messages of every role, with the share a person typed. |
| User messages | Prompts a person typed, with the average per active day. |
| Avg conversation length | Messages per conversation, with the prompts typed per agent or chat conversation. |
| Tokens | Tokens used as each provider reports them, with the output share named; the rest is input and cache. When some counts are estimates, the card says how many, and it reads All tokens estimated when every count is. |
| Estimated spend | API-price estimate for the tokens used, with the share of tokens Hansard could price. The i button explains the estimate and links to the pricing details. |
| Spend per message | Estimated spend divided by the prompts a person typed, so each prompt carries the agent work it set off. |
| Heaviest session | The conversation with the most tokens that started in the range. Its title opens it. |
| Tool calls | Tool calls, with edit calls. |
| Tool calls per turn | Tool calls per typed prompt in a typical agent conversation (a geometric mean over conversations, so it sits below all tool calls divided by all prompts), with the share of agent-driven conversations, those above 16 calls per turn. The i button explains the rate. |
| Words per turn | Words of agent reply per typed prompt in a typical agent conversation, the same geometric mean. |
| Session time | Total session length and the average per conversation, counting only sessions with reliable start and end times. Conversations that ran at the same time each add their own time, so a range can hold more session time than it has hours; from a day up the total reads in hours. |
| Active days | Days with activity, with the longest and current run of consecutive days in the range. |
| Most active month and Most active day | The month and the local calendar day with the most typed prompts. |
| Busiest hour | The local hour with the most typed prompts, and its share. While more than half of the range’s prompts have no recorded time, it names no hour and says how many wait for hansard rebuild --since all. |
| Favorite provider and Favorite project | The provider and the project that received the most typed prompts, and their share. Projects carry their GitHub or folder mark. |
| Favorite model | The model with the most tokens in the range, and its share of them. |
| Projects | Projects with activity in the range, split into GitHub repositories and other folders. |
Estimated spend is what the recorded tokens would cost at published API prices. It is not your subscription bill. To track plan limits, see Accounts and plan usage.
What each page shows
Section titled “What each page shows”| Page | Sections |
|---|---|
| You | Last week against the week before, then Waiting: how long agents waited for your approvals; How the work goes: which kinds of work succeed and which agent work a later change redid; and What it costs: the estimated cost of a finished task and how often limits stopped work. See See how your work with agents went. |
| Overview | Activity calendars (combined, plus separate agent, subagent, chat, and SDK batch job calendars), Conversation shape, Activity by provider (or the grouping chosen under Show by): a leader timeline of which group received the most of your messages each week or month, and your messages or conversations per day by group, and Session rhythm: the local hour of the day and the hour of the week when you send messages, when your days usually start and end, and session time per week or month. |
| Changes | The selected range against the equally long period before it: totals, a running total by day, which groups gained ground, how your hours shifted, tool, command, and skill shares, what was new or went quiet, working rates, and work modes and outcomes. See Compare periods. |
| Projects | Where your attention went: when each project was active, which project received the most human messages each week or month, the projects with the most human messages, and when projects had their first archived session; What each project costs: which projects cost the most, which cost the most each week or month, and which cost the most per human message against the median; What each project produced: file edits, commits, and judged outcomes; and, with search by meaning on, Subjects: what your conversations in the range were about, ranked by Conversations, Session time, or Tokens. |
| Agents | Comparisons of coding agents, chatbots, personal agents, and background agents, followed by Agent to agent exchanges and delegation trees, and Subagent delegation fan-out and growth. |
| Models | Your top models ranked by conversations, tokens, or spend. Model settings shows recorded reasoning effort and speed. One lab selection scopes Model settings and Model family insights. Model frontier compares quality with median cost, active time, and output tokens per conversation. See Compare model quality and cost. |
| Tools & harnesses | Harnesses: Which harnesses do you use? over time, Which harness did most of your work each month?, How does each harness work?, a sortable comparison, Which harnesses cost the most?, and What do their tool calls do? by kind; Harness versions: How did versions roll out? and Did outcomes change with versions?, with every version’s figures under Show version details; and Toolbelt: Which tools do they reach for?, the most used tools, shell commands, skills, MCP servers, slash commands, and @ references. |
| Work | Activity & autonomy: Which models work most on their own? draws a bubble for each model with at least five rated conversations, placed by its tool calls per human turn against its output words per human turn, sized by its rated conversations, and marked with its brand, across the User-steered, Balanced, and Agent-driven bands. Sorting the rates table under the chart by another per-model rate moves that rate onto the vertical axis, and models with fewer than 10 rated conversations are faded. Then Kinds of work: work modes, the leading kind of work each month, and what the work involves. |
| Outcomes | How the work turns out: how labeled work ended, which kinds of work go well, judged success over time, the leading model each month, model success against spend, spend by kind of work, and the evidence behind the judgments. |
| Usage & cost | Token usage by the Show by grouping, token mix, and the leading group over time. Provider mix (or the grouping chosen under Show by): how the mix shifted between the halves of the range (groups with at least 1% of either half’s tokens), how each group trended week by week (groups with at least 1% of the range’s tokens), with the rest under Show, and tokens per day by group. Estimated spend compares model costs by day, month, and prompt. Output token work shows what output tokens went to. See Compare projects and agents. |
| Limits & accounts | Subscription value estimates what Claude and Codex plans are worth at API prices. See Estimate what a plan is worth. |
| Optimize | Fixes ranked by how much of your usage each one covers, each with its evidence and one button. See Find what to change. |
| Speed | Agent runtime compares approvals, waits, failures, retries, and recorded requests by agent, with provider credits in their own units. |
| Context & cache | Prompt cache: why the prompt cache was rewritten. See Find why the prompt cache was rewritten. |
| Reliability | API errors over time, per 1,000 requests, by kind, by hour or weekday, and how much Claude Code work ran into them; Tool reliability by harness, by project, by tool, and over time; Codex guardrails; and Cloud agent outcomes. |
| Memory | Memory footprint (archive growth, memory by provider, and project against global memory), Memory activity (reads, writes, and loads per month in one chart, idle months included), Memory age & staleness, and Memory churn for the memory archive. |
| All charts | Every page’s sections in one scroll, in sidebar order, each page under its own heading. |
What time of day do you work? and When in the week do you work? count
each typed prompt at the local hour it was sent, on this device’s clock, and
follow the activity metric. Hover a weekday’s share in the week chart for its
average per day. When does your day start and end? counts the days whose
first and last message fell in each hour, with each day running from the
quietest hour of the range so a late night stays with the day it continues; its
takeaway names the most common start and end hour, the two bars drawn darker.
Conversations archived before prompt times were recorded spread their messages
evenly over the hours between their start and end, or keep them at the start
hour when a conversation ran longer than twelve hours. In the time-of-day bars
those messages draw in a lighter tint and the note under the chart says how
many; the day chart draws lighter the days whose first or last hour holds only
such messages. The week chart’s color carries each cell’s count, so no cell is
partly tinted: while those messages are more than half of the range every cell
draws lighter, and otherwise a dot marks each cell holding mostly such messages.
While they are more than half of the range, no chart names a busiest hour, slot,
or usual start and end, because one long conversation would decide it.
hansard rebuild --since all records the time of each prompt without a
reimport. The count of days in When does your day start and end? follows its
own day boundary, so it can differ from the Active days tile, which counts
calendar dates. How much session time each week? reads week by week, with
weeks starting on Monday, on ranges up to about six months, and month by month
on longer ones.
With Show by set to Project, the Overview leader timeline groups home, temporary, scratch, and harness catch-all folders into one entry that never leads, as on the Projects page, and names the chat messages that belong to no project in its takeaway and in each week’s tooltip. On Usage & cost, the mix shift compares the two halves of the days with token activity in the range, and its takeaway and tooltips name each half’s dates, so Show by Provider and Show by Project can compare different halves when chats carry tokens before any project does. The takeaway names the group whose share moved most, including groups folded under Show, and says when one half holds a sliver of the other half’s tokens.
The page list stays in the sidebar while a page scrolls, and choosing another page opens it at its top.
What does one prompt cost? divides estimated spend by the prompts you typed, for the five largest groups under Show by, week by week for ranges up to about six months and month by month for longer ones. Subagent and delegated SDK work adds spend without adding prompts, so a prompt carries the cost of the work it set off. A point needs at least 5 prompts and 80% priced tokens, and the scale is logarithmic. Each line ends in its color, mark, and average over the range, a group with a single point shows its value beside its dot, and the time axis starts at the first week or month with enough priced prompts. When most groups have fewer than three such weeks or months, the chart ranks their averages over the range instead of drawing lines.
Under Model family insights, How do releases compare? lists each dated model in release order with its value; dot size shows sessions. A lab with no models for the chosen metric, such as one whose models fold reasoning into output, opens on the first metric it has data for, and the subtitle says so. On phone widths the section’s charts name models without the family word their mark already shows, such as Opus 5.5 for Claude Opus 5.5. Which model did most of your work each month? gives each model that produced the most output tokens in at least one month its own row, fills the months it led, and traces the months it was in use without leading; bars above show each month’s output on a log scale, and months without output keep their place. Models in use that never led open under Show, shaded by their share of each month. When did you pick up each model? marks each model’s official release date with a diamond and its first and last archived use with a thin line, with a bar for the sessions started each week or month on one square-root scale and the model’s session count at the right. A model without a published release date has no diamond and starts at its first use, and models used in fewer than 10 sessions open under Show. The section’s charts share one name column, and the two timelines lay out the same months, so a month sits at the same place in both. Both timelines date use by the day its conversation started, so a conversation that later switched to a newer model can date that model’s use before its release; that use is left out of the monthly leaders and drawn from the release date in When did you pick up each model? The i button beside each title explains the details.
On Tools & harnesses, the Harness, Model, Work mode, and Session kind filters scope every harness chart, and Reset clears them. The harness chart stacks human turns, or conversations under Activity metric, per day, week, or month depending on the range, and follows Chart scale. The period still in progress keeps each harness’s color under a diagonal hatch and is marked so far; the key beside the chart gives each harness’s share of the whole range, and harnesses under 2% of it and of every period share one Other band that opens in place. A harness too small in a period to draw at least four pixels there adds its height to that period’s largest harness, and the tooltip lists every harness. The leader chart marks the harness with the most output tokens each week, or each month for ranges past about seven months; when a range opens on periods with little or no recorded output, such as chat exports without token counts, its takeaway says from when output was recorded. The comparison, cost, and tool-kind charts list every harness up to twelve; past that, harnesses under a floor the chart names (5 sessions and runs, or 50 tool calls) open below. Which harnesses cost the most? ranks estimated spend, spend per session, and spend per human turn; both divide all included spend, subagent work included, by the sessions a person started and the turns typed in that harness. In the tool-kind bars, a kind too small to draw on its own joins the grey segment with Other tools for that row; hover the harness, or open Show calls by kind for each harness, for every kind. The comparison ranks harnesses by the sessions you started there, the base its per-session rates divide by, and names each harness’s SDK and subagent runs under it. Rates use human sessions only, tokens per session divide every token by the sessions a person started, and a success rate needs 10 judged outcomes; Success takes a column only when at least two harnesses have a rate, and otherwise each row’s tooltip gives its outcomes. Did outcomes change with versions? plots only versions with 10 judged outcomes, oldest first, each with its likely range (a 95% interval); overlapping ranges mean the versions may not really differ. Sessions that span several versions or models stay in their own group, and missing versions stay unknown. The comparisons are observational: a newer version may also have received harder tasks. Which tools do they reach for? covers all history and ignores the range and the harness filters, while Include and Filter still apply. Each list opens on its eight most used entries as bars and opens the rest in place by name and count, 25 at a time, up to the 100 most used; a longer list says where it stops. Tools that harnesses name differently, such as Web Search and WebSearch or Bash and Bash Tool, count as one entry here and in the Changes page’s tool shares. In a bar, a group too thin to draw on its own keeps its color when it is the only one, and several share one hatched segment that the key names. A list whose first entry dwarfs the others, such as slash commands, draws its bars on a square-root scale and says so beside its unit. Tool usage is recorded per provider, so when the lists are split by provider, a provider such as Codex covers every harness it runs. Shell commands count each command in a shell call, so one call can run several.
On Projects, a project counts as first seen in the period that holds its
first archived session, so work from before that session, or from a stretch
the archive holds no sessions for, makes a project appear later than it began;
the chart names such a stretch below it. Temporary, scratch, and harness
catch-all folders, the home folder, and the Documents, Downloads, and Desktop
folders are left out of that count, never lead the weekly or monthly leader
charts, and share one Catch-all and scratch folders row in the ranked
charts; hover that row to see what it holds. Ranked charts lead with the
projects that hold at least 1% of the measured total, such as human messages,
spend, or file edits, and open the rest in place under a toggle that names the
rule. Spend per human message also needs 10 human messages for a steady
average, and outcomes list every project with at least five judged sessions.
In the human-message chart, groups with little share of the bars shown fold
into Other; hover a bar to see every group. File edits count each file an edit tool changed once per session, so a
file edited in three sessions counts three times. Commits seen are read from
git commit confirmations in archived shell output, so a quiet commit
(git commit -q) is not counted. Line counts cover sessions archived after
they were recorded; hansard rebuild --since all fills in older ones without a
reimport.
Subjects groups conversations whose text is closest in meaning, using the
local embedding model, and names each group from the titles of the
conversations nearest its center. A conversation counts on the day it started.
Conversations without an embedding, such as SDK runs, are left out. The section
appears once search by meaning is on and the
first subjects are ready; hansard embeddings subjects lists them from the CLI.
Reliability follows the selected time range. API error rates count only
Claude Code and Claude SDK sessions, because only their transcripts record
provider API errors, and divide by the API requests those sessions made. Each
error is placed on the day and local hour its session started, not the moment
it happened, so the hour and weekday chart shows when error-prone sessions
start and draws an hour pale when its errors came mostly from one or two days.
When neither the hour nor the weekday view has a column spread across several
days, the chart says which days the errors came from instead. How much
Claude Code work ran into API errors? splits the estimated spend, tokens, or
session time of those sessions by whether each hit an API error and, when it
did, by the kind behind most of its errors; it shows the work that ran into
errors, not what they cost, since a failed request is generally not billed,
and Stats does not measure the time or tokens spent on retries. The error rate
chart compares the first half of its periods that have 100 or more requests
with the second half, naming the days each half covers, and it switches from
days to weeks when too few days reach that floor to draw a line. Thinner
periods draw as rings, one above the scale is left out of the plot, and a
period that meets the floor above the scale draws as a caret; hovering any
period gives its rate. Tool error rankings rank every harness, project, and
tool with 50 or more calls in the range by its rate, drawn as dots on one
linear scale from zero. The project table leads with the
projects that made at least 1% of the range’s tool calls, and the tool table,
which covers every tool that failed at least once in the range, leads with the
tools behind at least 1% of its failed calls. Show more opens the rest,
with rows under 50 calls last as rings; every group keeps the table’s scale,
and a rate past its end draws an arrow at the end of its row. The catch-all
row for home, scratch, and uncategorized folders closes whichever group it
lands in, and the project table leads the section under Show by Project. Calls from a
format that records no tool results, or that has not flagged a single failure
in 200 or more results, stay out of the rates; a harness with only such calls
is named under the harness table instead of ranked. The tool trend panels start where the busiest harnesses report calls,
harnesses whose calls come mostly before that start get a second row on a
longer axis, and a period with fewer requests or calls than the floor is drawn
as a ring instead of setting the trend. Delegation and Codex guardrails count
only the sessions you started, in the Stats page and in hansard stats --json:
subagent threads are left out, and delegation is grouped by the delegating
session’s harness.
Choose the right breakdown
Section titled “Choose the right breakdown”Pick the Show by option that matches your question:
| Show by | Use it to answer | Example |
|---|---|---|
| Provider | Which agent app, CLI, or chat product handled the work? | Compare Codex with Claude Code. |
| Lab | Which organization makes the harness or model? | Compare activity across model providers. |
| Model | Which recorded model produced the responses? | Compare GPT, Claude, and Gemini families. |
| Project | Where did the work happen? | Compare repositories, or find uncategorized work. |
| Device | Which Hansard installation recorded it? | Compare a laptop with a workstation. |
Start with Provider when you compare tools. Switch to Lab when the organization is the subject, and to Model when the recorded model matters. Web chats count in overall totals. The overall Stats project breakdown covers agent folders. In Projects, web chats appear under their account or organization portfolio, with a Not in a project folder and named provider-project folders. Selecting a web folder or account portfolio shows its conversation and usage statistics.
Every provider and harness keeps one color on every page, whatever else a chart shows or a filter removes: harnesses from one company, such as Claude Code, Claude.ai, and Claude SDK jobs, take clearly different shades of that company’s color. Outcomes, token kinds, tool kinds, and work modes each use one shared set of colors across pages, apart from the provider colors.
Model colors follow each company’s palette: OpenAI families use distinct blues, and Claude families use terracotta, orange, and amber. Within a family, newer versions are darker in light mode and brighter in dark mode. Model badges use the same colors regardless of filtering or order. On every Stats page, each provider, lab, model, project, and device keeps one color for the selected range: series take colors in order of their tokens in the range, a model whose color is too close to one with more tokens takes a distinct fallback color, and projects and devices take the slots of one project palette in the same order. A series keeps its color in every chart that shows it, legends and tooltips included, and no two series share one while distinct colors last.
Every Stats table writes dollar amounts one way, and so do the tooltips and
subtitles that name a row’s amount: whole cents below $1,000 ($335.96), one
decimal in thousands and millions above ($11.9k), and <$0.01 for an amount
under a cent. Headline tiles and chart axes use compact dollars ($17k), and
counts everywhere use one compact rule (14B, 5.8B, 660.6M, 12k).
What headline stats leave out
Section titled “What headline stats leave out”Some archived sessions were started by software rather than a person. Hansard keeps them archived and searchable but leaves them out of headline numbers so they do not inflate your activity:
- One-shot SDK jobs. Sessions started through an agent SDK that made no
tool calls and got fewer than two replies. They stay out of totals unless
you turn on SDK jobs or pass
--include-sdk-jobs. SDK sessions that use tools or run for two or more replies count as delegated agent work. - Machine probes. Sessions another app creates when it runs an agent CLI to
check status, such as a usage tracker that runs
claude /usage. Hansard treats a session as a probe when it ran inside another app’s data folder, or when its only prompt is a single slash command with no file changes and at most one reply. Probes never count in Stats or Behavior, and there is no toggle to include them.hansard statsreports how many were left out on its Machine probes line. - Codex action assessments. The automatic reviews Codex runs on its own actions. They are left out the same way as probes.
Conversations teammates share through a Team are left out too, until you turn on Show team conversations in Sessions, Stats, and Graph in Settings → General.
Measures of your own activity count only prompts a person typed. That covers the activity calendars, Messages, conversation depth, and per-turn rates. Sessions started by an SDK or a parent agent, messages delivered between agents, and Codex task messages add no human turns there. Token, spend, and tool totals still include delegated SDK sessions and subagents according to the Include toggles.
A session whose harness recorded no model counts under that harness, for example as Model not recorded by Devin, instead of joining one unknown model shared by every harness. Such sessions stay out of model rates and model comparisons.
If Hansard classifies a session the wrong way, override it by project or path in your configuration:
hansard config set stats.probeSessions.forceSession '["/home/user/code/app"]'hansard config set stats.sdkJobs.forceDelegated '["github.com/example/app"]'Each value is a repository key or an absolute working directory, and *
matches any text. Use stats.probeSessions.forceProbe to exclude an automated
path, and stats.sdkJobs.forceJob to treat a path’s SDK sessions as one-shot
jobs. Changes apply without a reimport.
Read how the work turns out
Section titled “Read how the work turns out”When Work classification has labeled more than 25% of eligible conversations, Work and Outcomes show detailed classification charts. Until then, the Classification coverage card shows progress for each label set and links to Open classification settings. Past that point, one line gives the labeled share and names any optional profile that has not run yet.
Every chart in the section covers all history unless it names a range. The section reads top to bottom as four questions:
| Card | What it shows |
|---|---|
| What kind of work is it? and How does it end? | What each labeled session was mainly about, and how it ended. A kind under 1% reads as <1%. |
| Which kinds of work go well? | One row per work mode with 10 or more sessions with a goal, ranked by the share that succeeded, with the succeeded, partially succeeded, and failed split and a 95% range; modes with fewer follow, muted, under their own heading. |
| What kind of work led each month? | A leader timeline across all history: the work mode with the most labeled sessions each month, a strip of monthly volume on top, and the number of months each mode led. A month with fewer than 5 labeled sessions, or a tie at the top, names no leader, and its tooltip lists what it held. Modes that never led open under Show. |
| Is success improving over time? | The work mix and the judged success rate by week or month across all history, or in the selected range after Use selected range, for coding agents or chats, with a one-line verdict on the trend. Only weeks or months with at least 10 sessions with a goal join the line; thinner ones show as hollow rings and faded columns, and a dashed segment skips them. The shaded band is a 95% range, and the strip under each chart shows how many sessions carry labels. Weeks switch to months when too few weeks qualify; when the selected range still has fewer than three, the card says so and offers Show all history. |
| What does the work involve? | Task family by horizon, collaboration and autonomy, task difficulty, primary output types, and O*NET work activities. Each card appears once its optional profile covers more than 25% of sessions. The task-family matrix rates a cell only once it holds 10 sessions with a goal, shows the count alone below that, and stacks each family’s horizons under its name on narrow screens. Collaboration modes rank by judged success with a 95% range, and difficulty facets rank by their mean on the 1 to 5 scale. Output types under 1% of sessions open under Show, and a card left without a neighbour takes the whole row. |
| Which model did most of your judged work each month? | A leader timeline across all history: the model with the most sessions with a goal each month, with the months each model led. A month with fewer than 5 such sessions, or a tie at the top, names no leader. A model without a lab mark shows the mark of the harness that recorded it. Models used without ever leading open under Show. |
| Do newer models finish more tasks? | One bar per model that splits its sessions with a goal into succeeded, partially succeeded, and failed, so the green part ends at the success rate, with a 95% range at that point and a dashed line at your average. Models carry their brand marks and are grouped by lab and tier, such as Claude Opus or GPT Pro, on one shared axis and listed oldest release first. Models last used more than six months before a group’s latest session open under Show. Each tier compares its newest release with the one before it and names a change only when their ranges do not overlap; the subtitle gives the verdict. Labs with a single model share one group at the end. When no tier has two models, the card reads Which models finish tasks most often? |
| Which models deliver, and at what price? | A ranked table of every model with priced spend: its judged success rate, the same rate as in the card above, with a 95% range, then the estimated cost per session and per success, where per success is the cost of a session divided by the success rate. Most reliable marks the model whose range starts highest, and its range runs down every row as a shaded band whose edges the key names. Models whose own 95% range reaches that band come first under Within the margin of error, the rest under Succeed less often. Each group ranks by cost per success, with models under 25 judged sessions listed after the others, and a group names itself only when it holds two or more rows. Best value marks the cheapest model with 25 or more judged sessions within the margin of error, and the subtitle names both success rates and both prices; a cheaper model with fewer sessions says so in its tooltip. A row carries one badge, such as Best value, most reliable. When every model has fewer than 25 judged sessions, Cheapest, small sample marks the cheapest one within the margin of error, and the subtitle says the samples are small. Work compares one kind of work, or one difficulty band once horizon and difficulty labels cover most sessions, when at least two models qualify, and Split by effort ranks each recorded reasoning effort as its own row. Table and subtitle costs keep whole cents, costs under a cent read as <$0.01, and tooltips spell them out. |
| Which kinds of work cost the most? | Work modes ranked by estimated spend on sessions with a goal, with the cost of a session, the cost of a success, which divides it by the success rate, the priced share, and the success rate. Only some sessions have every token priced, and that share differs by kind of work, so each mode’s spend is its mean cost per fully priced session times all of its sessions with a goal. Modes with fewer than 10 priced sessions follow the ranking as outlined bars, a rough estimate. |
| How strong is the proof of success?, How strong are the signs of failure?, and Which signals show up, and how often? | How strong the cited evidence was, as one row per grade from 5 down to 1 and one for no evidence observed, each a share of the labeled sessions, with the sessions not labeled yet named below; and how often sessions were verified successes, wrote code, hit trouble, or were abandoned. Verified success and wrote code count coding-agent sessions only, since chats cannot run tests or edit code; trouble and abandonment show every labeled session, then coding agents and chats. |
The timelines and the model cards compare coding agents and chats separately. What kind of work led each month? and Is success improving over time? each have their own Coding agents and Chats switch, and the switch in How do models compare? sets the four model cards under it. The success and cost cards need at least 10 judged sessions per model, the leader timelines count every model, and sessions with an unknown model are left out. When a group has labeled sessions but no model to compare, the chapter subtitle says why. A goal-bearing session is one classified as Succeeded, Partially Succeeded, or Failed; sessions with no clear goal are left out of success rates. Verified success means the session succeeded with strong cited evidence that recorded tool and test results agree with. The model cards cover all history. These charts describe your own archive and are observational: harder tasks often get stronger models and higher effort.
Compare model quality and cost
Section titled “Compare model quality and cost”Under Models, Model frontier places each model by its median agent conversation. Each chart puts quality on one axis and the median cost, active time, or output tokens of a conversation on the other, on a log scale. The dotted Pareto line joins the models that no other model beats on both axes.
- Your success measures quality as the share of conversations with a clear goal that classification judged successful. Intelligence Index uses the Artificial Analysis Intelligence Index instead.
- By effort gives each model and reasoning effort its own dot. By model gives each model one dot.
- How far below list price do you pay? compares your median conversation with the same conversation at list price, which bills every cached token as fresh input. Dots below the diagonal cost less than list price, mostly through caching.
Model frontier covers human agent conversations across all history, not the selected range. A model with fewer than 5 conversations is left off a chart, and the chart says so. Cost counts only conversations whose tokens are all priced.
To plot the Intelligence Index, select Intelligence Index, paste a free API
key from your Artificial Analysis account,
and select Save key. Hansard keeps the key on this computer, fetches the
index at most once a day, and matches it to your models by name. Remove key
deletes the key. You can also set ARTIFICIAL_ANALYSIS_API_KEY in Hansard’s
environment.
See how your work with agents went
Section titled “See how your work with agents went”You, the first page in the sidebar, compares the 28 days that end on the range’s last day with the 28 days before them. It leaves out a project’s Stats, since it covers all of your work.
| Figure or card | What it counts |
|---|---|
| Waiting on you and How long did agents wait for approval? | How long each agent waited for an approval you gave, from its transcript or, where that records none, the telemetry agents report to this computer, with your approvals and the median wait. Agents that waited at the same time each count, so from a day up the total reads in hours. |
| Tasks that succeeded and Which kinds of work succeed? | Conversations with a goal by work mode, and how many succeeded, partly succeeded, or failed, from classification. |
| Which agent work was redone? | Commits an archived conversation made that a later commit reverted, and test failures that a later conversation fixed, from Hansard’s experience index of your repositories. Open opens the conversation. |
| Cost per finished task and What does a finished task cost? | The estimated spend of the conversations that succeeded, divided by their number, by work mode. Only conversations whose tokens are all priced count. |
| Limit hits and How often did limits stop work? | Plan limit windows the account router saw reach 100%, by account, and the conversations that ran into rate limits. |
Last week reads the same measures for the last full week, Monday to Sunday, against the week before. Recurring lists what held two or more weeks in a row, such as one agent waiting 15 minutes or more for your approvals, or one account’s limit filling, with the number of weeks. Up to three fixes follow: Open Notifications when agents waited 30 minutes or more, Copy rule when Fix tasks succeed less often than the rest, and Open conversation for the latest revert.
A conversation counts as one task, and a task that succeeded after you steered it counts as succeeded. The page does not show how long agents waited for your replies or questions, or how much agent time an hour of your attention bought, because Hansard does not record turn timing yet. While the page header scopes Stats to a team, You shows one line instead of its figures.
Find what to change
Section titled “Find what to change”Optimize, the first page in the Telemetry group, lists fixes ranked by how much of your usage each one covers. Its name in the sidebar counts the open fixes, for example Optimize 3. A share says how much of the usage a fix applies to, not what it saves.
| Fix | When it appears | What its button does |
|---|---|---|
| Allow … in Claude Code | Claude Code asked you about the same tool or kind of command 20 or more times in the last 14 days, and you allowed 95% or more. | Do it for me has your agent propose the narrowest allow rule. Copy rule copies the rule when Hansard knows it exactly. |
| Turn off the … plugin in Claude Code | Claude Code loaded a plugin in 10 or more conversations in the last 14 days, and its skills and MCP tools ran in fewer than 5% of them. | Do it for me has your agent propose turning the plugin off. |
| Keep each conversation on one account while its cache is warm | The account router sent requests to another account while the cache was warm, and they re-sent $10 or more in the last 14 days. | Open Accounts. Ask why asks your agent about the finding. |
| Keep Claude Code’s prompt cache for an hour | Claude Code came back after a pause of 5 to 60 minutes and wrote its 5-minute cache again, for $10 or more in the last 14 days. | Do it for me has your agent propose the promptCacheTtl setting. Copy setting copies it. |
| Start a new conversation when a task is done | In the past 4 weeks, 3 or more conversations averaged 150k tokens a request or more, and they hold 20% or more of your prompt tokens. | Do it for me has your agent propose a line for your agents’ global instruction files. Copy instruction copies the line. |
| Use … for … tasks | A model finished one kind of work about as often (within 5 points) as a model that costs twice as much or more, over 20 or more tasks each. | Compare models opens Outcomes. Ask why asks your agent about the finding. |
The first four read this computer’s Claude Code activity, so they appear on Stats itself and not on a project’s Stats. A fix never changes an agent’s settings by itself. Do it for me starts your own agent on the finding in a side panel. The panel shows each change the agent proposes, and nothing is written until you apply it. Dismiss hides a fix in this browser until its share doubles.
Find why the prompt cache was rewritten
Section titled “Find why the prompt cache was rewritten”Claude and OpenAI models read a repeated prompt from a cache at a fraction of its price. When the cache is gone, the next request processes the whole prompt again. Why was the prompt cache rewritten?, on the Context & cache page, counts the requests that processed most of a large prompt again, in Claude Code, the Claude SDK, and Codex, and ranks the causes by estimated cost:
| Cause | What happened |
|---|---|
| Idle | The cache expired between messages: after 5 minutes or an hour for Claude, depending on how the cache was written, and after 30 minutes for Codex. |
| Dropped while warm | The cache was gone before it expired. Behind the account router, that is mostly a request another account served. Otherwise a changed system prompt or tool list causes it. |
| Compaction | The conversation was compacted since the previous request. |
| Model switch | The conversation changed models. |
Transcripts do not name the account that served a request. Where Claude Code activity on this computer saw the request, Dropped while warm splits into Account switch and Other change. The split shows on Stats itself over a preset range without filters.
Conversations count once they are imported or rebuilt. To count older ones,
run hansard rebuild --since all.
The composer shows the same cache for one conversation: whether it is still warm, and once it expired, what your next message processes again. See Continue from a summary.
Claude Code activity
Section titled “Claude Code activity”Claude Code can send Hansard its own activity: each model request’s tokens, cost, and time, how long each tool call waited for your decision, compactions, and API errors. Claude Code that Hansard starts sends it while the watcher runs. Setup selects Claude Code activity by default and adds the local reporting settings only after you accept setup. Turning that row off writes nothing to Claude Code’s settings. Existing exports elsewhere remain unchanged. For terminal sessions after setup, use Settings → Handoff → Activity → Send Claude Code activity to Hansard. Turning it off removes only variables that still have Hansard’s values.
Under Speed → Agent runtime, Where did each agent’s requests go? includes Claude Code’s main conversation, subagents, compaction, and helper calls, such as title generation. It uses the last 14 days on this computer and ignores the selected range. How did each agent run? compares available waits and approvals across agents over the selected range.
The same activity shows Rate limited and compaction sizes on the Running view, and how fast a plugin’s MCP servers answered on its page in Skills & plugins.
Hansard keeps this activity on this computer for 14 days and never syncs it. Prompts, replies, tool input and output, commands, file paths, error text, and account identity are never kept. See Data boundaries.
Retained evidence in a conversation
Section titled “Retained evidence in a conversation”Finished live activity can also remain with an archived conversation on this computer. Its timeline and Details show approvals, errors, retries, hooks, and compactions. Each conversation retains at most 512 records and 256 KiB, with a notice when older detail exceeds the limit.
Rebuilding conversation data preserves this evidence. Personal Sync and Teams do not transfer it. No migration or reimport is required. Activity already expired from the collector cannot be recovered.
The activity switch also controls supported runtime hooks in Claude Code, Codex, Cursor, Copilot CLI, Factory, Augment, Antigravity, and OpenCode. They record available permission, compaction, subagent, and termination events. Cursor can include token counts, and OpenCode can include permission replies and retry notices. Unsupported fields remain absent.
Hansard adds no per-tool recorder hooks. Hook records stay on this computer and exclude prompts, tool input, and output. Integration repair installs missing hooks without turning activity on. Reload the client after repair. Cline transcript import remains supported, but setup does not configure its extension telemetry.
hansard rebuild --since all refreshes summaries from archived messages,
including recorded runtime details and provider credits. Credits keep their
provider’s unit and remain separate from USD. Automatic archive refresh also
updates these summaries without a provider reimport.
Estimate what a plan is worth
Section titled “Estimate what a plan is worth”Subscription value, on the Limits & accounts page, estimates what each Claude and Codex plan’s limits are worth at API prices for the way you use them. Hansard matches each rise in a plan’s limits with the requests that caused it and prices those requests for your mix of models and tokens. The requests come from the managed router’s own records while it routes a provider, and from archived transcripts and the 15-minute limit samples before then. The section covers its own last 45 days and ignores the selected range.
| Card | What it shows |
|---|---|
| What is each plan worth at your workload? | Each plan’s use as a monthly pace (Used / month) and the value of a fully used month (Full use), against the plan’s price. A plan that lists no price, such as a Team seat, offers Set price. |
| How do limits price each model and token type? | The credit cost of each model and token type relative to its API price. Above 1x, the limits charge more than the API price suggests. |
| How much of this is measured? | Where the requests behind the estimate come from. |
The estimate needs plan usage tracking. To print it from the CLI:
hansard account valuehansard account value --provider codexWhen a plan lists no price, hansard account value prints the command that
sets one, such as hansard account value price <provider:plan> 30 for $30 a
month. clear in place of the amount returns to the listed price.
Check how spend is estimated
Section titled “Check how spend is estimated”Select the question-mark icon beside Estimated spend (How spend is estimated) to open Spend estimation & pricing data. It has three sections:
- Method explains the calculation.
- Totals for the current filter splits total spend into harness-reported and estimated amounts, shows how much is low-confidence, and counts priced and unpriced tokens.
- Model release & pricing reference lists known models with their release dates and rates per million tokens. Model names in your archive that Hansard cannot yet match to a known model appear under Archive identities needing mapping, with their sessions and tokens.
When a session records the provider’s own cost, Hansard uses that value. A cost marked as estimated by the provider remains estimated spend; measured token counts are still reported as measured. This calculation updates without reimporting after restarting the viewer. If an aggregate estimate lacks per-model costs, model rows use published rates for their recorded tokens. Otherwise, it estimates each session by multiplying input, output and reasoning, cache-write, and cache-read tokens by the matching rates. Rates come from the session’s recorded route when one exists, then from dated price windows by session start day, then from a bundled price table. Retired models keep their last known rates. A model with no rate stays unpriced: its tokens count, it adds nothing to spend, and Your top models shows its spend as —. Spend is marked low-confidence when token counts were estimated, output tokens are missing, or the cache split is unknown. Pricing updates apply without a reimport.
Common questions
Section titled “Common questions”How do I see Claude Code and Codex token usage in one place?
Section titled “How do I see Claude Code and Codex token usage in one place?”Open Stats, or run hansard stats. Every harness in the archive counts in
the same totals, and the harness and model tables split tokens and estimated
spend by Claude Code, Codex, Cursor, and each other harness you use.
Is estimated spend what I actually paid?
Section titled “Is estimated spend what I actually paid?”No. It prices recorded tokens at API rates, or uses the cost a harness recorded when one exists. A subscription plan bills differently; to follow plan limits, use Accounts and plan usage. To see what a plan’s limits are worth at API prices, see Estimate what a plan is worth.
Why is my activity lower than my message count?
Section titled “Why is my activity lower than my message count?”Activity counts only prompts a person typed. Sessions started by an SDK or a parent agent, messages between agents, and machine probes add no human turns. See What headline stats leave out.
Troubleshoot missing data
Section titled “Troubleshoot missing data”Hansard shows gaps rather than guessing. Use the symptom to choose the next step:
| Symptom | Meaning | Next step |
|---|---|---|
| A model setting shows Not recorded, or a model, token count, or tool argument is missing | The source does not store that field. | Check Known limitations in the harness guide. |
| Tokens appear without spend, or a model shows — for spend | No provider cost or price match was found for that model. Its tooltip reads Unpriced token usage. | Open Spend estimation & pricing data and look for the model under Archive identities needing mapping. Unpriced tokens still count. |
| SDK work is missing from totals | One-shot SDK jobs are excluded by default. | Turn on SDK jobs, or run hansard stats --include-sdk-jobs. |
| A real project is missing entirely | Its sessions matched the machine-probe rule. | Add its path to stats.probeSessions.forceSession. |
| Classification cards show coverage instead of charts | Fewer than 25% of eligible conversations have current labels. | Run hansard classify status, then classify the needed label set. |
| A field added in a newer Hansard is still missing | Older imports did not record it. | Run hansard rebuild --stale-parsers --provider <provider> while the original harness files still exist, or follow the harness guide’s reimport steps. |
| Numbers change right after Stats opens | Stats shows the last saved numbers while it recalculates. | Wait for the refresh to finish. |
hansard rebuild without options refreshes views from the existing archive. It
cannot recover a field that the original import never recorded.