Skip to content
Sign up

flagon-io/g1t

public
00

Where people and agents ship software together. The open-source git platform for the whole job: issues, agents, checks and deploys to the edge.

gitagentscloudflarerustforgemcp
186 linesCodeBlameRaw
1# Agent spend and agent work per engineer
2
3Research, 2026-10-18. The question: how g1t should show agent spend and agent work per engineer, across every coding agent a team uses, as well as or better than the products that already do it.
4
5Sources are vendor docs, API references and changelogs, linked inline. **UNCONFIRMED** marks a claim that rests on a secondary source, on vendor source code without docs, or on nothing I could open. Vendor marketing claims are labelled as such.
6
7## Where g1t stands
8
9These are the integration points the plan below changes.
10
11**Claude Code** is the only agent that reports spend.
12
13- `plugins/claude-code/hooks/g1t.mjs` reads `message.usage` from the transcript on `Stop`, `SubagentStop` and `SessionEnd`, and removes duplicates by message id plus `requestId`.
14- It sends the totals with `g1t session heartbeat --usage-json` (CLI 0.4.1 or later).
15- That becomes `PATCH /repos/{owner}/{name}/sessions/{id}`, then the work service's `update_session`.
16- The work service prices it with `crates/contracts/src/model_prices.rs`, which has Anthropic list prices only, and writes `sessions.usage` and `sessions.cost_usd` (`services/work/migrations/0053_session_cost.sql`).
17
18**Other agents** report no cost.
19
20- Codex (`plugins/codex`) has no hooks.
21- Other agents use `plugins/other-harnesses/AGENTS.snippet.md`, manual CLI calls without usage.
22- Even if they sent tokens, `model_prices.rs` would price a non-Anthropic model as unknown (`cost_partial`).
23
24**Provider reports**, in `services/billing/src/provider_costs.rs`:
25
26- Anthropic `cost_report` is grouped by `workspace_id` and `description`.
27- OpenAI `organization/costs` is grouped by `line_item` and `project_id`.
28- Neither is grouped by user. Both are stored in `provider_costs` for the whole organization.
29
30**How work is attributed** (`services/work/src/signals/base.sql`, `weeks.sql`, `pulls.sql`):
31
32- A pull request is an agent's when its `pulls.agent` is `agent`, its author is the g1t agent, or a session is tied to it through `session_links`, `sessions.pull`, or a branch touch.
33- A session's cost is split evenly across the pull requests it is tied to.
34- Direct pushes count only as `agent_commits`, built from the plugin's `session touch` reports.
35
36**Why a solo workspace sees almost nothing** (`apps/web/app/lib/signals.ts`):
37
38| Finding | Why it stays hidden for flagon-io |
39| ------- | ------------------------------------------------------------------------------------------------------------- |
40| spend | `quarterSpend` counts only per-pull cost, so direct-push sessions drop out. It also needs two or more groups. |
41| review | Needs a non-owner review. |
42| shipped | Needs a closed milestone. |
43| stalled | Needs open pull requests. |
44| rework | Needs 5 or more agent merges a week. |
45| silos | Needs 3 people. |
46| spread | Needs 2 people who adopted a harness later. |
47
48The pulse tiles do count direct-push spend (`signals.server.ts:140`).
49
50**Who can see what:**
51
52- There is no per-person spend view. Billing `Usage` has no actor dimension (`crates/contracts/src/billing.rs:311`).
53- Session visibility is own, member or outside (`services/work/src/coordination/shown.rs`). Members see each other's `cost_usd`, but not each other's prompts.
54
55**Open question.** The sessions this research ran, including its own subagents, came back from coordination with `cost_usd: null` after they ended. Either the coordination payload leaves cost out, or subagent usage doesn't arrive. Step 0 checks which.
56
57## 1. What each agent exposes
58
59### Comparison of coding agents
60
61| Agent | Telemetry (OTel) | Admin / usage API | Local session files | Hooks | Per-user? | Cost field |
62| ---------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------ | --------------------------------------------------- |
63| **Claude Code** | Metrics `claude_code.cost.usage` (USD) and `claude_code.token.usage`; event `claude_code.api_request` with `cost_usd`, tokens, `request_id`. Attributes `session.id`, `user.email`, `user.account_uuid`, `organization.id`, `query_source` (main/subagent). Set centrally through the `env` of managed settings. [docs](https://code.claude.com/docs/en/monitoring-usage) | `GET /v1/organizations/usage_report/claude_code`: per actor (email or API key) per day: sessions, lines, commits, PRs, tool accept/reject, `model_breakdown[].tokens` and `estimated_cost` (cents). Admin key; no individual accounts; not Bedrock or Vertex. [docs](https://platform.claude.com/docs/en/manage-claude/claude-code-analytics-api). Enterprise (claude.ai) has a separate Analytics API (`read:analytics`). `usage_report/messages` and `cost_report` **cannot group by user**. [docs](https://platform.claude.com/docs/en/manage-claude/usage-cost-api) | `~/.claude/projects/<slug>/<session>.jsonl`, `message.usage` (input, output, cache read, cache write 5m/1h). No cost. Subagents in `<session>/subagents/agent-<id>.jsonl`. | 33 events; `session_id`, `transcript_path`, `prompt_id`; `SubagentStop` adds `agent_transcript_path`. No cumulative tokens in any hook. [docs](https://code.claude.com/docs/en/hooks) | Yes (email) | Yes: OTel, Analytics API. Estimate at list price. |
64| **Codex** (CLI, IDE, app-server) | `[otel]` in `config.toml`: `exporter`/`metrics_exporter` `otlp-http\|otlp-grpc`. Events carry `conversation.id`, `user.email`, `user.account_id`, `model`; `codex.sse_event` has token counts. **Source-only:** `codex.turn_cost` event and `codex.turn.cost_microusd` metric; whether they fire in the plain CLI is UNCONFIRMED. Central defaults in `/etc/codex/managed_config.toml`. [docs](https://learn.chatgpt.com/docs/config-file/config-advanced) | API-key use: `GET /v1/organization/usage/completions` and `/v1/organization/costs` with `group_by=user_id\|api_key_id\|project_id` (no Codex-specific `api_source`). ChatGPT workspaces: Codex Enterprise Analytics API, scope `codex.enterprise.analytics.read`; paths UNCONFIRMED. [docs](https://learn.chatgpt.com/docs/enterprise/analytics-api) | `~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl`, compressed to `.jsonl.zst` when cold (2026). `event_msg` `token_count` holds `info.total_token_usage` (input, cached, cache write, output, reasoning) and `rate_limits`. No cost. | **Codex now has hooks** (on by default): `SessionStart`, `SessionEnd`, `SubagentStart`, `SubagentStop`, `Stop`, `PostToolUse`, …; payload `session_id`, `transcript_path`, `model`, `turn_id`. No tokens. Legacy `notify` still exists. [docs](https://learn.chatgpt.com/docs/hooks) | Yes (email in OTel; `user_id` in Platform costs) | OTel (source-only); Platform costs API |
65| **Cursor** | **Server-side** OTel export (beta; turned on in Team Settings, nothing installed on machines). Metric `cursor.cost.usage` (USD), `cursor.token.usage`; logs `cursor.api.request`, `cursor.cloud_agent.pull_request`, `cursor.cloud_agent.commit`. Resource `cursor.user.email`; logs `cursor.conversation.id`. Plan likely Enterprise (UNCONFIRMED). [docs](https://cursor.com/docs/enterprise/opentelemetry-export/wire) | Admin API (`api.cursor.com`, Basic auth): `POST /teams/daily-usage-data` (per user per day: lines, accepts, requests); `POST /teams/filtered-usage-events` (per request: `userEmail`, `model`, `tokenUsage`, `chargedCents`, `conversationId`); `POST /teams/spend` (`spendCents` per member). AI Code Tracking API `GET /analytics/ai-code/commits` (per commit: tab, composer and non-AI lines; Enterprise alpha). [docs](https://cursor.com/docs/account/teams/admin-api) | `state.vscdb` SQLite (`cursorDiskKV`); undocumented, changes between versions. Token fields UNCONFIRMED. | `hooks.json` (MDM, team, project or user): `sessionStart`, `sessionEnd`, `stop`, `afterFileEdit`, `subagentStop`, …; payload `conversation_id`, `generation_id`, `model`, `user_email`, `transcript_path`. No tokens. [docs](https://cursor.com/docs/agent/hooks) | Yes (email) | Yes: `chargedCents`, `spendCents`, OTel |
66| **GitHub Copilot** | VS Code Copilot Chat: `github.copilot.chat.otel.*` / `COPILOT_OTEL_*`. `gen_ai.client.token.usage`, `gen_ai.conversation.id`; identity only with `captureIdentity`. No cost. [docs](https://code.visualstudio.com/docs/agents/guides/monitoring-agents) | The legacy `/copilot/metrics` API was retired 2026-04-02. New usage metrics reports: `GET /orgs/{org}/copilot/metrics/reports/users-1-day?day=` (download links). Per-user fields include `ai_credits_used`, `used_copilot_coding_agent`, `totals_by_cli.token_usage`. Billing: `GET /organizations/{org}/settings/billing/ai_credit/usage?user=` (`netAmount`). Seats: `/orgs/{org}/copilot/billing/seats`. [reports](https://docs.github.com/en/rest/copilot/copilot-usage-metrics), [billing](https://docs.github.com/en/rest/billing/usage) | Copilot CLI `~/.copilot/session-state/<id>/events.jsonl`; undocumented, token data UNCONFIRMED. | CLI and cloud agent: `.github/hooks/*.json`; `sessionStart`, `sessionEnd`, `agentStop`, …; payload `sessionId`, `cwd`, sometimes `transcriptPath`. No model, user or tokens. [docs](https://docs.github.com/en/copilot/reference/copilot-cli-reference/cli-hooks-reference) | Yes (login) | AI Credits per user (since 2026-06-01, usage-based) |
67| **Windsurf, now Devin Desktop** (renamed 2026-06-02) | None documented. | Enterprise: `POST server.codeium.com/api/v1/UserPageAnalytics` (`promptCreditsUsed` per email); `/CascadeAnalytics` (`cascadeRuns[].promptsUsed` per `cascadeId`). [docs](https://docs.devin.ai/windsurf/plugins/accounts/api-reference/user-page-analytics) | `~/.windsurf/transcripts/{trajectory_id}.jsonl`, only with a transcript hook. | Cascade hooks (`post_cascade_response`, …) with `trajectory_id`, `model_name`; Devin Local/CLI hooks with Claude Code names (`SessionStart`, `Stop`, …) and `session_id`. No tokens. [docs](https://docs.devin.ai/desktop/cascade/hooks) | Yes (email) | Credits, not dollars |
68| **Gemini CLI** | `telemetry` in `settings.json`, or the `/etc/gemini-cli/settings.json` system file. `gemini_cli.token.usage`, `gemini_cli.api_response` (token counts), `session.id`, `user.email`. **No cost metric.** `logPrompts` defaults to **true**. [docs](https://github.com/google-gemini/gemini-cli/blob/main/docs/cli/telemetry.md) | No per-user API. Code Assist metrics in Cloud Monitoring are aggregate; Vertex billing is per project. [docs](https://docs.cloud.google.com/gemini/docs/codeassist/generate-metrics) | `~/.gemini/tmp/<hash>/chats/session-*.json[l]`; each `gemini` message has `tokens{input,output,cached,thoughts,tool,total}` and `model`. No cost. | `SessionStart`, `SessionEnd` (best effort), `AfterModel`, …; `session_id`, `transcript_path`. [docs](https://github.com/google-gemini/gemini-cli/blob/main/docs/hooks/reference.md) | Via OTel only | No; tokens × price |
69| **Amp** | None found. | `GET ampcode.com/api/v2/workspace/analytics/daily-usage` (per user per day: `metrics.usage` USD, lines); `GET /threads/{id}/usage` (USD per thread, per model, with tokens). [docs](https://ampcode.com/api/external) | Threads live on the server; `--stream-json` has `usage` and `session_id`. | TypeScript plugin API (`session.start`, `agent.end` with `thread.id`). No usage. [docs](https://ampcode.com/docs/plugin-api) | Yes | Yes: USD per user and per thread |
70| **Devin** (cloud) | None documented. | Enterprise v3: `GET /v3/enterprise/sessions` (each with `user_id`, `pull_requests[]`, `acus_consumed`); `/v3/enterprise/consumption/daily/users/{id}`; `/metrics/prs`. [docs](https://docs.devin.ai/api-reference/v3/sessions/enterprise-sessions) | None (cloud). | No outbound session webhooks (UNCONFIRMED absence). | Yes | ACUs per session and per user |
71
72**How agents mark their own commits:**
73
74- **Copilot coding agent:** Copilot is the commit author, the assigning person is co-author, and an `Agent-Logs-Url` trailer links the session. [changelog](https://github.blog/changelog/2026-03-20-trace-any-copilot-coding-agent-commit-to-its-session-logs)
75- **Cursor:** the CLI adds a `Made with Cursor` trailer. Cloud agents commit as "Cursor Agent" with the person as `Co-authored-by`. [docs](https://cursor.com/docs/cli/reference/configuration)
76- **Devin:** the author is configurable: Devin, co-authored, or the person.
77- **Claude Code:** `Co-Authored-By: Claude`.
78
79**What "cost" means differs by plan.**
80
81- On a subscription (Claude Max, ChatGPT plans, Cursor seats), every token-based figure is a list-price **estimate**: what the work would cost through the API, not what was paid.
82- Swarmia calls this "token value". LinearB's docs say "Estimated consumption is not a billing total."
83- g1t should say so too. The Claude Code Analytics API's `customer_type` (`api` or `subscription`) tells the two cases apart.
84
85### Products that sell this
86
87| Product | Ingests | Attributes spend to work by | Per-engineer view, and its privacy stance | Pricing (public) |
88| --------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------ |
89| **LinearB** | Git/PR; Copilot, Cursor, Amazon Q; Claude Code via Admin API, Enterprise Analytics API or OTel. [docs](https://linearb.helpdocs.io/article/p4lfzfnu0w-ai-tools-integrations) | Tokens and cost in Metrics Builder; "estimated consumption … not a billing total". | User-level data; no stated stance. | $29 / $59 per contributor per month, 50+ and 100+ developers. [pricing](https://linearb.io/pricing/) |
90| **Jellyfish** | Claude Code, Cursor, Copilot, Windsurf, Gemini, Amazon Q, review bots; Git, Jira. [press](https://jellyfish.co/newsroom/end-to-end-ai-impact-sdlc/) | "AI Spend Insights" by tool, team and initiative (marketing). | Says it offers individual ROI; no safeguards found. | Quote only. |
91| **DX** (Atlassian since 2025-11-10) | Surveys; Git; Claude Code via analytics, Console or OTel; a **local daemon** that records commits for AI Code %. [docs](https://docs.getdx.com/reports/ai-code-percentage/) | AI % per commit, weighted by size; developers can correct their own commits. | Advises against individual evaluation: AI % is "a lens, not a KPI". Person-level drilldown still exists for admins. | Quote only. |
92| **Swarmia** | Copilot, Cursor, Claude Code (OTel and Analytics API), Codex, custom tools. [docs](https://help.swarmia.com/features/ai-tools/ai-adoption) | **Each person's daily usage is split across the issues and PRs they touched that day, weighted by effort.** AI-PR detection in tiers: trailer, co-author or label (high confidence); used AI in the last 24 hours (low confidence). The UI shows which rule fired. [cost](https://www.swarmia.com/changelog/2026-08-05-ai-cost-beta/) | Team first, with drilldown to individuals. "Not a tool for spying on developers"; no stack ranking. | **Free under 10 developers**; AI cost add-on **$5 per developer per month**. [pricing](https://www.swarmia.com/pricing/) |
93| **Faros AI** | Claude Code Analytics API (daily pull), OTel (real time), Git. [blog](https://www.faros.ai/blog/claude-code-analytics) | "Token Intelligence" by team and tool; headline metric "average estimated cost per commit". | Per-user data underneath; stance UNCONFIRMED. | Quote; free open-source Community Edition; Faros Desktop is free and local. |
94| **Span** | span-detect-1, a code classifier that needs no telemetry; an **agent trace recorder on the developer's machine** (rolled out by MDM): prompts, tool calls, tokens, dead ends. [site](https://www.span.app/) | AI Code Ratio per PR; wages and token spend mapped to project and work type. | **"Reports are about repos and teams, never people … An engineer's sessions are theirs."** Private by default; engineers see their own; PII stripped. One help snippet lists a "person" breakdown (UNCONFIRMED which is current). | Quote, per contributor. |
95| **Weave** (weaveos.com) | Git, PRs, reviews, AI telemetry from "50+ tools" (marketing); an LLM scores each PR in "Weave Hours". | "Connect token spend to engineering output" (marketing). | Explicit per-engineer table (AI Usage, AI Score, Output). | **$0 for teams under 5**; $50 per engineer per month for individual stats. [pricing](https://weaveos.com/pricing) |
96| **Waydev** | Git, tickets, CI; Copilot, Claude Code, Cursor; "Agent Registry". | "AI Feature Spend" and "Human and Agent Spend" (marketing; per-PR cost UNCONFIRMED). | Contributor pages and benchmarks. | Quote; no free plan. |
97| **Exceeds AI** | GitHub, GitLab, ADO; Cursor, Claude Code, Copilot. | AI share per commit and PR; an ROI example (marketing). | **Most individual-focused:** profiles, coaching, leaderboards. | Per manager seat: $65 per month. |
98| **Opsera** | Copilot, Amazon Q, Cursor, Windsurf dashboards. | Acceptance and adoption only. | "Top users" panels. | Not public. |
99| **Cursor dashboard** | First-party client telemetry. [docs](https://cursor.com/docs/account/teams/analytics) | Per-line signatures matched to later commits, on the device ("AI Share of Committed Code"). | Admins see everyone; members see themselves plus a top-10 leaderboard. | Included with Teams. |
100| **Copilot dashboards** | First-party. [changelog](https://github.blog/changelog/2026-07-22-new-copilot-usage-metrics-impact-dashboard) | Repo report counts Copilot-authored and merged PRs. The impact dashboard shows **cohorts** (`ai_adoption_phase`), not individuals. | Dashboards are org-level; a per-user API exists for admins. | Included. |
101| **RadiantOps** (radiantops.com; launched 2026-09-16, $7M seed; GitPrime founders) | GitHub, Bitbucket, Jira, Linear, Slack, Cursor, Codex, Bedrock. [site](https://radiantops.com/) | "Episodes of work" (code, tickets, chat, agent runs); agent share per change set, "spend per agent merge", rework, stalled, silos, "new ground", a daily Signal. | "A team is not a percentile": compares each team only with its own history. Individual view policy UNCONFIRMED. | Not public; enterprise. |
102
103RadiantOps' feature set is closest to what Home and the Workspace overview already show. Its inputs are other vendors' APIs; g1t's inputs are the forge and the sessions themselves.
104
105### Solo developers and small teams
106
107What these tools have in common is that they're about **limits, not ROI**:
108
109- **Claude Code `/usage`:** session cost, per-model tokens, a cache-hit line, plan-limit bars, and which skills, subagents and MCP servers used the most. `--max-budget-usd` caps a run. Anthropic's benchmark is about $13 per developer per active day. [docs](https://code.claude.com/docs/en/costs)
110- **ccusage:** daily, weekly, monthly and per-session reports, plus five-hour `blocks` with burn rate and a projection against the plan window. Covers Codex, Amp, Copilot CLI and Gemini CLI from their local files. [repo](https://github.com/ryoppippi/ccusage)
111- **tokscale, Sniffly, Grafana stacks** (claude-code-otel and others): top sessions by cost, cost by model, clusters of errors.
112- **Langfuse:** traces from Claude Code and Codex Stop hooks and from Copilot's OTel; cost per developer, project and session. [docs](https://langfuse.com/resources/engineering/coding-agent-tracing)
113- **Git AI:** agent hooks write per-line authorship to `refs/notes/ai` (agent, model, session); the paid tier adds cost per PR. [docs](https://usegitai.com/docs/get-started/how-git-ai-works)
114
115**The gap:** almost no small-team tool connects spend to **work that landed**. They answer "how fast am I burning?", not "what did that buy, and what was thrown away?".
116
117## 2. Patterns worth copying
118
1191. **Three paths in, one number out.** Every vendor joins vendor APIs (daily, backfilled, per user), OTel (real time, any provider) and git. g1t has a fourth path, its own session record, which the others lack.
1202. **Call it an estimate and show the basis.** Show "list-price value" next to the actual invoice when a provider report exists. Use `customer_type` and plan to say which case applies.
1213. **Detection in tiers, with the rule shown** (Swarmia). For every change g1t calls an agent's, say why: a session trailer, a session tie, a co-author, the Copilot author, or a `Made with Cursor` trailer.
1224. **Cloud-agent work belongs to the person who started it** (DX, Swarmia, Devin's `user_id`, Copilot's co-author).
1235. **Spend split over work by effort** (Swarmia). Use it only where g1t lacks a direct link, and label it as a split.
1246. **Idle-seat and plan-fit checks first** (Swarmia, Cursor): token value against what the seat costs.
1257. **Self-view softens individual data** (Span, Cursor members, DX's AI % override, Devin Personal Analytics).
1268. **Compare only with your own history** (RadiantOps). This is also the only baseline that works for one person.
1279. **A daily digest that says what to act on** (RadiantOps Signal, Waydev and Swarmia assistants). g1t already has one: Home.
12810. **For solo users, lead with limits:** burn rate, cache efficiency, and which skill, subagent or MCP server is eating budget.
129
130## 3. What g1t can do that they can't
131
132The products above infer the link between spend and work from outside. g1t is the forge and records the sessions, so it has the link directly:
133
134- **An exact join, not a heuristic.**
135 - A session's `G1t-Session` trailer ties its cost to the commit, the pull request, the work item it closes, and the merge-queue result.
136 - Swarmia needs a 24-hour window, Cursor needs line signatures, and Span needs an ML model to approximate this.
137- **Spend that never landed.** g1t sees sessions that cost money and produced no commit, sessions whose branch was abandoned, and pull requests closed unmerged. Only Span claims to see dead ends, and only through a recorder on each machine.
138- **Direct pushes.** Products built on PR data miss them. g1t receives the push and reads its trailers (`services/work/src/coordination/pushed.rs`).
139- **Rework from the history itself.** g1t sees follow-up commits and reverts on the same lines, failed checks and fix-ups in the merge queue, and their cost. It doesn't depend on a "rework" heuristic over someone else's API.
140- **One identity.** The session actor, the git pusher and the reviewer are one g1t account, so there is no email-to-author mapping page. Provider APIs keyed by email (Anthropic, Cursor, Amp, Windsurf) map onto g1t users through their verified addresses.
141- **The cost of reviewing agent work.** Review time, the number of rounds, and the CI minutes an agent's pull request used are all g1t's own data.
142- **Every harness in one place.** Claude Code, Codex, Cursor, Copilot CLI, Gemini CLI and Devin all have hooks now with `session_id` and `transcript_path`, so one plugin pattern covers them. The products above each need vendor admin keys, and most of those need Enterprise plans.
143- **Useful for one person.** Because g1t records sessions itself, a solo developer gets cost per landed change and spend that never landed without any admin API. No product above serves that case.
144
145## 4. Privacy model
146
147The stance: **spend is information about work, not a score for people.**
148
1491. **Your own, in full.** Every engineer sees all of their own sessions, spend, prompts and what landed, on every harness. This matches what `Sight` already does for sessions.
1502. **Others' spend appears as the cost of work, not ranked by person.**
151 - Teammates see what a change, work item, milestone or team cost, and who worked on it, as today.
152 - There is no view that lists or sorts people by spend, and no leaderboard, score or percentile.
1533. **Per-person totals are a workspace setting.** "Spend by person" can be shown to:
154 - **Owners and billing managers**, the default;
155 - all members;
156 - or nobody.
157
158 It is for budgets and plan fit, and the page says so. Every person always sees their own.
159
1604. **A minimum group of 3 for rollups.** A team or group total is shown only when it covers three people or more. Otherwise it would reveal a person's figure to people the setting doesn't allow. A one-person workspace is all self-view, so this never hides anything there.
1615. **Provider data waits for consent.** Per-user rows from a provider API, such as an Anthropic Analytics email or a Cursor `userEmail`, are joined to a g1t user only through an address that person has verified. Unmatched rows stay in "not linked to anyone".
1626. **Content stays with its author.** Prompts and tool output stay own-only for members (as today) and are never sent to an outside service.
163 - g1t's own OTel receiver should drop `user_prompt`, `assistant_response` and body events unless the person turns them on.
164 - Gemini CLI's `logPrompts`, which defaults to true, should be set to false in the settings g1t writes.
1657. **Opting out is visible.** `G1T_RECORD=0` and `G1T_RECORD_PROMPTS=0` already exist. Totals show how much they leave uncounted ("3 sessions unrecorded"), so nobody mistakes missing data for zero.
1668. **No outside benchmarks.** Compare only with the workspace's own history, as Home already does.
167
168## 5. Ranked plan
169
170The order is by what it unlocks for flagon-io first: one person, direct pushes, Claude Code and Codex. Then it widens to teams and other agents. Effort is S (a day or less), M (a few days) or L (a week or more).
171
172| # | Step | Effort | Unlocks | Exact integration |
173| --- | -------------------------------------------------------------- | --------------- | -------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
174| 0 | **Check that subagent and ended-session cost arrives** | S | Trust in every number below | Compare `sessions.cost_usd` and `usage` for this session and its subagents with the transcripts. Check `SubagentStop` → `agent_transcript_path` in `plugins/claude-code/hooks/g1t.mjs`, and whether coordination `session.ended` payloads leave out `cost_usd` (`services/work/src/coordination/shown.rs`). |
175| 1 | **Spend for direct pushes and solo workspaces** | S–M | Spend findings show for flagon-io. | `quarterSpend` in `apps/web/app/lib/signals.ts`: add sessions not tied to a PR, grouped by the work item their push closed (`pushed.rs` `trailer_session`) or else by repo. Add a `direct` arm to `services/work/src/signals/pulls.sql`, keyed by `session_touches` with kind `commit`. Allow one group when the workspace has one person. |
176| 1b | **Findings for one person** | M | Home says something useful with no reviews or milestones. | New findings in `signals.ts`, each compared with your own last 4 weeks: **spend that never landed** (sessions with cost and no commit or merged PR); **cost per landed change**; **cache efficiency** (cache read ÷ input from `sessions.usage`); **most expensive session this week**, with what it produced. Scale `rework` and `stalled` thresholds to the workspace's volume instead of fixed counts. |
177| 2 | **"Your spend" and spend by person, with the privacy setting** | M | Per-engineer view (section 4). | Work service query over `sessions(actor_id, started_at, cost_usd, usage, harness, model, item_id)` joined to `session_links` and `session_touches`, by day, repo, work item, harness and model, split into landed and not landed. Route `apps/web/app/routes/workspace/spend.tsx` (`-/spend`, defaulting to you). CLI `g1t spend [--by repo\|item\|harness] [--since 7d]`. Workspace setting `spend_visibility = owners\|members\|nobody`, plus the 3-person rollup rule. |
178| 3 | **Codex cost** | M | flagon-io's other harness reports spend. | `plugins/codex/hooks.json` for `SessionStart`, `Stop`, `SubagentStop` and `SessionEnd`, running a shared `g1t.mjs` that reads `transcript_path`: the last `event_msg` of type `token_count`, `info.total_token_usage` (`input_tokens`, `cached_input_tokens`, `cache_write_input_tokens`, `output_tokens`, `reasoning_output_tokens`), reading `.jsonl.zst` too. Send `model` from the hook payload with `g1t session heartbeat --usage-json`. Add OpenAI list prices to `crates/contracts/src/model_prices.rs`. Mark ChatGPT-plan sessions (`rate_limits.plan_type`) as an estimate. |
179| 4 | **Check per-person spend against the provider** | M | "g1t saw 82% of your Claude Code spend". Per-engineer spend for people who don't use the plugin. | In `services/billing/src/provider_costs.rs`: daily `GET /v1/organizations/usage_report/claude_code?starting_at=` with paging, keeping `actor.email_address`, `model_breakdown[].estimated_cost.amount`, `customer_type`, `commits_by_claude_code` and `pull_requests_by_claude_code`. OpenAI `GET /v1/organization/costs?group_by=user_id`. Add an `actor` column to `provider_costs` (new migration) and match it to verified emails. |
180| 5 | **Cursor** | M | The first agent g1t doesn't run itself. | `plugins/cursor`: `.cursor/hooks.json` or `~/.cursor/hooks.json` for `sessionStart`, `stop` and `sessionEnd`, sending `conversation_id`, `user_email`, `model` and `g1t session start --harness cursor`. Cost comes from the Admin API, as Cursor hooks carry no tokens: `POST /teams/filtered-usage-events` (`tokenUsage`, `chargedCents`, `conversationId`), joined on `harness_session_id`; `POST /teams/spend` for per-person totals. Key stored like the provider keys. Enterprise: `GET /analytics/ai-code/commits` for AI lines per commit. |
181| 6 | **OTLP receiver** | L | Coverage with no plugin installed, rolled out through managed settings; real time; Cursor's server-side export with nothing on the machines. | `POST /v1/logs` and `/v1/metrics` (OTLP/HTTP, protobuf and JSON) on the API, authenticated with a workspace token in `OTEL_EXPORTER_OTLP_HEADERS`. Read Claude Code `claude_code.api_request` (`cost_usd`, tokens, `session.id`, `user.email`, `query_source`); Codex `codex.sse_event` tokens and `codex.turn_cost` (`usage.estimated_usd`, `conversation.id`); Cursor `cursor.cost.usage` and `cursor.api.request` (`cursor.conversation.id`, `cursor.user.email`); Copilot `gen_ai.client.token.usage`; Gemini `gemini_cli.api_response`. Join `session.id` or `conversation.id` to `sessions.harness_session_id`. Drop content events by default. Docs: a managed-settings snippet for each harness. |
182| 7 | **Gemini CLI and Copilot CLI hooks** | M | Two more local harnesses. | Gemini: `~/.gemini/settings.json` hooks `SessionStart` and `SessionEnd`; read `transcript_path` (`session-*.jsonl`, `tokens{input,output,cached,thoughts}` per `gemini` message); add Gemini prices. Copilot CLI: `.github/hooks/*.json` `sessionStart` and `agentStop` (`transcriptPath`). Its tokens are UNCONFIRMED, so cost comes from `GET /organizations/{org}/settings/billing/ai_credit/usage?user=`. |
183| 8 | **Cloud agents' work credited to whoever started it** | M | Copilot coding agent, Cursor cloud agents and Devin PRs count as the launcher's agent work, with cost. | In `services/work/src/signals/base.sql`'s `tied`, count as agent work: commits with the `Agent-Logs-Url` trailer or Copilot as author (credited to the co-author); `Co-authored-by` the Cursor Agent, or a `Made with Cursor` trailer; Devin's `GET /v3/enterprise/sessions` `pull_requests[]` with `acus_consumed` and `user_id`. Show the rule that matched. |
184| 9 | **Amp and Devin Desktop spend** | S each, after 4 | Per-person dollars from their APIs. | Amp `GET /api/v2/workspace/analytics/daily-usage` (`metrics.usage` per `user.email`) and `/threads/{id}/usage`. Windsurf/Devin Desktop `POST /api/v1/UserPageAnalytics` (`promptCreditsUsed`, shown as credits). |
185
186**Recommended first slice:** steps 0, 1, 1b and 2, then 3. That gives flagon-io per-engineer spend, cost per landed change and spend that never landed, with the privacy setting in place, before any vendor integration.