Usage attribution

An agent usually starts in the main checkout, runs har env launch to open a slot, works there, and may move to another slot or back to main. Mission Control books each piece of LLM usage on the slot the agent was working in when that usage happened, from evidence only. The design is recorded in ADR 0003.

Two kinds of facts

Usage observations. The smallest billable unit each agent reports, with its own timestamp:

Agent	Unit	Live	Offline (har control sync)
Claude Code	one API request	native OTLP event claude_code.api_request	session transcript, plus one transcript per subagent
Codex	one turn	@osfactory/otel-hook	rollout token_count events, per turn
Cursor	one generation	@osfactory/otel-hook	none

A live and a harvested copy of the same unit share one key (Claude’s request id, Codex’s turn id), so they never count twice.

Bindings. “From time t, this session works in slot N.” HAR creates them from:

• launch / release — har env launch, teardown or complete ran inside the session, from the CLI or from the HAR MCP tools. Each agent gives every command it runs its session id (CLAUDE_CODE_SESSION_ID, CODEX_THREAD_ID, CURSOR_CONVERSATION_ID); HAR records it next to the slot in .har/session-bindings/ (git-ignored) and har control sync sends it. Teardown also releases the session that launched the slot, even when a human runs it from a terminal.
• workspace — the session itself was started inside the slot’s work dir.
• observed — a tool event of the session edited or read a file, or ran a command, inside the slot’s worktree (reported by otel-hook as a branch name and a salted hash, never a path).

Each observation goes to the binding in force at its timestamp. Usage with no binding is the session’s unassigned usage: planning in the main checkout before a launch shows as such and is never booked on a slot. Bindings that arrive late (a sync after the spans) re-attribute usage already received.

What HAR configures

With telemetry on (har telemetry on, or at launch), HAR writes to the env block of ~/.claude/settings.json:

• CLAUDE_CODE_ENABLE_TELEMETRY=1
• the OTLP logs exporter (OTEL_LOGS_EXPORTER, OTEL_EXPORTER_OTLP_LOGS_*, OTEL_LOGS_EXPORT_INTERVAL) to {HAR_CONTROL_API_URL}/api/otel/v1/logs, for the per-request claude_code.api_request events;
• the OTLP metrics exporter to /api/otel/v1/metrics, kept as a cross-check of session totals.

HAR never sets OTEL_LOG_USER_PROMPTS or OTEL_LOG_TOOL_DETAILS, so prompt text and tool content stay out of these events. A key you set to another value is left alone (HAR warns when it sets telemetry up), and har telemetry off removes only the keys HAR wrote.

Codex’s native OpenTelemetry ([otel] in ~/.codex/config.toml) is not configured: its log events carry no turn or response id and its metrics no session id, so they could not be matched with harvested turns. Codex usage comes from otel-hook and the rollout harvest.

Which sessions har control sync harvests

Only sessions the repository has evidence for: sessions with a binding in this repository (recorded locally or known to Mission Control), and sessions whose transcript or rollout was started inside an active slot’s work dir. A session that never touched HAR is not read.

When Mission Control is older than this attribution model (no /session-bindings or /usage-observations route), sync warns once and falls back to per-slot totals for sessions that ran inside a slot’s work dir.

Limitations

• Interleaved worktrees in one context. An agent alternating between two worktrees within a turn has each observation attributed to the worktree it most recently touched.
• Subagents launching slots. A Claude subagent’s commands carry its parent’s session id, so a launch run by a subagent binds the parent session; the subagent’s own file events move it to its slot. A Codex subagent is its own thread; HAR reports it as a subagent of the root thread.
• Shell-only work in Claude Code. Claude’s hook cwd is the session’s, so Bash-only work in a worktree is attributed through launch bindings and file events.
• Codex granularity. A Codex observation is a whole turn, timestamped at its last response; a turn that launches a slot is booked on that slot.
• Cursor reports no cost and keeps no offline record: its cost is estimated from model prices, and usage missed while otel-hook was not installed is lost. Its session env was confirmed from the IDE bundle, not at runtime.