imbue-mngr-usage

mngr usage -- agent-agnostic CLI for rolling-window usage / quota data.

What it does

Provides a single mngr usage command that surfaces rolling-window usage data in human / json / jsonl / format-template output. The command itself knows nothing about any specific agent type or provider; it walks events files on disk and renders whatever it finds.

Architecture

This package contains:

• The mngr usage CLI command and its rendering helpers.
• The UsageSnapshot, WindowSnapshot, and CostSnapshot data types.

Discovery is by path convention. The CLI walks <host_dir>/agents/*/events/<source>/usage/events.jsonl (the same shape mngr transcript uses for events/<source>/common_transcript/...), scans every event line, and aggregates per <source>: rate-limit windows collapse freshest-wins (the account-level counter), while cost is grouped per session_id and filtered to a recency window (--since, default 24h). The <source> segment is free-form -- whatever the writer plugin chose.

When multiple writers contribute, each renders as its own [source] section in human output and as an entry in the JSON sources array.

Destroyed agents

A destroyed agent's usage still counts. Before an agent's (or its whole host's) state directory is deleted, its events/<source>/usage directories are copied to <local_host_dir>/preserved/<agent-name>--<agent-id>/; for remote agents the files are pulled to the local machine so they survive host destruction. This is on by default and controlled by the preserve_on_destroy option on the usage plugin config (set it to false to discard usage on destroy).

mngr usage reads these preserved files back by default, so a destroyed agent's accumulated cost and rate-limit history still counts toward the totals. Pass --no-preserved (on mngr usage and mngr usage wait) to consider only live agents.

Filtering by event age

Pass --since DURATION (e.g. --since 1h, --since 7d) on mngr usage or mngr usage wait to restrict the per-session cost aggregation by age. Sessions whose last event is older than --since are dropped from sessions[] and from the per-mode aggregates (subscription_cost.* / api_cost.*) computed off them. Default is 24h, configurable via the since_seconds option on the usage plugin config.

--since only shapes the cost surface. Rate-limit windows always reflect the freshest reading across all agents -- they track an account-level counter, so there's no older reading to drop.

--stale-after is not an age filter: it's only the stale-warning threshold for the snapshot file (controls whether the human output prints a "snapshot last updated X ago" warning). It does not change which events are aggregated.

Output formats

• mngr usage (human summary: per-mode cost line(s) + window lines)
• mngr usage --detail (human + per-session breakdown lines)
• mngr usage --format json (summary JSON: per-mode aggregates subscription_cost and api_cost, session_count plus per-mode counts, windows)
• mngr usage --format json --detail (JSON with sessions[] per source; each session carries a cost_mode tag)
• mngr usage --format jsonl
• mngr usage --format '5h:{five_hour.used_percentage}%/{seven_day.used_percentage}%'

The --detail flag is independent of --verbose (which controls log level). It toggles only the per-session breakdown surfaces.

Waiting on a predicate

mngr usage wait --until <CEL> blocks until at least one source's CEL context satisfies every --until expression, then exits 0. Composable with shell:

mngr usage wait --until 'five_hour.elapsed_percentage > 75 && five_hour.used_percentage < 50' \
  && mngr message my-agent "ok, kick off the next batch"

The CEL context per source mirrors one entry of mngr usage --format json's sources array. Each window exposes the writer-emitted fields (used_percentage, resets_at, window_seconds, label, ...) plus the reader-derived seconds_until_reset, elapsed_seconds, and elapsed_percentage (the last two require window_seconds from the writer; absent on variable-duration windows like Claude's overage). Source-level fields are exposed too. Cost is split by auth mode and never lumped: subscription_cost.* aggregates sessions whose Claude Code process was on a Claude.ai Pro/Max subscription (cost is imputed by Claude Code -- the user actually pays a flat subscription), and api_cost.* aggregates sessions whose process was on a direct ANTHROPIC_API_KEY (cost is real billable spend). Pick the mode you care about, e.g. api_cost.total_cost_usd > 20.0 means "I've spent more than $20 of real API money across recent sessions". session_count is the total across both modes; subscription_session_count and api_session_count break that down per mode. since_seconds and sessions[] (every session in the window, newest-first, each carrying a cost_mode of "SUBSCRIPTION" or "API_KEY") are available as well. Use --since to tighten or widen the aggregation window from its default (24h). For a per-session predicate, index sessions[] directly (e.g. sessions[0].cost.total_cost_usd > 5 for the most recent session's own contribution).

Exit codes mirror mngr wait: 0 matched, 1 error, 2 timeout. Default poll interval is 30s; use --interval for tighter cadence. To restrict matching to a specific writer, use the top-level source field in CEL (e.g. --until 'source == "claude" && five_hour.used_percentage < 50').

Polling from cron (check mode)

For recurring automation, let cron own the cadence: poll the plain mngr usage --format json snapshot on a schedule and branch in the shell. See cron automation recipes for worked examples: using up an about-to-expire 5h window (with a weekly pace check), warming a fresh window the moment the last one elapses, and working through a queue of task files a couple at a time.

Implementing a writer plugin

A writer plugin is responsible for producing cost_snapshot events at the conventional path. The minimal contract is just the JSONL line shape:

{"source":"<your-source>/usage","type":"cost_snapshot","event_id":"evt-<hex>","timestamp":"<ISO 8601>","session_id":"<uuid>","cost":{"total_cost_usd":<float>},"rate_limits":{"<window-key>":{"used_percentage":<float>,"resets_at":<unix-ts>}}}

Every event must carry a session_id (non-empty string) plus at least one of rate_limits or cost. Events missing session_id are dropped with a WARNING log naming the source and event_id -- so if you're writing a usage plugin and don't see your data in mngr usage, check the log first. Any plausible statusline source has a session identifier handy, so this is rarely a constraint; if a future tool doesn't, synthesize a stable per-session id on the writer side. The type field is informational -- the reader doesn't validate it -- but "cost_snapshot" is the current convention.

Append one line per refresh to:

<agent_state_dir>/events/<your-source>/usage/events.jsonl

mngr usage will pick it up automatically -- no plugin registration with this package required.

The writer chooses both the window keys and (optionally) per-window labels. Keys are used by format templates ({<key>.used_percentage}) and should be identifier-safe if you want format-template support; the per-window label controls human display (e.g. "5h" vs the literal key "five_hour"). Render order is the writer's insertion order in the JSONL.

Writers may also include "window_seconds": <int> per window to declare a fixed window duration. When present, mngr usage derives elapsed_seconds and elapsed_percentage, which mngr usage wait CEL predicates can use to express "75% of the window has elapsed" without callers hardcoding window durations. Omit window_seconds for windows without a fixed length (e.g. Claude's overage indicator).