map-tokenreport · diff
git:20260615.564a64f to git:20260618.1acadce
18 added, 4 removed. Audit A to A.
---
name: map-tokenreport
description: |
Show per-subtask/agent token accounting (input, output, cache read/creation, cost, cache-hit ratio) for the current branch. Use when asked for token usage, run cost, or a token report. Do NOT use to plan or run work; use map-efficient.
effort: low
disable-model-invocation: true
argument-hint: "[branch]"
---
# /map-tokenreport - Token Accounting Report
Purpose: surface how many tokens (and how much money) the current branch's MAP
run spent, attributed to the subtask, phase, and agent that spent them.
Read-only reporting — this skill does not plan, implement, or run quality
gates.
+ ## Constraints (NEVER)
+
+ This skill is strictly read-only reporting.
+
+ - **NEVER** edit code, state, or git from this skill — it renders an existing rollup and stops.
+ - **NEVER** run or resume a MAP workflow here; if work is needed, hand off to `/map-efficient`.
+ - **NEVER** present `est_cost_usd` as a billing source of truth — it is a per-model estimate.
+
The numbers come from the `map-token-meter` hook (wired on `SubagentStop` and
`Stop`), which reads each Claude Code transcript's per-turn `usage` block and
appends attributed rows to `.map/<branch>/token_log.jsonl`, rolled up into
`.map/<branch>/token_accounting.json`. This skill just renders that rollup.
## What it shows
- **input / output** tokens, **cache_read** (cheap cache hits) and
**cache_creation** (cache writes) — tracked separately because they bill at
very different rates.
- **est_cost_usd** — priced per model via `MODEL_TOKEN_PRICES` in
`map_step_runner.py` (estimate, not a billing source of truth).
- **cache_hit_ratio** = `cache_read / (input + cache_read)` — how well prompt
caching is paying off.
- Breakdowns by subtask, by agent (actor / monitor / research-agent /
orchestrator / ...), and by phase.
- **research_roi** — research-agent / Codex researcher tokens and cost compared
with downstream Actor/Monitor tokens, so users can see whether delegated
exploration paid for itself.
## Steps
### Step 1: Resolve the branch
```bash
BRANCH=$(git rev-parse --abbrev-ref HEAD | sed -E 's|/|-|g; s|[^a-zA-Z0-9_.-]|-|g; s|-{2,}|-|g; s|^-||; s|-$||')
```
If the user passed an explicit branch argument, use it instead of the detected
one.
### Step 2: Render the report
```bash
python3 .map/scripts/map_step_runner.py token_report "$BRANCH"
```
`token_report` re-reads `token_log.jsonl`, rebuilds `token_accounting.json`,
and prints a per-subtask table, a per-agent table, research ROI, and a TOTAL
line with the cache-hit ratio and estimated cost. It is idempotent and read-only
against your code.
### Step 3: Optional — inspect the raw rollup
For the per-phase split or per-subtask research ROI details, read the JSON
directly:
```bash
jq '{aggregate, by_agent, by_phase, research_roi}' ".map/${BRANCH}/token_accounting.json"
```
### Step 4: Summarize
- Report the totals (input / output / cache), the cache-hit ratio, the research
- token share, and the estimated cost. Call out the most expensive subtask or
- agent, a low cache-hit ratio, or research cost that looks high relative to
- Actor/Monitor. Then STOP — do not change code from this skill.
+ Emit exactly this report shape, then STOP (this skill never changes code):
+
+ ```text
+ Tokens (<branch>): in <I> / out <O> / cache_rd <CR> / cache_cr <CC>
+ Cache-hit ratio: <X>% · Est. cost: $<C> · Research share: <R>%
+ Most expensive: <subtask|agent> ($<amount>)
+ Flags: <low cache-hit | high research cost | none>
+ ```
+
+ Fill every field from the `token_report` output. If a value is unavailable, write `n/a` — never omit a line or invent a number.
## Examples
Report token usage for the current branch:
```text
/map-tokenreport
```
Report for a specific branch:
```text
/map-tokenreport feat/add-multiply
```
Typical output:
```text
Token accounting — feat-add-multiply (93 assistant turns)
subtask input output cache_rd cache_cr $cost
-------------------------------------------------------------------------
ST-001 183 150,879 14,085,841 792,780 47.31
-------------------------------------------------------------------------
TOTAL 183 150,879 14,085,841 792,780 47.31
cache hit ratio: 100.0% est cost: $47.31
```
## Troubleshooting
- **`token_accounting.json` does not exist / report is empty.** The
`map-token-meter` hook has not fired for this branch yet. It records on
`SubagentStop` and `Stop`, so a brand-new branch with no completed turns has
nothing to show. Confirm the hook is wired in `.claude/settings.json`
(`SubagentStop` and `Stop` entries) and that `.map/<branch>/token_log.jsonl`
exists.
- **Everything is attributed to `orchestrator` / `unattributed`.** Those turns
ran outside a MAP subtask (e.g. direct edits, or before `step_state.json` had
a `current_subtask_id`). Per-subtask attribution appears once a
`/map-efficient` run sets the active subtask.
- **Cost looks high relative to raw input.** That is expected when most input
is cached: `cache_creation` is the dominant cost on cache-heavy runs. Compare
`cache_creation` vs `cache_read` and the cache-hit ratio to see where the
spend goes.
- **Unknown model in cost estimate.** `MODEL_TOKEN_PRICES` falls back to the
default model price for unrecognized model ids; update that table in
`map_step_runner.py` when a new model ships.