headroom/docs/metrics-technical-guide.md
Tejas Chopra 250ede2f7f
fix(reporting): show net vs gross savings, real skip thresholds, and the effective profile (#3123)
## Description

Six reporting/config defects found while investigating a user reporting
~1% savings on Claude Code. **None of these changes how much Headroom
compresses** — all of them change whether an operator can tell what it
did. Every one was found by reading that user's own 227,777 lines of
proxy logs against the code.

## Changes Made

- **`perf/analyzer`: parse and render `tok_inflated`.** Every PERF line
carried it; nothing downstream read it. The report could print
`321,239,562 -> 313,274,727` directly above `8,455,763 saved` — two
figures that differ by exactly the 490,928 tokens of inflation it
omitted.
- **`content_router`: report the real skip thresholds.** The routing
summary hardcoded `skipped (<50 words)` regardless of what was in force.
Wrong number (the message gate is `min_tokens`, 10–250 by profile),
wrong unit (tokens and characters, never words), and it merged two
different gates under one label.
- **`perf/analyzer`: disclose that Transform Effectiveness is partial.**
It is built only from `pipeline.py`'s `Transform NAME:` lines.
`compression_units.py` / `compression_batches.py` contain zero logging
calls, so the table read `content_router: 189,783 saved` against a PERF
total 44x larger. Reports the divergence rather than a coverage ratio —
the two are different populations and neither contains the other (those
lines carry no request_id, fire per stage, and are emitted before the
forwarder decides).
- **`perf/analyzer`: disclose the routing denominator.** Percentages
were taken over 4 of the router's 17 outcome buckets, silently dropping
buckets larger than several it displayed.
- **`savings_tracker`: stop dropping tool-schema dollars.**
`estimate_request_savings_usd` prices four buckets; `record_request`
read three. `tool_schema` was computed and discarded, so a quarter of
the token headline never reached "Cost saved". The two inputs are
disjoint (verified at the call site), so this is additive, not
double-counting.
- **`agent_savings`: an unknown profile no longer degrades to
`balanced`.** `balanced` is a different product posture from the default
`coding`: cache→token mode, dedup off, tool-search off, user messages
uncompressed, message floor 25x higher, block floor 20x higher. A typo
in `HEADROOM_SAVINGS_PROFILE` silently reconfigured the whole proxy. Now
degrades to `DEFAULT_PROFILE` and names the resolved profile in the
warning.
- **`agent_savings`: give `min_chars_for_block` a config-object path.**
Every other router pipeline kwarg travels on the config object; this one
alone was env-only, so an unseeded proxy applied every sibling `coding`
knob while this floor stayed at 500 instead of 25.
- **`server`: log the resolved compression posture at startup**, reading
cross-turn dedup off the constructed router rather than the environment
(the router resolves it as `config OR env`, so reading env alone would
be a guess).

## Testing

- [x] Unit tests pass, [x] ruff, [x] mypy, [x] new tests added

```text
uv run pytest tests/ -k "content_router or agent_savings or perf or analyzer or savings or proxy_server or cli_perf or prometheus"
620 passed, 25 skipped
uv run mypy headroom  # Success
uv run ruff check . && ruff format --check .  # clean
```

## Real behavior proof

- **Setup:** macOS arm64, Python 3.12, this branch. Input: 60 MB /
227,777 lines of real proxy logs from the reporting user (6 rotated
files, 2,792 PERF lines, 2026-08-17 → 2026-08-19).
- **Steps:** pointed `headroom.perf.analyzer.LOG_DIR` at that directory
and rendered the report before and after the patch.
- **After-fix output (real data, unmodified):**

```text
Requests:     2792
Tokens:       321,288,161 -> 313,323,326 (2.6% messages)
Tokens saved: 11,158,901 (3.4% reduction)
  · inflated      490,928 (net message reduction 7,964,835)
  · messages       8,455,763
  · tool schemas   2,703,138
  ! stage-level total 190,641 != PERF message total 8,455,763 — this table sees only
    engines that emit a Transform line, counts per stage, and does not check whether
    the mutation shipped
  Skipped:     44641 (77%) — below size floor
  (shares are of these 4 buckets only, n=58319; see `[router] route_counts=` for the
   full outcome space)
```

The arithmetic now closes on the page: `8,455,763 - 490,928 =
7,964,835`, matching the token delta exactly. Before the patch none of
the three annotated lines existed and the `Skipped` line claimed `<50
words`.

- **Profile resolution verified by execution**, not inspection —
subprocesses with controlled env:

```text
vanilla (nothing set)          mode=cache dedupe=1 tool_search=1 min_tokens=10  min_chars=25
HEADROOM_SAVINGS_PROFILE=coding  mode=cache dedupe=1 tool_search=1 min_tokens=10  min_chars=25
unknown profile name (before)  mode=token dedupe=0 tool_search=0 min_tokens=250 min_chars=500
unknown profile name (after)   -> resolves to `coding`, warning names it
coding, seeding never runs     min_chars=25 (was 500 before this patch)
```

- **Not tested:** live paid Anthropic traffic. These are
reporting/config surfaces; the wire path is untouched by this PR.

## Review readiness

- [x] Self-reviewed. Three overclaims in my own first draft were
corrected before this PR: a false subset claim in the Transform
Effectiveness note, a comment asserting `min_chars_for_block` was the
*only* env-only field (it is the only env-only *router pipeline kwarg*;
`cross_turn_dedup`, `tool_search`, `protect_reads`, `code_aware`,
`effort_router`, `lossless` remain env-only via a different mechanism
and are **not** fixed here), and a money-path expression that relied on
`a + b if c else d` grouping.

## Known remaining (deliberately out of scope)

- `Requests: N` still overcounts: the Codex WS forwarder reuses one
`request_id` across every turn (one observed 156x), plus ~18 duplicate
PERF emissions.
- `compression_units.py` / `compression_batches.py` remain unlogged —
this PR *discloses* the blind spot rather than closing it.
- The headline stays **gross**. True net is `11,158,901 - 490,928 =
10,667,973` (3.3%, not 3.4%). Making net the headline lowers every
user's reported savings ~4.4%; that is a product call, not mine, so the
inflation is surfaced beside it instead.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Tejas Chopra <tejas@Tejass-MacBook-Pro.local>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
2026-08-19 00:16:37 -07:00

12 KiB
Raw Blame History

Headroom Metrics — Dashboard Guide

What each metric shows, so you can build panels against it.

Two endpoints. Both are on the proxy (default :8787).

Surface How to get it Use it for
PrometheusGET /metrics Always on, no config Everything below. Start here.
OpenTelemetry — OTLP/HTTP HEADROOM_OTEL_METRICS_ENABLED=1 + pip install "headroom-ai[proxy,otel]" Same data, dotted names, plus per-tenant labels

Names differ between them: Prometheus uses headroom_tokens_saved_total (milliseconds for timings), OTel uses headroom.proxy.tokens.saved (seconds). Both are listed below.


The savings panel — start here

headroom.proxy.tokens.saved is the headline number. It already combines compression + tool-schema deferral — no need to add anything to it.

Metric What it shows
headroom.proxy.tokens.saved (OTel) Total input tokens Headroom kept out of the request. Compression + tool savings, combined. This is your hero number.
headroom.proxy.savings.usd{source} (OTel) Dollars saved, split by layer: compression, tool_schema, output_shaping, provider_cache. Sum for the total.
headroom_persistent_savings_tokens_saved_total Same tokens-saved number, but survives proxy restarts. Use for "lifetime saved" tiles.
headroom_persistent_savings_compression_savings_usd_total Lifetime dollars saved, durable across restarts.
headroom_tokens_input_total Input tokens actually sent upstream (post-compression). The denominator for a reduction %.
headroom_tokens_output_total Output tokens returned by the provider.
# Hero tile: tokens saved per second
rate(headroom_tokens_saved_total[5m])
  + sum(rate(headroom_savings_attributed_tokens_total{source="tool_search",realized="true"}[5m]))

# Context reduction %
100 * rate(headroom_tokens_saved_total[5m])
    / clamp_min(rate(headroom_tokens_input_total[5m]) + rate(headroom_tokens_saved_total[5m]), 1)

# Lifetime tiles (survive restart)
headroom_persistent_savings_tokens_saved_total
headroom_persistent_savings_compression_savings_usd_total

One catch on the Prometheus side. headroom_tokens_saved_total is compression only — it leaves out tool-schema deferral. The OTel headroom.proxy.tokens.saved includes both. That's why the query above adds the tool_search term back in. On tool-heavy workloads the gap is large.


Latency panel

All Prometheus timings are in milliseconds, exposed as _sum / _count / _min / _max. Build means with rate(sum)/rate(count).

Metric What it shows
headroom_overhead_ms_* Latency Headroom itself adds. Handler entry → end of compression. Excludes the LLM call. This is the "what does this cost us" number.
headroom_latency_ms_* Total request duration, including the provider.
headroom_ttfb_ms_* Time to first byte from upstream. Streaming requests only.
headroom_stage_timing_ms_*{path,stage} Where time went inside the handler — compression_first_stage, upstream_connect, memory_context, etc.
headroom_transform_timing_ms_*{transform} Time per compression transform. Use to find a slow transform.
# Headroom's added overhead, mean ms
rate(headroom_overhead_ms_sum[5m]) / rate(headroom_overhead_ms_count[5m])

# End-to-end, mean ms
rate(headroom_latency_ms_sum[5m]) / rate(headroom_latency_ms_count[5m])

# Slowest stages
topk(5, rate(headroom_stage_timing_ms_sum[5m]) / rate(headroom_stage_timing_ms_count[5m]))

No percentiles are available. There are no histogram buckets on /metrics, and the OTel histograms ship with default buckets that put every request into one bucket, so histogram_quantile() returns nonsense. Means work fine. For real p95/p99 today, use the headroom perf CLI.

Also: divide each _sum by its own _count. Overhead and TTFB are only sampled when > 0, so their counts are smaller than the latency count.


Cache panel

Metric What it shows
headroom_provider_cache_hit_requests_total{provider} Requests that read from the provider's prompt cache.
headroom_provider_cache_requests_total{provider} Requests with any cache activity. The correct denominator for hit rate.
headroom_cache_read_tokens_total{provider} Tokens served from cache (the discounted ones).
headroom_cache_write_tokens_total{provider} Tokens written into cache (these carry a premium).
headroom_cache_write_ttl_tokens_total{provider,ttl} Cache writes split by TTL — 5m vs 1h.
headroom_uncached_input_tokens_total{provider} Input tokens that missed cache entirely.
headroom_cache_bust_total Requests where compression broke a cached prefix. Should stay near zero.
headroom_cache_miss_attribution_total{provider,reason} Why a cached prefix missed — ttl_expiry, prefix_change, unknown.
# Cache hit rate by provider
sum by (provider) (rate(headroom_provider_cache_hit_requests_total[5m]))
  / sum by (provider) (rate(headroom_provider_cache_requests_total[5m]))

# Compression breaking cache — alert if this rises
rate(headroom_cache_bust_total[5m])

Don't use headroom_requests_cached_total as a hit rate. It mixes the provider's prompt cache with Headroom's own response cache into one boolean, so it measures neither.


Traffic & health panel

Metric What it shows
headroom_requests_total Requests handled. Unlabelled.
headroom_requests_by_provider{provider} Traffic split by provider — anthropic, openai, gemini, bedrock
headroom_requests_by_model{model} Traffic split by model. Capped at 1024 distinct; overflow lands in model="other".
headroom_requests_failed_total Upstream 5xx errors.
headroom_requests_rate_limited_total Requests Headroom rejected via its own rate limiter (not upstream 429s).
headroom_compression_failed_total{reason} Compression failures — timeout or error. Fails open, so traffic keeps flowing but savings quietly stop. Worth an alert.
headroom_compression_quarantine_total{event} Compression disabled after repeated timeouts — activated, skipped, released.
headroom_inbound_requests_active In-flight requests, gauge. Counts all HTTP including /metrics.
headroom_active_ws_sessions Live Codex WebSocket sessions, gauge.
# Failure rate
rate(headroom_requests_failed_total[5m])
  / clamp_min(rate(headroom_requests_total[5m]) + rate(headroom_requests_failed_total[5m]), 1)

# Savings silently stopped
sum by (reason) (rate(headroom_compression_failed_total[5m]))

# Traffic mix
sum by (provider) (rate(headroom_requests_by_provider[5m]))

Anthropic subscription panel

Only if you're on an Anthropic OAuth/subscription plan. OTel only, gauges, no labels.

Metric What it shows
headroom.subscription.5h_utilization_pct How much of the 5-hour rate-limit window is used (0100).
headroom.subscription.7d_utilization_pct Same for the 7-day window.
headroom.subscription.5h_seconds_to_reset Seconds until the 5-hour window resets.
headroom.subscription.7d_seconds_to_reset Seconds until the 7-day window resets.
headroom.subscription.overage_usd Extra-usage credits consumed, in dollars.

Attribution — where savings came from

Metric What it shows
headroom_savings_attributed_tokens_total{source,realized} Tokens saved, broken out by named source. source="tool_search" is tool-schema deferral.
headroom_savings_attributed_usd_total{source,realized} Dollars saved by source. Gauge, can go negative — don't rate() it.
headroom_savings_attribution_events_total{source,realized} How often each source contributed.
headroom_waste_signal_tokens_total{signal} Wasteful patterns detected in the input — json_bloat, base64, repetition, reread… This is diagnosis, not savings.

These rows explain the headline total — they are never added to it.


Compression internals

Metric What it shows
headroom.compression.tokens.input (OTel) Tokens going into the compression pipeline.
headroom.compression.tokens.output (OTel) Tokens coming out.
headroom.compression.tokens.saved (OTel) The difference. Pipeline-level view of compression only.
headroom.compression.runs (OTel) Pipeline executions. Note: per pipeline run, not per request.
headroom.compression.pipeline.duration (OTel, seconds) How long the pipeline took.
headroom.compression.transforms{transform} (OTel) Which transforms fired. High cardinality — drop or aggregate at the collector.

Five things that will break a dashboard

  1. Only savings counters survive a restart. 55 of 60 Prometheus families reset to zero when the proxy restarts. Only headroom_persistent_savings_* is durable, and it needs HEADROOM_WORKSPACE_DIR on a persistent volume — otherwise it resets on every deploy.

  2. No percentiles anywhere. Use means. See the latency section.

  3. headroom_latency_ms measures differently for streaming. On streaming requests the timer starts after compression, so end-to-end is latency + overhead. On non-streaming it's just latency. Don't mix both in one panel.

  4. A 5xx erases its own savings. Requests that fail upstream are dropped from every savings and token counter. During a provider incident, savings rates look artificially clean while throughput falls.

  5. /metrics needs auth if you set a proxy token. With HEADROOM_PROXY_TOKEN set, any non-loopback scraper must send Authorization: Bearer <token>. Loopback is always exempt.


Metrics the docs mention that don't exist

If panels came back empty, this is probably why. These names appear in the published docs but not in the code:

headroom_compression_ratio · headroom_latency_seconds (and _bucket) · headroom_cache_hits_total · headroom_cache_misses_total · headroom_cost_usd_total · the mode="optimize" label on headroom_requests_total

The shipped examples/grafana/headroom-dashboard.json also filters every panel on pool and hook labels that no metric emits — the dropdowns will be permanently empty. Its metric names are otherwise correct.


Setup reference

# Prometheus — nothing to do, GET /metrics is always on

# OpenTelemetry
pip install "headroom-ai[proxy,otel]"
export HEADROOM_OTEL_METRICS_ENABLED=1
export HEADROOM_OTEL_METRICS_ENDPOINT=https://otel.corp.example/v1/metrics
export HEADROOM_OTEL_METRICS_HEADERS="authorization=Bearer XXX"
export HEADROOM_OTEL_RESOURCE_ATTRIBUTES="service.instance.id=$HOSTNAME"
Variable Default Notes
HEADROOM_OTEL_METRICS_ENABLED 0 Master switch
HEADROOM_OTEL_METRICS_EXPORTER otlp_http Or console. No gRPC exporter exists.
HEADROOM_OTEL_METRICS_ENDPOINT unset Passed verbatim — /v1/metrics is not appended
HEADROOM_OTEL_METRICS_HEADERS unset k=v,k2=v2
HEADROOM_OTEL_METRICS_EXPORT_INTERVAL_MS 10000
HEADROOM_OTEL_SERVICE_NAME headroom-proxy
HEADROOM_OTEL_RESOURCE_ATTRIBUTES unset Set service.instance.id here — Headroom doesn't, and replicas will collide

Verify with curl -s localhost:8787/stats | jq .otel.

Multi-tenant labels: register_otel_metric_attribute_provider() adds request-scoped attributes (tenant, team, cost centre) to every OTel datapoint. Max 16 attributes, 256 chars each.

Air-gapped deployments: HEADROOM_OFFLINE=1 disables all outbound traffic — the anonymous usage beacon (which is on by default), the update check, and model downloads.