Traceway

Drive a Traceway instance from the terminal with the traceway CLI. The first word of the argument decides the flow:

The CLI is under active development. If a flag documented here does not appear in traceway <command> --help, trust the binary. If a traceway MCP server is connected, prefer its tools over shelling out to the CLI: they wrap the same API with the same semantics, and this skill's knowledge is available as its resources. The server is this same binary (traceway mcp).

Ground Rules (All Flows)

• Reads are safe: any list / show / query subcommand may run freely; they never mutate server state.
• Writes require explicit user instruction: exceptions archive / unarchive are the only mutating data commands; only run them when the user asks by name, with --yes in non-interactive contexts. "Look at this error" means read it, not archive it.
• Output: piped output defaults to JSON (table on a TTY). Prefer JSON + jq, and --fields a,b,c to trim responses. Keep --page-size at 10 to 20 for triage.
• Time windows: always bound queries, default --since 1h for "now" questions, --since 24h otherwise. --since accepts s, m, h, lowercase Nd (no 1w, no 7d2h). Absolute windows via --from / --to (RFC3339).
• Exit codes: 0 ok, 1 generic/API, 2 usage, 3 connection, 4 auth, 5 not found, 6 rate limited, 7 server 5xx. Errors emit {"error":"<stable_id>","message":"...","hint":"...","exit_code":N} on stderr; branch on the error field.
• On exit code 4 (auth), do not run traceway login yourself; switch to the Login flow and let the user enter credentials.
• Never guess a fix: if the telemetry does not explain the failure, do not change behaviour on a hunch. Add the instrumentation that would explain it next time, and say what is still unknown. See "When the evidence is not enough" in the Debug flow.

Resolving Dashboard URLs

Users paste dashboard URLs (https://<instance>/<route>) as references in any flow. Resolve by route family:

Time window: most dashboard URLs carry ?preset=<p> or ?from=<iso>&to=<iso> (sticky across pages); honor them instead of the default window.

• preset values 5m 30m 60m 3h 6h 12h 24h 3d 7d map directly to --since; the CLI has no month unit, so map 1M to --since 30d and 3M to --since 90d.
• from/to are ISO timestamps; pass via --from/--to, appending Z (or the correct offset) when missing, since the CLI requires RFC3339.
• No time params means the page was on its default; pick --since per the ground rules.

preset/from/to set the window for list/group views. Detail URLs additionally carry ?t=<iso> — the single record's timestamp, URL-encoded. That t value is exactly what the by-id commands need as --recorded-at (or --started-at for sessions). See "Fast by-id lookups" next.

Fast by-id lookups (always pass the timestamp)

The by-id detail commands — exceptions occurrence, endpoints show, tasks show, ai-traces show, sessions show, traces show — require the record's timestamp (--recorded-at, or --started-at for sessions). Telemetry tables are partitioned by day: with the timestamp the lookup is bounded to a small window and ClickHouse prunes to a few partitions; without it the server scans every partition (slow cold load). The flag is mandatory for exactly this reason — never omit it. It can be approximate (within ±24h), and you can recover or estimate it when it isn't handed to you; see "When you don't have the timestamp" below.

Where the timestamp comes from, in order of preference:

1. A dashboard URL — the ?t=<iso> param is the record's recordedAt; URL-decode it and pass it verbatim. (Sessions have no t; use the session start, from=, or a linked occurrence's recordedAt.)
2. A list/group you already fetched — every exceptions show occurrence carries recordedAt. Capture the id and its recordedAt together, then drill in.
3. A notification — see below.

Query order when you hold an id: resolve its recordedAt first (URL, group, or notification), then call the by-id command with it.

When you don't have the timestamp

The flag is required, so you must supply something — but it can be approximate. The lookup window is ±24h around what you pass (±48h for traces show), and if the record isn't in that window the server falls back to an unbounded scan. So a timestamp within a day of the truth stays fast; a wrong guess still returns the right record, just slower. Resolve it in this order:

1. Recover it from an API. For an occurrence whose hash you know (e.g. /issues/<hash>/<occurrenceId> pasted without ?t=), run traceway exceptions show <hash> and read that occurrence's recordedAt — the hash endpoint needs no timestamp. A group's firstSeen/lastSeen from exceptions list bound when its occurrences happened (lastSeen ≈ the most recent one).
2. Estimate from context. A notification's send time, the issue's firstSeen/lastSeen, or the URL's preset/from window all put you inside ±24h — good enough for a fast lookup.
3. Ask the user. If nothing pins it down (e.g. a bare occurrence/endpoint id with no hash, no time, and no list to recover from), ask roughly when it happened — "around when did this fire? within a day is enough" — and pass that. Don't invent a placeholder like "now" when the issue is old; that defeats the pruning and can miss the ±24h window entirely.

Resolving a notification

Traceway notifications (email / Slack / webhook) come in two shapes. Read the body to tell them apart: an exception notification carries a Hash: and Exception ID:; a performance notification (latency, apdex, impact) names an endpoint and a threshold and has no hash. Webhook payloads also carry a ruleType field that names the rule exactly. Resolve each differently.

Exception notification

The body embeds everything for a direct, fast lookup. It contains:

• Hash: <16-hex> — the exception group → traceway exceptions show <hash>.
• Exception ID: <uuid> — the specific occurrence.
• Occurred at: 2006-01-02 15:04:05 UTC — the occurrence timestamp. Convert to RFC3339: replace the space with T and UTC with Z (→ 2006-01-02T15:04:05Z).
• View details: /issues/<hash> — the deep link.

So from a notification, go straight to the occurrence (fast), then pivot reusing the same timestamp:

traceway exceptions occurrence <Exception ID> --recorded-at <Occurred at → RFC3339> --output json
# the result carries distributedTraceId and sessionId → traces show / sessions show below

Performance notification (an endpoint became slow or critical)

These fire from latency/throughput rules, not from an error. There is no hash, no exception id, no occurrence to fetch. The webhook ruleType (or the subject wording) tells you which:

What the body gives you: the endpoint name (parse it from the subject/body; it is not a separate field), the metric/percentile, the current value, the threshold, and the lookback window. The /endpoints?from=&to= link is only a ~3-minute window around the fire time, so treat the fire time as the onset hint, not the incident window.

To resolve, hand straight to the Performance flow scoped to that endpoint, using the fire time as the time anchor:

# 1. confirm the alert against current stats for that endpoint
traceway endpoints list --search "<endpoint>" --since 1h --output json | jq '.data[0] | {p50, p95, p99, count, impact, impactReason}'
# 2. is this an accepted baseline? check whether an operator marked it slow
traceway endpoints slow "<endpoint>" --output json    # {offsetMs, reason}; offsetMs 0 = not marked
# 3. find when it crossed the threshold, then a slow trace (see Flow: Performance)
traceway endpoints chart --metric-type p95 --since 6h --interval-minutes 15 --output json

Step 2 matters: an impact_score_* or apdex_drop alert is already offset-aware, but endpoint_p95_threshold / endpoint_p99_threshold fire on raw latency regardless of any slow-marking, so a marked-slow endpoint can alert on latency its operators have already accepted. See "Account for user-marked slow endpoints" in performance.md.

Flow: Login

1. Check for an existing install

traceway version

If it prints a version, skip to authentication.

2. Install if missing

On POSIX systems (Linux, macOS), use the install script:

curl -fsSL https://cli.tracewayapp.com/install.sh | sh

If the script cannot be used, install manually: prebuilt binaries are on the tracewayapp/traceway releases page under cli/vX.Y.Z tags (the latest release may be a Backend release, so filter for CLI tags):

OS=$(uname -s | tr '[:upper:]' '[:lower:]')
ARCH=$(uname -m); [ "$ARCH" = "aarch64" ] && ARCH=arm64
URL=$(curl -s "https://api.github.com/repos/tracewayapp/traceway/releases?per_page=20" \
  | grep -o "https://[^\"]*traceway_[^\"]*_${OS}_${ARCH}\.tar\.gz" | head -1)
TMP=$(mktemp -d)
curl -sL "$URL" | tar -xz -C "$TMP"
install -m 755 "$TMP/traceway" ~/.local/bin/traceway && rm -rf "$TMP"

Make sure ~/.local/bin is on PATH (or install to /usr/local/bin). Fallback, build from source (requires Go):

git clone https://github.com/tracewayapp/traceway && cd traceway/cli
go build -o bin/traceway ./cmd/traceway && install -m 755 bin/traceway ~/.local/bin/traceway

Verify with traceway version.

3. Authenticate

Login prompts for the password interactively, so ask the user to run it themselves (in Claude Code, suggest typing ! traceway login --url https://<instance> so the output lands in the session):

traceway login --url https://<traceway-instance>

Non-interactive alternative when the password is in a secret store (never echo a password into the command line or shell history):

printf '%s' "$TRACEWAY_PASSWORD" | traceway login --url https://<instance> --username you@example.com --password-stdin

Multiple instances or accounts coexist via profiles: traceway login --url ... --profile work, then traceway profiles list / traceway profiles use work.

4. Select a project and smoke-check

traceway projects list
traceway projects use <project-id>
traceway exceptions list --since 24h

The selected project is used implicitly by all subsequent commands.

Flow: Debug

/traceway debug issue X or /traceway debug <free-form bug description>.

1. Resolve the issue reference

X can be several things; resolve it to an exception hash (16 hex chars):

When a search returns multiple groups, show a shortlist (hash, count, lastSeen, first stack line) and ask the user which one before drilling in.

traceway exceptions list --since 7d --search "checkout" --output json \
  | jq '.data[]? | {hash: .exceptionHash, count, lastSeen, top: (.stackTrace | split("\n")[0])}'

2. Drill into the issue

traceway exceptions show <hash>

This is the high-value call: full stack trace, occurrence list with recordedAt, attributes (user IDs, app versions, request context), and optional distributedTraceId / sessionId per occurrence. firstSeen correlates with deploys: a group that first appeared right after a release points at that release's diff. A bogus hash exits 5 with not_found; fall back to search.

When the user gave an issue URL (or hash), fix the LAST occurrence — not "the group". A single hash can bundle several distinct errors: the hash is computed from a normalized stack trace with the message stripped, so two unrelated failures that share their top frames (e.g. both captured at the same middleware/recovery frame) collapse into one group. The group's representative stack trace and firstSeen may belong to a different, now-dormant error than the one the user is looking at. Anchor on the most recent occurrence and fix that specific failure path:

# The occurrence the user actually wants: the latest one. Pin its exact message + attributes + trace.
traceway exceptions show <hash> --output json \
  | jq '.occurrences | sort_by(.recordedAt) | last
        | {recordedAt, message: (.stackTrace | split("\n")[0]), attributes, traceId, distributedTraceId}'

# Then confirm whether the group is homogeneous or mixed — distinct first lines = distinct bugs:
traceway exceptions show <hash> --output json \
  | jq -r '[.occurrences[].stackTrace | split("\n")[0]] | group_by(.)
           | map("\(length)x \(.[0])") | .[]'

If the group is mixed, scope the fix to the last occurrence's error only; mention the other clusters in the report but do not fix them unless asked. The last occurrence's message and attributes (not the group's representative stack) define the failure to reproduce and fix — and its recordedAt is the time hint to pass to the by-id traces show / sessions show lookups below.

3. Triage and correlate (also the entry point for free-form bug descriptions)

From the description extract symptom, affected endpoint/feature, and time window, then read several signals before forming a hypothesis:

traceway exceptions list --since 24h --order-by lastSeen        # what is erroring (firstSeen for regressions, count for volume)
traceway logs query --since 24h --min-severity 17               # errors and worse
traceway logs query --since 24h --search "payment declined"     # search log bodies
traceway logs query --since 24h --service checkout-api --min-severity 13
traceway endpoints list --since 24h --search "checkout"         # latency p50/p95/p99 and error counts, --order-by impact|count|p95|lastSeen

Severity is an OTel number, not a name: 1 TRACE, 5 DEBUG, 9 INFO, 13 WARN, 17 ERROR, 21 FATAL. The flag is --min-severity 17, never --severity error.

Correlate by trace: when an occurrence or log line carries a trace ID, pull the whole request timeline; this is usually the fastest route to a root cause:

traceway exceptions show $HASH --output json | jq -r '.occurrences[0].distributedTraceId' \
  | xargs -I{} traceway logs query --trace-id {} --output json

Pull the whole cross-service trace and the user's session, reusing the occurrence's recordedAt as the (mandatory) time hint so both lookups stay partition-bounded:

OCC=$(traceway exceptions show $HASH --output json | jq -c '.occurrences[0]')
TS=$(jq -r '.recordedAt' <<<"$OCC")
DT=$(jq -r '.distributedTraceId // empty' <<<"$OCC")
SID=$(jq -r '.sessionId // empty' <<<"$OCC")
[ -n "$DT" ]  && traceway traces show "$DT" --recorded-at "$TS"      # every endpoint/task/ai-trace/exception node across services
[ -n "$SID" ] && traceway sessions show "$SID" --started-at "$TS"    # the session + the exceptions that fired in it

traces show is usually the single highest-value RCA call: it stitches one logical request together end to end across services.

Check metrics for systemic causes (spikes lining up with firstSeen suggest saturation rather than a code bug):

traceway metrics query --name system.cpu.utilization --aggregation max --since 24h
traceway metrics query --name <name> --aggregation avg|sum|count|min|max [--tag key=value] [--group-by <tag>]

The CLI also accepts p50|p95|p99, but the server has no quantile aggregation for metric points and silently computes avg for them — never present those as percentiles. Latency percentiles come from traceway endpoints list, computed from raw request durations. There is no metrics list; a bogus name returns an empty series: {} cleanly, so probing names is safe. Host metrics from the Traceway OTel Agent live under system.* names, and OTLP histogram metrics are stored as two series, <name>.avg and <name>.count.

4. Correlate with the code

1. Open the files and lines named in the stack trace and read the failing path.
2. If the issue started at a known time, check what shipped then: git log --since "<firstSeen>" --until "<firstSeen + 1h>" or the deploy history.
3. Form a hypothesis that explains the targeted occurrence's full observation set (its message, affected endpoint, timing, volume) — not just the first stack frame. When a URL/hash anchored you to the last occurrence (step 2), explain that error; do not stretch one hypothesis to cover sibling clusters that merely share the hash.
4. Propose or implement the fix per the user's instruction, scoped to the targeted occurrence's failure path.

5. When the evidence is not enough, instrument instead of guessing

Before writing a fix, answer one question honestly: does the telemetry actually explain this failure? You have enough when you can name the failing line, the input that reached it, and why that input was wrong. A stack trace on its own is usually not enough. Neither is a plausible story that the data neither confirms nor contradicts.

If you cannot explain it, do not ship a speculative fix. A guess that changes behaviour hides the symptom, destroys the next reproduction, and costs another incident cycle. Instead, add the instrumentation that would answer the open question, and let the next occurrence do the diagnosing.

Pick the signal that closes the specific gap:

Rules for what you add:

• Emit it inside the active span of the request or task. That is what stamps the trace id on the log and makes it show up on the trace.
• Put identifiers in attributes, never in the message text. Messages are stripped before grouping, and attributes are filterable.
• Keep cardinality low. User ids and request ids belong in span attributes, never in a metric tag or a span name.
• Never log secrets, tokens, or personal data.
• Ship it as permanent instrumentation, not a temporary debug print. Something you remove tomorrow answers nothing.
• If the area has no instrumentation at all, set it up first with the traceway-setup skill rather than bolting on one log line.

Then close the loop with the user. State what you could not determine, what you added, and the exact command to run when it fires again:

traceway logs query --trace-id <id> --output json
traceway exceptions show <hash> --output json | jq '.occurrences | last | .attributes'

Adding instrumentation is a legitimate outcome of a debug session, not a failure to fix. Report it as the deliverable it is.

6. Report and clean up

Summarize: symptom, evidence (exception hashes, log excerpts, metric anomalies), root cause, fix. Include traceway exceptions show <hash> references so the user can verify. After a fix is deployed and verified, archive only when the user asks:

traceway exceptions archive <hash> --yes

Flow: Performance

/traceway perf <endpoint or symptom>, or any request about slowness, high latency, p95/p99, or "why is X slow". Approach it as a senior performance engineer: quantify first, localize the cost to a span, confirm the cause against the checklist, separate code from saturation, then root cause. Reads only; never archive or mutate.

For the full methodology and the bottleneck checklist (Select N+1, missing index, chatty IPC, connection-pool exhaustion, GC pauses, and ~20 more, each tied to the Traceway signal that confirms it), read performance.md in this skill directory (over MCP: the traceway://knowledge/performance resource). It is the authoritative reference.

The loop, using the read commands documented in this skill:

1. Quantify and localize. traceway endpoints list --since 24h --order-by p95 (or impact). Read the latency shape first: p50 already high means every request pays it (systemic: query, index, algorithm); p99 much greater than p50 means a tail (contention, pool exhaustion, GC, retries, a flaky dependency). Note that impact/impactReason are offset-adjusted for marked-slow endpoints but the raw p50/p95/p99 are not.
2. Check the accepted baseline. Before calling an endpoint slow, run traceway endpoints slow "<endpoint>". A non-zero offsetMs means an operator accepted that much latency (with a reason), so only latency beyond the offset is a real regression; offsetMs 0 means default thresholds apply. Skip this only for a clearly systemic, multi-endpoint slowdown.
3. Pinpoint when it started: adjust the window down. Do not investigate the default window blindly; narrow it until it brackets the onset. traceway endpoints chart --metric-type p95 --interval-minutes <n> returns latency over time for the top endpoints ({timestamp, endpoint, value} in ms): read down one endpoint's buckets for the step where p95 jumps. Confirm the cause with traceway metrics query --name <metric> --interval-minutes <n> (the infra/runtime curve). For a specific route not in the top 5, bisect traceway endpoints list --search <name> --from <a> --to <b> over adjacent windows. See performance.md, "Pinpointing when the slowness started".
4. Get a representative slow trace. traceway endpoints show <endpointId> --recorded-at <t> for one request's span waterfall, or traceway traces show <distributedTraceId> --recorded-at <t> for the cross-service timeline. Take the trace from inside the slow window you just found; find the long pole.
5. Match the long pole to the checklist. Is the dominant span a database call, an external call, in-process compute, or a gap before work starts (queueing / lock wait / pool exhaustion)? Look it up in performance.md.
6. Separate code from saturation. Pull infra/runtime metrics over the same window: traceway metrics query --name system.cpu.utilization --aggregation max, then mem.used, go.gc_pause. A spike that lines up with the latency onset points at saturation, not a code bug.
7. Correlate with code and deploys. With the onset time from step 3, check what shipped then: git log --since '<onset - 30m>' --until '<onset + 30m>' or the deploy history. A jump at a release is a regression; a gradual ramp with no deploy is data growth (N+1, missing index, no pagination).
8. Report. The bottleneck, the evidence (endpoint percentiles, the onset time, the dominating span, any correlated metric anomaly), the root cause mapped to a checklist item, and a concrete fix, with the endpoints show / traces show references so the user can verify. If the endpoint is marked slow, state whether the latency exceeded the accepted offset.

If the waterfall does not localize the cost, stop before proposing a fix. An unexplained gap, one fat span covering the whole handler, or a tail you cannot reproduce all mean the same thing: the trace is too coarse to act on. Add child spans around the suspect steps inside that handler, then wait for the next slow request to point at the real one. See "When the evidence is not enough" in the Debug flow.

Flow: Query

For free-form requests ("what's broken in prod?", "is /api/checkout slow?", "show errors for service X"), use the read commands directly. A request specifically about latency or slowness routes to the Performance flow above.

Command reference

The by-id show/occurrence commands take their id from a dashboard URL, a notification, or an exceptions show occurrence — and require the record's timestamp (--recorded-at / --started-at); see "Fast by-id lookups" above.

Not implemented yet (do not fabricate flags; point the user at the web UI): list verbs for tasks / sessions / ai-traces / traces (only by-id show exists for those), and metrics list/discover.

Recipes

# What's broken right now
traceway exceptions list --since 1h --order-by lastSeen --page-size 10 --output json \
  | jq '.data[]? | {hash: .exceptionHash, count, lastSeen}'

# Did anything NEW break since a deploy at 13:00 UTC
traceway exceptions list --from 2026-06-11T13:00:00Z --to "$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
  --order-by firstSeen --output json \
  | jq '.data[]? | select(.firstSeen >= "2026-06-11T13:00:00Z") | {hash: .exceptionHash, firstSeen, count}'

# Worst endpoint by latency
traceway endpoints list --since 1h --order-by p95 --page-size 1 --output json | jq '.data[0]'

# Errors for one service (exceptions --search is free text, not a service filter; use logs)
traceway logs query --service checkout-api --min-severity 17 --since 1h --output json \
  | jq '.data[]? | {timestamp, body, traceId}'

Empty results (data: null or data: []) are not errors: widen the window, re-check the active project (traceway projects list), and if the app was never connected to Traceway, set it up first (the traceway-setup skill).