analysing-har

name

analysing-har

description

Use when analysing or debugging a HAR (HTTP Archive) network capture. Triggers include "analyse har", "analyze har", "debug har", "wiretrail", a `.har` file appearing in the conversation or working tree, or questions about what's wrong in a capture, request storms, wasteful/duplicate API calls, retries, 4xx/5xx errors, slow requests, auth/JWT/token-refresh problems, rate limiting, redirects, searching/extracting fields from request or response bodies, comparing a capture against a baseline, or what differs between repeated calls. wiretrail is the recommended CLI — prefer it over manually grepping the HAR JSON or loading it into a browser.

Analysing HAR files with wiretrail

Overview

wiretrail is a fast, deterministic, agent-friendly CLI for post-mortem analysis of HAR (HTTP Archive) captures. It answers narrow, repeatable questions in a single command — storms, duplicates, retries, errors, slow calls, auth flows, diffs — with structured terminal output and a stable --json schema. It redacts secrets by default, so its output is safe to quote back to the user.

Reach for wiretrail instead of grepping the raw HAR JSON or describing the file by hand: it parses a 143 MB capture in ~0.5 s and each command runs in well under a second.

Source: https://github.com/johnneerdael/wiretrail · cargo install wiretrail

When to use

• A .har file appears in the conversation or working tree.
• The user says "analyse/analyze this HAR", "debug this HAR", "wiretrail", or hands you a network capture exported from Chrome/Edge DevTools, HTTP Toolkit, Charles, Proxyman, Fiddler, or mitmproxy.
• The user asks: "why is the app making so many requests at startup", "which calls are wasteful", "what's failing", "what's slow", "is the auth/token refresh broken", "are we getting rate-limited", "what differs between these repeated calls", "give me a curl to reproduce this".

Step 0: ensure wiretrail is installed

command -v wiretrail >/dev/null && wiretrail --version || cargo install wiretrail

If cargo is missing, install Rust first (curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh), or build from source: git clone https://github.com/johnneerdael/wiretrail && cd wiretrail && cargo build --release (binary at target/release/wiretrail).

Invocation basics

wiretrail <FILE.har> [COMMAND] [OPTIONS]

• No command → summary (the default).
• For "what's wrong with this capture?", run auto first — it prints the summary, ranks the likely problems, and inlines the relevant deeper analysis (errors, retries, auth, …) scoped to exactly where the trouble is, in one shot. Use summary when you just want the oriented overview; summary now ends with a ranked recommended next steps section naming the exact follow-up commands.
• Global options: --json (machine output), --top N (list size, default 10), --filter "<expr>", --config wiretrail.yaml, --unsafe-include-secrets.
• Filter language: host:api.foo.com status:>=400 method:POST path:*login* time:>1000ms has:req.header.authorization.
• Exit codes: 0 clean, 1 findings reported, 2 invalid HAR. (A 1 is normal — it just means the command found something.)
• Entry IDs (e000123) are stable across commands — pivot from a finding to show-entry/curl on the same id.

Redaction is on by default. Quote wiretrail output directly. Only add --unsafe-include-secrets when the user needs a replayable request (auth headers, tokens, URL secrets) — and warn them the output then contains live credentials.

The fast path (recommended)

For "analyse this HAR / what's wrong?", run auto — one command does the whole triage: summary + ranked recommendations + the relevant deeper analysis inlined, each scoped to the offending host/route, with the reproducing command printed above it. Then read down the report; everything actionable is already there.

wiretrail capture.har auto                      # summary + auto-drilled HIGH+MED findings
wiretrail capture.har auto --all                # also drill LOW findings
wiretrail capture.har auto --min-severity high  # only HIGH/CRITICAL

diagnose is the same ranked findings without the inlined drill-downs (a compact "what's wrong" list, each with a suggested follow-up command + evidence IDs). Use diagnose when you want the shortlist; auto when you want it already investigated.

The manual triage workflow

When you want to go deeper on one thread (or auto pointed you somewhere), work top-down; each step narrows the next. Don't dump every command — follow the findings.

1. Orient — summary (note totals, error count, biggest duplicate group, and the recommended next steps). validate first if you suspect a sanitized/partial capture (it reports timings/bodies/auth coverage and which commands are limited).
2. Group — subsystems (named integrations) or hosts (per-host latency/bytes/ errors). Answers "who is this app talking to, and how much?"
3. Wasteful traffic — duplicates (repeated calls), then diff on a suspicious group to see what varies (volatile-only = truly redundant; meaningful = real differences). storms for time-clustered bursts; pagination for loops / N+1; rate-limit for 429 handling.
4. Failures — errors (4xx/5xx grouped, with parsed message/code/correlation), retries (failures with backoff), cascade (first failure + downstream blast radius), transitions (401→200, 429→429, 5xx→2xx), redirects (chains/storms), slowest (timing-phase breakdown + bottleneck). startup for boot-time concurrency + critical path.
5. Auth — auth (failures, token rotation, refresh flows) and jwt (decode tokens, expiry/skew — sub hashed, never the raw token).
6. Dig into bodies — search "<pattern>" (grep request/response bodies, --regex /--ignore-case, redaction-safe snippets), extract '$.path' (pull a JSON-path field across entries). export --format ndjson|csv to pull metadata into jq/DuckDB.
7. Regression / governance — compare <baseline.har> (diff vs a known-good run; --fail-on <severity> gates CI), rules --pack auth,security,… (enforce conventions / config rules), checks (required headers + content-type).
8. Hand off / reproduce — handoff (backend blocks: correlation IDs + sanitized curl), report (markdown dossier), curl <id> (one replay; add --unsafe-include-secrets if they need to actually run it), show-entry <id> (full redacted detail).

Command reference

Command	Use when you want…
auto	the whole triage in one shot — summary + ranked findings, each drilled inline and scoped (--all, --min-severity). Start here for "what's wrong?".
diagnose	a compact ranked "what's wrong" shortlist (no inlined drill-downs), each with evidence IDs + a suggested command.
summary (default)	the one-screen overview + hints + ranked recommended next steps.
validate	to check capture quality/sufficiency (timings/bodies/auth coverage; sanitized? anomalies) before trusting findings.
hosts	per-host request count, methods, status mix, p50/p95/max latency, bytes, dup count.
subsystems	hosts grouped into named integrations (vendor heuristics + wiretrail.yaml).
endpoints	a normalized API catalog (method, {id} path, statuses, content types).
timeline	a chronological view (raise --top, e.g. --top 5000, for the full list).
duplicates	repeated calls grouped by fingerprint; retry-pattern flagged.
retries	repeats following a failure, with trigger statuses + backoff gaps.
storms	bursts to one host/endpoint in a window (--window-ms, --min-count).
pagination	pagination loops + N+1 fan-out (--max-pages, --fanout-min).
rate-limit	429s, Retry-After, X-RateLimit-*, cooldown violations.
errors	4xx/5xx grouped by endpoint+status, with message/code/correlation/body.
redirects	redirect chains/storms, cross-host hops.
transitions	status sequences (auth-recovered, rate-limit-persisted, recovered-5xx).
slowest	top-N slow calls + timing-phase breakdown + bottleneck label.
startup	boot/startup profile: max concurrency, critical path, slow dependencies.
cascade	the earliest failure and the downstream failures it triggered.
jwt	decode JWTs (claims/exp/skew), redacted (hashed sub, no signature).
auth	401/403 patterns, inconsistent auth, token rotation, refresh-flow problems.
handoff	backend hand-off blocks for failed + slowest requests.
show-entry <id>	the full redacted request/response/timings for one entry.
diff	what query/headers/body vary across repeated calls to one endpoint.
search <pattern>	grep request/response bodies (--regex, --ignore-case); redaction-safe snippets.
extract <jsonpath>	pull a JSON-path field ($.errors[0].code, [*]) from bodies; --target req|resp.
export	flatten entries to NDJSON/CSV (--format) for jq/DuckDB/spreadsheets (metadata only).
compare <baseline.har>	regression diff vs a baseline: new errors, latency/payload regressions; --fail-on gates CI.
rules	enforce conventions: wiretrail.yaml rules + built-in --pack auth,caching,payments,security,rest,graphql.
checks	required-header rules (config) + content-type mismatches.
report	a shareable markdown dossier.
curl [id]	sanitized, safety-labeled replay command(s).

Worked example — an app-startup capture

A 353-request startup capture. wiretrail capture.har auto would do this entire walkthrough in one command — below is what it surfaces (and how to reproduce each piece by hand). summary immediately points the way:

$ wiretrail capture.har summary
...
hints:
  - 29x duplicate calls: POST youtubei.googleapis.com /youtubei/v1/visitor_id prettyPrint=false
  - 18 error responses (4xx/5xx/failed)
next useful commands: duplicates · errors · slowest

Group the traffic:

$ wiretrail capture.har subsystems
Google  [-]  (165 req, 0 err, 71 dup)
Supabase  [-]  (38 req, 3 err, 33 dup)
TMDB  [-]  (25 req, 0 err, 15 dup)
jackettio.nexioapp.org  [-]  (21 req, 0 err, 21 dup)

Characterise the waste — are the 27 Supabase calls redundant?

$ wiretrail capture.har duplicates
  29x  POST youtubei.googleapis.com/youtubei/v1/visitor_id
  27x  POST yjyuomfgkqwmjvnoxurn.supabase.co/rest/v1/rpc/sync_resolve_account_secret

$ wiretrail capture.har diff
POST ....supabase.co/rest/v1/rpc/sync_resolve_account_secret  (27 calls, body: volatile-only)

volatile-only → the 27 calls differ only in noise: genuinely redundant work.

The auth story (the 3 Supabase errors):

$ wiretrail capture.har auth
auth failures:
  1x [401] yjyuomfgkqwmjvnoxurn.supabase.co /rest/v1/rpc/sync_pull_profiles
token refreshes:
  e000007 yjyuomfgkqwmjvnoxurn.supabase.co [200]

$ wiretrail capture.har jwt
7a236a4aaf8a2cae (1x, req.header.authorization) [EXPIRED]
  exp: 1779568690 (expired 60342s ago)

The 401 was caused by an expired JWT in the Authorization header; a refresh followed and a fresh token appears in later responses. Hand it off:

wiretrail capture.har report > capture-dossier.md     # shareable, redacted
wiretrail capture.har handoff                          # blocks w/ correlation + curl

Output for agents (--json)

Every command supports --json and emits a stable envelope. Pipe through jq:

wiretrail capture.har duplicates --json | jq '.result.groups[] | select(.count > 10) | {count, host, norm_path}'
wiretrail capture.har errors --json     | jq '.result.groups[] | {status, norm_path, count, message: .error_message}'

Envelope shape: { tool, schema_version, command, capture, result, warnings, next_commands }.

Quick reference

wiretrail c.har auto                         # one-shot smart analysis (start here)
wiretrail c.har                              # summary only (overview + next steps)
wiretrail c.har diagnose                     # ranked "what's wrong" shortlist
wiretrail c.har duplicates ; wiretrail c.har diff   # repeats + what varies
wiretrail c.har errors --json                # failures as JSON
wiretrail c.har slowest                      # slow calls + bottleneck
wiretrail c.har auth ; wiretrail c.har jwt   # auth/token story
wiretrail c.har search "timeout" --ignore-case      # grep bodies (redacted)
wiretrail new.har compare baseline.har --fail-on high   # regression gate
wiretrail c.har show-entry e000123 --unsafe-include-secrets   # raw detail

Common mistakes

• Dumping every command. For "what's wrong?", run auto (one shot) or start with summary/diagnose and follow the ranked recommendations. Three or four targeted commands usually tell the story — don't run all 32.
• Using report as the "tell me what's wrong" command. report is a static markdown dossier; the smart synthesizers are auto (drilled) and diagnose (shortlist). Reach for those first.
• Reaching for --unsafe-include-secrets by reflex. Default output is already safe and complete for analysis; only use the flag for a replayable request, and flag to the user that the result contains live credentials.
• Treating exit code 1 as an error. 1 means "findings reported"; 2 is the real "bad HAR" code.
• timeline looking truncated. It honours --top (default 10) — raise it.
• Hand-parsing the HAR JSON. Long base64/encoded path blobs (which often hide API keys) are collapsed to {blob} and redacted by wiretrail; reading the raw JSON re-exposes them.

What HAR (and wiretrail) can't tell you

Packet loss, full TLS certificate chains, client/OkHttp call stacks, DNS resolver internals, and service-worker/proxy behaviour aren't in a HAR unless a tool wrote them into custom fields. JWT analysis is structural only (claims/expiry, no signature verification). N+1 detection is a best-effort heuristic. wiretrail analyses what's present and won't invent the rest.