Spoiler

Spoiler reads session recordings and reports what users tried, how it ended, and what blocked them.

Code measures what happened. A model explains it. Code checks the explanation.

• In: any rrweb recording, from a file or fetched from PostHog. Web, native iOS and Android.
• Out: versioned JSON: a deterministic trace, and an analysis that cites it.
• Model: any OpenRouter model, one call per visit. Compiling needs no model and no network.

Example

An admin invites a teammate, but every seat is taken. From the 22-second recording, Spoiler reports:

Task: Invite priya@example.com to the workspace. Outcome: workaround.

Send invite was refused three times because all 5 seats were in use. Deactivating Ben Ortiz did not free his seat. Removing him did.

Hypothesis: the message doesn't say that deactivated members still hold seats.

The why comes from the vocabulary term seat, drafted from members.server.ts:4. It reads: "Every member holds a seat until removed, deactivated members included."

Each claim cites refs into the trace, which code compiles from the recording with no model:

The answer also claimed "Rage-clicked Send invite after the first refusal", citing e3 and e5. Code dropped it: the clicks were 3.5 s apart, and neither carries a rage flag.

The session is synthetic and this narration is hand-written. With --model, a model writes it.

How it works

source files ─▶ vocab build ─▶ vocabulary ─┐
                (model, per release)       ├─▶ compile ─▶ trace ─▶ analyze ─▶ analysis
recording ─────────────────────────────────┘   (code only)         (model, then code)

1. Vocabulary. vocab build has a model read source files you name. It drafts surfaces, controls, domain terms and rules, each citing its source line. What the sources don't cover is listed under gaps, not guessed. The snapshot pins every source by SHA-256 and is never regenerated implicitly. --candidate packages a vocabulary you wrote instead, with no model call.
2. Compile. compile replays the recording's DOM log, one mirror per tab. Each click resolves to an element, a vocabulary feature, what changed, and how fast. Rules raise flags: dead, unresponsive, rage, slow, error_after, error_shown, thrash. The same recording, vocabulary and compiler version always give the same trace. The trace also carries a timeline: when the user was there (posthog-js's own activity rule), which tab was in front, the gaps nobody acted in, and per tab how much of the page the recording could show (exact, blind before its first snapshot, dropped while the recorder was idle or paused, stale until it caught up). Players and narration read it instead of re-deriving it.
3. Narrate. analyze sends the trace as TSV, plus the vocabulary it touched. The recording itself is never sent. run makes one call per visit. The model returns tasks with a goal, outcome, obstacle and friction, all citing refs. Outcomes are done, workaround, gave_up or unclear.
4. Gate. Code checks every answer before accepting it:
   • Off-schema, or over 15% of cited refs missing: rejected. A live call gets one corrected retry.
   • dead_click, rage_click, error and slow friction is dropped without a matching flag.
   • Timestamps, durations, paths and data changes are written from the trace, never by the model.
   • check lists everything dropped, unexplained, or uncited.

Install

curl -fsSL https://spoiler.sh/install | sh

• Installs the binary for your OS and CPU to ~/.local/bin, after checking its sha256.
• Linux gets the static musl build, which runs on any distribution.
• … | SPOILER_VERSION=v0.1.0 sh pins a release. SPOILER_INSTALL_DIR picks the directory.

From PyPI. The wheel only puts the spoiler binary on PATH; there is no Python API.

pip install spoiler

uv tool install spoiler works the same way. With Rust 1.88+, from crates.io:

cargo install spoiler --locked

Releases has an archive per target, each with a .sha256:

• aarch64-apple-darwin, x86_64-apple-darwin
• x86_64-unknown-linux-gnu, aarch64-unknown-linux-gnu (glibc 2.28+)
• x86_64-unknown-linux-musl, aarch64-unknown-linux-musl (static; any Linux)

From a checkout: cargo install --path crates/cli --locked.

Quick start

Offline, from a checkout, using committed fixtures. No account or API key.

Build, and put the binary on PATH:

cargo build --release --locked
export PATH="$PWD/target/release:$PATH"
mkdir -p artifacts

Compile a recording into a trace. No model, no network:

spoiler compile \
  --recording corpus/click_changes_text.json \
  --vocab corpus/vocabulary.yaml \
  --app demo \
  --out artifacts/trace.json

Read the trace the way a model does:

jq -r .tsv artifacts/trace.json

Validate a stored answer against the trace:

spoiler analyze \
  --trace artifacts/trace.json \
  --vocab corpus/vocabulary.yaml \
  --response examples/click_changes_text.response.json \
  --out artifacts/analysis.json

Swap --response FILE for --prepare-only to write the exact model request instead. Neither sends anything.

On your product

Set credentials, which Spoiler reads only from the environment, and two variables used below:

export POSTHOG_API_KEY=phx_…
export OPENROUTER_API_KEY=sk-or-…
MODEL=…       # any OpenRouter model id
SESSION_ID=…  # a PostHog recording id

Describe each app in product.json. project is its PostHog project id:

{
  "apps": {
    "web": { "project": 123, "host": "app.example.com", "audience": "workspace admins" }
  }
}

Build a vocabulary from the source files that name your routes and controls:

spoiler vocab build \
  --config product.json \
  --source app/routes.ts \
  --source app/routes/members.tsx \
  --source app/members.server.ts \
  --source-revision "$(git rev-parse HEAD)" \
  --model "$MODEL" \
  --out vocab.json

Check it, and review the drafted matchers before relying on them:

spoiler vocab check --vocab vocab.json

Fetch, compile and narrate one PostHog session:

spoiler run \
  --project 123 \
  --session "$SESSION_ID" \
  --vocab vocab.json \
  --app web \
  --model "$MODEL" \
  --out session.json

• run narrates every visit with user gestures. --visit N picks one.
• Rebuild the vocabulary when you ship. Each trace records the digest it was compiled against.
• After upgrading spoiler, rerun with --previous session.json: a visit whose model request is unchanged reuses its analysis ("via": "reused", or "regated" when a newer gate judged the stored answer again), so only changed questions reach the model. With --prepare-only, what is left as a request is exactly what would be paid for.

Commands

• One JSON artifact per command, to stdout or --out (written atomically).
• Any input path accepts - for stdin, so commands pipe.
• Configuration is flags only. Credentials come only from the environment.
• No model is ever chosen for you.
• spoiler <command> --help lists every flag.

Each stage also runs alone, and they pipe:

spoiler recordings fetch --project 123 --session "$SESSION_ID" \
  | spoiler compile --recording - --vocab vocab.json --app web \
  | spoiler analyze --trace - --vocab vocab.json --prepare-only

Failures are JSON on stderr: {"error": "…", "retryable": false}.

Reference

Vocabulary

version: 1
apps:
  workspace: { project: 1, host: app.test, audience: "workspace admins" }
surfaces:
  - { id: workspace.members, app: workspace, route: /settings/members, name: "Members" }
features:
  - id: members.invite.send
    surface: workspace.members
    name: "Send invite"
    matchers: { testid: [send-invite] }
    source: "members.tsx:11"
terms:
  - term: seat
    means: "Every member holds a seat until removed, deactivated members included."
    source: "members.server.ts:4"
statuses:
  - { kind: member, value: deactivated, label: Deactivated }
gaps: ["Billing page: billing.tsx is not among the sources, so seat purchases are unnamed."]

• Matchers, most specific first: testid, data_attr, aria, title, placeholder, href, text, role, class_contains. aria, text and title also take *_template forms.
• surface: "*" marks app chrome and requires app.
• Optional: statuses, grid (row identity), telemetry (URLs to ignore), error_text, thresholds.
• Defaults: English error patterns, common monitoring requests ignored, slow_ms: 1000.
• vocab check reports invalid matchers and inert entries.

Recordings

• Accepted: rrweb event arrays, JSONL, compressed input, and PostHog snapshot lines.
• Native iOS and Android: taps resolve against PostHog wireframes. Native routes are screen names such as SettingsScreen, matched case-sensitively.
• Screenshot-mode frames, including Flutter and React Native, are opaque images. A tap on one is reported as screen (x,y), with no invented control.
• A visit ends after 30 minutes without actions (thresholds.visit_gap_ms).

PostHog

• recordings list returns one page. Pass its next_cursor back as --cursor.
• It lists recordings that started over 24 hours ago and ended by --until. Use settled windows, and overlap consecutive syncs.
• Segments more than 7 days outside the window can make a recording look complete.
• The default host is https://eu.posthog.com. Set --host for US or self-hosted.
• Listings and snapshot ranges share PostHog's per-key throttle (paid: 60/min, 300/h). A paid key fetches about 75 typical recordings an hour.
• On 429, fetch honours Retry-After for up to --max-wait seconds (60), then exits 75.
• --max-requests caps requests per fetch (default 50, at most 59).

Data and security

• Recordings are untrusted input. --max-input-mib (512) bounds decoded size. HogQL and model responses are capped at 32 MiB.
• Recordings are sensitive. Review what you send to PostHog and OpenRouter.
• Model calls request zero-data-retention routing. Each analysis records model, tokens and cost.
• Credentials travel only over HTTPS, and redirects are disabled.
• Report vulnerabilities privately through GitHub security advisories.

Limits

• The DOM mirror can't see iframe documents, canvas pixels, or shadow-root internals. Events from unrecognized rrweb plugins are not interpreted. coverage counts all of these.
• rrweb records no key presses, so a keyboard-shortcut change has no gesture.
• Masked inputs reveal only their length.
• Tests cover the compiler and the gate, not narration quality or live PostHog behaviour.

Development

cargo fmt --all --check
cargo clippy --workspace --all-targets --locked -- -D warnings
cargo test --workspace --locked

• corpus/ holds synthetic recordings with golden traces. They define compiler behaviour.
• Edit cases in scripts/corpus.py, then run python3 scripts/corpus.py.
• Accept an intended output change with SPOILER_BLESS=1 cargo test --test corpus. Review every golden diff.
• A compile-rule change bumps COMPILER_VERSION. Readers reject traces from other versions.
• Fixtures stay synthetic: no real recordings, credentials, or customer data.
• Releasing is documented at the top of .github/workflows/release.yml.

License

Apache-2.0