pcaps

PCAP Fixtures

Local test pcaps for Heron development (capture playback + turn grouping). The .pcap files themselves are not checked in (see .gitignore). This README documents what each fixture contains so tests and local runs can assert against known ground truth.

All captures are loopback (127.0.0.1:8317) plaintext HTTP — matching Heron's post-TLS server-side deployment model.

Fixtures

File	Provider / Endpoint	Client	Size	Turns (Complete/Incomplete)	Notes
claude-cli-messages.pcap	Anthropic /v1/messages?beta=true	claude-cli	3.8 MB	1 (1/0)	Single connection, single tool-calling turn
claude-cli-messages-multi.pcap	Anthropic /v1/messages?beta=true	claude-cli	5.4 MB	3 (≥1 Complete)	Long multi-turn session, single session_id; auto title-gen (empty tools) filtered as auxiliary, Task sub-agent calls attach to parent turn
codex-cli-messages-multi.pcap	OpenAI /v1/responses	codex-cli	18 MB	2 (1/1)	Multi-turn session, single session_id (see note); 2nd turn cut off mid-roundtrip by EOF
openclaw-openai.pcap	OpenAI /v1/chat/completions	OpenClaw (OpenAI/JS SDK + GLM)	1.4 MB	4 (4/0)	Two distinct user sessions on openclaw; client echoes assistant.tool_calls[].id without the underscore (calld9c1...) — exercises canonicalize_tool_id. Without it the 4 turns would shatter into many single-call turns
openclaw-anthropic.pcap	Anthropic /v1/messages	OpenClaw (Anthropic/JS SDK + GLM-5)	1.0 MB	4 (4/0)	One main session on openclaw (compaction-summarizer calls filtered as auxiliary by OpenClawProfile::is_auxiliary before turn assembly — pre-profile they appeared as two extra gen-* synth-id sessions because their first-user/first-assistant boilerplate hashed identically). GLM-5 emits parallel tool_use blocks where every content_block_start arrives before any input_json_delta — exercises the index-keyed SSE accumulator. Without per-index tracking, parallel tool_use.input collapses to "" or attaches to the wrong block
hermes-openai.pcap	OpenAI /v1/chat/completions	Hermes Agent (Nous Research) via OpenAI/Python SDK + GLM-5	553 KB	2 (2/0)	One user-facing 4-call conversation classified as hermes by body fingerprint (HermesProfile matches ≥2 of skill_view/skill_manage/skills_list/delegate_task/session_search/cronjob in tools[]), plus a 1-call chat-title-generation one-shot. The title-gen call has no tools and no Hermes markers, so HermesProfile does not match it and it falls through to generic — by design, since on the wire it is an independent turn (own session, own user-start, own finish_reason=stop).
gemini-cli-apikey.pcap	Gemini AI Studio /v1beta/models/{m}:streamGenerateContent?alt=sse	Gemini CLI (API-key mode) via @google/genai SDK	1.3 MB	2 (2/0)	7 LlmCalls split 4+3 across one shared session_id (form tu-<16hex> — Gemini has no protocol-level tool ids, so first_assistant_sig_* synthesizes a stable opaque id by FNV-1a hashing the canonical model-turn sig string). Turn A (4 calls): initial prompt + 3 tool roundtrips, closes on call-4 pure-text response (no functionCall ⇒ wire STOP not rewritten to synthetic TOOL_USE ⇒ is_turn_terminal true). Turn B (3 calls): user follow-up prompt at call 5 re-arms is_user_turn_start. No dedicated gemini-cli profile yet → all calls land on generic. Exercises (a) GenericProfile::matches() covering gemini-aistudio, and (b) first_assistant_sig_* returning ToolId(_) (not Text(_)) for tools-bearing model turns so generic's helper-shape one-shot gate doesn't spuriously fire on call 1 (which historically split call 1 into its own session).

Turn counts are ground truth verified against the current implementation and are intended as assertions for turn-grouping tests (server/h-turn/tests/integration.rs). Incomplete turns reflect streams that did not close cleanly within the capture window — the grouping is still deterministic across runs and shard counts.

Note on codex-cli-messages-multi.pcap: Codex's X-Codex-Turn-Metadata header reuses one turn_id across a whole codex invocation. The implementation honors the protocol, so a capture spanning multiple user-interactive sessions can still report a small number of turns.

Usage

Run Heron against a fixture via the pcap-file capture backend:

# example — adjust flags to match current CLI/config
cargo run -p heron -- --pcap testdata/pcaps/claude-cli-messages-multi.pcap

Obtaining the files

These (legacy, hand-distributed) pcaps are not in git. Ask @timmy.yuan for a copy or capture your own loopback traffic against 127.0.0.1:8317. They feed the turn-grouping tests in server/h-turn/tests/integration.rs, which skip when absent.

---

Committed regression corpus (corpus/)

Separate from the legacy fixtures above: testdata/pcaps/corpus/ holds a curated, secret-scrubbed, git-LFS-committed corpus that runs as a golden regression gate in CI (server/h-turn/tests/corpus_golden.rs, cargo test --workspace).

• corpus.toml — the single source of truth for the (backend × agent × wire_api × scenario) matrix. Each [[fixture]] is either status = "active" (a scrubbed corpus/<file>.pcap + golden/<id>.json are committed) or status = "pending" (a target cell whose capture isn't obtained yet — listed for visibility, skipped by the test).
• corpus/*.pcap — scrubbed fixtures, stored via git-LFS (.gitattributes). Run git lfs pull to materialize them; CI checks out with lfs: true. An unsmudged LFS pointer is treated as "absent" and skipped.
• golden/<id>.json — the deterministic extracted projection (no uuids / timing). Regenerate with just corpus bless and review the diff.
• <id>.scrub.json — per-fixture redaction audit (which rules fired, size), so reviewers can confirm scrubbing happened without seeing secrets.

Why "backend" is a label, not a detection output

vLLM / SGLang / Ollama / LiteLLM / CLIproxy all speak openai-chat / openai-responses on the wire — Heron does not distinguish them per-deployment. Their value as fixtures is the parsing quirk each one locks in (SSE framing, usage-block field names, finish_reason vocab, the CLIproxy mixed-format usage fallback). backend_label + [fixture.expect.quirks] document that; the hard detection assertion is wire_apis.

Adding / refreshing a cell

See scripts/pcaps/README.md for the capture → scrub → bless → commit workflow and the capture recipes. Quick:

scripts/pcaps/scrub_pcap.sh testdata/pcaps/raw/<id>.raw.pcap --id <id>
# flip status="active" in corpus.toml, then:
just corpus bless && just corpus test && just corpus lint

Never commit raw captures — testdata/pcaps/raw/ is gitignored.