Heron v0.7.3

[0.7.3] — 2026-08-17

Fixed

• Opening an agent turn made the timeline wait on every request and response
  body in it. The detail panel fetched the calls list body-bearing, and the
  timeline, stat cards, agent breakdown and collapsed cards are built from
  scalars — none of them reads a body. Measured against a production store, that
  list is 2–20 MB and 0.4–4.8 s where the same list without bodies is 16–60 KB
  and ~10 ms; the panel then parsed those megabytes on the main thread. It now
  paints off the small shape and pulls bodies afterwards as a background
  upgrade, so nothing the reader is waiting for is behind them.
• Turns over the 50-call threshold fetched the bodies anyway, then threw them
  away. Whether to ask for bodies was decided from call_count, which arrives
  from a different request — so on mount the answer defaulted to "yes" and the
  body-bearing fetch went out for every turn, to be abandoned a few ms later
  when the count came back over the threshold. apiFetch passed no
  AbortSignal, so "abandoned" still meant downloading, parsing and caching the
  full response: a 102-call turn pulled 20.6 MB it could never render. The span
  and body endpoints now pass the signal, and the query no longer fires before
  its own precondition is known — the same turn now costs 60 KB to open.

What's Changed

• ci(prod): chain deploy-prod off release.yml — the release trigger never fires by @vaderyang in #193
• soak: say whether the load gate is enforcing, because the log said it wasn't by @vaderyang in #194
• fix(console): the turn timeline waited on every body in the turn (v0.7.3) by @vaderyang in #195

Full Changelog: v0.7.2...v0.7.3

Heron v0.7.2

[0.7.2] — 2026-08-15

Added

• sglake (sglog) storage backend. [storage] backend = "sglake" runs the
  whole console against a Splunk-compatible log platform, alongside DuckDB and
  ClickHouse — writes over the HTTP Event Collector, reads in SPL, all thirty
  StorageBackend methods. Unlike the SQL backends, stored request and response
  bodies are full-text searchable (search index=heron_bodies "SELECT * FROM users"). Retention is pushed as per-index TTLs rather than DELETE, since
  the store is append-only. heron sglake-props prints the index-time
  extraction stanzas that keep aggregates on the columnar path; they are
  generated from the event structs themselves, so they cannot drift from the
  schema. See docs/design/10-sglake.md, which also
  documents the known divergences — proxy pairing does not annotate traces on an
  append-only store.

Fixed

• Pagination could return the same row on two pages and skip another. Every
  paginated list ordered by a non-unique key (request_time most often) with no
  tie-break, and LIMIT/OFFSET issues one query per page, so the engine was
  free to order tied rows differently between them. Walking 26 spans that shared
  a timestamp five at a time returned 18 distinct rows. All paginated queries in
  the DuckDB and ClickHouse backends now end their sort with the row id.
• DuckDB truncated duration_ms on HTTP exchanges to whole milliseconds, so
  every sub-millisecond exchange showed 0.0. ClickHouse already computed it
  from microseconds.
• /api/metrics/timeseries could be asked for an unbounded number of
  buckets — a 7-day window at 10s granularity materialises 60,480 points per
  series. It now refuses above 20,000 and names a granularity that fits.

Changed

• The agent-trace panel switches to lazy per-call body loading above 50 calls
  rather than 200, measured against the real cost of answering with bodies.

What's Changed

• docs(readme): swap hero GIF to the live dashboard view by @vaderyang in #178
• feat(storage): sglake backend — full-text searchable bodies, and five defects it exposed by @vaderyang in #191
• ci(prod): production tracks releases, gated on a read-path smoke by @vaderyang in #192

Full Changelog: v0.7.1...v0.7.2

Heron v0.7.1

[0.7.1] — 2026-06-24

Fixed

• Console blank-page regression from the v0.7.0 OTel rename. The traces API
  serializes span_ids (renamed from call_ids), but the AgentTurnDetail
  type and the agent-breakdown component still read call_ids, so opening an
  agent trace detail threw TypeError: undefined is not an object (evaluating '…call_ids.length') and — with no React error boundary — blanked the entire
  console. Aligned the type + component to span_ids and guarded the access so
  a missing field degrades to "0 calls" instead of a white screen.

Changed

• UI vocabulary aligned with the OTel model. Renamed user-visible labels:
  "Agent Turns" → "Agent Traces" and "HTTP Exchanges" → "HTTP Logs" (nav, detail
  titles, the overview gauge, the "Trace ID" metadata row, empty/error states,
  the SFT export tooltip, and the distribution chart). A trace is one agent
  interaction; a step within it remains a turn/span. Backend metric names, route
  paths, and code identifiers are unchanged.

What's Changed

• v0.7.1: fix console blank page (span_ids) + UI rename (Agent Traces / HTTP Logs) by @vaderyang in #177

Full Changelog: v0.7.0...v0.7.1

Heron v0.7.0

[0.7.0] — 2026-06-24

Added

• eBPF capture discoverability — an experimental, opt-in advisory. When no
  eBPF source is active, the server now logs a one-line notice (Linux only) that
  on-host TLS capture exists and how to enable it (ebpf cargo feature + a
  type = "ebpf" source — there is no --ebpf runtime flag). The Settings
  page's eBPF capture toggle is now tagged experimental and its description
  spells out the opt-in requirements.

Changed

• OpenTelemetry-aligned rename of the storage entities and HTTP API. The
  agent_turns table is now traces, llm_calls is now spans, and
  traces.call_ids is now span_ids; a new forward-looking spans.kind
  column (always 'llm' today) leaves room for wire-visible tool spans. The
  Rust domain/query types (AgentTurn→Trace, Turn*/Call* DTOs →
  Trace*/Span*) and StorageBackend methods (write_calls→write_spans,
  query_turns→query_traces, …) follow suit. New canonical routes
  /api/traces* and /api/spans*; the pre-rename /api/agent-turns* and
  /api/llm-calls* keep working as deprecated aliases (RFC 8594 Deprecation
  header). Retention config keys calls/turns are accepted as serde aliases
  for spans/traces. Existing DuckDB/ClickHouse databases auto-migrate in
  place on init() with no data loss (idempotent detect-then-rename).

Docs

• README rebuilt for the launch — GIF-first, conversion-driven layout.
  Headline tagline "The Wireshark for AI Agents", a hero demo GIF up top, a
  30-second pcap-replay quick start, and a three-up "What Makes Heron Different"
  table (agent-turn reconstruction · service topology · SFT trajectory export),
  while preserving the existing technical accuracy (passive positioning across
  wire and on-host TLS boundary, opt-in eBPF, docs/configure.md links,
  just-based contributor flow). API examples updated to the canonical
  /api/traces routes.
• Corrected launch collateral added under launch/. Paste-ready Product
  Hunt / Hacker News / Twitter copy aligned to the shipped product: install via
  the curl … install.sh one-liner (no npm package), canonical /api/traces
  endpoint, and v0.7.0 framing.

What's Changed

• ci: repoint agent-loop wrappers to Netis/olympus (agent-ops renamed) by @vaderyang in #165
• Upgrade Olympus mechanism to v0.4.0 (full post-rename migration: .olympus.json + olympus_ref) by @vaderyang in #167
• fix(turn): keep eBPF-sourced turns whose opening call was missed by @vaderyang in #168
• fix(ebpf): detach uprobe links for dead inodes to stop bpf_link fd leak by @vaderyang in #172
• chore(olympus): align .olympus.json — current model + Greek-deity agent names by @vaderyang in #173
• refactor: OpenTelemetry-aligned rename (agent_turns→traces, llm_calls→spans) by @vaderyang in #174
• Launch v0.7.0 → main: README rebuild + collateral + eBPF advisory by @vaderyang in #176

Full Changelog: v0.6.0...v0.7.0

Heron v0.6.0

[0.6.0] — 2026-06-16

Changed

• eBPF on-host SSL-uprobe capture is now a first-class, soak-gated
  capability (experimental since 0.5.1). It lifts the plaintext of
  TLS-encrypted LLM calls directly at the in-process SSL_read / SSL_write
  boundary — covering dynamically-linked OpenSSL/BoringSSL (Python openai /
  anthropic SDKs, curl, Node, most CLIs) and statically-linked, symbol-stripped
  BoringSSL single-executable runtimes (Claude Code's / opencode's Bun binaries,
  located by byte-signature offset) — and stamps every call with its owning
  process (pid · command · executable). A new staging ebpf-soak gate replays
  real TLS traffic through the freshly deployed binary and asserts the uprobe
  attaches, traffic is captured, and a process-attributed LlmCall is parsed and
  persisted end-to-end; both prod promotion and release now require a passing
  ebpf-soaked status alongside staging-soaked, so on-host capture can no
  longer silently regress into prod or a cut release.

Fixed

• eBPF SSL uprobes attach under the non-root staging service. The staging
  unit granted only CAP_BPF + CAP_PERFMON, but the kernel gates uprobe
  perf_event_open (perf_uprobe_init) on CAP_SYS_ADMIN specifically — which
  those caps don't cover and perf_event_paranoid doesn't relax. The SSL_write
  uprobe attach therefore failed with perf_event_open failed, and because a
  failed capture source is non-fatal (the co-located packet tap keeps the
  pipeline healthy) the symptom was a silent ebpf_uprobes_attached = 0 rather
  than a crash — which is why the ebpf-soak gate had never passed. The
  committed staging unit now carries CAP_SYS_ADMIN (matching what the prod
  deploy already injects), so the gate goes green and on-host capture works under
  a non-root service.

Docs

• README repositioned from "network wire" to passive capture. The headline
  "Agent observability from the network wire" became inaccurate once on-host eBPF
  landed — those bytes never touch the wire (on a client they're the
  pre-encryption plaintext). It now leads with the durable differentiator
  (passive; no SDK, sidecar, or proxy; never in the request path) and names both
  capture surfaces: off the network wire, or lifted from the host's TLS boundary
  by eBPF.

What's Changed

• ci(ebpf): staging eBPF verification gate before prod by @vaderyang in #148
• fix(ebpf): follow inode rotation & reach already-running sessions (v0.5.4) by @vaderyang in #157
• fix(wire/anthropic): don't record control-plane telemetry as model=unknown LLM calls (v0.5.5) by @vaderyang in #158
• fix(agents/claude-cli): drop security-monitor sidecar from agent-turns (v0.5.6) by @vaderyang in #159
• docs(readme): elevate eBPF on-host capture to a first-class capability by @vaderyang in #160
• fix(ebpf): grant CAP_SYS_ADMIN to staging unit so SSL uprobes attach by @vaderyang in #161
• docs(readme): reposition from "network wire" to passive (wire or eBPF) by @vaderyang in #162
• release: v0.6.0 — eBPF on-host capture promoted to first-class (soak-gated) by @vaderyang in #163

Full Changelog: v0.5.3...v0.6.0

Heron v0.5.3

[0.5.3] — 2026-06-15

Fixed

• eBPF capture of large and multi-connection TLS traffic. A big SSL_*
  buffer is split into several ring-buffer events; the synthesizer used to append
  each at a running sequence counter, so a silently-dropped or reordered chunk
  shifted every later byte and spliced the next request's bytes into the previous
  body (invalid JSON → no captured call). The BPF program now stamps each event
  with its absolute per-(connection, direction) stream offset and the
  synthesizer places every chunk at that offset, so a dropped chunk leaves a gap
  at its true position instead. Connections also finalize properly now — an
  SSL_free uprobe (dynamic libssl) and an SSL_read-returns-0 EOF (static
  targets) emit a close, and a per-connection generation folded into the
  synthetic tuple gives a reused SSL* pointer a fresh, non-overlapping flow
  key.
• Claude Code conversations form agent-turns. The Anthropic
  mid-conversation-system feature appends a trailing role=system notice after
  the user's prompt, so the messages array often ends with system on a fresh
  turn. is_user_turn_start / extract_user_input now skip trailing system
  messages and evaluate the last non-system one, so fresh turns are recognized
  (and their prompt preview populated) instead of being discarded as
  "no user start".
• In-progress turn timestamps rendered as the year ~58423. Active-turn
  registry rows emitted start_time/end_time in microseconds, but the field is
  milliseconds (the DB path returns epoch_ms, the console renders with
  new Date(ms)). A µs value read as ms is 1000× too far in the future. Both
  active-registry conversions now divide to milliseconds, matching finalized
  turns. Added a timestamp-unit test suite across the eBPF clock and the API
  boundary.

What's Changed

• feat(prod): build the eBPF-enabled binary in deploy-prod by @vaderyang in #150
• fix(prod): raise container fd limit for the eBPF build by @vaderyang in #151
• feat(console): eBPF capture status card on pipeline-health by @vaderyang in #152
• fix(ebpf): correct large/multi-connection capture + form Claude Code agent-turns by @vaderyang in #155
• release: v0.5.3 — eBPF capture correctness + timestamp fix & test suite by @vaderyang in #156

Full Changelog: v0.5.2...v0.5.3

Heron v0.5.2

[0.5.2] — 2026-06-11

Added

• eBPF capture is now a managed Settings source. The console's pipeline
  Settings lists, adds, edits, and removes the on-host eBPF SSL-uprobe source
  like any other ingress, with an availability guard: a binary built without the
  ebpf feature reports ebpf_available = false on /api/runtime-config, and
  the API rejects an eBPF source on such a build so a stray config can't wedge
  the next boot. (The capture engine itself shipped in 0.5.1; this makes it
  operable from the UI.)
• eBPF metrics in pipeline-health. A new ebpf metric group surfaces the
  capture path's health — uprobes attached, events received / dropped, bytes
  captured, frames synthesized, active connections, and process-cache size — in
  the debug pipeline-health view and on /api/internal-metrics.

Changed

• Console theme aligned to the Kami parchment design system. Replaced the
  too-dark palette with the canonical parchment surfaces (paper canvas, lifted
  ivory, ink-blue accent), matching the heron-ai.pages.dev landing site.

Docs

• Refreshed the README screenshots in the corrected Kami theme.

What's Changed

• feat(ebpf): Settings enable/disable toggle + pipeline-health metrics by @vaderyang in #147
• fix(console): align Kami theme to the parchment design system by @vaderyang in #145
• docs: refresh README screenshots in the corrected Kami theme by @vaderyang in #146
• release: v0.5.2 — eBPF Settings source + metrics, Kami theme by @vaderyang in #149

Full Changelog: v0.5.1...v0.5.2

Heron v0.5.1

[0.5.1] — 2026-06-11

Added — eBPF on-host TLS capture (experimental, Linux)

• New ebpf capture source: a fourth ingress alongside the packet taps
  (pcap / pcap-file / cloud-probe). It attaches uprobes to the target's
  SSL_read / SSL_write and reads plaintext at the in-process TLS boundary —
  so Heron can observe TLS-encrypted LLM calls on the host that makes them,
  with no proxy, TLS terminator, or MITM, and nothing on the request path.
  Plaintext chunks are dressed as synthetic Ethernet/IP/TCP frames
  (FlowSynthesizer) and fed through the existing dispatcher → reassembler →
  HTTP/SSE parser → wire-API decoder → turn tracker unchanged.
• Process attribution. Every eBPF-captured call carries its owning process
  (pid · comm · resolved executable), threaded end-to-end through
  RawPacket → ParsedPacket → TcpFlow → LlmCall into the
  process_pid / process_comm / process_exe storage columns (DuckDB Phase-7
  migration + ClickHouse mirror) and surfaced in the console's LLM-calls list and
  call detail. Packet-tap sources leave it null.
• Target coverage. Dynamically-linked OpenSSL/BoringSSL by exported symbol
  (Python SDKs, curl, …), and statically-linked, symbol-stripped BoringSSL —
  e.g. Claude Code's Bun runtime — located by byte-signature → ELF file offset →
  offset uprobe. A built-in flavor = "bun" ships read-anchored prologue
  signatures, so stock Bun / Claude Code works with zero manual derivation.
• Linux-only and off by default: built behind the non-default ebpf cargo
  feature on h-capture (absent from prebuilt release binaries). Needs
  CAP_BPF + CAP_PERFMON (kernel ≥ 5.8) or root, plus kernel BTF; heron doctor reports a capture.ebpf check. HTTP/1.x only, like every source. See
  docs/design/02-capture.md and docs/design/03-ebpf-static-targets.md.

Added

• SFT trajectory export from reconstructed agent turns and sessions:
  OpenAI-style messages JSONL with tool calls, tool results, and assistant
  reasoning preserved and tool-call arguments rehydrated to objects. Export a
  single turn/session from its detail view, or batch-export the current Agent
  Turns filter as one-line-per-turn JSONL (Anthropic + OpenAI-chat wire formats;
  unsupported formats reported and skipped).
• Three-theme console, switchable from the sidebar and persisted per browser:
  Kami (warm washi-paper, the new default), Dark, and Light —
  charts, topology graph, and timeline gantt all re-theme.

Fixed

• ClickHouse SQL literal escaping (dialect-aware): a backslash in a
  dimension-filter value could break out of the quoted literal in the ClickHouse
  backend. Escaping is now dialect-aware across both backends.

Security / CI

• Self-hosted CI runners are gated to same-repo PRs, closing fork-PR code
  execution; a release may only be cut from a commit that passed staging-soak,
  and prod deploys are gated on the load soak.

What's Changed

• test(chaos): fault injection under sustained concurrent write load (PR3) by @vaderyang in #121
• feat(test): pcap regression corpus framework (scrub + manifest + golden, git-LFS) by @vaderyang in #120
• feat(staging): nightly longevity soak — the 102 GB endurance catcher (PR4) by @vaderyang in #122
• ci(perf-gates): enforce load soak before prod + require staging-soaked before release by @vaderyang in #123
• docs: quality/release pipeline + v0.5.0 refresh (ClickHouse, perf gates) by @vaderyang in #124
• feat(test): cliproxy mixed-format (#96) — precise OpenAI-shape + Anthropic-usage cells by @vaderyang in #125
• feat(triage): warm maintainer-voice replies on every verdict; reproduce before auto-queuing by @vaderyang in #126
• dogfood: delegate triage to the agent-ops reusable workflow by @vaderyang in #127
• dogfood: delegate issue-implement (wiwi) to agent-ops@v0.1.0 by @vaderyang in #129
• dogfood: delegate pr-review + pr-revise to agent-ops@v0.2.0 by @vaderyang in #131
• dogfood: delegate leakage guard + bump triage/implement to agent-ops@v0.2.0 by @vaderyang in #132
• dogfood: remove in-tree agent-bot fallbacks (all surfaces delegated + proven) by @vaderyang in #133
• feat(console): three-theme redesign — Dark Observatory, Light Atelier, Kami by @vaderyang in #134
• feat(export): SFT trajectory export from agent turns/sessions by @vaderyang in #135
• fix(staging-soak): measure steady-state RSS growth, not warm-up by @vaderyang in #136
• chore(pr-review): bump agent-ops pin to v0.3.1 (diff-injection review fix) by @vaderyang in #137
• fix(deploy-prod): rebuild console bundle before cargo build (stale embedded UI) by @vaderyang in #138
• fix(deploy-prod): resolve bun via ~/.bun/bin fallback (non-login deploy shell) by @vaderyang in #139
• feat(console): default to Kami theme + refresh README screenshots in Kami by @vaderyang in #140
• fix(ci): gate self-hosted CI to same-repo PRs by @vaderyang in #143
• fix(storage): dialect-aware SQL literal escaping for the ClickHouse backend by @vaderyang in #142
• feat(capture): native eBPF SSL-uprobe capture (process attribution + static-binary targets) by @vaderyang in #141
• release: v0.5.1 — eBPF capture docs + VERSION/CHANGELOG by @vaderyang in #144

Full Changelog: v0.5.0...v0.5.1

Heron v0.5.0

[0.5.0] — 2026-06-04

Added — ClickHouse storage backend

• New h-storage-clickhouse crate: a drop-in StorageBackend implementation
  backed by ClickHouse (HTTP interface, async serde RowBinary via the
  clickhouse crate), selected with storage.backend = "clickhouse". Mirrors
  the DuckDB backend's full read + write surface. Fact tables use MergeTree;
  agent_turns uses ReplacingMergeTree(_version) (FINAL reads; version-bumped
  full-row re-insert for update_turn_metadata). Timestamps are
  DateTime64(6, 'UTC') mapped to i64 micros. Retention runs via lightweight
  DELETE on the shared [storage.retention] schedule.
• Backend-neutral logic (dimension-filter SQL builders, header/token-estimate
  converters, the serving-software classifier) extracted from
  h-storage-duckdb into h-storage (dialect / convert / classify) and
  shared by both backends — single source of truth.
• storage_bench binary + scripts/bench-storage.sh (just bench-storage):
  ClickHouse-vs-DuckDB write-throughput + read-latency comparison through the
  identical workload. Findings + methodology in
  docs/design/bench-clickhouse-vs-duckdb.md.
• ClickHouse query_services optimised 46× (5129 ms → 112 ms at 1M rows): the
  body-sample ROW_NUMBER window (which read every row's ~2 KB body columns)
  became an id IN (… LIMIT N BY …) two-phase fetch; arrayDistinct(groupArray)
  → groupUniqArray(N); quantileExact → quantileTDigest.

Fixed

• Agent-sessions agent_kind multi-select returned nothing. Selecting more
  than one agent kind sent a CSV (claude-cli,codex-cli) that the sessions
  query exact-matched as a single literal, so the list went empty; selecting one
  kind worked. Fixed in both the DuckDB and ClickHouse backends. Root-caused
  beyond the symptom: agent_kind is now CSV-parsed once at the API boundary
  into a Vec (like every other multi-select filter), so no storage backend
  ever sees a raw CSV to mis-handle.
• Agent classifier no longer flags unrecognized tool names as "suspicious".
  Tools from non-Claude-Code agents (e.g. web_search, read_file,
  memory_get) were tagged suspicious purely for being absent from a hardcoded
  registry. Any named tool is now classified as a function-call surface; the
  registry treadmill is gone.

Internal

• Perf/reliability gate for the release pipeline: a rate_pps-throttled
  pcap-file load-soak (tara --load) and criterion hot-path micro-benchmarks.
• Deploy pipeline hardening: staging-soak stamps its staging-soaked status via
  the REST API (the runner has no gh), and deploy-prod supersedes older
  waiting approvals instead of letting them pile up and wedge the queue.

What's Changed

• chore(rebrand): scrub residual TokenScope brand strings → Heron by @vaderyang in #72
• fix(storage): Phase-5 migration adds agent columns nullable (DuckDB compat) by @vaderyang in #76
• ci: move runner label tokenscope → heron by @vaderyang in #73
• pr-review: post_review.py downgrades APPROVE → COMMENT on heading-format drift by @vaderyang in #77
• check-leakage: also flag machine-specific username paths by @vaderyang in #78
• heron panics on shutdown: JoinHandle polled after completion (supervisor double-poll, exit 101) by @vaderyang in #80
• fix(llm): register AgentClassifier{Unknown,Mixed}Count on llm worker — stops silent capture loss (#81) by @vaderyang in #82
• docs: launch assets — README badges, technical blog, launch posts by @vaderyang in #87
• docs: warn that a bare cargo build ships a blank console (--features console) by @vaderyang in #83
• feat(staging): continuous auto-deploy to heron-stage VM (Quality Infra L6) by @vaderyang in #84
• ci(staging): add workflow_dispatch manual deploy to deploy-staging by @vaderyang in #89
• ci: bump test-job timeout 25→60m for the on-main artifact build by @vaderyang in #90
• chore(deps): bump DuckDB 1.5.1 → 1.5.3 (fixes checkpoint/ART bug behind the prod outage) by @vaderyang in #92
• feat(agent-bot): mara — prod observer that auto-files heron incidents by @vaderyang in #86
• fix(classifier): expand known-tool registry to current CLI inventory (#85) by @vaderyang in #93
• feat(h-llm): cap stored request/response bodies for 1M-token contexts (#68) by @vaderyang in #95
• feat(staging): tara soak runner — replay-invariant gate after deploy (L7) by @vaderyang in #94
• feat(prod): gated auto-deploy to production (deploy-prod — quality-chain tail) by @vaderyang in #97
• docs: add issue-filing guide + triage-gate explainer by @vaderyang in #98
• fix(classifier): unrecognized tool names are function calls, not "suspicious" (#85) by @vaderyang in #101
• feat(capture): pcap-file loop/duration replay mode (load-soak primitive) by @vaderyang in #99
• docs: scrub internal deployment specifics from pr-review-agent doc by @vaderyang in #100
• feat(mara): confirm-debounce — a deploy/restart blip no longer files a phantom incident by @vaderyang in #104
• chore(ci): run CI on dedicated ci-pool runner, off the agent pool by @vaderyang in #112
• feat(perf): tara load + memory soak — the pre-deploy perf/reliability gate (PR1) by @vaderyang in #106
• CI hygiene (§5): scrub internal CI backend model identifier from workflow + script comments (follow-up to #100) by @vaderyang in #109
• feat(storage): ClickHouse backend (drop-in) + storage benchmark by @vaderyang in #103
• CLIproxy 混合格式问题：OpenAI 请求 → Anthropic 响应，usage token 解析不到 by @vaderyang in #107
• feat(agent-bot): wiwi auto-revises on vivi CHANGES_REQUESTED (bounded loop) by @vaderyang in #113
• feat(perf): criterion hot-path micro-benchmarks + CI bitrot gate (PR2) by @vaderyang in #110
• chore(agent-bot): move auto-merge allowlists out of committed source by @vaderyang in #102
• ci(staging): stamp staging-soaked commit status — prod go/no-go signal by @vaderyang in #114
• agent-sessions: agent-kind 过滤把整段 CSV 当单值精确匹配，选中后列表清空 by @vaderyang in #111
• fix(staging-soak): stamp staging-soaked status with curl, not gh — unblocks ALL prod deploys by @vaderyang in #115
• refactor(filters): parse agent_kind CSV at the API layer (Vec), not in storage — root-cause fix by @vaderyang in #116
• fix(deploy-prod): supersede older waiting runs (cancel-in-progress: true) — kills the pile-up wedge by @vaderyang in #117
• chore(release): v0.5.0 by @vaderyang in #119

Full Changelog: v0.4.0...v0.5.0

Heron v0.4.0

[0.4.0] — 2026-05-29

Changed — Rebrand to Heron

• Project renamed TokenScope → Heron. The binary is now heron
  (was tokenscope); the 10 internal library crates moved from the
  ts-* prefix to h-*; the GitHub repo is now Netis/heron (the old
  URL redirects). Console title, logo (a new heron mark), and all
  install/docs URLs updated.

Added — Quality infrastructure

• Deterministic fault-injection harness for the DuckDB backend
  (feature-gated) plus recovery tests that drive the FATAL → reopen →
  every-surface-works path without relying on real load pressure.
• Schema-migration tests over synthesized legacy DB shapes, locking the
  auto-migration paths against silent regression.
• CI lint gates: referenced-secret provisioning, secret-value sanity,
  validated-constructor scoping, and an infra-leakage gate that fails on
  any non-allow-listed private IP or private-key block in tracked files.

Security / privacy

• Removed the demo deploy tooling, which hard-coded a server address, a
  jump-host username, and a plaintext password. Demo setup is now an
  AI-agent prompt in the docs instead.
• Scrubbed internal infrastructure identity (private IPs, hostnames)
  from source comments, docs, scripts, and test fixtures; tests now use
  RFC5737 documentation ranges.

Added — Agent-era observer (H002)

• Agent traffic classification: every LlmCall carries is_agent_request,
  tool_surface, agent_topology, tool_call_count, tool_names. Every
  AgentTurn rolls up tool_surfaces, tool_call_total, agent_topology,
  suspicious_skills. New tool_surface dimension on llm_metrics.
• Console: agent-aware columns and filters on Agent Turns; Agent breakdown
  section on turn detail; tool-surface facet on Performance.
• Config: [agent_classifier] block in default.toml for tool taxonomy.
• Internal metrics: agent_classifier.unknown_count,
  classifier_mixed_count.

Capture

• Default live-capture configuration now covers common LLM-serving ports,
  reducing the need for explicit CLI capture filters in quickstart flows.

LLM wire-API support

• OpenAI Chat streaming now captures delta.reasoning_content and
  delta.reasoning, with console rendering before normal content.
• OpenCode agent profile detection added for clients that expose a stable
  x-session-affinity anchor.

Agent turn tracking

• Generic fallback turn grouping now requires a tool/function-call anchor, so
  text-only SDK calls stay on the LLM Calls page instead of producing
  synthetic one-call Agent Turns.

Metrics

• TTFT handling now distinguishes streaming and non-streaming calls, with
  stream-only TTFT charts and backfilled rollups from stored call data.
• Dashboard active-resource history added for TCP connections and agent
  turns.
• Long-range chart axes use date-aware labels for multi-day windows.

Console

• Settings page added for capture sources, including interface discovery,
  source editing, grouped source-type controls, and restart flow.
• LLM Calls gained stream/non-stream filtering.
• List pages persist the selected item in the URL.
• Agent-kind filter options are derived from observed data in the active
  window instead of a fixed list.

API

• GET /api/capture/interfaces lists available capture interfaces.
• PUT /api/capture/sources updates capture-source configuration and
  restarts the process when needed.

Documentation

• README reframed around agent observability with refreshed screenshots.
• README quickstart now uses the default live-capture command and no longer
  includes an explicit capture-filter example.
• Removed the LLM call detail screenshot and its README reference.
• Removed project-origin/company copy from public docs.

Development

• Headless PR review workflow added for CI.
• Repository instructions now require PR text to scrub private environment
  details before publication.

What's Changed

• feat(settings): in-app capture configuration + self-restart by @vaderyang in #12
• feat(dashboard): TTFT stream/non-stream split + rollup backfill + filters by @vaderyang in #13
• feat(dashboard): Active TCP Connections + Active Agent Turns charts by @vaderyang in #14
• fix(ts-llm/chat-sse): capture delta.reasoning_content + delta.reasoning by @vaderyang in #15
• feat(filters): dynamic agent_kind dropdown from actual data in window by @vaderyang in #16
• feat(ts-llm/agents): opencode profile keyed off x-session-affinity by @vaderyang in #17
• feat(console/list-pages): persist selected item id in URL by @vaderyang in #18
• feat(ci): headless PR review agent (phase 1) by @vaderyang in #28
• docs(README): reframe as agent observability + Playwright-driven screenshots by @vaderyang in #30
• Clean up README capture docs by @vaderyang in #32
• fix(pr-review): :4000 LiteLLM + auth secret + NO_PROXY + auto-merge trusted PRs by @vaderyang in #31
• fix(pr-review): drop dup no_proxy env key by @vaderyang in #33
• fix(pr-review): override inherited http_proxy so curl reaches LiteLLM by @vaderyang in #34
• fix(pr-review): keep http_proxy; only set no_proxy by @vaderyang in #35
• fix(pr-review): overlay reviewer scripts from default branch by @vaderyang in #36
• fix(pr-review): never APPROVE on agent failure / unstructured body by @vaderyang in #37
• fix(pr-review): feed prompt via stdin to claude --print by @vaderyang in #38
• feat(console): selected_at anchor recovers item window on stale shared link by @vaderyang in #19
• fix(pr-review): brand as "vivi"; never post failures to PR by @vaderyang in #39
• feat(console): TPS instead of TPOT, agent-turns column reorder, sidebar logo by @vaderyang in #20
• ci: opt all JS actions into Node.js 24 runtime by @vaderyang in #40
• feat(console/charts): x-axis tick adapts to time-range duration by @vaderyang in #21
• fix(console/agent-sessions): keep prior data during refetch (no flash) by @vaderyang in #24
• feat(api): lite mode for /calls — unblock mega-turn detail page by @vaderyang in #23
• feat: fold llmproxy duplicate turns by passive pair detection by @vaderyang in #22
• feat(services): Path view + Overview agent charts (deploy roll-up) by @vaderyang in #27
• Fix dynamic agent kind filters by @vaderyang in #29
• Fix generic agent turn grouping by @vaderyang in #41
• Revise unreleased changelog by @vaderyang in #42
• Update repository agent guidance by @vaderyang in #43
• Add local agent workflow skills by @vaderyang in #44
• Add agent-bot: issue triage + wiwi dev agent + auto-merge by @vaderyang in #45
• post_review: use ADMIN_GH_TOKEN for the admin-merge fallback by @vaderyang in #46
• Reject out-of-range time params; clamp SQL timestamp formatter by @vaderyang in #47
• pair_sweeper: CHECKPOINT after each batch + reopen on FATAL by @vaderyang in #48
• agent-bot: use AGENT_GH_TOKEN for the agent:try label fan-out by @vaderyang in #51
• ci: P0 lint gates for secret-ref + validated-constructor classes by @vaderyang in #53
• DuckDB FATAL recovery: rebuild every connection (closes #50) by @vaderyang in #52
• wiwi: drop PAT from actions/checkout; embed in push URL instead by @vaderyang in #54
• wiwi: bump timeout to 120m + stream claude output live by @vaderyang in #55
• wiwi: clear extraheader before push so PAT (workflow scope) wins by @vaderyang in #56
• ci: lint secret VALUES, not just references (catches the '-' bug) by @vaderyang in #57
• wiwi: enforce commit + add auto-commit fallback by @vaderyang in #58
• Rebrand → Heron (Phase 1: rename app binary tokenscope → heron) by @vaderyang in #59
• agent-bot: tolerate transient LiteLLM outages — pre-flight wait + mid-run retry by @vaderyang in #63
• ci: deterministic recovery + schema-migration tests (P1 quality infra) by @vaderyang in #64
• ci(pr-review): bump vivi agent + job timeout to 7200s/120min by @vaderyang in #67
• Rebrand → Heron Phase 2: rename ts-* library crates → h-* by @vaderyang in #66
• chore(rebrand+privacy): repo→heron, scrub infra leakage, leakage gate, heron logo by @vaderyang in #69
• fix(privacy): scrub internal IP left in test fixture (lost in #69 race) by @vaderyang in #70
• chore(release): 0.4.0 — Heron rebrand + quality infra + privacy by @vaderyang in #71

**Full Chang...