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 Changelog: v0.2.0...v0.4.0