@hybridaione/hybridclaw 0.22.0 → 0.24.0

package/AGENTS.md

@@ -171,8 +171,11 @@ should verify no broken references.
 
 ### Prerequisites
 
-- Node.js 22 (matches CI and `engines` field)
-- npm
+- Node.js 22 (matches CI and the `engines` field)
+- npm 11.10+ — run `corepack enable` so repo commands use the `packageManager`
+  pin (`npm@11.10.0`). Contributors need this version because npm's
+  `min-release-age` supply-chain gate (see `SECURITY.md`) only takes effect on
+  npm 11.10+; it is deliberately not enforced on end users via `engines.npm`.
 - Docker when working on container-mode behavior or image builds
 
 ### Common Commands
@@ -347,11 +350,16 @@ Skill resolution order (first match wins):
 When the user says "bump release":
 
 1. Bump the requested semantic version (if unspecified, default to patch).
-2. Update version strings in:
+2. Update `package.json` to the new version, then run `npm run version:sync`
+   to propagate it through product package metadata:
    - `package.json`
-   - `package-lock.json` (root `version` and `packages[""]`)
+   - `package-lock.json` and `npm-shrinkwrap.json` (root `version`,
+     `packages[""]`, and product workspace entries)
+   - `console/package.json`
+   - `desktop/package.json`
    - `container/package.json`
-   - `container/package-lock.json` (root `version` and `packages[""]`)
+   - `container/package-lock.json` and `container/npm-shrinkwrap.json` (root
+     `version` and `packages[""]`)
    - any user-facing version text (for example `src/tui.ts` banner)
 3. Move `CHANGELOG.md` release notes from `Unreleased` to the new version
    heading (or create one).

package/CHANGELOG.md

@@ -2,6 +2,211 @@
 
 ## Unreleased
 
+## [0.24.0](https://github.com/HybridAIOne/hybridclaw/tree/v0.24.0) - 2026-06-11
+
+### Added
+
+- **Human distillation (R72)**: New `hybridclaw coworker` command group and
+  bundled `human-distill` skill that distills a real person's source material
+  into a coworker agent. Collectors normalise Slack exports, mbox email,
+  meeting transcripts, chat JSONL, documents, and gap-driven interview
+  questionnaires into an agent-scoped corpus with quality weighting, stable
+  provenance ids, and third-party PII masking at ingest. A resumable
+  ingest → analyse → build → merge → correct pipeline writes the persona into
+  the standard identity files (`IDENTITY.md`, `SOUL.md`, `USER.md`, `CV.md`)
+  and a generated work-module skill, with `/admin/distill` managing subjects,
+  consent, source uploads, corpus documents, and runs from the browser. Every
+  claim cites corpus documents (uncited claims are flagged, not written),
+  every merge is an F4-versioned reversible edit, and conflicting evidence is
+  surfaced as operator review items. Distilling a real, named human is
+  hard-blocked until a consent artefact is recorded; all lifecycle actions
+  emit hash-chained `distill.*` audit events; `coworker forget` erases corpus,
+  persona, work module, runs, and revision snapshots as one identifier set.
+  Includes a leakage/fidelity eval (`coworker eval`), conversational
+  corrections (`coworker correct`), and one-bundle multi-host export/install
+  for Claude Code, Codex, OpenClaw, and HybridClaw. Manifesto: Principle VII -
+  A coworker you can trust with real responsibility.
+- **Cloud memory sync**: Agents now sync local memory files (`MEMORY.md`,
+  `USER.md`, and recent daily memory notes) with the HybridAI cloud and receive
+  shared installation- and company-scoped memory back for prompt context. Sync
+  runs at conversation start and periodically every five minutes with
+  per-agent rate limiting, requires `HYBRIDAI_API_KEY`, `HYBRIDAI_BASE_URL`
+  (HTTPS only), and `HYBRIDAI_CHATBOT_ID`, and stays disabled when those are
+  unset. Shared memory appears read-only in the agent file editor under a
+  "Shared memory" group.
+- **A2A operator pairing**: Added `/admin/a2a-trust` for pairing two HybridClaw
+  instances: operators fetch a peer Agent Card by URL or canonical DNS
+  identifier, preview its identity and key fingerprint, and trust it with an
+  optional peer-side approval prompt. Incoming pairing requests arrive through
+  a rate-limited `/a2a/pairing/requests` endpoint and can be approved or
+  declined from the console with audit-trail decision reasons.
+- **Inbound A2A envelopes**: Added a JSON-RPC Agent Card envelope receiver for
+  cross-instance A2A delivery, including canonical sender/recipient metadata,
+  idempotent persistence, signed bearer-token validation, read-only admin
+  inbox visibility, and audit events for malformed or rejected envelopes.
+- **HybridAI proxy agents**: Agents can now proxy conversations to hosted
+  HybridAI chatbots via per-agent `proxy` config, SecretRef-backed API keys,
+  HTTPS-only upstreams, selectable channel- or user-scoped conversation ids,
+  streaming response forwarding, and `/status` visibility for proxy mode.
+- **Explicit agent addressing**: Chat channels and web chat can address
+  specific agents inline, with mention autocomplete, avatar-backed mention
+  pills, canonical recipient handling, and fanout hardening for local
+  proactive delivery.
+- **`byd-battery` skill**: Added read-only monitoring for BYD Battery-Box
+  Premium HVS/HVM/LVS/LVL home-storage systems over local Modbus TCP or
+  Fronius inverter delegation, covering state of charge, pack telemetry, cell
+  extremes, tower/module inventory, decoded alarms, firmware info, and energy
+  counters, with allowlisted register ranges and no write operations.
+- **`mailchimp` skill**: Added Mailchimp Marketing and Mailchimp
+  Transactional/Mandrill workflows for credential checks, audience/member
+  reads, guarded subscriber and tag mutations, campaign draft/content/report
+  operations, automations, journeys, and approval-gated campaign or
+  transactional sends through the gateway HTTP proxy.
+- **Amber approval cards**: Web chat approval prompts now render as structured
+  confirmation cards with an approval-tier badge, parsed action/tool/reason
+  detail rows, and separated confirm/deny and trust-scope button groups.
+- **Named local model endpoints**: Local provider config now supports
+  additional named Ollama, LM Studio, llama.cpp, and vLLM endpoints with
+  endpoint-prefixed model ids, CLI setup via `--name`, and per-endpoint model
+  behavior flags for Qwen thinking markup and Gemma tool-call formats.
+- **Auxiliary model testing**: Added `/aux test <task> <prompt>` so operators
+  can exercise configured auxiliary model routes directly and see the
+  provider/model used for the request.
+- **Response rating forwarding**: Thumbs up/down ratings on web chat responses
+  are forwarded to the HybridAI feedback API when HybridAI authentication is
+  active, alongside the existing local rating store and audit events.
+  Forwarding is non-blocking and skips silently when auth is unavailable.
+
+### Changed
+
+- **README and skill docs positioning**: The public README and skills docs now
+  lead with HybridClaw's validated business-skill workflow: helper-backed
+  production skills, eval scenarios, approval tiers, credential boundaries, and
+  `Qwen/Qwen3.6-27B-FP8` as the small-model validation baseline.
+- **Multi-agent and credential-isolation positioning**: The README and docs
+  now call out multi-instance A2A workflows, hosted proxy agents, explicit
+  addressing, and SecretRef-backed execution that keeps raw credentials out of
+  model context.
+- **HybridAI Cloud launch path**: The README, docs landing page, and
+  installation guide now link to the managed HybridClaw cloud offering at
+  `hybridclaw.io`.
+- **Desktop release order docs**: The macOS desktop release guide now
+  recommends building and notarizing from the exact version tag, uploading
+  desktop assets to a draft GitHub Release, then publishing after assets are
+  verified.
+- **Provider request payloads**: Empty tool definitions are omitted from
+  HybridAI, OpenAI-compatible, Codex, and Ollama provider requests instead of
+  sending empty `tools` arrays.
+- **Local vLLM tool behavior**: OpenAI-compatible local providers infer and
+  remember native-tool fallback support, count prompt-side tool overhead in
+  context guards, refresh named-endpoint metadata, and parse Gemma text tool
+  calls emitted before or after Markdown wrappers.
+- **Structured actors in audit data**: Audit/event records now carry unified
+  actor identities across A2A envelopes, board cards, adaptive-skill
+  observations, scoreboards, and structured audit queries.
+- **Coworker liveness scan**: Skill scans during coworker liveness checks use
+  set-based agent filtering, speeding up gateways with many agents.
+- **Release version sync tooling**: `npm run version:sync` now keeps root,
+  console, desktop, container, lockfile, and shrinkwrap package versions in
+  sync, with `release:check` validating the same invariant.
+- **Dependency updates**: Routine minor and patch dependency updates across the
+  gateway, console, container, and desktop packages, plus overrides that lift
+  transitive `shell-quote` and `tmp` to patched releases flagged by npm audit.
+
+### Fixed
+
+- **Gateway token timing safety**: Gateway API and bearer token checks use
+  constant-time comparison to avoid timing side channels.
+- **Delegation identifiers**: Delegation session and batch job identifiers use
+  cryptographic UUIDs instead of seeded pseudo-random strings.
+- **Containment check stalls**: Media and artifact path containment checks
+  resolve real paths asynchronously so large directory validation no longer
+  blocks the gateway event loop.
+- **TUI ANSI truncation**: Terminal output truncation handles incomplete ANSI
+  escape sequences and wide glyphs without corrupting styled text.
+- **Duplicate hatching after onboarding**: Switching agents after onboarding
+  no longer re-triggers the workspace bootstrap kickoff, and bootstrap job
+  detection recognizes both bulleted and numbered job lists.
+- **Host runtime dependency detection**: Source-checkout host runtime checks
+  no longer misreport container dependencies as missing when npm hoists a
+  package whose `exports` map does not expose `package.json` (e.g.
+  `dompurify`).
+- **Installer and setup hardening**: The one-line installer, postinstall
+  container setup, Node version guard, and Homebrew/source-checkout paths handle
+  no-sudo installs, user npm prefixes, and container dependency setup more
+  reliably.
+
+## [0.23.0](https://github.com/HybridAIOne/hybridclaw/tree/v0.23.0) - 2026-06-09
+
+### Added
+
+- **`blink` skill**: Added Blink camera and video-doorbell workflows for
+  OAuth v2 login with PIN handoff, device/network/camera inventory, motion
+  clip listing and artifact-backed downloads, guarded arm/disarm and motion
+  control plans, and fresh-thumbnail refresh handling that keeps Blink media
+  bytes out of model context.
+- **`hue` skill**: Added Philips Hue Bridge support for local CLIP v2 reads,
+  bridge-link credential capture, self-signed bridge TLS handling, guarded
+  light/group/scene/behavior changes, Remote API token refresh support, and
+  LAN policy diagnostics for Hue Bridge requests.
+- **Fleet topology admin UI**: Added `/admin/fleet-topology` for viewing the
+  local A2A instance identity, checking trusted child instances through their
+  Agent Card URLs, and adding, editing, or removing A2A trust-ledger peers from
+  the console.
+- **Admin secrets console**: Added `/admin/secrets` as a write-only secret
+  manager that lists set and declared-but-empty secrets by metadata, supports
+  overwrite and unset actions, and never returns cleartext secret values to the
+  browser.
+- **Chat code rendering**: Web chat now syntax-highlights completed code
+  blocks, shows language labels, and provides a touch-reachable copy button
+  with success feedback while skipping highlighter work for actively streaming
+  messages.
+- **Sentry error reporting**: `hybridclaw env set SENTRY_DSN <dsn>` enables
+  optional gateway Sentry reporting for startup failures, uncaught exceptions,
+  unhandled rejections, and errors recorded through shared gateway/agent spans,
+  with default `production` environment, automatic
+  `hybridclaw@<package-version>` release naming, secret redaction, and graceful
+  shutdown flushes.
+- **Scheduler heartbeat polling action**: Config-backed scheduler jobs can use
+  the explicit `heartbeat_poll` action kind so empty `HEARTBEAT.md` files are
+  skipped before any model turn is started.
+
+### Changed
+
+- **Installer and npm policy alignment**: The bootstrap installer works more
+  cleanly on fresh Debian/Ubuntu hosts and no-sudo system Node setups, while
+  contributor docs, CI, Docker, and package metadata pin npm 11.10+ without
+  forcing consumer-facing `engines.npm` warnings.
+- **Second-opinion TUI formatting**: `/second-opinion` output in the terminal
+  wraps long model-comparison and validation responses instead of spilling past
+  the viewport.
+- **Skill setup guidance**: Skill authoring docs now require chat-friendly
+  `/env` and `/secret` setup alternatives alongside local `hybridclaw env` and
+  `hybridclaw secret` commands.
+- **Web agent hatching kickoff**: Switching to an agent with an active
+  `BOOTSTRAP.md` in web chat now sends a hidden kickoff turn so hatching starts
+  immediately while the visible slash-command response remains local command
+  output.
+
+### Fixed
+
+- **Interrupted agent shutdown output**: Container/runtime shutdown output now
+  distinguishes expected interrupted-run signal errors from real runtime
+  failures.
+- **Slash autocomplete flags**: TUI slash-command autocomplete keeps literal
+  argument and flag completions instead of rewriting them.
+- **Chat delete guard**: Deleting browser chat sessions is blocked while a run
+  is still active for that session.
+- **Malformed Unicode in prompts**: Container utilities sanitize malformed
+  Unicode so provider prompts and local OpenAI-compatible requests do not fail
+  on invalid surrogate data.
+- **Trace export identity preservation**: Session trace exports preserve trace
+  hash identifiers while applying secret redaction.
+- **Web approval buttons**: Approval buttons now emit gateway-supported
+  commands: `Allow once` sends `/approve yes`, `Allow always` sends
+  `/approve all`, and scoped buttons send their matching `session` or `agent`
+  approvals.
+
 ## [0.22.0](https://github.com/HybridAIOne/hybridclaw/tree/v0.22.0) - 2026-06-05
 
 ### Added