Release 0.4.0

[kostandinang]

TL;DR

Ships side-chat: unified mutation surface, shared patch list, impact-tier routing, and an agent classifier that triages high-impact downstream review.

📹 Walkthrough: https://www.loom.com/share/de8f107613084a868df8b7697d64308f

@hashintel/brunch@0.4.0 — already on npm (latest); tag 0.4.0 on origin. This PR lands the package.json bump that direct-push-to-main was blocked from delivering.

Subject to evolve. This is the last shape where the side-chat is a separate surface. The side-chat popover, Pending review section, and top-bar patch list will fold into a unified chat surface — sub-conversations and the review queue rendered inline as collapsibles, reconciliation as a target-grouped, async-by-default thread with a "Reconcile Now" trigger. See docs/design/CONVERSATIONAL_WORKSPACE_RUNTIME.md.

Side-chat — what landed

Surface

• Popover anchored to rows; draggable into the right-side panel.
• Item-level pins (chat-with on a row), span-level pins (text selection → floating 💬 Chat / 📝 Annotate), and multi-pin reasoning.

Explore mode

• Free-form chat against pinned context. Streamed reply, no staging. Volatile (doesn't persist across sessions).

Annotate

• Span-level durable note. Stages into the shared patch list; persists with inline content tint, action-rail badge, hover excerpt. (Ignore until the context injection from the unified chat happens, right now it doesn't make any sense)

Shared patch list & Apply

• Direct row edits and chat-driven edits funnel into one canonical patch list (N change(s) · Undo · Apply) routed through one apply endpoint.

Impact-tier routing

• Soft (≤2 downstream, no active-review-set member): lands directly with a refresh-hint toast.
• Hard (3+ downstream, or any anchor/affected item in an active review set): opens the Pending review section.
• Active-review-set override forces Hard routing even on small edits to protect un-actioned review cards.

Edit mode

• Side-chat Edit mode toggle (placeholder flips to "Suggest an edit…").
• Word-level before/after diff on staged rows (warm tint removed, cool blue added).
• Frozen source snapshots on Pending review rows — ↗ view source diff popover stays stable under further source edits.
• Resolve is idempotent — clears the row without mutating the target.

Agent classifier on Pending review

• Run agent walks each row: queued → classifying → classification.
• Four classes with per-row actions:
  • auto-confirm → Confirm (resolve only)
  • auto-edit → View proposal → Apply (edit + resolve) / Skip (resolve only)
  • substantive → Open side-chat with target pinned
  • failed → Re-run
• Bulk header actions: Confirm all (N) over auto-confirm rows, Apply all suggested (N) over auto-edit rows (serial).

Re-entrant cascade

• Edit target on a Pending review row stages a fresh edit; if itself high-impact, opens new Pending review rows for its downstream items. Queue absorbs new rows without duplicating existing open ones.

Propose-edge

• In Edit mode with two items pinned, model can propose typed relations: depends_on, derived_from, constrains, refines, verifies.
• Edge applies persist a relation chip without manual refresh.
• Invalid relation kinds (e.g. context↔constraint) return policy feedback and do not persist — surfaced rather than silently dropped.

Drill-down [Deferred]

• In Edit mode with a single open-phase item pinned, model can persist a detail-focus intent attached to the anchor's reference code.
• Doesn't mutate the spec; steers the next frontier turn in the relevant open phase toward the anchor.
• Rejected with policy feedback when the anchor is in a closed phase.

Other changes shipped in 0.4.0

• Multi-chat substrate (FE-691) — chat table + reconciliation_edge foundation that the next-shape conversational runtime builds on.
• Decouple turns from specifications (FE-697) — interview turns no longer 1:1 with specs; enables multiple chats per spec.
• Prompt + context substrate (FE-698) — prompt registry, observer context-pack foundation, scenario runner + observer scenario composition, agent mutation audit / tool-surface design.
• Direct in-place editing on the structured-list view (FE-657) — row-level edit affordance that feeds the same shared patch list.
• Graph view — split-button chip toggle (FE-671), re-land of graph view PRs orphaned via graphite-base/73 (FE-655).
• Interview stability — grounding accepts option-only responses (FE-693), streamed questions + dev config stabilization (FE-650).
• Observer — improved capture of knowledge-graph items (FE-639).
• Bug fixes — restore tools-panel flicker fix lost in FE-616 refactor (FE-622).
• Design docs — conversational workspace runtime design (FE-708), agent mutation surface design.
• Dev experience — /planning-pr skill encoding planning-PR convention (FE-703), branch prefix derived from gt user branch-prefix in workflow docs (FE-651), experimental warning banner on README (FE-696).
• Dependencies — vitest 4.1.5, oxlint 1.61.0.
• Schema — 7 new Drizzle migrations (0013_annotation → 0019_reconciliation_need_agent_columns).

Test plan

• CI status checks pass
• npm view @hashintel/brunch@0.4.0 resolves with expected bin + drizzle files

🤖 Generated with Claude Code

[cursor]

PR Summary

Low Risk
Low risk: this PR only bumps the package version in package.json and package-lock.json with no functional code changes.

Overview
Release bump only. Updates @hashintel/brunch version from 0.3.0 to 0.4.0 in package.json and package-lock.json.

^{Reviewed by Cursor Bugbot for commit f4ef577. Bugbot is set up for automated code reviews on this repo. Configure here.}

[augmentcode]

🤖 Augment PR Summary

Summary: Prepares the @hashintel/brunch 0.4.0 release by landing the release-it generated version bump that couldn’t be pushed directly to main due to branch protection.

Changes: Updates package.json version from 0.3.0 to 0.4.0 (lockfile updated separately).

_{🤖 Was this summary useful? React with 👍 or 👎}

[augmentcode]

Review completed. No suggestions at this time.

Comment augment review to trigger a new review at any time.