[Bindy-lbb]

Relay recipients currently receive generic preceding answers, without an inspectable record of their assigned step or the exact material used. This adds optional operator-authored step requirements and saves the same bounded handoff payload that enters each sequential group turn's model input. The client exposes that saved payload in an expandable note.

Closes #617.

Behavior and acceptance criteria

Acceptance criterion	Implementation and evidence
Explicit per-step requirements	Relay composer, HTTP/IPC/storage contract; selected recipients only, 4 KiB each / 32 KiB total. API and renderer invalid-input tests.
Traceable, persisted input	Versioned payload with current/next step, task excerpt, explicit instruction, completed source identities/timestamps/excerpts and canonical digest. Controlled-driver test checks running, completed and reopened records against injected bytes. Parent-record validation rejects mismatched execution identities even with a valid digest.
Deduplication and bounded context	Exclude relay predecessors before history compaction; share the existing 64 KiB context, 256 KiB final-input and 12-source limits. Prioritize the immediate predecessor; test escaped data, UTF-8 and 32-member relays.
Inspectable client	Collapsible, lazily rendered saved notes with source references and missing-instruction, truncation, omission and redaction notices. Completion is explicitly distinguished from approval.
Failure and isolation	Invalid handoff/budget fails durably before model execution; subsequent steps remain blocked. Existing cancel, owner-stop, workspace-switch, restart and parallel tests cover sequencing.
Compatibility	Personal and parallel execution retain their existing paths. New fields are optional for new readers; runtime packaging allowlists include both new modules.

This is phase one of collaborator context: sequential handoff notes. It does not add delegation, parallel dependency scheduling, cross-task memory, permission transfer or extra model summarization calls.

Validation ledger

Expected outcome: all required checks pass before merge. All published GitHub checks now pass, including Node 24 on Ubuntu/macOS, native staging/installers, layout parity and security checks. The local dependency-audit result below remains recorded as observed. Independent review is still required.

Check	Actual result
Build, updater preparation, UI and renderer type checks	Passed.
Script/runtime inventory tests	233 passed.
UI package tests	53 passed.
Full server suite	968 passed, 4 skipped in an isolated network namespace with DAC bypass capabilities removed; command below.
Server affected suites	38 passed; after adding parent-identity assertions the group suite passed again, 16/16.
Renderer suite	Final full run: 1,249 passed across 87 files. Initial run exposed a wrong lingering-dropdown selector in the new test; corrected and reran both affected files (38/38) and the full suite.
Desktop main + renderer production build	Build passed; 568/569 tests passed initially. Existing PID-reuse sentinel cleanup test failed once; the entire packaged-smoke file passed on rerun (39/39).
Security audit	Fails: existing Vite 6.4.3 dependency, 1 high-severity advisory aggregate. Dependency manifests/lockfile for Vite are unchanged by this PR.
Browser QA	Passed on real GroupsPanel and production CSS with a deterministic mocked IPC fixture, Chromium at 1440×1000 and 480×900. Requirement submission, saved-note expansion/collapse and source display checked. Zero browser errors; narrow viewport scroll width equals viewport width. This is renderer QA, not a live-provider or packaged OS acceptance claim.
Patch whitespace	git diff --check passed.

The raw npm run check stopped at the server suite: the shared container exposes only ports 48800–49100, causing 165 EADDRINUSE failures; its root privileges also defeat an existing chmod-based hire failure fixture. Lower concurrency alone did not resolve this. Without changing shared host settings, the complete server suite passed with:

unshare --net sh -c 'ip link set lo up && setpriv --bounding-set=-dac_override,-dac_read_search node --test --test-concurrency=2 --test-timeout=120000 apps/server/dist/test/*.test.js'

Remaining checks were run separately after the raw check stopped. No coverage percentage is claimed; behavioral assertions were added for the changed execution, persistence, validation and UI paths. Native macOS/Windows packaging checks are delegated to repository CI.

Compatibility, risks and rollback

• Deploy server, desktop IPC and renderer together. Older strict readers can reject records with new fields; the client does not silently retry by dropping requirements.
• Excerpts remain untrusted public output. Filtering removes recognizable credentials/private reasoning but is not a general DLP guarantee. Hashes establish consistency, not authenticity or business acceptance.
• A request that leaves insufficient space for required handoff data fails explicitly; it does not run with silently missing requirements.
• Revert the implementation to roll back code. Use a pre-upgrade workspace backup or a separate workspace when downgrading strict readers; do not rewrite historical evidence to simulate compatibility.

Design and operational notes: relay-handoff-notes-v1.md, thread-context-and-collaboration.md. Changelog updated.

Review focus: accepted-source identity checks, exact saved/injected payload, budget fitting and omission accounting, durable pre-execution failures, and old-reader compatibility.