Historical record

This page records what happened on a date, with its observed failures and limits. It is not a requirement for the current implementation. Current behaviour is in Status.

Headless collaboration acceptance

Observed: 2026-10-16. Result: headless collaboration accepted. This does not accept browser UI, native input, accessibility or a signed product release. The machine-readable record freezes the four installed candidates, source/tree/input/lock/package identities, selected owners, accepted Core artifact and evidence digests. The protocol guide defines the implemented contracts and budgets.

Fixed composition

• Rho implementation: 3f406d346909779be49f8335af5bde4fdaf83590.
• Plugins: 3040ef43cddaf55d15b750bc5eb558e0a0d954f9, published on owner main before the application reference.
• Accepted public Core: d0159da6feb01d1be9e3b4b34b67953241fedfcb, reused without rebuilding. Binary SHA-256: eee223026f7e0b4dd1ee14a409c023919626c4175be5728048a02af739af4ce1.
• Environment: macOS arm64, Node 22.22.3, npm 10.9.8, Python 3, external R 4.5.2, Claude Code 2.1.302, MCP SDK 1.29.0.

All packages retain version 0.4.0; commit and artifact digests, not that version alone, distinguish these candidates from the historical eight-owner 0.4 baseline. The current Core checkout is not selected. No dirty build-input override was used. Later acceptance-only documents do not replace or rebuild these fixed tarballs.

Candidate	Physically installed owners	Result
Full	Documents, Files, R, Requirements, Tables, Media, Charts, Agent, Annotations	Deterministic and actual Claude collaboration pass
Collaboration-minimal	Documents, Files, Tables, Charts, Agent, Annotations	Both collaboration variants pass without R/Requirements/Media plugins
Documents/Files	Documents, Files	Reduced installation, Unicode draft/export and unavailable-owner checks pass
Files/R	Files, R	Exported-object and unavailable-action checks pass without a Host

Every installation is a real, non-symlink npm installation outside all source checkouts. Each includes Contracts and only the dependencies of its selected owners. All nine owner tarballs also pass independent installation/import with Contracts and their declared external dependencies. Shipped operation declarations and Skill digests are checked from the installed entries. No source-directory fallback, hidden aggregate package, global Skill installation or registry publish is part of this result.

Executed checks

Source regressions were run sequentially: Plugins 14/14 and Rho 24/24 pass. The v1 schema and fixtures remain unchanged. Checks cover declarations and Skills, per-project capabilities, live MCP filtering, authorization/project boundaries, conditional writes, retained receipts, composition protection/undo, scoped selection, exact UTF-16 and stable-entity anchors, budgets and archive validation.

For both Full and Collaboration-minimal, the installed packaged acceptance script ran twice: deterministic SDK stdio MCP and actual Claude Code. Each scenario uses an independent HTTP human and the real shared MCP adapter, then checks owner state:

1. The Agent discovers shipped Skills, reads ordinary typed CSV and prepares a scatter plot, then composes code/table/plot/notes views with stable IDs.
2. The human edits the shared buffer, rearranges views, protects the table and submits document-range and table-entity feedback from Documents drafts.
3. The source changes after feedback preparation. The Agent fetches the frozen context, conditionally updates the same notes and submits explicitly linked replies. Server authors, fetched Skills, source snapshots and reply relations are asserted independently of the model's success text.
4. A service restart preserves versions/content/protection. A retained-cursor gap is explicit; two observation clients read matching owner snapshots. Format-3 restore preserves acknowledged documents, feedback, replies and protection.
5. Instrumentation asserts Host starts = 0, native dispatches = 0, including restart and restore. No native operation record is created.

Actual Claude candidate	Compose	Respond	Native counts
Full	Exit 0; 73,949 ms	Exit 0; 67,938 ms	0 Host / 0 dispatch
Collaboration-minimal	Exit 0; 72,026 ms	Exit 0; 56,068 ms	0 Host / 0 dispatch

All four model phases use the installed MCP adapter and Claude Code 2.1.302. Each phase has a 600-second timeout and a 24-turn limit. Model unavailability or failure is a failed run, never substituted with a scripted response.

A supplemental installed HTTP/SSE driver also passes: two separately authenticated live observers converge after a Unicode edit; publication failure forces both to resnapshot acknowledged content; retained scoped credentials reconnect after restart; an actual HTTP cursor-gap rejection causes snapshot-first recovery. An injected writable-backlog metric exercises subscription termination at its byte limit over a real HTTP stream. This is a deterministic boundary check, not a kernel/socket slow-reader performance benchmark. Host/native counts remain zero.

Separate native branch

test/native-collaboration.test.mjs passes in the source suite and against the installed Full candidate, using a new disposable project and an owned Host. The external driver changes only imports to installed absolute module URLs; both driver digests are recorded. It asserts:

• One explicit r.execute produces exactly one native dispatch and one entry in the script's output file.
• Closing its code view does not cancel the task. Repeated submission retains the same Core request identity and adds no dispatch.
• Killing only the owned Host leaves the original operation unknown; delivery of the same request does not create or execute a replacement.
• The retained owned R process group is cleaned up. No user Host/session is restarted or stopped.

Native execution is not evidence of the collaboration path and is not included in its zero-dispatch count.

Recovery evidence and limits

Format 3 is explicitly selected. Existing format-1/2 readers remain; lossy exports to older formats fail. Missing owners and conflicting content are reported with the source archive retained. Restore imports no credentials/grants, installs no plugin and replays no native task. A hash-valid archive containing an invalid reply relationship is rejected before any document import; owner restore also rejects malformed relationships without replacing its existing state.

Restore is not one cross-owner filesystem transaction. A later operational I/O or budget failure can leave earlier successful imports in place; keep the source archive, inspect owner facts and reconcile explicitly. No universal rollback or crash-atomic multi-owner import is claimed. Existing native crash/continuity and external-write tests retain their own scope; this run is not an exhaustive fault/performance campaign.

Development failures remain recorded locally: an early SDK resolver used a symlink location instead of its real package path; a concurrent Plugins rebuild temporarily removed linked declarations; additional tests exposed acceptance of a malformed reply archive. Fresh sequential regressions pass after those fixes. Supplemental observation-driver attempts also failed with an invalid undo-group ID and a synthetic corked-response assertion; the final driver uses a valid ID and an explicitly injected backlog metric. Those failed attempts are not counted as passes or represented as realistic network-load evidence.

Local evidence and reproduction

Ignored work/collaboration-candidate-* directories held the fixed tarballs and composition inventories; that superseded evidence was removed on 2026-10-26 at the user's request, and the surviving work/collaboration-install-locations.json still records the external install roots those trials observed. Re-running the packaged driver below produces the trial-deterministic/report.json and trial-agent/report.json records again. Source test logs and installed-run logs use the work/collaboration- prefix. The published JSON contains sanitized projections and hashes; private connection files, credentials and raw model transcripts stay local and are not published.

From a new directory outside the checkouts, with the selected Node and Claude clients available, install one exact candidate and run its packaged driver:

npm install --ignore-scripts --no-audit --no-fund /absolute/candidate.tgz
PACKAGE="$PWD/node_modules/@rho-workspace/local"
env -u RHO_PACKAGE_ROOT -u RHO_CORE_BINARY \
  node "$PACKAGE/scripts/collaboration-acceptance.mjs" "$PWD/trial-scripted"
env -u RHO_PACKAGE_ROOT -u RHO_CORE_BINARY RHO_COLLAB_MODEL=1 \
  node "$PACKAGE/scripts/collaboration-acceptance.mjs" "$PWD/trial-agent"

The final conclusion is limited to headless collaboration accepted. Browser rendering, input/accessibility, native windows, remote multi-user operation, model-loop hosting, npm registry publication and signing remain unaccepted or out of scope. The existing uncommitted design work is not part of the accepted build.