Updated: 2026-10-27. Agent 0.4.10 connects Claude and Codex through shared ACP v1 identity, permission and recovery handling. External clients can select the stdio relay to expose bounded observations while retaining their input and runtime. Advertised session close releases native writers; process-group escalation survives wrapper exit. Standard diff paths supply file references without retaining contents. All 89 owner checks pass. Rho's installed Westlake HPC acceptance passed eight turns for each Agent plus external Codex load, live observation across Rho restart and same-session reconnect; the record keeps native-job cancellation limits, failed attempts and installation evidence. Core and the native collector are unchanged; uninstrumented clients still need the transcript fallback.
Updated: 2026-10-25. Agent 0.4.5 adds rho-agent-observe: selected Claude Code
JSONL and dsh read-only native persistence feed Core's public activity CLI.
Immutable selection, original call IDs, prefix verification and retained cursors
allow late attachment and observer restart. Deterministic summaries preserve
tool return, native exit and unknown outcome separately; optional bounded output
inspection supplies current hashes without claiming historical production.
Four actual WestlakeHPC runs covered both clients: analysis backfill yielded
14/28 events, live attachment yielded 2/6 final events, and collector/owned Host
restart preserved the original invocation (one run marker each). Repeated scans
added zero events. Build and all 74 owner tests passed after the explicit native
dependency correction; actual evidence and limitations are in
Agent observation acceptance.
Rho browser composition, user installation and Linux observation acceptance remain
separate. Existing application observation/feedback entries are unchanged.
Updated: 2026-10-24. Annotations 0.4.3 adds an optional table-question UI using the single Documents buffer, exact source/row anchors, original submission identities and explicit Agent replies. Replies may name up to 16 exact output refs, resolved through Files; these author links do not establish domain correctness. UI intent retention is capped at 16 KiB; exceeding it preserves the draft and refuses a new submission. A submitted question folds to one status control; a host may place the question's actions in the view's own header and keep only the open thread below. Assigned task feedback keeps its existing delivery entry. Agent 0.4.4 adds document-open identity, resource id/version and output-stream metadata to its bounded activity log, without tool inputs, code or row/output contents. Owner checks pass 67/67, including retry/archive links and bounded metadata after restart. Rho owns the composed browser/real-model acceptance; S6 remains human unverified.
Updated: 2026-10-23. Agent now defaults to four tools: owner discovery, selected contract resolution, invocation and retained-result reading. Large results have scoped references with bounded JSON-pointer and Unicode text pages, explicit release, fresh authorization checks and no replay. Files streams hashes/snapshots and reads ranges through a bounded digest/fingerprint cache. Media negotiates the owner page size, enforces a display byte cap and cancels inactive reads. Documents supplies metadata-only summaries, bounded record/aggregate reads and conditional browser saves that flush pending edits and preserve retry identity. Media description leaves browser rendering unobserved; it does not claim the UI is unavailable. Owner checks pass 65/65; focused media checks pass 5/5. Rho records the composed real browser, actual R execution and Claude Code trials separately.
Updated: 2026-10-22. Documents and Tables now draw their own bodies in an
application's item frame, following the Paper page
04 · 插件形态.
Both ./ui entries export a card body and form declarations (reference, card,
summary, compare); compare is reached by naming a change or a recorded comparison,
not by zooming. The Agent names show; without it Tables shows a count when an
earlier table is given, else rows, and Documents shows a passage for an anchor, a
change for a named group, else the outline. People may switch show on the card.
Size caps are form parameters with maxima (Tables 3×3 card, 12×6 summary, 50 rows
and 12 columns; Documents one section, three changes, ten at most), and each body
reports what it left out, counted from the whole resource. A missing comparison
reads "No comparison recorded yet". Null cells read "empty", since a null may be a
genuine NA. Source checks pass 53/53, with seven new DOM-free card-model tests on
real owners. Browser files named ui*.tsx may now use the optional UI peers.
Rho's composed browser check is recorded in Rho; versions are unchanged.
Updated: 2026-10-22. Files 0.4.1 provides an optional finite source-folder resolver; Documents 0.4.2 uses the injected file-path port for real reads, conditional saves and restoration checks. R 0.4.1 can freeze a saved script through the same file service. Unknown folders, traversal, internal source state and symlinks refuse; equal filenames remain distinct and overlapping aliases share a buffer within one workspace. Source checks pass 46/46, including actual two-folder reads/saves, external-write conflict and saved-script planning. Rho selects and persists folders; these owner checks do not accept its UI or installation. No Core changes or scientific refits were needed. Annotations 0.4.2 accepts both selected Documents 0.4.1 and 0.4.2 as optional peers; its behavior is unchanged.
Updated: 2026-10-21. Application phase 5 now accepts the selected external offline
grades candidate with Agent 0.4.3 and the existing Documents/Files/Tables/Annotations
owners. Actual Claude Code 2.1.306 / Westlake HPC DeepSeek completed two original
business sessions and two restricted presentation sessions: 77 tool calls,
1,098,713 input tokens including cache and 29,705 output tokens. The application
records 32 installed checks, real browser/worker/save-ACK faults and frozen
AC01–AC21 coverage. Earlier interrupted-trial calls/usage lower bounds are retained;
no provider invoice or exactly-once native execution claim is made. Core remains
unchanged; same-project multi-task, full product acceptance and registry publication remain
unselected. Local evidence belongs to Rho's
work/phase5-installed-dual-agent-20261021/results.html.
Agent 0.4.3 adds explicit three-tool MCP profiles, bounded complete-body cache,
closed schema checks and original managed-intent reuse. The optional read-only
rho.agent-note displays at most 600 Unicode characters; Rho writes its author
and binding basis through the assigned guarded path. Generic rho_call feedback
fetches retain exact native return attribution. Headless compatibility and optional
React isolation remain; the catalog now lists the actual note view. Source owner
checks pass 45/45, with Rho 63/63 and real two-project SDK/browser evidence recorded
by the application. No model ran in those source/SDK checks; the separate phase-5
model and installed evidence is summarized above. See layered MCP.
Agent 0.4.2 adds exact original submissions, frozen targeted contexts, one durable bounded delivery file, whole feedback reads/status and optional Claude hook/worker entries. ACK loss permits only original lookup; persisted hook attempts never re-emit after ambiguous crashes. Native MCP returns carry exact fetch attribution. Annotations 0.4.1 adds the contextual question entry and exact retained thread contract. It reuses Documents 0.4.1 buffer-readiness notification and Tables 0.4.2 confirmed-selection action; no duplicate editor or domain workflow is introduced. Private schema refinement retains ACK/hook-boundary facts without changing frozen public profiles. Component regressions pass 44/44; Rho 57/57 and actual two-project browser checks have separate application evidence. This phase used zero model tokens. Real Claude consumption, presentation enhancement and installed acceptance remain pending. See Agent feedback.
Tables 0.4.1 adds an optional DOM-observation callback and pure coverage schema. It reports exact visible keys/columns, typed-value digests, query and actual viewport/hit-test coverage after data paint. Virtual buffers, sticky-header-covered rows and native hidden tabs do not claim visible cells. Headless readers remain unchanged; no Agent model or Core runtime is required. Owner checks pass 43/43; Rho records actual split/refresh/restart/hidden/occlusion checks separately. Reliable feedback, presentation enhancement and the full installed combination remain work.
Agent 0.4.1 adds pure/native/Claude stream-observer entries, registered attachment and bounded original-source observations. Separate adapter authority and the application project writer protect ingestion; no model loop, queue or execution journal is introduced. Existing Files/Tables supply artifact validation unchanged. The Agent guide records limits and scope. Owner source checks pass 43/43. Application source acceptance ran two real Claude Code 2.1.306 / Westlake HPC DeepSeek projects: an explicit native resume repaired one manifest path without rerunning either Python computation. Application/usage evidence and remaining phases belong to Rho; no independent installation or publication of the candidate is claimed.
Documents and Tables now ship separate ./ui entries with exact optional peers;
headless imports load no React or DOM. The existing confirmed design uses
CodeMirror serial identified input, authenticated tab author labels, owner undo,
individual Agent-change revert, recovery copies and separate exact-version reads.
AG Grid Community requests 200-row blocks, caches five blocks, supports one sort
and typed filter, and preserves explicit keyed selections through cache changes.
npm test passes 40/40. Rho’s 40/40 source tests pass; its
U2 acceptance
records composed browser/install results and remaining gates. U3/U4 and manual
macOS candidate-window input are separate from the Chrome composition check.
Documents now retains per-change author/delta history, exact before/after reads, authenticated own-group undo and targeted single-edit revert. Later overlapping edits are refused; unrelated text is preserved. Legacy global undo remains compatible. History is bounded by entries/deltas and the shared undo text budget; format-3 restore validates its versions, ranges and digests. Stable receipts, pending-save resume/abandon and live-document binding prerequisites remain. The 31 owner tests and Rho's 40 tests pass, including real stdio MCP/HTTP author checks and zero native attempts in the new flow. That backend prerequisite batch did not build UI or extend browser grants. The Documents guide and Rho handoff record API limits and frontend integration still required.
All nine descriptors now provide icons, body copy, ordering, Skill descriptions, action/view descriptions and developer/license/link/dependency information for Rho's U1 catalog. R and Requirements expose bounded executable-file observations without running Rscript or starting a Host. Missing Rscript leaves exported object reads available. These fields grant no access and add no functional renderer. See catalog presentation.
npm test passes 17/17, including real non-executing executable observations;
all nine isolated owner installation/import checks pass. Application composition,
browser acceptance and installation evidence remain with Rho.
Nine owner packages now include the minimal annotations capability. All ship
versioned operation declarations and digest-checked Skills in the actual compiled
entry; Rho checks business-schema parity. The shared MCP adapter intersects the
live catalog with grants and optional plugin filters before listing/calling tools.
Annotations freezes Documents UTF-16 excerpts or Tables stable-key entities and
submitted Documents drafts; replies explicitly link messages. Agent contexts
separately record preparation, fetch and explicit reply observations. No UI or
new execution journal is introduced. See Collaboration.
Fourteen owner tests and all nine isolated tarball installation/import checks pass, rechecked on 2026-10-17. They cover annotations, versioned entries and Skill digests as well as the existing owner boundaries. The two native example regressions also pass. Source Rho integration and actual Claude Code 2.1.302 collaboration passed without Host starts or native dispatches. The README now lists all nine owners and their shipped entry/feedback scope separately from planned UI. Installed application evidence remains the application's responsibility: its collaboration acceptance records fixed Full and Collaboration-minimal installations. The earlier eight-plugin installation below is historical.
Eight independent headless plugins 0.4 supply one generic r.execute action,
explicit R requirements and preparation, shared documents, declared file/table/
media/object resources, bounded readers and keyed comparisons, typed composition
descriptors and project-scoped stdio MCP. The npm packages ship compiled
capabilities; no project recipe, native R script or scientific checker is bundled.
The plugin model defines ownership. Documents, files, R, requirements, tables, media, charts and Agent each own a package, descriptor and implementation. A thin contracts package has no catalog, capability state, format registry, native plan, document budget, CSV parser or MCP dependency. Format/requirement/script schemas live in pure owner contract entries; import and export guards keep those boundaries explicit. Rho selects and wires services; plugins do not import one another. Only Tables installs csv-parse and only Agent installs the MCP SDK. Root is a private source workspace. See package boundaries and selection. Five registered view descriptors report UI unavailable. Those descriptors validate props/bindings without implementing renderers. There is no marketplace or dynamic loader.
The public interface reference enumerates the eight Node APIs, dependency ports, schema subpaths and missing capabilities. Browser consumers use Rho's versioned HTTP/SSE contract; no browser SDK or renderer is supplied.
- Twelve focused capability tests pass: document/undo budgets, immutable MCP intents,
typed strings/nulls and entity namespaces, joins/provenance, resource freshness,
script/argument snapshots, requirement probes, byte completion and containment.
R owns
maxScriptBytes; saved-script execution does not borrow document limits. Only the R execution action requires Requirements; exported objects remain readable with Files/R alone, without Documents, Requirements, R or a Host. - All eight plugin tarballs passed isolated installation/import with only their contracts dependency and required external dependencies. No sibling plugin, aggregate package, native R script or example is installed.
- Two separate example tests pass: native singleton histogram behavior and exact script identity under a spaced Chinese path.
- The ordinary R project uses base R:
four sample rows, real linear-model coefficients, PNG/PDF and an exported model
snapshot. Identifiers
001and literalNAremain strings. - The paired analysis is an explicit scientific consumer. Rho's actual installed 0.4 trial completed five DESeq2 fits, 16,139-row typed results, independent arithmetic/native checks, apeglm shrinkage and a comparison recording both input versions.
- Both projects use the same action, requirements, declaration, reader and composition contracts. Rho's full 0.4 installation passes both consumers and recovery. Two reduced installations physically include only Contracts/Documents/ Files or Contracts/Files/R. Unicode documents or exported objects work without R or a Host; missing action dependencies remain unavailable before dispatch. Removal/reselection retains bytes, drafts/undo, selection intent, view references and original requests.
- Actual Claude Code 2.1.301 through MCP SDK 1.29.0 fetched v2 context and changed the same unsaved Chinese/emoji buffer, without starting a Host.
The eight-owner application results and installation evidence are in Rho's plugin boundary review. Earlier evidence and retained failures remain in Rho's extraction acceptance. The earlier scientific record remains a separate 200-simulation result: five saved models, 121/121 native observations, 12/12 original/retrieved arithmetic checks and rejected altered counterexamples. The new bounded integration uses one simulation and does not replace that record.
Agent evaluation remains a separate development tool. Its recorded acceptance includes 43 regressions, eight actual CLI baseline cases, six optional library-fixture cases and one actual-model execution trial. Retained full-model evidence remains 15/16, including its failure. No evaluator is required by any capability package.
Rho owns v2 HTTP/client/fixtures and the retained v1 observation bridge. Old requests keep original parameters, versions and Core identity; observation and restoration do not rerun recipes. The Plugins collector retains legacy resource/check facts without selecting old source or requiring the old scientific package set. Incompatible new work uses v2, rather than redefining the accepted v1 actions.
The first product direction remains npm + local service + a single-window browser WebUI. React/TypeScript, Vite, Dockview, CodeMirror 6, AG Grid Community and ECharts are selected UI choices; their views and browser assembly are not implemented. Electron and cross-window movement follow later. No UI work is part of this extraction. See Roadmap.
R execution is batch Rscript. Objects are explicit exported snapshots, with provenance and bounded JSON previews; there is no live variable inspection or R-memory recovery after Host replacement. A Core Host is not an R interpreter. Document/resource reads start no Host or R process. Media returns original bytes and metadata, without image decoding, PDF rendering or browser interaction.
Execution success, byte integrity, domain checks and model judgments remain separate facts. Generic integrity checks do not validate scientific methods. Native tools retain the Host's OS authority; this is not an execution sandbox. Candidate optimization, blind-set management and hillclimbing remain outside the implemented evaluator.
Source publication, accepted application composition, npm installation and signed distribution have separate evidence. Local ignored trial artifacts are retained at their selected paths, rather than published as a permanent remote archive.