Browser-based graphical shell and control plane for a Linux server you own — real terminal, file manager, live metrics and BYOK AI in one mobile-first workspace. Next.js 16 · React 19 · node-pty · Tailwind v4 · shadcn/ui. Public alpha, self-hosted.
MSO testing handoff
- Live-build regression gate:
bun run buildmust refuse before.nextmutation when the same checkout is being served;mso buildis the safe out-of-tree verification path andmso deployowns production rebuild/replacement. Post-deploy acceptance must verify every JS/CSS asset referenced by root HTML plus authenticated browser rendering of/settings; one sampled chunk is not sufficient.
Update this file at the end of every manual or automated test session. Keep it factual; never include tokens, cookies, prompts containing secrets, or hidden transcripts.
Current status
The current release target is MSO CLI 1.12.1, MCP server 1.10.0, toolset 2026.09.05.1; use mso --version plus /api/health for the exact Git build because documentation-only release commits also advance that identity. The v1.10/v1.11 findings and the v1.12 sectioned-terminal/recoverable-error work below are resolved and retained as historical test evidence. Do not re-open a resolved finding from prose alone: reproduce it against the current build first. There is no known 1.12 release blocker in this handoff.
Current verified behavior:
2026-09-04 Cognitive Runtime public-claim audit: source/main/origin + loopback/public production were revalidated on the latest concurrent baseline. Fresh reruns pass
bench:cognitive,bench:memory, P9 200-correction calibration, P10 lifecycle, and P10 retrieval calibration. Current deterministic evidence: 100% required-tool recall, 96.6% active-schema reduction, memory 8/8, P9 6/6, P10 lifecycle 8/8, lexical/local-semantic 4/6 vs 3/6, graph fixture 3/3. Canonical raw comparison artifacts were re-hashed: P6e0b5e1e30a9f1e5b34624e42d4bb5195792e50129b74de7d4006906a99716ba5, P7 two-runa0c12961d142420f6ac334ff021bde7ab82faa89f9294107d3bd46823606a530, P8 three-runf70e2f5e8132f0808976ed69e2cb6769d2cec861d58612a8cca2ef5d80a7d838, P8 cacheb4a2b1be5dc6edfb2ce8282eae64e3a4fc206cd043ecb825a1eae6a617c9f871. The canonical P7 file ismso-p7-repeat-full-2x9.json; the similarly namedmso-p7-repeat-full-2runs.jsonis a different, non-canonical run. Recomputed P7+P8 evidence is 5/5 perfect for both MSO and Hermes with 74.7% fewer normalized tokens/attempt, 12.8% lower mean average latency, and 34.1% lower mean per-run p50 for MSO on that exact matched nine-scenario corpus. P4/P5 exact numeric tables are historical-only because their exact raw artifacts were not independently revalidated in this audit. Cost/cache/OpenClaw remain unranked. Tool Forge source/non-Docker tests pass, but the local Forge Docker image was absent, so three Docker-only tests were correctly skipped and executable evaluation fails closed until explicitly provisioned.Fresh 3 MCP presentation now has three canonical resources (List + Block + Page). The Page widget origin is now deployment-derived from
OS_MCP_UI_ORIGIN/OS_PUBLIC_ORIGIN; portable source carries no operator domain. External Page apps default to none and are opt-in through validated per-installationMSO_SURFACE_APPS_JSON.workflow_startremains headless; Block has no frame domains and cannot execute a mutation directly.MCP App acceptance: starting a workflow must produce structured workflow identity without opening an App. A Block must render bounded validation/CRUD data and return a user-approved follow-up when its action is pressed; it must not call a mutation itself. A Page must render native
/monitor//project//diffviews throughtools/call; any external app must come only from the current installation's validated catalog. Compatibility URIs remain non-advertised aliases.Alfa Cockpit acceptance: Assistant → Chat shows a compact project/model/activity bar; Cockpit opens as a desktop panel and mobile drawer; project search can select a project outside the first bounded page; selected project metadata reaches the next Alfa system prompt but raw
KNOWLEDGE.md/repo-memory bodies do not; private typed-memory keys/values render asPrivate memory; Host Skills lazy-load from/api/skills; model switch stays within the configured provider; Activity & Runs shows local Alfa tool state alongside MCP workflow activity without treating the browser run as MCPworkflow_finish.Alfa Automation acceptance: a saved recipe's Run action switches to its configured persona and submits the ordered recipe through normal Alfa execution. Read steps may execute directly; write/exec steps must still park on the existing approval card. Playbook starter prompts appear as bounded empty-chat quick actions.
P10 memory retention/lifecycle: isolated fixtures now archive only cold finished claims before the hard 2,000-record / 2 MiB live-ledger limits. Trigger/target are 1,400 -> <=1,000 live records (or ~1.4 MiB -> ~1 MiB); archive segments are content-addressed, owner-only and archive-before-ledger atomic.
bench:memory:lifecycleis 8/8 deterministic with a 1,400-correction seed ending at 1,000 live + 401 archived / two segments and a 56-byte resolved projection. It verifies explicit archived-history retrieval, historical temporal resolution, aggregate-only no-content telemetry,0700/0600permissions, interrupted-commit live-wins semantics, backdated replacement over immutable archive, and frozen session A / fresh session B / resume-copy behavior. The P10 retrieval gate is 4/6 lexical (synonym + Indonesian/English misses remain explicit); P10D also calibrates the existing localmso-local-hybrid-v1encoder on the same rows and gets 3/6 with zero network calls / zero persisted memory vectors, so production vector storage stays blocked by measured non-improvement rather than assumption. 3/3 two-hop relationship fixtures keep graph storage blocked. The exact-history crowding bug (Runtime-1vsRuntime-10...) is fixed by exact phrase/token scoring; legacybench:memoryremains 100%.P9 memory calibration: isolated temporary-memory fixtures now cover repeated corrections, explicit-vs-observed authority, confidence/recency, future validity, forget, and history integrity. P9 fixed a real temporal privacy bug where
forgetcould leave a future-scheduled claim that resurrected later; forget now closes all current/future-capable intervals while preserving already-ended history. Final 200-correction stress from sourced249da4: 6/6 calibration scenarios, existing memory benchmark 100%, Cognitive Runtime memory gates 100%; 200 ledger rows / 135,180 bytes resolve to 1 row / 58-byte compatibility projection (99.96% smaller) and history retrieval remains capped at 100. No calibrated case requires graph memory.P8 calibration: repeat aggregation now reports descriptive per-run spread without inferential claims; three new complete nine-scenario runs from source
51ae31fwere 27/27 task + 27/27 scenario-policy and 3/3 perfect runs for both agents, with 54/54 row identity/provider/model/token-arithmetic checks passing. Combined with P7, both agents are 5/5 perfect; five-run means are MSO vs Hermes 16,907.6 vs 66,791.1 normalized tokens/attempt (~74.7% lower for MSO), 18,985.9 vs 21,779.8 ms average latency, and 14,130.8 vs 21,457.2 ms per-run p50. Newbench:cache-calibrationfinal 4-round/12K-prefix run was 8/8 success; MSO cache-read was zero on 4/4 while Hermes naturally reported positive cache reads on 2/4 (9,728 + 19,968), but cache frequency is unranked because request envelopes differ. P8 also fixes false token non-comparability when independently valid normalized proof shapes vary across cache/no-cache rows. Cost remains non-comparable because attribution is asymmetric.P7 repeatability gate:
bench:corpus:repeatadds explicit/deterministic seeds, alternating starting-agent order, and strict per-run/per-agent scenario + provider/model identity checks before any cross-run ranking. The first two complete 9-scenario runs from source5bd5788onopenai-codex/gpt-5.6-terrawere 18/18 task + 18/18 scenario-policy for both agents and 2/2 perfect runs each. Aggregate MSO vs Hermes: 17,271.3 vs 65,073.3 normalized tokens/attempt (~73.5% lower for MSO), 18,922.6 vs 21,077.4 ms average latency, and 13,761 vs 20,513.5 ms p50. All 36 token-proof arithmetic checks passed. This is a bounded two-run baseline, not statistical reliability proof; cost remains non-comparable.Historical MCP 1.7 hardening baseline: ChatGPT was fail-closed to a 29-tool compact MSO-only profile (~31.8 KiB descriptor JSON / ~8k rough tokens), full MSO remains 73 transport tools, and project-specific MCP names stay dynamic behind
project_mcp_tools/project_mcp_call. OAuth resource binding + rotating refresh credentials, Streamable HTTP protocol handling, and legacy registered ChatGPT-client profile detection have focused regression coverage. Deterministic security gates (Trivy/OSV/Gitleaks/Semgrep/ShellCheck) pass; Codex Security AI scanning is explicitly incomplete because executor/cost guards stopped it, with 0 findings in the targeted partial reports.P6
mso-agent-quality-v2expands the matched corpus to nine scenarios with repo-debug, schema migration, and transactional rollback. Finalopenai-codex/gpt-5.6-terrafull run: MSO 9/9 task + 9/9 scenario-policy, Hermes 8/9 + 9/9; normalized tokens/attempt 16,683.4 vs 67,433.1 (~75.3% lower for MSO), p50 15,285 vs 23,294 ms, while MSO average latency was ~2.6% higher. The Hermesmulti-readmiss did not reproduce in a 3/3-vs-3/3 alternating-order follow-up, so it is not treated as a universal reliability claim. P6 also fixes over-broad manual-test routing, restores explicit exec validation for repo tasks, and prevents duplicateworkflow_startinside long single-user turns. Cost remains non-comparable.P5 provider-usage normalization preserves Codex/OpenAI cache/reasoning details without double counting, keeps unknown fields unknown, and proof-gates canonical token semantics. The P5 numeric comparison below is historical release-reported evidence only; the 2026-09-04 claim audit did not recover its exact raw corpus artifact, so do not use it as the current public baseline without a fresh matched rerun. The historical run reported 6/6 task + 6/6 scenario-policy for both and 3,510.8 vs 59,743.2 normalized tokens/success; cost semantics remained non-comparable. MSO's proof was
explicit-inclusive-contract; Hermes wasexact-exclusive-cache-sum.P4
bench:corpususes six seeded, per-agent isolated 0700 scratch scenarios, exact scenario-tree scoring (policyObservationScope=scenario-tree, not whole-host sandbox proof), and rotating scenario-major runner order. The exact P4 numeric table is historical release-reported evidence only; the 2026-09-04 audit found retained P4 artifacts with different timing values and therefore did not approve the old table as the current public baseline. The release report recorded 6/6 task + 6/6 scenario-observable policy for both onopenai-codex/gpt-5.6-terra; OpenClaw remains unranked until an equivalent model/provider path exists.every durable session has a unique short public
name(milo,luna,nara, …) independent from its longertitle;/renamechanges the handle and/titlechanges the description;human
@namemention routing resolves active agents only; internalagent-aaliases remain compatibility data for explicit APIs;composer drafts wrap dynamically to terminal width/height, repaint on resize, and ↑/↓ navigate visual rows before history;
local_agents_listexposes lease status separately fromconsumerConnected/consumerCount;local_agent_inbox(wait_ms=0..20000)can make one foreground MCP request a real Local Agent receiver; it returns early on peer delivery, closes the read→subscribe race against the durable mailbox, and never starts a background listener;local_agent_request_waitprovides bounded foreground outcomes without resend/background polling;local_agent_requesthas MCP-dispatch coverage and remains a fresh bounded worker, never a claim to wake/control another terminal/ChatGPT process;standalone Local Agent MCP tests isolate the A2A worker import boundary instead of failing on Vitest's unresolved
server-only;read_pipelinebatches up to six eligible read-only MCP calls and applies bounded declarative transforms server-side; child calls retain their normal guards/rate limits, are forced to read scope, and cannot carry a foreignworkflow_id;exact write/exec approval defaults to one compact safe line; Enter opens redacted-safe details plus the unchanged canonical digest, and a separate explicit
allow/denydecision is required.the terminal separates
Assistant,Agent work,Local agent,Error, andInput · @namewith full-width dividers; the composer identity is@name ›and permission lives in the bottom footer asmode ask|auto|yolo;recoverable HTTP/API failures preserve the interaction and correlation state, classify mutation outcome as
not_started,completed, oruncertain, redact safe error summaries, and never auto-retry an uncertain write/exec mutation.
Historical copy prompt for Agent Alpha (do not use as current baseline)
You are taking over MSO testing work in the current repository root (`git rev-parse --show-toplevel`).
Read docs/TESTING-HANDOFF.md first. Current verified source baseline: CLI 1.12.0, MCP server 1.7.3, toolset 2026.09.03.4; resolve the exact live Git build from `mso --version`/`/api/health`. Preserve existing work and do not re-open resolved v1.10-v1.12 findings unless you can reproduce them against the current build.
For a new issue: record the exact reproduction, distinguish durable state from live receiver/process state, avoid duplicate write/exec retries when outcome is uncertain, add the smallest focused regression, run typecheck plus the relevant contract tests, then update this handoff with factual results.
Resolved historical takeover prompt
The prompt below is retained because it records the acceptance criteria that produced 1.11/1.12. It is not current open work.
You are taking over MSO testing work in the current repository root (`git rev-parse --show-toplevel`).
Read docs/TESTING-HANDOFF.md first. Preserve existing uncommitted changes unless they directly conflict with the requested fix.
Implement and verify the open items below:
1. Expose and exercise local_agent_request end-to-end through the MCP tool catalog. It must run a fresh bounded worker from a same-owner durable session even when its live terminal lease is offline. It must never claim to wake/control the original ChatGPT conversation or terminal process.
2. Diagnose local-agent request/reply delivery. A target shown `idle` only means its lease is current; it does not prove a consumer is subscribed or will reply. Add explicit delivery/consumer observability and a bounded timeout/result state so requests cannot silently wait forever. Preserve durable mailbox and correlation semantics.
3. Replace the verbose exact-approval prompt with the compact interaction below. Keep canonical full-payload approval binding unchanged.
4. Improve the terminal conversation layout, drawing from Claude Code/Codex CLI ergonomics:
- Render a full-width horizontal divider before each distinct section: agent work/progress, normal assistant response, and local-agent discussion/inbound or outbound messages. Dividers must make the three streams visually unambiguous without duplicating content.
- Keep the input composer in a dedicated bottom area separated from transcript output by a full-width horizontal divider.
- Move the current permission mode (`yolo`, auto-write, ask, etc.) into the bottom composer/status area below that input divider, not among transcript headers.
- Put the current agent's short public name/handle in the header slot currently used for permission mode, matching the identity-forward layout familiar from Claude Code/Codex CLI. Preserve clear indication of the active permission mode in the new bottom location.
- Preserve keyboard accessibility, resize redraw behavior, compact output, exact approval semantics, and existing local-agent correlation/inbox behavior.
Exact-approval UX requirement:
- Show only one changing status line while a write/exec approval is pending, e.g. `Approval needed: exec_run — run typecheck`.
- The line must describe the current task/tool and update for the next call; do not print `[WRITE · YOLO]`, payload JSON, hash, byte count, or a multi-line confirmation by default.
- `Tab` focuses/selects the pending approval action; `Enter` opens exact-call details (tool, redacted-safe args summary, canonical digest, scope, and allow/deny choices).
- From details, the user can explicitly allow or deny; no approval is implied by Tab or Enter alone.
- Preserve accessible keyboard behavior, cancellation/interrupt behavior, and the existing exact full-payload server-side digest check.
Validation minimum: focused unit tests for success, offline target, absent subscriber/no reply, self-target rejection, principal isolation, transcript section dividers, input divider, agent-name header placement, permission-mode bottom placement, compact default approval line, Tab-to-details, Enter-to-details, explicit allow/deny, and canonical approval binding; then npm run typecheck. Report changed files and exact test results.
Session log
2026-09-03 — ChatGPT MCP scanner/auth/project boundary hardening
Goal
- Make ChatGPT action refresh reliably scan a compact, standards-aligned MSO-only catalog while keeping full MCP capability for other clients.
- Keep every project-owned MCP tool/config/credential dynamic/private behind generic MSO primitives.
Changed
- ChatGPT client profile: 29 transport tools / 28 model actions + app-only
workflow_status; direct calls to hidden full-catalog names fail closed. - Descriptor normalization adds title, all required safety annotations and mirrored OAuth security schemes to every full/profile tool.
- Added
project_mcp_tools+project_mcp_call; project.mcp.jsonnever becomes global tool definitions. - Added OAuth resource binding,
iss, rotating refresh tokens, grant-family revocation, protocol-header validation and Streamable HTTP 405 behavior for the unimplemented SSE listener. - Removed current cross-repository MCP contract/coupling and sibling-repo push-gate dependency.
Verified so far
- TypeScript: PASS after OAuth/project/profile changes.
- Focused OAuth/route/store/profile/project-MCP contract tests: PASS.
- ChatGPT descriptor metric: 29 tools / 31,908 bytes / ~7,977 rough tokens / 2,604 max tool bytes.
node scripts/check-docs.mjs: PASS at 72 model MCP tools + one app-only bridge.- Full repository verify/build and live deployment verification are the remaining release gates for this session.
2026-09-03 — ChatGPT-session Local Agent foreground two-way MVP
Goal
- Make two separate ChatGPT conversations connected to the same MSO MCP communicate directly in both directions while each destination is actively holding a bounded MCP receive call.
- Reuse the existing file mailbox + in-process Local Agent subscriber; no DB, webhook, broker, WebSocket, daemon, or worker spawn.
Changed
- Added optional
wait_ms=0..20000to read-scopelocal_agent_inbox; default zero preserves immediate reads. - Added race-safe foreground receive: read mailbox → subscribe → read again → wait for event/timeout → final read, always unsubscribe.
- MCP dispatcher keeps
local_agent_inboxpresence receivable/idlewhile the call is open instead of marking itbusy; normal tool calls are unchanged. - Advanced toolset schema identity to
2026.09.03.2; upstream P3's newread_pipelineremains intact, so the catalog stays at 70 model/operator tools + app-onlyworkflow_status.
Verified before full release gate
- Targeted typecheck + ESLint: PASS.
- Targeted Local Agent/MCP bundle: 5 test files / 50 tests passed.
- New end-to-end MCP-dispatch test proves A→B request delivery and B→A correlated reply delivery while each destination is in
local_agent_inbox(wait_ms), with sender statusdeliveredand receiver subscriber cleanup after return. node scripts/check-docs.mjs: PASS after merging the concurrent P3 read-pipeline toolset.- Full release gate PASS with exit 0:
bun run verifyfollowed bybash scripts/verify-build.sh; lint remains 0 errors / 8 existing max-lines warnings, dependency audit is clean, and the out-of-tree production build completes successfully.
Client test requirement
- ChatGPT uses a frozen MCP action snapshot. Refresh/re-scan MSO actions after deployment so
local_agent_inbox.wait_msis visible to the model; then use two new ChatGPT conversations for the live A↔B test.
2026-09-03 — repository docs + unused Agent surface audit
Goal
- Make every current MSO document consistent with the shipped 1.12 Agent/runtime contract without deleting still-valid historical evidence.
- Remove only demonstrably unused recent Agent code/surface and strengthen machine checks against future documentation drift.
Changed
- Audited all 71 repository Markdown files; current references were synchronized while
PROGRESS, generated changelog, dated audits/plans and older release evidence were preserved as historical records. - At the 1.12 documentation-cleanup checkpoint, corrected MCP scope/count catalogs to 66 model/operator tools (32 read / 22 write / 12 exec) plus app-only
workflow_status, and corrected current A2A docs to include authenticated outbound credentials, streaming and optional authenticated inbound serving. - Updated 1.12 terminal/provider/install/Local Agent/architecture/troubleshooting guidance and replaced the stale takeover prompt with a current-baseline prompt while retaining the resolved old prompt as historical acceptance evidence.
- Strengthened
check-docs.mjsso architecture/operator markers plus MCP, ChatGPT and connectors scope partitions must match source exactly. - Removed dead
coreToolNames()and made implementation-only Agent helpers private instead of exporting unused internal API.
Verified before release gate
- Markdown inventory: 71 files; structural issues: 0.
- Relative Markdown links: 125 checked / 0 broken.
- Stale current-claim sweep for obsolete 31/29/59/56-tool counts, old toolset sample and anonymous-only A2A wording: clean.
node scripts/check-docs.mjs: PASS — 53 checked project docs, 66 model MCP tools + 1 app-only bridge, 22 slices, 10 AppShell feature dirs.npm run typecheck: PASS after the documentation/checker and unused-export cleanup.
Repository-wide verification
- Targeted Agent/docs regression: 10 test files / 45 tests passed.
- Full release verification PASS with exit 0:
bun run verify(typecheck, lint, coverage/full test suite, docs/architecture checks, dependency audit) followed bybash scripts/verify-build.shfor the out-of-tree production build. - Lint remains 0 errors / 9 existing max-lines warnings; this audit adds no warning-only file.
- Dependency audit:
bun auditreports No vulnerabilities found. - Final Agent dead-code scan: 0 dead top-level helpers; the removed
coreToolNames()was the only definition-only function found before cleanup. - Final Markdown validation remains 71 files / 125 relative links / 0 structural issues / 0 broken links; strengthened
check-docs.mjspasses.
Open work / handoff
- No known documentation, Agent-contract, dependency-audit, or build blocker. Future tool additions/renames should update source first and let the strengthened docs gate identify every stale MCP scope/count representation instead of manually copying counts.
2026-09-02 — sectioned terminal + recoverable API failures
Goal
- Separate Assistant, Agent work, Local Agent, Error, and Input streams without buffering model output or duplicating content.
- Move permission mode to the bottom composer area while keeping the short public
@nameas the terminal identity. - Make HTTP/API transport failures recoverable and explicit without automatically repeating an uncertain mutation.
Changed
- Added full-width
Assistant,Agent work,Local agent,Error, andInput · @namesection dividers. - Replaced the old
[ask] ›identity prompt with@name ›; the bottom footer now showsmode ask|auto|yoloand Tab cycles that mode in place. - Preserved wrapped composer editing, vertical cursor navigation, resize repaint, and incoming-event draft redraw.
- Added structured Agent API errors plus a per-turn mutation journal with
not_started,completed, anduncertainoutcomes. - A write/exec transport failure with uncertain delivery now stops the turn before the model can retry the mutation. A later assistant/API error after a completed mutation reports the mutation as completed and tells the client not to repeat it.
- Recoverable errors are redacted, rendered as a bounded
Errorsection, persisted as safe durable context, and keep session/correlation state. Pending exact-approval metadata is retained if the approval interaction itself fails before a decision. - Moved turn/tool orchestration out of
mso-agent.mjs; removed obsolete internalpermissionPromptandrenderStatusBarhelpers rather than keeping compatibility shims.
Verified
- Focused Local Agent/layout/error/approval/dispatcher regression bundle: 13 test files / 91 tests passed.
- Additional composer/layout/error focused suites passed after extracting composer row-reservation helpers to keep lint line budgets unchanged.
- Pre-release PTY smoke PASS:
@smoke-ashowedmode yolo, oneAssistantsection for model output, oneAgent worksection for/spawn, oneLocal agentsection for outbound collaboration, and a boundedErrorsection for an unknown mention.@smoke-breceivedLOCAL_LAYOUT_SMOKEunder aLocal agent · [smoke-a]divider and its Input area/draft was redrawn correctly. Temporary PTYs were closed. - Full release gate PASS with exit 0:
npm run typecheck,npm run lint, fullnpm run test,npm run check, productionnpm run build, andgit diff --check. - Lint: 0 errors / 9 existing max-lines warnings; no new warning-only file remains after the composer extraction.
- MCP toolset remains
2026.09.02.7because this release changes Agent/TUI runtime semantics but adds no MCP tool.
Release/live verification
- Feature release commit
eeaaa27was pushed tomain; deployed service reportedbuildSha=eeaaa27and CLI1.12.0. - Post-deploy PTY smoke PASS: source
@live-ashowed theInputdivider,mode yolo, oneAssistantsection withLIVE_ASSISTANT_112, one outboundLocal agentsection, and a boundedError · local agentsection for an unknown mention. Target@live-breceivedLIVE_LOCAL_112underLocal agent · [live-a]while its bottom composer remainedmode ask. - Both temporary PTYs were closed after the smoke. No runtime regression was found.
Open work / handoff
- No release blocker remains for 1.12.0. Future work should start from new reproducible observations rather than re-opening the resolved section-layout or recoverable-error findings above.
2026-09-02 — named sessions, responsive composer, consumer observability, compact approval
Goal
- Apply the user's v1.10 test findings without losing the uncommitted local-agent worker/subagent changes.
- Make session identity easy to type and separate it from session description.
- Make long terminal input editable as a responsive wrapped draft.
Changed
- Added per-session familiar short names with same-principal uniqueness locking and backward-compatible legacy fallback names.
- Added
/rename <name>; kept/title <text>for the longer session topic/description. - Restricted human
@namementions to active sessions while preserving explicit offline queueing in the lower-level mailbox API. - Added receiver subscription observability and
local_agent_request_waitbounded outcomes. - Preserved/finished
local_agent_requestand configurable subagent max-turn policy from the previous test session. - Reworked the composer into a terminal-size-aware wrapped input viewport with vertical cursor movement and resize repaint.
- Replaced verbose exact-call approval output with compact status → details → explicit allow/deny UX while preserving canonical digest verification.
Verified so far
- Focused regression bundle: 13 test files / 97 tests passed.
npm run typecheckpasses after the changes.node scripts/check-docs.mjsreports 53 Markdown files, 66 model MCP tools + 1 app-only bridge and is current.local_agent_requestis exercised through the real MCP dispatcher with its worker boundary mocked, not by direct tool invocation only.
Observed
- The old standalone
tools-local-agents.test.tsserver-onlyfailure was reproduced and fixed by isolating the A2A worker boundary in that test. idleremains a presence-lease state by design; receiver subscription is now a separate field instead of overloadingidle.
Repository-wide verification
- Full release gate PASS:
npm run typecheck,npm run lint, fullnpm run test,npm run check, productionnpm run build, andgit diff --check. - Lint: 0 errors / 9 existing max-lines warnings; the responsive composer tests were split into a dedicated file so this change adds no new warning-only file.
- The first full-load run exposed an overly broad per-principal session-name lock; it was replaced with per-candidate-name locking, after which unrelated conversation/session creation remained concurrent and the full suite passed.
Live 1.11.0 smoke finding
- Fresh PTYs received familiar names
@omarand@cali;/rename qa-omarchanged only the handle while the title stayedMSO Agent session;consumerConnected=true/1was correct while the receiver was attached. - A 40×16 PTY rendered one long logical draft across four visual rows; resizing to 72×20 repainted the same draft across two visual rows without submission/data loss.
- One real race remained: immediately after closing the target PTY, the presence lease was still
idlefor its grace window whileconsumerConnected=false, so@qa-omarcould still enqueue. The 1.11.1 patch tightens human mention eligibility to lease current + receiver subscribed and race-proofs the server send boundary with the same rule. Lower-level explicit mailbox sends keep offline/no-consumer queue compatibility.
1.11.1 verification
- Targeted receiver-required mention patch: 24/24 tests passed.
- First full-load gate exposed two test-timing races unrelated to the mention behavior. Session-store concurrency passed immediately when isolated; the gateway exclusion fixture still used a two-second lock-holder despite claiming explicit synchronization. The fixture now holds the lock until the test terminates it.
- Final full release gate PASS with exit 0:
npm run typecheck,npm run lint, fullnpm run test,npm run check, productionnpm run build, andgit diff --check. - Lint remains 0 errors / 9 existing max-lines warnings. Full-load session-store concurrency and gateway exclusion both pass in the final run.
1.11.1 release/live verification
- Released as commit
3e06c70withHEAD == origin/main; livemso.servicereportedbuildSha=3e06c70and CLI1.11.1. - Close-target race smoke PASS: after target PTY disconnect the lease could still report
idle, butconsumerConnected=false; human@namewas rejected immediately and the target mailbox marker count stayed unchanged. - Temporary PTYs were closed. Durable smoke session records were not manually deleted because MSO does not yet expose a supported AgentSession delete API and direct store-file deletion could leave indexes/mailbox references inconsistent.
2026-09-02 — local agent wake + subagent capacity
Goal
- Let a same-owner durable local-agent session perform work even if its live terminal receiver is offline.
- Increase focused subagent model/tool rounds for testing.
Changes made (uncommitted)
- Added
local_agent_requestinlib/mcp/tools-local-agents.ts(exec scope). - Added
handoffOwnerLocalSession()inlib/a2a/local-session.ts. - The request creates a fresh bounded worker using target durable session context; it does not wake or control the original terminal/ChatGPT session.
- Added
OS_SUBAGENT_MAX_TURNS: default12, configurable up to absolute cap48. Callers still passmax_turnsper invocation.
Verification
npx vitest run lib/a2a/local-session.test.ts lib/agent/local-agent-messaging.test.ts→ 15 passing.npm run typecheck→ passing.lib/mcp/tools-local-agents.test.tscould not load in standalone Vitest because dependencyserver-onlywas unresolved via the full MCP tool graph. This is a test-environment/import issue, not treated as a passing test.
Observed local-agent issue
agent-kis currently reportedidlewith a fresh lease, but no reply arrived.- This status proves presence heartbeat only. It does not prove the ChatGPT-side consumer is subscribed, has read the inbox, or will execute
local_agent_reply. - Do not send duplicate request messages while diagnosing. Inspect delivery/inbox state and add observability/timeout behavior first.
Open work (resolved by the 1.11 candidate)
local_agent_requestnow has MCP dispatcher coverage.- Directory rows expose receiver subscription state and
local_agent_request_waitgives bounded outcomes. - Product UX is documented:
@name/mailbox for active-session collaboration;local_agent_requestfor an immediate fresh bounded worker from durable context;local_agent_request_waitfor bounded mailbox observation.
2026-09-02 — compact exact-approval UX requirement
Requested behavior
- One changing status/task line only for a pending exact write/exec approval.
Tab, thenEnter, opens the exact-call details; details contain the approval controls.- Default display must not expose YOLO/auto-write labels or payload/hash noise.
Handoff (resolved by the 1.11 candidate)
- Compact approval runtime and focused regression coverage are implemented; canonical exact full-payload digest binding is unchanged.
2026-09-02 — terminal HTTP/API error resilience
Resolution
- Shipped in MSO CLI 1.12.0 (
eeaaa27) and live-verified before the final testing-doc sync (be0e436). The requirements below are retained as the original acceptance evidence.
Observed
- During a documentation-only update, the file write succeeded, but the next tool/UI step surfaced an HTTP 400 and interrupted the visible flow before the assistant could confirm completion.
Required follow-up
- Treat recoverable transport/API failures (including HTTP 400) as a visible, bounded UI state rather than silently terminating the active interaction.
- Preserve the user draft, session identity, transcript, pending permission state, and correlation IDs across a recoverable failure; do not duplicate a mutation automatically.
- Render a concise error section/divider with a safe error summary, whether the preceding operation completed, and a clear retry/continue action. Never expose secrets, raw request bodies, or hidden transcripts.
- The assistant/client must still emit a user-facing completion, partial-result, or failure message after a tool/API error. If status is uncertain, say so and verify before retrying any write/exec mutation.
- Add focused tests for: HTTP 400 after a successful mutation, HTTP 400 before dispatch, draft preservation, no duplicate write on retry, pending approval preservation, and a visible user-facing error/result message.
Historical session-note template (reference only)
### YYYY-MM-DD — short title
**Goal**
- ...
**Changed**
- files / behavior
**Verified**
- exact commands and results
**Observed**
- reproducible facts only
**Open work / handoff**
- ...