Repository navigation
feat: add Strands TUI harness with bounded foreground cancellation - #8207
milesatao-svg wants to merge 53 commits into
Conversation
Build the real daemon in PR CI for isolated VPS runtime verification without local validation suites or dependency installations. Record source revision, Go build info, and executable hash.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An orchestrator that ran `ao spawn --agent <x>` overrode the project's configured Worker agent, so a project set to a non-Claude worker still got claude-code workers when the orchestrator delegated. A worker an orchestrator spawns in a project now runs that project's configured Worker agent regardless of the --agent it named: the orchestrator coordinates, the project owner chooses the worker agent. Guarded like inheritedSpawnPermissions (parent must be an orchestrator in the same project) and only when a Worker agent is configured, so standalone spawns, non-orchestrator callers, and unconfigured projects keep the requested harness. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
# Conflicts: # frontend/src/docs/content/plugins/agents/index.mdx
…nds non-refresh After merging main, the sessions-constraint repair checked strands before memcode, so a pre-strands (v202) schema failed with "missing strands" instead of the expected memcode assertion. Order the memcode check first (its migration is older). Strands pins its CLI and declines unattended version refresh, so list it among the non-refresh harnesses in the support matrix. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Copied source evidence 1/3Historical material from original PR #6517, copied for this account migration. Source PR head when copied: Original comment by @nikhilachale on 2026-10-10T12:38:40ZRevision-pinned conformance evidence for commit c3c834e0e7ef. Full report: c3c834e0e7ef/strands-tui-conformance.md. The complete report is reproduced below, including the actual native-only screenshots pinned to this revision. Strands remains unregistered. These are not AO screenshots. Strands CLI terminal conformance — blocked candidateStrands is not registered in AO. The candidate adapter remains isolated in Release and execution environment
Native result
The cancellation command was exactly: sleep 30; printf done > CANCEL_SHOULD_NOT_EXISTEach cancellation was sent approximately two seconds after accepting the native The shipped SDK's Actual native screenshotsThese are live xterm captures from the separate At 10:56:28 UTC, five seconds after Escape, Strands shows both “Cancelled” and At 10:56:56 UTC the composer has settled. The run's delayed artifact now exists The image SHA-256 values are respectively Candidate boundariesThe unregistered package tests the intended thin adapter boundaries:
This is a candidate, not a completed integration. Native in-TUI session switching, Reproduce and evidenceThe opt-in, standard-library-only python3 backend/internal/adapters/agent/strands/testdata/native_cancel_probe.py \
--binary /absolute/path/to/strands \
--home /absolute/path/to/disposable-profile \
--report /absolute/path/to/new-report-directory \
--interrupt escape
# Repeat with a fresh report directory and --interrupt ctrl-c.VPS evidence root:
Audit tokens are nonsecret fixtures. The model echoed them in its final answer; ValidationThe candidate's unit tests, race test and All probe-owned processes were terminated and native evidence retained. Do not Original comment by @nikhilachale on 2026-10-10T12:49:16ZExact-head CI verification at 2026-10-10 12:48 UTC for c3c834e: 21/21 check runs completed successfully. Every inspected check and workflow was tied to this head SHA.
No candidate source change or test rerun was needed. The five package failures from the earlier VPS full-backend run remain recorded and untriaged; this CI result does not relabel that local run as passing. The conformance blocker remains: native Escape and Ctrl+C did not stop the approved shell tool or its delayed write. Strands remains unregistered, with AO lifecycle/UI evidence BLOCKED / NOT RUN. Repository CI passing is not native Strands qualification. Full conformance report with revision-pinned native-only screenshots. Original comment by @nikhilachale on 2026-10-10T21:28:19ZStrands now runs inside AO as a registered TUI harness. Audit The complete result remains diagnostic FAIL: 24 PASS / 1 authentication BLOCKED / 1 AO catalog FAIL / 2 catalog NOT_RUN. AO's auth probe still reports configured credentials rather than verified authorization; native listing is unavailable and AO's catalog is empty. Real provider execution used native CLI 0.2.0 / harness 0.1.1 / SDK 1.20.0 and real Z.ai AO003 used AO The public SDK plugin now propagates cancellation into Strands' native sandbox. Final native Ctrl+C reaped an already-started shell in 97 ms, settled in 210 ms, and produced no delayed write over 32 seconds. Independent AO Kill003 reaped the target shell in 389 ms with no late artifact; the unrelated restored control survived with the same generation immediately and later. Capture 2026-10-10T21:08:21.227532Z; UI source The complete report preserves released-native 28-second late writes, the invalid top-level config flag, AO001's actual pre-spawn stale-readiness stop, Kill002's fetch failure and the blank Electron renderer-startup failure. Only the worker's own service TasksMax was raised from 256 to 512; host RAM was unchanged. Scope is foreground TUI. Model-callable Final HEAD: The migration-ledger CI failure at1b9caed was corrected in cf448df and its focused ledger/migration tests passed. Owned final audit sessions were killed (HTTP200) and the daemon stopped (HTTP202); isolated evidence/profiles remain. Original comment by @nikhilachale on 2026-10-10T21:42:03ZFinal CI: all 23 checks pass at the current evidence head Permission clarification: AO003 used bypass-permissions / native Auto. The native Escape/Ctrl+C controls used separate one-time approvals. AO auth probe remains configured, rather than verified authorization, despite successful real provider calls. The strict audit remains diagnostic FAIL for the documented auth/catalog gaps. Published real AO proof. Original comment by @nikhilachale on 2026-10-12T16:53:08ZHarness-specific reports, JSON receipts, screenshots and provenance now live on this integration PR branch, rather than the skill PR #6509. Evidence index with embedded screenshots and JSON links. Historical successes, failures and scope limitations remain as recorded. Moving these files does not rerun native/AO testing or upgrade any verdict. This is evidence relocation only. Existing tested source/native hashes and failures are preserved; the documentation commit is not a fresh runtime pass. Applicable checks run in PR CI. Original comment by @nikhilachale on 2026-10-12T18:07:48ZCanonical main conflicts reconciled at Validation is PR CI only; current-head checks are pending until GitHub reports them. This merge does not constitute a new native/AO run or promote historical fresh/restore/provider-role proof. Harness evidence stays on this integration PR; PR6509 remains reusable guidance and linked summaries. Original comment by @nikhilachale on 2026-10-12T18:16:43ZCanonical main conflicts reconciled at Validation is PR CI only; current-head checks are pending until GitHub reports them. This merge does not constitute a new native/AO run or promote historical fresh/restore/provider-role proof. Harness evidence stays on this integration PR; PR6509 remains reusable guidance and linked summaries. Original comment by @nikhilachale on 2026-10-12T20:58:13ZCanonical main conflicts reconciled at Validation is PR CI only; current-head checks are pending until GitHub reports them. This merge does not constitute a new native/AO run or promote historical fresh/restore/provider-role proof. Harness evidence stays on this integration PR; PR6509 remains reusable guidance and linked summaries. Original comment by @nikhilachale on 2026-10-15T06:35:44ZPublished, incomplete qualification — AO UI evidence BLOCKED; not merge-ready. Head Publication report · All ordered screenshot checkpoints Every applicable latest UI checkpoint remains NOT_RUN because this stage had no genuine isolated Electron dependencies and insufficient disk preserving512MiB. API/native receipts are auxiliary, never screenshot substitutes. No image was fabricated. Exact publication-head CI is being checked separately below; old green heads do not qualify edited/merged source. No routine VPS tests, lint, typecheck, gofmt or CI emulator ran. Linux/provider evidence does not establish other OS/provider coverage. Original failures/profiles/logs remain intact privately. No skill6509 writes, production changes, merge, release or deployment. The initial publication live-feeds check failed: a pinned install spec was sent as an npm package name (HTTP400). The new source separates unversioned package metadata from pinned0.2.0 installation/reinstallation and refuses automatic updates to unqualified releases. Added regression cases are remote-CI-only; all original runtime/UI non-pass gates remain. Exact-head CI (2026-10-15T07:12:38.660620+00:00): complete at Original comment by @nikhilachale on 2026-10-18T17:46:00ZACP queue status: BLOCKED / incomplete, not merge-ready. Disk below 256MiB safety floor; preserve all source/evidence The supervisor retained partial source and moved to the next harness after two bounded attempts. No routine VPS suites were run. New detailed reports/evidence belong in this PR comments; no success claim is inferred from worker exit. Original comment by @nikhilachale on 2026-10-18T22:05:13ZReconciliation only: refresh the live PR branch and canonical main, resolve the existing merge preserving both Strands and main’s memcode/shared architecture changes, reconcile pending migration allocation, regenerate API contracts from merged source, then refresh both heads again before a normal push. No ACP feature implementation, native/UI qualification, routine local suites, or new repository evidence files in this pass. Historical evidence and failures remain untouched. Current starting PR head: Original comment by @nikhilachale on 2026-10-18T22:08:25ZRECONCILED — ACP qualification remains incomplete; draft retained. Pushed head: Conflict decisions: preserved both Strands and main’s memcode registrations, selectability tests, install/auth metadata, CLI help, API enums, product pickers, documentation and schema repair paths. Preserved all main shared Chat/nativeacp/persistent-host changes. Moved the pending PR-only Strands migration from 0201 to fresh 0213 after main’s 0212; main migrations and burned ledger are unchanged. Updated its regression test to exercise the 0212→0213 boundary and inverse. Regenerated OpenAPI and TypeScript with the actual generators (openapi-typescript 7.4.4); no query/schema source changes required sqlc regeneration.
The initial staged diff check reported a trailing blank line in main’s Cleanup: parent removed only verified unused ignored historical node_modules, disposable npm cache, and two unused superseded Codex vendor versions, checking code status/lockfiles unchanged. This worker deleted no additional source/evidence/history/credentials/cache; no native sessions, daemons, Electron profiles or production processes were started. One stalled owned conflict-script process was stopped; generation commands have completed. Historical reports/screenshots remain byte-identical. New findings are inline here; no evidence files or branches were added. PR6509 received no modifications/comments. No merge/release/deploy. ACP implementation and real current-head qualification continue in the separate ACP pass; reconciliation itself has no external blocker. Original comment by @nikhilachale on 2026-10-18T23:26:59ZContinuing existing PR #6517 from Plan: (1) repair concrete current-head CI regressions after inspecting failing logs (install inventory count, pre-Strands schema repair ordering, documentation helper lint); (2) inspect the released CLI ACP permission/cancellation and instruction/load paths and current official upstream, preserving CLI identity/defaults; (3) implement only defensible candidate bindings/guards with regression coverage, leaving Chat unregistered until native approval/cancel/load/instruction gates pass; (4) refresh both heads before any push, inspect current-head CI, and publish an explicit gate table. Routine suites run in CI only; no local suites or dependency installs. Existing historical failures remain authoritative until separately reproduced and fixed. Resource observation: 924 MiB available at start, above 512 MiB reserve but insufficient for speculative Electron dependency installs. Parent cleanup remains preserved; no additional deletion. Real AO Chat UI/direct attachment and independent review remain required readiness gates, not substitutes for source work. Original comment by @nikhilachale on 2026-10-18T23:38:27ZBLOCKED — native ACP contract fails; draft retained, Chat remains unregistered. Source/CI repairs and opt-in real-native regression coverage pushed. No merge/release/deploy. Head: Source progress:
Native provenance: stock installed
The first native startup failures ( Supported-route investigation: official CLI 0.2.1, commit Exact-head CI observation:
External/native work remaining: a supported regular-CLI permission broker plus correct cancellation, fresh-on-restore instructions and durable workspace rejection, followed by real AO nativeacp/persistent-host qualification, actual Chat UI/direct attachment, full independent acceptance and completed current-head CI. These failures prevent Cleanup: every owned native process was reaped; executing shell and unrelated owned control are gone after provenance checks. No AO daemon/Electron or production session was started/restarted, no real profile was written, and no broad process kill occurred. Parent cleanup removed only verified unused ignored historical node_modules, disposable npm cache and two superseded Codex vendor versions after status/lockfile checks. This worker deleted no further code/cache/evidence/history/credentials; free space remains about 909 MiB, above the 512 MiB reserve. Historical evidence files are untouched. New Git files are executable test source only; no report/JSON/PNG evidence committed. PR6509 received zero modifications/comments. Original comment by @nikhilachale on 2026-10-19T15:50:24ZContinuing source/CI implementation on PR6517 only. Live head was Next source slice: resolve the pending Strands This worker performs source edits, formatting, Git and remote-CI inspection only. Native/provider/AO/UI qualification is NOT RUN here. No local suites/builds/installations or additional disk cleanup; historical evidence remains untouched. Detailed progress stays inline in this PR. Original comment by @nikhilachale on 2026-10-19T15:54:07ZIndependent read-only review of The reviewer checked shared ACP admission on fresh initialization and reconnect, disabled fresh fallback, reconnect-only no-spawn behavior, timestamp migration/ledger preservation, upgrade/downgrade/legacy-QM/unknown-harness/integrity coverage and explicit production registry exclusion. No tests/builds/native/provider calls were performed by the reviewer. This is a ports-compatible, deliberately unavailable candidate, not a functional Strands ACP integration. Real Send/permission/cancel/replay/Stop lifecycle is not exercised because native admission is closed. Approval bridging, executing-tool cancellation and additive instructions on exact restore remain unresolved native boundaries. Original comment by @nikhilachale on 2026-10-19T15:54:10ZPushed
Source closure is not claimed: the regular first-party server exposes no requestPermission client to plugins. Released regular server and current inspected main call only Current-head CI: checks, pending at publication. No local tests/builds/lint/typecheck/provider/AO/UI work ran. Independent source review: #6517 (comment). Historical native failures and screenshots remain untouched; no new evidence/report assets committed, no deletions beyond the pending migration rename, no production changes or processes launched. Disk about 537 MiB; no additional cleanup. Original comment by @nikhilachale on 2026-10-19T15:56:48ZStrands source snapshot: NOT IMPLEMENTED / NOT READY. Safe candidate and main reconciliation are pushed at Implemented source progress: nativeacp/persistenthost-based
Precise implementation boundary: released first-party regular server and current official source inspected at 1adcaf9f give Actual fingerprints: AO source Exact-current-head CI (previous-head green is historical):
No current-head failure was reported at this observation; queued/pending checks are not passed. First failures remain retained in prior ACP report: missing native permissions, executing child survival, restored instructions and workspace ownership. No failed attempt or historical attachment was removed/replaced. The independent review’s test-host cleanup finding was fixed in Cleanup: this source-only worker launched zero native/provider/AO/Electron processes, touched no real profile/history/production data, ran no routine suites/builds/installations/generation, and removed no cache/evidence/user code. The pending migration was renamed only; main’s already-merged migrations are unchanged. Parent cleanup before handoff removed only verified unused ignored node_modules, disposable npm cache and superseded unused Codex vendor versions after status/lockfile checks; historical evidence/credentials/Git metadata were preserved. No additional disk cleanup occurred. New deliverables are source/tests only; reports remain inline. PR6509 and other workers’ PRs/worktrees received no modifications/comments. |
|
Important Draft PR not reviewedDraft PRs are not automatically reviewed by default.
To automatically review draft PRs, update your CodeRabbit configuration: reviews:
auto_review:
drafts: true
|
Copied source evidence 2/3Historical material from original PR #6517, copied for this account migration. Source PR head when copied: Original comment by @Meraj-08 on 2026-10-20T10:16:43ZScreenshots
Original comment by @nikhilachale on 2026-10-21T06:17:18ZFresh testing-only retest starting for Original comment by @nikhilachale on 2026-10-21T06:34:29ZFresh bounded testing-only audit: FAIL / incomplete integration. AO UI attachment evidence: BLOCKED. No full integration PASS. Tested/current head: Actual native package 0.2.0, harness 0.1.1, SDK 1.20.0. Native launcher SHA-256
Independent api-001 (06:32:39.079–06:32:47.260 UTC), PASS limited to nonprovider reads and negative admission, not a diagnostic lifecycle run: disk recovered to964MiB, permitting one cache-backed serial rebuild003 (06:32:00.298–06:32:06.096, exit0) with the same AO SHA-256. Native witness command: node <staged-skill>/scripts/audit_ao.mjs --agent strands \
--repo <tested-clean-checkout> --ao <runtime>/bin/ao \
--local-contract <runtime>/contract-001.json \
--report <runtime>/strict-001/report \
--startup-timeout-seconds 40 --timeout-seconds 80Original contract SHA-256 {"schemaVersion":1,"agents":{"strands":{"binary":"/path/to/native/strands","versionArgs":["--version"],"integration":{"args":["--help"],"expect":"--session-id"},"session":{"args":["--print","--session","off","--builtin-tools","write","--builtin-plugins","","--skills","off","--prompt","{prompt}"],"proofFile":"LOCAL_AUDIT_PROOF.json"},"interrupt":{"input":"\^[","settleTimeoutMs":5000}}}}The help probe proves availability, not auth. No Completion matrix:
Screenshot index was created before actions. Zero actual Electron captures or published screenshots. At setup gate, free disk 722MiB left 22MiB above the 700MiB floor; official pinned Electron44.7.0 Linux archive alone is 122,998,223 bytes (117.3MiB), before extraction/dependencies. Dependencies absent; npm ci/Electron NOT_RUN. Initial RAM total3,906MiB / available2,509MiB; disk blocked setup. A later capacity observation was below floor after strict had already finished; reclaiming only the finished owned AO binary restored 720MiB free. Binary hashes/build metadata retained. No old frontend, symlinked dependencies, terminal mirror, synthetic pixels or API images substituted. No approved GitHub attachment uploader; private queue has zero images. Every capture/attachment below is unavailable for these reasons, regardless of functional status.
Current source scope: TUI adapter, unregistered gated ACP candidate, capability catalog, manual auth plan, model-entry mode. Source is not runtime/UI proof. Installed native README states print/plain/ACP have no permission UI and use stock harness tool behavior; installed regular ACP server has no requestPermission bridge. Live AO Chat rejection is now proved independently; unfinished Chat/reviewer lifecycle and UI qualification remain non-pass. CI at 2026-10-21T06:34:07.616101+00:00: 25/25 current-head check runs completed/success, no failing/pending check jobs. Local routine unit/race/lint/vet/typecheck/frontend-build/runner-fixture suites NOT_RUN by CI-only policy. Native OS/container/renderer coverage below belongs to CI, not authenticated provider/live Electron proof: build-test, renderer-smoke, mac-update-helper, test, sqlite-race (3), windows-workspace, sqlite-race (0), chat-race (1), sqlite-race (2), backend-core (1), lint, chat-race (0), sqlite-race (1), session-store-race, api-drift, backend-core (0), sqlc-drift, linux-proof-artifact, live-feeds, cloud-build-test, native (ubuntu-latest), native (macos-latest), native (windows-latest), container, scan. Current PR Cleanup verified 2026-10-21T06:33:29.951526+00:00: PASS for owned test processes. Strict daemon PID1469987 exited0; /proc entry absent. Exact version/help/provider/build/runner PID+kernel-start identities no longer existed and children were reaped. API daemon PID1479248/startticks161011189 received shutdown HTTP202 and exited0/reaped; no API session IDs were created. Native ACP PID1477807/startticks160975238 exited0/reaped. Invoking worker/control-host ancestry excluded and never signaled. No UI or AO session created. No production service/profile/account modification. Reports/receipts/logs, minimal isolated profile and private strict SQLite state retained outside Git and unshared; source clean, no evidence commits. Parent final owned-process verification and clean-worktree removal pending. Next actions/unresolved: provide sufficient scratch capacity; use the same authorized provider route once quota permits a successful native file witness; run a NEW strict attempt preserving these failures, requalify auth/readiness and genuine restore cues. Independently qualify actual current-head UI/model/provider controls and any supported Chat/reviewer journeys. Parent must capture/attach safe real PNGs with provenance via an approved attachment path; there are presently none to attach. Resolve conflicts only through separately authorized source work, then recheck affected CI/runtime. PR-description link to this new evidence comment: pending parent action; description not edited. |
Copied source evidence 3/3Historical material from original PR #6517, copied for this account migration. Source PR head when copied: Original comment by @nikhilachale on 2026-10-21T06:35:45ZPublication correction: the first evidence comment is preserved. Readback altered its displayed Escape notation, so that displayed contract example is invalid. This new full evidence comment reconstructs Escape with String.fromCharCode(27); functional results and all original failures are unchanged. Fresh bounded testing-only audit: FAIL / incomplete integration. AO UI attachment evidence: BLOCKED. No full integration PASS. Tested/current head: Actual native package 0.2.0, harness 0.1.1, SDK 1.20.0. Native launcher SHA-256
Independent api-001 (06:32:39.079–06:32:47.260 UTC), PASS limited to nonprovider reads and negative admission, not a diagnostic lifecycle run: disk recovered to964MiB, permitting one cache-backed serial rebuild003 (06:32:00.298–06:32:06.096, exit0) with the same AO SHA-256. Native witness command: node <staged-skill>/scripts/audit_ao.mjs --agent strands \
--repo <tested-clean-checkout> --ao <runtime>/bin/ao \
--local-contract <runtime>/contract-001.json \
--report <runtime>/strict-001/report \
--startup-timeout-seconds 40 --timeout-seconds 80Original contract SHA-256 {"schemaVersion":1,"agents":{"strands":{"binary":"/path/to/native/strands","versionArgs":["--version"],"integration":{"args":["--help"],"expect":"--session-id"},"session":{"args":["--print","--session","off","--builtin-tools","write","--builtin-plugins","","--skills","off","--prompt","{prompt}"],"proofFile":"LOCAL_AUDIT_PROOF.json"}}}}The displayed JSON above omits only the control-character field. Before serializing the runtime contract, add The help probe proves availability, not auth. No Completion matrix:
Screenshot index was created before actions. Zero actual Electron captures or published screenshots. At setup gate, free disk 722MiB left 22MiB above the 700MiB floor; official pinned Electron44.7.0 Linux archive alone is 122,998,223 bytes (117.3MiB), before extraction/dependencies. Dependencies absent; npm ci/Electron NOT_RUN. Initial RAM total3,906MiB / available2,509MiB; disk blocked setup. A later capacity observation was below floor after strict had already finished; reclaiming only the finished owned AO binary restored 720MiB free. Binary hashes/build metadata retained. No old frontend, symlinked dependencies, terminal mirror, synthetic pixels or API images substituted. No approved GitHub attachment uploader; private queue has zero images. Every capture/attachment below is unavailable for these reasons, regardless of functional status.
Current source scope: TUI adapter, unregistered gated ACP candidate, capability catalog, manual auth plan, model-entry mode. Source is not runtime/UI proof. Installed native README states print/plain/ACP have no permission UI and use stock harness tool behavior; installed regular ACP server has no requestPermission bridge. Live AO Chat rejection is now proved independently; unfinished Chat/reviewer lifecycle and UI qualification remain non-pass. CI at 2026-10-21T06:34:07.616101+00:00: 25/25 current-head check runs completed/success, no failing/pending check jobs. Local routine unit/race/lint/vet/typecheck/frontend-build/runner-fixture suites NOT_RUN by CI-only policy. Native OS/container/renderer coverage below belongs to CI, not authenticated provider/live Electron proof: build-test, renderer-smoke, mac-update-helper, test, sqlite-race (3), windows-workspace, sqlite-race (0), chat-race (1), sqlite-race (2), backend-core (1), lint, chat-race (0), sqlite-race (1), session-store-race, api-drift, backend-core (0), sqlc-drift, linux-proof-artifact, live-feeds, cloud-build-test, native (ubuntu-latest), native (macos-latest), native (windows-latest), container, scan. Current PR Cleanup verified 2026-10-21T06:33:29.951526+00:00: PASS for owned test processes. Strict daemon PID1469987 exited0; /proc entry absent. Exact version/help/provider/build/runner PID+kernel-start identities no longer existed and children were reaped. API daemon PID1479248/startticks161011189 received shutdown HTTP202 and exited0/reaped; no API session IDs were created. Native ACP PID1477807/startticks160975238 exited0/reaped. Invoking worker/control-host ancestry excluded and never signaled. No UI or AO session created. No production service/profile/account modification. Reports/receipts/logs, minimal isolated profile and private strict SQLite state retained outside Git and unshared; source clean, no evidence commits. Parent final owned-process verification and clean-worktree removal pending. Next actions/unresolved: provide sufficient scratch capacity; use the same authorized provider route once quota permits a successful native file witness; run a NEW strict attempt preserving these failures, requalify auth/readiness and genuine restore cues. Independently qualify actual current-head UI/model/provider controls and any supported Chat/reviewer journeys. Parent must capture/attach safe real PNGs with provenance via an approved attachment path; there are presently none to attach. Resolve conflicts only through separately authorized source work, then recheck affected CI/runtime. PR-description link to this new evidence comment: pending parent action; description not edited. Original comment by @nikhilachale on 2026-10-28T09:01:08ZNew fixes pushed — current head25bbd4386Native terminal Setup, unknown-auth setup control, first-party native auth verification, freeform model guidance and official branding are implemented. Frontend135/135PASS; both affected backend packages and actualnative offline smokePASS. Fresh UI after captures and new-head CI are running. No live provider authorization/inference is claimed. Previous30/30 results below belong to e255a13. Work in progress: user authorised fixes for the AO-side setup/auth/models/branding gaps after this review. Previous30/30CI results below belong to e255a13; new fixes will receive fresh verification. Current review: real Electron findingsCurrent head: Verdict: NOT merge-ready under skill7081. UI review and all30 official GitHub checks are complete; additional VPS suites are still running. Provider lifecycle qualification remains blocked. Real Electron44.7.0/Linux Xvfb, native Strands0.2.0, isolated data. Official daemon build and real npm ci passed. Daemon SHA256
Unchanged UI skill runner: 3PASS/4FAIL/1NEEDS_USER/1NOT_RUN. Catalog expectation conflicts with truthful freeform mode; two New Task checks expect a legacy dialog. Raw failures preserved; manual current-route findings above do not weaken assertions. Screenshot provenance / attachment gapActual unchanged screenshots were visually reviewed, retained privately on VPS, and not added to product diff: Attachment gap resolved: three original before-fix screenshots are attached directly in this PR comment. Official GitHubCLI2.103.3 --attach was installed privately on VPS; all image URLs return HTTP200 and SHA256 matches originals. The earlier blocker came from old CLI2.45.0. No evidence branch or local computer/browser access was used for the upload. New after-fix captures are pending the newly authorised fixes. Concrete outstanding issues: missing advertised Setup/docs action (repro in table), missing real provider-auth verification, generic branding. Native ACP additionally returns-32601 for set_mode/set_model and restores stale private instructions; keep it unregistered. Exact-head GitHub CI: 30/30 SUCCESSVerified 2026-10-28T09:26:10.494504+00:00 on
Setup-button root cause confirmed in source
Parent-head native protocol, regression and preflight evidenceStrands review against harness skill PR #7081 — current evidenceSource: Reviewed native release: official npm {
"@strands-agents/cli": "0.2.0",
"@strands-agents/harness": "0.1.1",
"@strands-agents/sdk": "1.22.0",
"@agentclientprotocol/sdk": "1.3.0",
"ink": "7.1.1",
"react": "19.2.8"
}
Fix applied and verified: Node preload leaked into project subprocessesBefore, AO's inherited Now the Strands launch enables a one-use flag and the preload consumes it. The parent keeps its loader workaround; child ESM/CJS hook APIs and inherited options remain native. Exact green results: Native CLI0.2.0 plus AO plugins/preload also reached the composer: {
"fixtureProfile": true,
"providerRequests": 0,
"exit": 130,
"readyComposer": true,
"pluginLoaderError": false,
"privateInstructionsVisible": false,
"bytes": 5218
}This used a fixture profile with zero provider requests, so it is bootstrap proof only. Frozen-head preflightThe unmerged Strands migration was regenerated after current main's migration maximum; SQL content retained. Installer cardinality now checks the domain harness set and exact membership. Generated API drift check passes. Existing incoming mobile/cloud paths match main; no unrelated net change was introduced. Actual auth API result{
"installation": "installed",
"authentication": "unknown",
"reasonCode": "auth_check_inconclusive",
"effectiveReadiness": "unknown"
}This is not proof the user is signed out, and configured credentials would not alone establish Connected. Freeform model entry is intentional because the released CLI has no public model-list command; no catalog was invented. Remaining workComplete the real UI checkpoint report and full applicable CI, report remote checks and unavailable-runner gaps, then refresh this comment. Authenticated behavior remains unverified until provider setup and a truthful auth check exist. New reports/logs/screenshots are kept outside the product diff; verified evidence is recorded here in PR comments. Original comment by @nikhilachale on 2026-10-28T09:34:50ZReal Strands UI screenshots — before the new fixesActual isolated Electron44.7.0 on VPS/Linux with native CLI0.2.0. Runtime source8e5461b902eef6ee7f531c2094105d3b2c279264; e255a13 only changes a test. No mock, credentials or provider requests. These show the gaps being fixed, not a passing final implementation. Before fixes: installed Strands0.2.0, no setup action, generic avatarBefore fixes: freeform model contract shown as an empty catalogBefore fixes: New Task readiness gate with no provider authentication |
🏆 Review leaderboardNov 5–12, 2026 · UTC
Ranked by distinct external PRs reviewed, then review rounds, then PR comments. Self-activity and bot activity are excluded. Show 6 more reviewers
|
Account migration verification
|
Resolve the eight conflicts against main adf9f6a by keeping main's fields and behavior and re-adding only the Strands entries: - dto.go, systeminstall.go, schema.ts: keep main's new fields (branch, baseBranch, install progress, cua-driver) and add strands to the enums. - chatdriver registry test: Hermes now has native chat on main, so only Strands stays in the TUI-only list. - agentauth plans: keep main's Claude login text and the Strands setup plan. - migration ledger: keep main's entries and append the Strands migration. - HarnessAccountTab: take main's showSetup condition, which already covers setup actions with unknown auth. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
settings.harness.models.manualTitle and settings.harness.models.manual were English-only, which failed the locale coverage test and crashed the interpolation alignment test on the missing zh-CN value. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The daemon returns its manual policy catalog, which is text mode with no models for Codex, Claude Code, Cursor and other direct-entry harnesses, as a placeholder while a first load is pending and while the harness is signed out. The Models tab then told those users the harness accepts model IDs instead of providing a list, and hid the sign-in prompt. Require an idle, warning-free, empty text catalog before showing that copy. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Reverts 25d44c6 and dabac61. They make an orchestrator-spawned worker ignore an explicit --agent in favor of the project's Worker agent, and update the using-ao spawn guide to match. That changes spawn behavior for every harness, has no PR of its own, and is not on main. Keep this PR to the Strands harness; the change can be proposed and reviewed separately. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
20261028082940187 sat below 40 migrations main shipped since, so a fresh database ran it mid-history and an upgraded one applied it out of order through WithAllowMissing. Take a new generator-format ID above main's 20261114072016988, and run the round-trip test from main's latest schema so it isolates the Strands step instead of reversing every later migration. Also list strands in `ao spawn --harness` help, as every other harness is. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ut of the diff The harness integration guide keeps audit reports, evidence and screenshots out of the PR diff (they belong in PR comments) and rules out unregistered drivers. Remove the docs/research Strands reports, JSON evidence and PNG captures, the unregistered strandsacp Chat driver, and its opt-in native ACP probe and live test. Strands stays TUI-only. The historical evidence remains reachable at the revision-pinned links in the PR description and comments. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Current-head review and ao-agent-harness receipt —
|
| Commit | Change |
|---|---|
8069adaad3, 2e55727e57 |
Merged current main and resolved all 8 conflicts. Main's fields and behavior are kept (DelegateTask branch/baseBranch, install progress, cua-driver, Hermes native chat, Claude "setup token" text, main's showSetup rule). The only additions are Strands enum, plan and ledger entries. |
3158aae834 |
CI fix: added settings.harness.models.manualTitle and settings.harness.models.manual to de, es, fr, ja, ko, pt-BR, ru and zh-CN. This fixes both the locale-coverage failure and the interpolation crash in vitest (2), which also turned the test gate red. |
4e798d90ce |
Defect fix: the new free-form Models panel also fired on the daemon's manual placeholder catalog (text mode, no models) for Codex, Claude Code, Cursor and similar harnesses, both during a first load (refreshState: refreshing) and while signed out (warningCode: auth_required). Those users saw "accepts model IDs instead of providing a model list" instead of the sign-in prompt. The panel now requires a settled, warning-free, empty text catalog. Regression tests added. |
b66ee720b1 |
Scope fix: reverted 25d44c6981 and dabac6135f. They made orchestrator-spawned workers ignore an explicit --agent for every harness, had no PR of their own and are not on main. |
c19163a5e9 |
Migration fix: the Strands CHECK migration moved from 20261028082940187 (below 40 later main migrations) to 20261114173932630, above main's latest 20261114072016988. The round-trip test now starts from main's latest schema. |
3d6224ec32 |
Scope fix (skill §3): removed the committed docs/research Strands reports, evidence JSON and PNG captures, the unregistered strandsacp Chat driver, and its opt-in ACP probe and live test. Historical evidence stays reachable through the revision-pinned links in the description. |
CHECK migration vs main's later migrations: no migration after the Strands version rebuilds sessions or rewrites its harness CHECK. Later migrations only ADD/DROP COLUMN. The ('claude-code','codex','kimi') CHECK in 20261103030401452 belongs to a separate usage table. goose runs WithAllowMissing and reconcileHarnessConstraint repairs Strands as well. Static review only; CI must confirm.
Skill step receipt
| Step | Result | Evidence / gap |
|---|---|---|
| 0 Intake | PASS (manual) | Only open Strands PR; continues closed #6517. harness_intake.mjs not run (no script execution on this host). |
| 1 Compatibility | PASS (historical) | The table and unsupported-surface proof come from the earlier native qualification linked in the description (CLI pinned to 0.2.0; ACP lacks mode/model/permission behavior, so Chat is unregistered). Not re-run at this head. |
| 2 Build / touchpoints | PASS (static) | Checked by equivalent greps, not by running the script. Required TUI, install, auth, models and branding touchpoints are all wired: adapter, domain, agent registry, SQLite repair + migration, 2/2 DTO enums, OpenAPI, schema.ts, product label, GetAgentHooks, activity dispatch, terminal refresh (explicitly unsupported), refresh matrix, install target and plan, auth plan, model catalog, avatar, licence. Optional: spawn help was added in this update; doctor probe has no ExpectedVersionPrefix because --version prints only 0.2.0 (the adapter enforces the pin at launch); no docs/harnesses/strands.md. Chat and reviewer are out of scope (TUI-only). |
| 3 Contract | BLOCKED | A local-contracts.json entry plus doctor needs a host with the pinned CLI and credentials. Per the skill it is never committed. |
| 4 Live audit | BLOCKED | Needs an authorized macOS/Linux host with Strands 0.2.0 and provider credentials. Covers TUI spawn, first prompt, system prompt and AGENTS.md, hook status, follow-up, usage, Kill, exact restore, and V1 roles. |
| 5 UI audit | BLOCKED | Needs the real desktop app on an authorized host. Checkpoints 1–9, including Connected state, Models free-form panel and New Task. |
| 6 Hand-off | PARTIAL | AO-side static findings are fixed (above). The live and UI Fix sections can't be produced until steps 4–5 run. |
Open risks for live testing
- Windows launch: private instructions are passed as a multi-line
--instructionsargv value ahead of--prompt, and the Windows binary is the npmstrands.cmdshim. Batch argument handling of newlines or long lines could truncate the instructions or drop the task prompt. Needs a Windows TUI launch check; the installer is currently offered on all OSes. - Restore errors: a missing native ID, history or locator returns an untyped adapter error. That matches Hermes but may surface as a 500 rather than a fresh launch or typed 4xx.
- Free-form copy scope: the Models panel message now applies to every settled free-form catalog (for example Aider, Goose, Hermes), not just Strands. Product should confirm its "configure a provider and default model in its CLI" wording for those harnesses.
CI
GitHub is holding all workflow runs for this head as action_required (fork workflows need maintainer approval). No current-head check results exist yet. Earlier results belong to 25bbd438 only.
Not merge-ready: it still needs current-head CI, a fresh review, maintainer approval and the BLOCKED live/UI gates above.
nikhilachale
left a comment
There was a problem hiding this comment.
Changes requested.
I reviewed the Strands TUI adapter (launch/restore, checkpoint validation, activity and permission bridge, auth probe, Node preload), the registry, domain, migration and install wiring, and the Settings Models and Account changes. The pinned install and refused updates, the fail-closed native config validation, restore checkpoint checks and auth status handling look careful. Two problems need fixing before this is ready (see inline comments):
safeIDrejects valid AO session IDs that contain.or uppercase letters, so Strands cannot launch in projects such asnext.js.- The
--importpreload is injected throughNODE_OPTIONS, so every Node process the agent starts from the project inherits it. Older Node versions (<18.18 / 16) cannot run at all with it.
Optional, non-blocking:
backend/internal/cli/doctor.go: the newstrandsentry is inserted between the Command Code comment and itscommand-codeentry. The comment now sits above the wrong harness, so movestrandsbelow.HarnessModelsPanel: the newfreeformOnlybranch applies to every harness with a settled, empty text-mode catalog, not just Strands. Please check that the "Configure a provider and default model in its CLI" copy fits the other manual/text harnesses that hit it.AuthStatusruns a Node subprocess plus a network catalog call (up to 15s) each time it is invoked. Consider caching if readiness polls it often.
safeID was applied to the AO session id, but AO session ids carry the project prefix, which keeps dots and case (a `next.js` project yields `next.js-1`). Every Strands spawn in such a project failed and SessionInfo returned a hard error. Validate the AO id with the alphabet `ao hooks` accepts, bounded in length, which still rules out separators, a leading dot and `..` in the launches/ file name. The native session id keeps the strict check. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The --import preload stayed in NODE_OPTIONS for every Node process the agent started, so a project using a Node without --import support could not run node or npm at all. The preload now removes its own token from NODE_OPTIONS, as it already consumes the launch-only workaround flag, and leaves the user's options intact. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Keeps Command Code's comment beside its own entry. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
nikhilachale
left a comment
There was a problem hiding this comment.
Approved. Both findings from the previous review are fixed at d4f23b87d4:
- AO session IDs: the new
validAOSessionIDaccepts dotted and mixed-case project prefixes (next.js-1,MyApp-2). It still rejects leading dots,.., a trailing., path separators and IDs over 200 characters. The native UUID keeps the strictsafeIDcheck. The new test covers launch,SessionInfoand restore for the accepted IDs, and rejection for the unsafe ones. - NODE_OPTIONS leak: the preload now removes its own
--import "file:.../agents/strands/ao-node-hooks.mjs"token fromprocess.env.NODE_OPTIONS. It keeps any other options and deletes the variable when nothing is left. The pattern matches the URL thatwithNodePreloadproduces on both Unix and Windows. Spaces are percent-encoded, so the quoted value never contains a quote. The new test checks that a child process sees exactly the remaining options.
The doctor entry has also moved below Hermes. The rest of the delta since the last review comes from main (Cloud agent switch, handoff note extraction) and does not touch the Strands adapter. The Strands migration is still the newest one.











Code changes: +1175 -9
Tests: +1641 -8
Others: +3139 -4
Binary assets: 5 historical PNGs, excluded from line counts. New official SVG assets include upstream license/notice. No new reports, logs or screenshots are in the product diff.
What
Add the Strands TUI harness and fix the setup/auth/models/branding gaps found during the PR7081 skill review. Current source:
25bbd4386ea9d4cf3aed5fc2d14054e4c4adf739, reconciled with main49951ca.strands --setupwizard in AO's terminal.The PR remains draft while fresh UI/CI verification completes and actual provider lifecycle qualification remains unavailable on this VPS.
Current review/evidence · Actual before screenshots
Why
The previous branch offered a setup plan that unknown-auth users could not reach, displayed an empty catalog for a freeform harness, used a placeholder avatar and could not verify provider authorization. Native Strands exposes the setup/provider discovery APIs needed for truthful checks.
Before / After
Before captures are attached in the linked comment. Fresh after-fix captures are being taken from the current source and will be attached directly using official GitHubCLI --attach; no local computer access or evidence branch is used.
How
The auth probe loads the installed official CLI0.2.0's native configuration/discovery APIs in memory, without load()'s disk rewrites or custom modules/plugins. It emits only a validated status and suppresses raw native errors. Probe execution is bounded. Anthropic/OpenAI/Google first-party authenticated catalogs are supported; static, custom or inconclusive inventories never prove authorization.
Setup guidance explicitly states native wizard keys last only for that terminal; persistent provider credentials are required for future AO sessions. No user credentials were read/injected or provider login completed by this review.
The SDK activity/cancellation hooks, exact checkpoint validation, private instructions, domain/installer/migration/API wiring and current-main reconciliation remain. Automatic installation stays pinned to0.2.0. The native0.2.2 comparison did not qualify an upgrade.
Intentional scope: TUI only. Native ACP lacks required mode/model APIs and permission behavior and restores stale private instructions; its candidate stays unregistered. No Chat/reviewer or unsupported permission claims are made.
Testing
Checklist
Preserved historical PR description — prior source snapshots, not current qualification
Code changes: +998 -13
Tests: +1222 -5
Others: +2980 -3
Latest source snapshot:
104316c03b6b34d8debec7f5d4f42a58204414b6includes actual mainff6e7b3a424e8e0765f0fc521572e52d2aefdfaf. Unregistered safety-gated candidate; Strands ACP is not implemented or ready. Native permission/cancel/restore contracts remain unresolved; current-head CI pending. Implementation snapshot.Current ACP status:
cb740a387cb7142ea97b36a7d0950141f7e4d0a4; actual main55b02b848dd3f05f4d010fc5d603d5e9e6217669integrated. Draft; native approvals, executing interrupt, fresh restore instructions and workspace ownership fail. Chat stays unregistered; current-head CI pending. Final ACP report.Current reconciliation:
0f659657247b427f7bdd0bb304593f826e0f3f4b; actual main55b02b848dd3f05f4d010fc5d603d5e9e6217669integrated. Draft; ACP qualification incomplete and current-head CI pending. Reconciliation comment. Historical status below remains preserved.Binary assets: 5 (excluded from line counts).
Current publication, 2026-10-15:
4638c82aca17efa7d7d4148af37d6f2578a4637f. Incomplete qualification; AO UI evidence BLOCKED; not merge-ready. Main conflicts reconciled. Exact-head CI is recorded below; no local routine suites.Published current-main reconciliation and new exact7ebd0dca diagnostic evidence:24PASS/1authBLOCKED/1catalogFAIL/2NOT_RUN. Independent ordinary file edits worked, but fresh/restored private/project/native instruction receipt effects were absent. Assisted runner success does not replace that independent failure.
Current report and scoped evidence. Historical runtime/source attribution is retained there.
The initial publication live-feeds check failed: a pinned install spec was sent as an npm package name (HTTP400). The new source separates unversioned package metadata from pinned0.2.0 installation/reinstallation and refuses automatic updates to unqualified releases. Added regression cases are remote-CI-only; all original runtime/UI non-pass gates remain.
Exact-head CI (2026-10-15T07:12:38.660620+00:00): complete at
4638c82aca17efa7d7d4148af37d6f2578a4637f— 25 SUCCESS (25 reported; 25 expected including the merged Go aggregate gate). Current checks. MergeabilityMERGEABLE, merge stateBLOCKED; draft handoff. CI does not qualify missing native/UI/provider/OS gates. The initial head293493b9 live-feeds HTTP400 failure remains preserved; the fix is tested only by this newer head’s CI. Canonical main subsequently advanced to48c4ef5e with only two non-overlapping preview-process files; GitHub reports this head mergeable. These completed checks belong to the stated publication head, not a newly rebuilt merge with that later base.Preserved earlier description (historical status, validation and evidence snapshots)
Binary assets in the current PR diff: 5 (excluded from line counts).
Strands is now registered as a TUI-only harness using the supported public SDK shell replacement. Exact native restore, additive private instructions and foreground tool cancellation are demonstrated inside real AO Electron. Custom tools/sandboxes/subagent modules and unsupported native reconfiguration fail closed or are explicitly out of scope. No Chat or handoff support.
Detailed recovery report and real screenshots. The functional audit passed all lifecycle gates; its overall verdict remains diagnostic FAIL (24 PASS / 1 auth BLOCKED / 1 catalog FAIL / 2 catalog NOT_RUN). Current evidence HEAD:
022084e99b25e5aff7e31c46ebc1b6e869bf320f; all 23 current-head CI checks pass (workflow). AO003 used bypass-permissions/native Auto; native one-time-approval controls were separate. AO authentication probe remains configured rather than verified authorization, despite successful real provider calls.The diff includes 5 PNG assets (3 actual AO Electron captures and 2 historical native-only captures); binary assets have no line counts.
Original native-only checkpoint and preserved failure history
Strands CLI 0.2.0 displays “Cancelled” after Escape or Ctrl+C while an approved shell tool continues running and writes its delayed artifact about 28 seconds later. This draft records that blocker and an isolated, unregistered terminal adapter candidate. Strands is not available in AO; AO lifecycle and UI conformance remain BLOCKED / NOT RUN.
The candidate preserves native profile instructions/model/plugins, separates the visible task from standing instructions, rejects missing or malformed restore history, and reports native activity through a supported SDK plugin. It does not modify domain/registry/activity dispatch, migrations, APIs, installation, pickers, Chat, or reviewers. Unsupported permission mappings and model overrides shadowed by native model modules fail explicitly.
The full conformance report records pinned upstream revision, native initial/private-instruction and exact-restore evidence, both cancellation failures, runner limitations, reproduction instructions, and two actual native-only terminal screenshots. A PR comment provides revision-pinned report and image links. The screenshots are from a separate native reproduction and do not show AO.
Validation on the Linux VPS:
c3c834e0e7ef, all 21 GitHub checks passed, including Go build/tests/race/lint/drift, CLI E2E on Linux/macOS/Windows and in the container, and gitleaks. These repository checks do not establish native Strands cancellation or AO lifecycle/UI support. No API regeneration is needed.Do not register Strands until a pinned upstream fix or a separately reviewed existing containment boundary proves bounded cancellation without late side effects, followed by independent AO lifecycle and actual AO UI evidence. This draft is not a merge-ready integration.
Binary PNG assets in current PR diff: 5 (binary line counts excluded from the header).
Account migration verification
25bbd4386ea9d4cf3aed5fc2d14054e4c4adf739.milesatao-svg/agent-orchestrator:codex/copy-pr-6517-strands. Base:main.