Repository navigation
fix(billing): say Free includes every supported agent locally - #6963
Conversation
The Free plan card listed "Claude Code, Codex and Cursor", which read as if Free were limited to three agents. The local daemon has no plan checks on harness choice; plan limits apply only to cloud sandboxes. Say "Any supported coding agent, including Claude Code, Codex and Cursor" in all eight locales. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🏆 Review leaderboardOct 13–20, 2026 · UTC
Ranked by distinct external PRs reviewed, then review rounds, then PR comments. Self-activity and bot activity are excluded. Show 8 more reviewers
|
AgentWrapper
left a comment
There was a problem hiding this comment.
Approved. Reviewed head 2db7e60. This is a copy-only change to settings.billing.freeAgents, the only key that changes. It is updated in all 8 locale catalogs the renderer ships (en, de, es, fr, ja, ko, pt-BR, zh-CN), and the Free-card assertion in BillingSection.test.tsx matches the new English string. The claim is accurate: local harness selection is validated only against the known harness set, and the local daemon has no plan or entitlement checks; plan limits are enforced only on billed cloud sandboxes. The translations reuse each catalog's existing term for coding agent and read naturally.
Optional, out of scope: the same Free card still says "Same-network connectivity (no cloud relay)" and "2 devices logged in at once". The first undersells the free Tailscale and managed HTTPS tunnel access, and nothing in the code enforces the second. The docs PRs (#6912, #6915) already dropped both. Worth aligning the card in a follow-up once the founder settles the relay and device lines.
ru.json landed on main after this PR, so the Russian Free card still
Rishet11
left a comment
There was a problem hiding this comment.
Approved. No required changes found at 5f69ae7.
The diff updates only the Free-plan agent copy in all nine supported locale catalogs and the matching BillingSection assertion. All nine catalogs parse, and the assertion matches the English copy.
Read-only review. Tests and the app were not run; CI is still pending.
Rishet11
left a comment
There was a problem hiding this comment.
Checked the claim in code: the local daemon validates harnesses only against AllHarnesses (backend/internal/domain/harness.go:53) and has no plan checks; plan entitlements only gate cloud sandboxes. Added the missing ru.json string. CI green.
* fix(desktop): respawn the daemon after an unexpected exit (OrchestratorInc#6171) (OrchestratorInc#6197) * fix(desktop): respawn the daemon after an unexpected exit When the daemon child exited while the app stayed open, the supervisor only recorded an "exited" failure status and relied on the manual Restart daemon banner. Nothing respawned it, so sessions stalled and activity tracking degraded until the user intervened (OrchestratorInc#6171). Respawn an app-owned daemon on an unexpected exit with bounded exponential backoff (up to 6 attempts in a 5-minute sliding window, 1s-30s). A deliberate stop is left alone: the Go daemon removes its liveness run-file and exits 0 on a graceful shutdown (including ao stop), so only a left-behind run-file, a non-zero exit, or a terminating signal counts as a crash. The pending respawn is cancelled on explicit stop/restart and on app quit. Co-authored-by: CommandCodeBot <noreply@commandcode.ai> * fix(desktop): bind daemon auto-restart to the crashed child The crash test keyed only on "a run-file exists", which both undid a deliberate ao stop and blamed a successor daemon's run-file on the exited child. - Classify by ownership: respawn only when the run-file still names the dead child's PID. A removed file (ao stop, self-stop) or a file naming a different PID (a successor) is left alone. Drop the exit-code heuristic: a requested `ao stop` that outlives its 10s drain deadline exits non-zero after removing the file, and must not be respawned. - Re-read the run-file before the scheduled respawn fires and stand down if it is gone or names someone else; remove the known-stale marker before spawning so a reused PID cannot be mistaken for a live daemon. Co-authored-by: CommandCodeBot <noreply@commandcode.ai> --------- Co-authored-by: CommandCodeBot <noreply@commandcode.ai> Co-authored-by: codebanditssss <diwankhushi428@gmail.com> Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> (cherry picked from commit 1ead720) * fix(report): never press Ctrl-C in terminal orchestrators for stuck reports (OrchestratorInc#7435) * fix(report): never press Ctrl-C in terminal orchestrators for stuck reports A stuck report made the coordinator interrupt the target orchestrator before submitting. In a terminal (TUI) orchestrator that Ctrl-C cancels the running turn, or a prompt that was submitted moments earlier, and the following submit is refused. The batch is then retried at the end of the 3-minute interrupt window with another Ctrl-C, so the loop never ends: on 28 Oct it cut six orchestrator turns, cancelled one delivered session message and lost one report that AO had already acknowledged. Terminal orchestrators queue typed input during a turn, so stuck reports now take the normal submit path there, with no interrupt reservation and no deferral. Chat-mode orchestrators keep the rate-limited interrupt. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MquzzN9pQgHv18U6ti9Sz2 * fix(report): retry a refused submission behind reports already due Schedule selection takes the earliest pending report across all projects. A submission refused because the composer holds the operator's draft already backs off for OrchestratorBusyRetry, but any other refusal (for example "prompt submission was not observed") was released with its original availability. The same batch stayed at the front, and every retry picked it again, so one orchestrator that kept refusing blocked report delivery for every other project. Stuck reports used to escape this through the interrupt-window deferral, which the previous commit removes for terminal targets. Other refusals are now deferred by the retry delay. Reports that are already due go first, and the refused batch keeps its batch identity and its refusal reason in last_error. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MquzzN9pQgHv18U6ti9Sz2 --------- Co-authored-by: David Wang <wym163@hotmail.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit 17eb960) * fix(session): return CHAT_CONTROLLER_NOT_READY from send without a live controller (OrchestratorInc#7476) Sending to a chat session whose agent controller stopped or never started surfaced chatsvc.ErrNoController as a bare 500. Map it, and ErrNotProvisioning, to the same 409 steer-or-send returns. Fixes OrchestratorInc#7359 (cherry picked from commit e060567) * fix(workspace): preserve quoted Windows project script commands (OrchestratorInc#7246) (cherry picked from commit 3671d4f) * fix(remote): stop a reset remote tunnel from crashing the main process (OrchestratorInc#7515) * fix(remote): stop a reset remote tunnel from crashing the main process Over HTTPS the remote proxy runs its HTTP/2 session inside a WebSocket (createWebSocketStream). When that connection is reset (sleep, a Wi-Fi or VPN change), ws closes the WebSocket without its TCP socket finishing, so the stream's end never completes. The session is never told its socket is gone, and every frame it then writes (a stream's RST_STREAM, the keep-alive's GOAWAY) fails Node's assert in JSStreamSocket.doWrite with an uncaught ERR_INTERNAL_ASSERTION: one "JavaScript error occurred in the main process" dialog per frame. Destroy the stream when the WebSocket closes, so the session closes with it and the next request opens a new tunnel. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(remote): end a reset tunnel without dropping what it read, and retry its GETs Review follow-ups on the tunnel reset fix: - Destroy the tunnel stream at its end, not on the WebSocket's close. While a write is in flight the session stops reading, so frames the host sent before closing can still sit in the stream; destroying it on close dropped them, and a finished answer came back as a 502. The stream ends only after the WebSocket closed and everything it read was handed on. - Share that stream (webSocketTunnelStream) with the test hosts, whose own tunnel had the same half-close bug: a host writing after a reset threw ERR_INTERNAL_ASSERTION in the test process. - A host that can't be reached to check its identity (fetch's network failure or timeout) is answered 502 UPSTREAM_UNAVAILABLE, not HOST_IDENTITY_UNVERIFIED, and logged as unreachable. - A GET or HEAD whose session was lost before its answer goes once more, over a new session, unless the proxy is closing. POSTs and DELETEs still answer 502. - Log a tunnel the host or the network closed, with its close code (1006 is a reset). - The reset test waits fast.pingMs * 2.5 instead of a fixed 500 ms. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(remote): count only network failures as unreachable, and keep a retry from opening sessions after close or forcing HTTP/1.1 - unreachable() now matches fetch failures by their cause's network code (refused, reset, DNS, unreachable, undici socket and connect timeout) or a timeout. A host that answered (a redirect, a TLS certificate) is again an identity failure, and the log names fetch's cause, not only "fetch failed". - A retry whose session open fails no longer switches the proxy to HTTP/1.1 for a minute. It goes the moment its session died, and a LAN host still restarting refuses it. - close() now stops every request from opening a session, not only retries. A request that was waiting on a session that turned out closed (a GOAWAY) opened a new one, which nothing closed. - A request sent again logs its first try's failure. - The retry tests send body-less DELETE/POST requests, so the method check is what stops them going twice. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit 3f514ac) * fix: preserve literal tabs in command cues (OrchestratorInc#6353) * fix: preserve command bytes when invoking cues * ci: install shells required by Cue regression tests Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> * ci: isolate Zsh completion setup for Cue shell fixtures Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> --------- Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> (cherry picked from commit 962baa3) * fix: exclude unvisited bookmarks from Firefox history import (OrchestratorInc#6380) Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> (cherry picked from commit 26ea26f) * fix(billing): say Free includes every supported agent locally (OrchestratorInc#6963) * fix(billing): say Free includes every supported agent locally The Free plan card listed "Claude Code, Codex and Cursor", which read as if Free were limited to three agents. The local daemon has no plan checks on harness choice; plan limits apply only to cloud sandboxes. Say "Any supported coding agent, including Claude Code, Codex and Cursor" in all eight locales. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(billing): add Russian Free agents copy ru.json landed on main after this PR, so the Russian Free card still --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Rishet Mehra <rishetmehra11@gmail.com> (cherry picked from commit 0b85db4) * fix(billing): say Free connects from anywhere via Tailscale or the tunnel (OrchestratorInc#6964) * fix(billing): say Free connects from anywhere via Tailscale or the tunnel The Free plan card said "Same-network connectivity (no cloud relay)", which undersold Free: Connect Mobile also advertises the computer's Tailscale address and a managed HTTPS tunnel on every plan. Say "Connect on the same network, or from anywhere with Tailscale or the HTTPS tunnel" in all eight locales, matching the docs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(billing): add Russian Free network copy ru.json landed on main after this PR, so the Russian Free card still said same-network only. Match the new English line. --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Rishet Mehra <rishetmehra11@gmail.com> (cherry picked from commit 818a2ba) * fix(workspace): move orphaned worktree folders aside before add (OrchestratorInc#7483) A folder left at a fixed worktree path (such as a project's orchestrator worktree) after git pruned its registration made every later `git worktree add` fail with "already exists". When the folder is provably an orphaned worktree of the project's repo, rename it to a sibling <name>.orphaned-<timestamp> and continue; anything else at the path fails with ErrWorkspacePathOccupied (409 WORKSPACE_PATH_OCCUPIED) and is left untouched. Fixes OrchestratorInc#7358 (cherry picked from commit 4255167) * fix: preserve newer files when removing cancelled downloads (OrchestratorInc#6377) Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> (cherry picked from commit 9041e10) * chore(sync): ledger + settle миграции — батч 2026-11-01 Co-Authored-By: Claude Code <noreply@anthropic.com> * fix(sqlite): дополнить леджер миграций записями 9017/9018 из fork/main PR-мержи #193-#202 принесли 9017_session_review_only и 9018_pr_last_seen_at без записей в shippedMigrations — TestMigrationVersionLedger ловит пропуск. Co-Authored-By: Claude Code <noreply@anthropic.com> * fix(report): вырезать chat-прерывания застрявших репортов У форка нет chat-оркестраторов (чаты выпилены/отключены), поэтому rate-limited Ctrl-C путь из upstream OrchestratorInc#7435 для них мёртвый код. Застрявшие репорты во всех режимах идут обычным submit-путём: терминальные оркестраторы очередят ввод сами, а отказ submit бэкаффится retryLater. Вырезано: ветка interrupt в deliverProject, ReportInterruptWindow, AcquireReportInterrupt (интерфейс + sqlite-стор + фейки), stuckWorkers. Тесты: TestCoordinatorStuckInterruptRateLimitAndCoalescing удалён, NeverInterrupts теперь гоняет и chat-режим, DeferredBatchReclaimsAfterExpiry (issue #121) переведён на retry-бэк офф. sqlc-запросы Get/PutReportInterrupt и таблица в живых БД оставлены инертными (regen сломан, OrchestratorInc#4621). Co-Authored-By: Claude Code <noreply@anthropic.com> * fix(migrations): burn migration number 9016 The review_run_github_review_state migration shipped to fork main builds (PR #189, 2026-10-31) as 9016 and was then renumbered to 20261028124744442. Installs from those builds have 9016 recorded as applied, so a future file reusing the number would be skipped silently there; add it to burnedVersion next to 9003/9004. Review finding on #203. Co-Authored-By: Claude Code <noreply@anthropic.com> * fix(ci): green the test, lint, and sqlc-drift gates - coordinator.go: drop the unused PreparedBatch.hasState left behind by the interrupt removal (golangci: unused). - sqlite: the 9016 -> 20261028124744442 migration rename moves the github_review_state column earlier in sqlc's schema, so reorder the review_run SELECT column lists to match the regenerated model and commit the sqlc output (sqlc-drift). - daemon-auto-restart integration test: mock the daemon-log API (openDaemonLog, writeDaemonLog, closeDaemonLog, daemonLogPath) that the non-keep startup path now calls; without them every start test died on ReferenceError before spawning. Co-Authored-By: Claude Code <noreply@anthropic.com> --------- Co-authored-by: Yash Maheshwari <yashmaheshwari8983@gmail.com> Co-authored-by: CommandCodeBot <noreply@commandcode.ai> Co-authored-by: codebanditssss <diwankhushi428@gmail.com> Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> Co-authored-by: rtd-davidwang <david.wang@run-true.com> Co-authored-by: David Wang <wym163@hotmail.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Pardeep Singh <pardeep.shergill17@gmail.com> Co-authored-by: OllieinCanada <73385593+OllieinCanada@users.noreply.github.com> Co-authored-by: Dhruv Sharma <dhruvcoding67@gmail.com> Co-authored-by: prateek <karnalprateek@gmail.com> Co-authored-by: Rishet Mehra <rishetmehra11@gmail.com> Co-authored-by: axisrow <axisrow@users.noreply.github.com>
Summary
The Free plan card on the Upgrade page listed "Claude Code, Codex and Cursor". That reads as if Free were limited to three agents, but local use isn't plan-gated at all. This changes
settings.billing.freeAgentsto a count-free line that won't go stale as harnesses are added:The translations reuse each catalog's existing term for "coding agent" (for example
settings.cloudAgents: Coding-Agent, agente de código, コーディングエージェント, 코딩 에이전트, 编码代理).Nothing else on the plan cards changes. The relay, device and limit lines are a separate founder decision.
Why the claim is true
domain.AgentHarness.IsKnown()overdomain.AllHarnesses(backend/internal/domain/harness.go).backend/internal/config/config.gosays the cloud offering, including billing, "is decided per organization by the cloud control plane".backend/internal/pricingonly tracks provider token costs.cloud/internal/postgres/billing_store.go:admitCreate,sandboxBilledTx), as fix(billing): show plan limits apply only to cloud agents #6907's description also states.Tests
BillingSection.test.tsx.vitest run src/renderer/components/settings/BillingSection.test.tsx src/renderer/i18non Node 24: 3 files, 37 tests passed, including the i18n locale-coverage check.Risks
Copy only. Translations were written in-house with each locale's existing terminology; a native-speaker pass is welcome but not required.
🤖 Generated with Claude Code