Demo videos

Feature demo videos recorded from the real TUI. Index: qa/demos/index/. SOP: qa/demos/README.md.

0.1.48

Corbanu Terminal 0.1.48

• Busy providers no longer end your turn. When OpenRouter reports that a
  model's shared upstream pool is briefly rate-limited and asks callers to retry,
  Corbanu now waits and retries (up to four times, honoring Retry-After) instead
  of stopping the task on the first 429. Your own quota and usage limits still
  stop immediately.
• MiMo V2.6 Pro on OpenRouter. xiaomi/mimo-v2.6-pro is now in /model.
• Better GLM 5.3 Flash results. GLM 5.3 Flash on Z.AI, OpenRouter and Vercel
  gets a short work pattern: batch independent calls, write complete
  implementations, and verify by running code before finishing.

Carries forward the 0.1.44 limitation that cancelling a long streamed answer can
leave later requests stalled; start a fresh session if this happens.

Released under explicit operator authorization. Full cross-platform
qualification and the competitive benchmark cycle remain incomplete.
Release record.

0.1.47

Corbanu Terminal 0.1.47

• Long Claude sessions with many screenshots work again. Claude rejects a
  request with more than 20 images if any image is wider or taller than 2000
  pixels, so every turn failed, even after "continue". Images are now kept
  within that limit, and existing sessions recover when resumed after updating.
• Stronger loop protection. On top of 0.1.46, Corbanu now also stops a model
  that cycles through the same few calls or flips an edit back and forth without
  getting new results.

Restart Corbanu after updating, then resume any session that was failing.
Carries forward the 0.1.44 limitation that cancelling a long streamed answer can
leave later requests stalled; start a fresh session if this happens.

Released under explicit operator authorization. Full cross-platform
qualification and the competitive benchmark cycle remain incomplete.
Release record.

0.1.46

Corbanu Terminal 0.1.46

Fixes several provider and agent-loop problems found in benchmarks against
Hermes.

• Runaway tool-call loops are stopped. A model that repeats the same call
  over and over (seen with GLM 5.3 Flash on Vercel, up to 250 times) now has
  repeats refused after three in a row, and the turn stops after eight.
• GPT-6 Sol works over OpenRouter. Its code-execution calls were rejected
  before; it now passes the benchmark tasks.
• No more runaway whitespace output. One malformed tool call could make a
  model emit tens of thousands of blank tokens; the tool format no longer allows
  it.
• GLM 5.3 and GLM 5.3 Flash work on Vercel accounts that require zero data
  retention.
• Fewer wasted Vercel requests. A rejected server-side continuation is no
  longer retried every turn, and Corbanu no longer adds a fake "Continue."
  message the model mistook for the user.

Known limitation: a model that cycles through several calls without making
progress is not yet stopped; a fix is in progress. Carries forward the 0.1.44
limitation that cancelling a long streamed answer can leave later requests
stalled; start a fresh session if this happens.

Restart Corbanu after updating. Released under explicit operator authorization.
Full cross-platform qualification and the competitive benchmark cycle remain
incomplete. Release record.

0.1.45

Corbanu Terminal 0.1.45

Fixes prompt caching for Claude and other explicitly cached models over
OpenRouter and Vercel.

• Cache markers now follow the newest turn in agent loops. Before, every tool
  turn re-billed the whole conversation at the full input price when Claude,
  MiniMax, or some Qwen models were used through the openrouter or vercel
  providers. In an Opus 5.5 test on OpenRouter, a debugging task dropped from
  $0.53 to $0.30 and cached input rose from 38% to 79%.
• Claude Plan, direct Anthropic, and GLM, Kimi, and DeepSeek V4 routes were not
  affected.

Carries forward 0.1.44 unchanged otherwise, including its known limitation:
cancelling a long streamed answer can leave subsequent requests stalled; start a
fresh session if this happens.

Restart Corbanu after updating. Released under explicit operator authorization.
Full cross-platform qualification and the competitive benchmark cycle remain
incomplete. Release record.

0.1.44

Corbanu Terminal 0.1.44

Updated models and restored GPT-6 access through existing ChatGPT sign-in.

• Fixes OpenAI compatibility negotiation: an outdated client-version header could reject GPT-6 Sol and Luna even when the same login worked in official Codex.
• Adds Claude Opus 5.5 Plan with exact model routing, separate from Anthropic API billing.
• Adds GLM 5.3 Flash to Z.AI Coding Plan and fixes persistence of its supported reasoning efforts.
• Adds OpenRouter Grok 4.7, DeepSeek V4.1 Flash, GLM 5.3 Flash, and nvidia/nemotron-3-ultra-550b-a55b:free.
• Removes legacy OpenRouter DeepSeek V4 Pro/Flash, Grok 4.5/4.6, and Owl Alpha from normal selection and automatic allocation. Explicit legacy IDs remain compatible.
• Hides GPT-5.5 from the selector while preserving explicit-ID compatibility; includes GPT-6 Luna.
• Preserves the provider-credential replacement fix shipped in 0.1.43. No account or billing fallback is introduced.

Known limitation: cancelling a long streamed answer can leave subsequent requests stalled. Start a fresh session if this happens. Normal Sol chat, tool use, and restart passed local checks; cancellation recovery did not.

FlashX is not included: Z.AI currently excludes it from Coding Plan, and a separate pay-as-you-go route was not authorized.

Restart Corbanu after updating. The operator explicitly authorized this release with the disclosed limitations. Competitive/model benchmarks and full cross-platform interactive qualification remain incomplete; no benchmark pass is claimed. Release record.

0.1.43

Corbanu Terminal 0.1.43 lets you replace a configured provider credential from the Providers menu, so a saved invalid token no longer strands a session.

• Configured providers keep their Activate/Deactivate controls and now also list their existing "Replace with …" account or API-key methods.
• Claude Plan uses credential-scoped reauthentication: renewing a known credential source, otherwise starting a full Replace.
• No authorization, vault, or credential-storage boundaries change; masked entry, cancellation, persistence, and cache invalidation are unchanged.

Restart Terminal after updating. This release preserves your existing profile and credentials.

The fix passed focused automated tests (codex-tui provider manager and Claude intent, 11 tests) with refreshed menu snapshots, plus the keyboard-driven terminal recovery checks recorded for the 2026-09-12 provider-replacement hotfix. Hosts that were running the local 0.1.42-provider-replacement candidate can return to normal corbanu update behaviour with this release. Release evidence.

After updating, open /providers, pick a configured provider, and confirm a "Replace with …" entry is listed alongside Deactivate.

0.1.42

Corbanu Terminal 0.1.42 fixes wallet startup and model-request recovery, and adds Team Context plus DeepSeek V4.1 Flash to Corbanu API.

• Adds shared collaborator work summaries through /tasknode team and corbanu tasknode team context --json, with refresh and report freshness indicators.
• Updates the bundled Task Node skill so agents can retrieve Team Context while preserving the active account and sharing permissions.
• Fixes wallet startup in standalone installations where the client looked for the old daemon filename.
• Keeps models in their correct provider tabs, including Claude Plan.
• Improves Campaign Tracker error handling and recovery.
• Adds DeepSeek V4.1 Flash with High reasoning under /model → Corbanu API.
• Recovers safely from confirmed released API attempts without the misleading duplicate-request 409.
• Omits the unsupported parallel-tool control on Flash requests while preserving tool use.

The matching gateway fixes are already deployed. Restart Terminal after updating. This release preserves your existing profile and credentials.

The fixes passed focused automated tests and live terminal checks. Cross-platform archives also run packaged wallet-client startup checks before publication. The broader competitor/model benchmark cycle remains incomplete; no benchmark pass is claimed. Release evidence.

After updating, start a new chat and ask: “Refresh my Team Context and tell me what each collaborator is working on.” Team Context summarizes shared rewarded work; it is not a live feed of every unfinished task.

0.1.41

Corbanu Terminal 0.1.41 fixes Task Node account recovery when using multiple named profiles on one machine.

• Each profile keeps its own independently linked session.
• Completing /tasknode link replaces a stale saved session when you run /tasknode status.
• Status shows the current profile, and CLI linking instructions keep the selected profile.
• Authentication errors give current Corbanu instructions.

The live Task Node linking service also now opens GitHub's account picker. If a saved session was revoked, restart Corbanu, run /tasknode link, choose the intended GitHub account and then run /tasknode status.

The repair passed 48 focused Rust tests and fresh 0.1.41 terminal checks in TensorCash and Isometric Game covering two-account isolation, relinking, restart, cancellation and concurrent status. This emergency release is authorized with the full competitor/model benchmark matrix incomplete. Release evidence.

0.1.40

This emergency release fixes agent-side Task Node commands selecting a different account from the active terminal profile. Agent commands now inherit the terminal's home and profile and reject missing or conflicting scope before opening account credentials. Unix launchers preserve the inherited home.

Provider misalignment_policy_violation responses now stop immediately instead of triggering reconnect attempts or transport fallback. Later stream tool output is discarded. The provider's refusal remains in force; upgrading does not clear a provider-side conversation stop.

Validation: 228 focused Rust regressions passed, plus five installer tests and the no-regex provider-path check. Actual TUI and helper account agreement passed for default and named profiles, including cold restart, using local model fixtures and read-only live Task Node status. The interactive refusal fixture issued exactly one request. Full competitor benchmarks and cross-platform interactive acceptance remain unqualified and are disclosed in the release record.

Restart existing terminal processes after upgrading to load the repair.