Polytoken 0.8.20

Polytoken 0.8.20 added the claude-haiku-5.5 model to its Anthropic provider, cleaned up rate-limiting logic, and made the rate-limit timer honor a provider's Retry-After header. The release also lets the ask_user_question tool fire in goal mode within a few minutes of the last prompt instead of being suppressed entirely, an experimental change that may be reverted, and fixes hyperlinks, clipboard copying, and built-in facets such as plan-reviewer respecting their designated modelgroups. The prior 0.8.19 release introduced a breaking change binding built-in facets (execute, plan, orchestrate) to their own model groups, so switching facets now selects that facet's model group rather than keeping the session's current model.

Changelog 0.8.20 Section titled “0.8.20” 0820 Polytoken 0.8.20 adds Claude Haiku 5.5, cleans up rate-limit handling, honors Retry-After, and lets ask user question fire in goal mode. - Added claude-haiku-5.5 to the Anthropic provider, added under protest. - Cleaned up some more rate-limiting logic. - If a provider sends a Retry-After header, the rate-limit timer will honor it. - ask user question can fire in goal mode within a few minutes of your last prompt, instead of being suppressed entirely. This is experimental and may be reverted. - Some fixes around hyperlinks and copying to clipboard. - Built-ins like plan-reviewer now respect their designated modelgroups. - Bugs. Fixed? Added? Who can say. 0.8.19 Section titled “0.8.19” 0819 Polytoken 0.8.19 binds the built-in facets to their own model groups, makes image handling in the read tools honest, sends explicit non-strict tool definitions to OpenAI-compatible providers, completes —model values from your actual model catalog, explains failed facet switches and deferred-catalog misconfigurations, improves MCP OAuth and herdr integration, sharpens delegation advice and goal-clear queuing, retries capacity limits smarter, and cleans up better after crashes. - Breaking change: switching to a built-in facet execute , plan , orchestrate , including through a plan handoff, now selects that facet’s model group polytoken:execute facet , polytoken:plan facet , polytoken:orchestrate facet instead of keeping whatever model the session was on. If you deliberately picked a different model and then switch facets, the session moves to the group’s model. To keep your own selection, author the group with your models for example modelgroups "polytoken:execute facet" = "your-model" , or define a project facet without a model pin; removing an authored override restores the shipped binding. On configurations where the group cannot resolve, facet switches and handoffs fail with an error naming the unresolved reference instead of silently keeping the old model, and GET /tools/effective reports the same failure. - Tool definitions sent to OpenAI Responses, Codex, and Chat Completions endpoints now state strict: false explicitly, so endpoints that default to strict mode stop inventing values for optional tool fields. - The read tools now say out loud what they always did: images are recognized by file extension only, up to 5 MB, and a renamed screenshot without an extension is a binary blob no matter what its bytes think. - --model completion now offers your actual configured models reasoning variants included instead of whatever files happened to be lying around in the current directory. - Denying tool search under a deferred-tools configuration now fails loudly instead of stranding the tool catalog where neither you nor the model can find it. - Tier defaults in model groups now resolve to the first usable member in authored order, so a group whose first member is disabled or unavailable keeps working instead of failing default inference. No member is ever added automatically. - A legacy defaults tier choice that matches an authored reserved tier group now loads cleanly instead of erroring, while a conflicting one fails the load with an error naming both values and how to remove either one. A raw defaults block in a current config file migrates like an older one instead of being silently used and then silently dropped on save. - A missing default with several eligible models now produces one error that lists the eligible models and names the reserved group and polytoken config ui , instead of two contradictory errors. - Provider Retry-After cooldowns now span prompts: a rate limit with a hint benches that provider for the rest of the daemon session, new prompts skip a cooling provider and start on the next usable group candidate, and a prompt whose whole route is cooling waits visibly and cancellably until the earliest cooldown expires instead of probing. Cooldowns last only for the daemon session. - Improved MCP OAuth handling. - Improved herdr integration a little. - Better subagent delegation advice non-spicy . - Adjustments to the /goal clear UX, to let you queue them up. - When a saved-session goal is active, the model can now actually ask you a structured question if you sent a prompt within the last three minutes: goal mode still answers its own questions autonomously when you’re away, but a question asked right after you’ve said something reaches you instead of being answered on your behalf. The window is per-session and resets on restart; goal continuations the daemon starts on its own never count as your presence. - Labeled Markdown links in the conversation are now real terminal hyperlinks OSC 8 : in terminals that support them Ghostty and others you can click a link label to open it, while the styled label and double-click-to-copy behave exactly as before. Polytoken emits the sequences safely — hostile URLs are encoded or skipped, links under popups and floating windows are never marked up — with no setting and no terminal detection; unsupported terminals ignore it entirely. - Better retry logic when Responses APIs return capacity limit errors. - You can’t wedge a session that requires tool search by disallowing it. - Daemon cleans up after itself better after a crash. - Some odds and ends. - @ file references to files that are not valid UTF-8 now deliver a short omission note instead of garbled replacement characters. - The Jobs viewer’s detail pane takes the mouse wheel across its full width when the content fits, instead of letting those scrolls reach the job list. - A subagent in a provider rate-limit wait no longer transiently shows a stalled marker. 0.8.18 Stable Section titled “0.8.18 Stable” 0818-stable Polytoken 0.8.18 makes facet tool lists closed, adds baseline herdr support and model-group classifier selection, and fixes fd-duplication shell opacity. - Breaking change: in facet frontmatter, polytoken.tools lists are now closed lists. Some undesired tools were sneaking in previously. It is a change to the intended behavior, but it might break something that was wrong in the past. - Baseline herdr support. Still figuring this out. - The permissions classifier can now use a modelgroup. - Shell commands that use fd duplication no longer force an opaque shell check. - Perf improvements for glob . - Assorted housekeeping bug fixes. 0.8.17 Section titled “0.8.17” 0817 Polytoken 0.8.17 adds GPT-6.1 Sol, headless continue, session repair, and smarter group failover. - gpt-6.1-sol now added to OpenAI and Codex providers. - polytoken continue --no-attach is now a thing. - Fixed a write race that was causing sessions to explode hilariously. - polytoken sessions repair <session-id can now fix sessions that have exploded hilariously. - The TUI might not go black when you edit prompts in a separate editor now. It also might not. I won’t make that promise a second time. - Squelched reasoning-loop restart cards. Now your GPT models will spiral endlessly, but more quietly. It doesn’t affect any actual behavior. - Model groups fail over a little smarter. - Goal driver doesn’t fail, a little smarter. - Experimenting with adding spiders instead of bugs. 0.8.16 Section titled “0.8.16” 0816 Polytoken 0.8.16 adds Claude Sonnet 5.5, routes TypeScript 7 projects to the native language server, and lands feedback fixes. - Added claude-sonnet-5.5 , mostly under protest. - Projects using TypeScript 7 or newer will use the native TypeScript language server. - Some feedback fixes. More later. - Added one bug, in particular, to upset specifically you. 0.8.15 Section titled “0.8.15” 0815 Polytoken 0.8.15 improves TUI performance, adds extra definition directories, fixes the MCP configurator, and introduces daemon.spice level. - Improved TUI performance substantially through the fixing of embarrassing oversights. - Added command-line flags and config fields for extra skills, facets, and subagents directories. - The config UI for setting up MCPs was unusable. It has been upgraded to usable. - Added daemon.spice level as a configuration option. Allowed values are 0, 1, and 2. It defaults to 1. Go to 2 if you want to opt into more thorough over time steering for models when used in Polytoken. Go to 0 if you’re scared. - Maintained bug homeostasis, probably. 0.8.14 Section titled “0.8.14” 0814 Polytoken 0.8.14 makes model groups rate-limit-aware, slims compaction’s skill handling, and polishes the group config UI. - Model groups now barrel-shift when they detect rate limiting in a provider. Unlike other errors, this doesn’t eject the provider from the group. - Model group entries now partially degrade more politely. - Model group config UI sucks less now. Still sucks. Working on it. Probably best to have the agent edit its config still though. - Compaction doesn’t reinject all used skills anymore, and instead provides a list of previously-used skills to the new context. Skills recently used are included as part of the verbatim context transfer. - Completion bell no longer fires on assistant turn as long as asynchronous jobs are still running. - Did you know that the term “bug” originates in Middle English and refers to a terrifying spirit? And now you people put them in your code. Not me though. 0.8.13 Section titled “0.8.13” 0813 Polytoken 0.8.13 adds new OpenAI, Codex, and Anthropic catalog models, and makes model-group failures visible. - Added gpt-6-sol , gpt-6-sol-1m , gpt-6-luna , and gpt-6-luna-1m to the OpenAI and Codex providers. - claude-opus-5.5 has also been added to the Anthropic provider, I guess. - Model groups keep usable candidates when another model is unavailable, and Polytoken now explains when a group cannot run. 0.8.12 Section titled “0.8.12” 0812 Polytoken 0.8.12 fixes the thinking drawer leak and adds Xiaomi MiMo 2.6 models. - Fixed the bug where a collapsed Thinking drawer streamed its accumulated reasoning into the main conversation while it ran. - New models: Xiaomi MiMo 2.6 Pro and MiMo 2.6 Flash. - The usual. 0.8.11 Section titled “0.8.11” 0811 Polytoken 0.8.11 moves feedback reporting off Sentry, adds GLM-5.3 catalog models, and speeds up large sessions. - We have stopped using Sentry because of issues with reporting. We’ve lost feedback submissions between September 4 and today because we hit a limit and dropped them on the floor after a flood of bad requests. Sorry about that. - Relatedly, license agreements have updated to reflect that we are no longer using Sentry. - GLM-5.3 is now on Mistral. GLM-5.3 and 5.3 Flash are now on DeepInfra. - Some pretty nice performance improvements with large sessions. - You can add TUI logs to /feedback now. - /goal resume now unblocks goals. - Compaction re-adds recently attached images. - Refilled 5-hour and weekly polytokens. - More bugs. Added or removed, who’s to say? 0.8.10 Section titled “0.8.10” 0810 Polytoken 0.8.10 fixes session resume for migrated and fresh sessions, makes compact context operator-directed, and polishes the config UI. - Bug fix for migrated old sessions. And also new sessions. I wasn’t lying about the “more bugs added” last time. - compact context is now described as operator-directed, so Terra should stop trying to compact itself when it hits 40%. - Minor improvements to the config UI. - A steady state of bugs was maintained. Probably. Polytoken 0.8.9 adds Parallel and You.com search, model groups with ordered failover, live background job output, and a daemon reap command. - New search providers: Parallel and You.com. Both seem pretty good, and they have pretty solid intro offers. - Model groups: list fallback models in order and Polytoken moves to the next one when a provider fails. Substitutable for a model name everywhere ideally? that a model name can be used. Can also be used with a subagent call for fanout. The robot should know how. - Background shell jobs now stream their output live in the Jobs viewer, and the output survives timeouts and cancellation. - polytoken daemon reap cleans up stray local session daemons. - DeepSeek V4.1 Flash is now on Ollama Cloud. - Subagent spawning respects the concurrency cap by queueing counted batches instead of rejecting the excess. - Feedback artifacts are now compressed, so you can send embiggened backfeed. - Turns can now end in an explicit Errored state instead of limping to a normal completion. This should be moderately less confusing. - Tool results are capped in the transcript with the full output spilled to the session directory, so one chatty tool cannot bloat history. - Failed edits now echo the anchor and the nearest match when the target text is not found, so the model can correct itself. - service tier: fast is now accepted as an alias for priority on OpenAI/Codex. - OpenCode Go’s GLM-5.3-Flash had some bugs because they do things other people don’t. This is a theme. Also if they go ahead and give money to dhh for a package registry https://x.com/thdxr/status/2092476954062557523 , we’ll be removing the provider, so keep your head on a swivel for that. - Some bugs removed. - More bugs added. Polytoken 0.8.8 overhauls compaction, makes ask user question a rewind point, and surfaces retries and failures you could not see before. - ask user question is now a rewind point. - Increased allowances for reasoning loops. - Converted subagent compaction to standard compaction dropped tool scavenging . - Reworked compaction behavior to significantly reduce both tokens used during compaction and the compaction output to your models. - Better feedback when compactions fail. - Provider retry backoffs should be more visible now. - When the TUI goes down unexpectedly, it might tell you why now. Might. - polytoken continue runs the auto-updater before letting you pick a session to reattach. - More bugs. Polytoken 0.8.7 adds DeepSeek V4.1 Flash, session HTML export, experimental native compaction, and a subagent concurrency limit. - New model: Deepseek v4.1 Flash is available on Deepseek, DeepInfra, and OpenCode Go. It has different names on each because Deepseek is being weird. Many such cases. - New command: polytoken export-session . Makes some HTML out of your session, if you’re inclined to share. - Native compaction support enabled for OpenAI and Codex, using GPT 5.6 and 6 models. Experimental. If it’s worse-quality, give me some /feedback . - Default iteration cap for sessions raised from 200 to 1000. - daemon.max concurrent subagents limits the number of subagents a session can have at once. Additional ones are queued. - compact context can now switch facets as part of compacting. - /feedback can attach some new diagnostics. Don’t unless you know what they are. - Some bug fixes to LSPs to fix some failures-to-launch. - The usual bug fixes and other nonsense. Polytoken 0.8.6 clarifies the instructions the model gets for writing files. - gpt-6-astra is a nitpicking weirdo who demanded changes to the system prompt. Polytoken 0.8.5 adds opt-in 1M-context Codex variants and improves turn-interruption reliability. - -1m variants added for the Codex provider. These will use more of your usage. Most of the time, you should not use these. - Some internal changes around turn management and accounting should reduce the situations where hammering escape does not work. - The daemon now watches its own performance and is able to dump some data when tokio gets sad. - Ro-bug fixes and polish. Polytoken 0.8.4 adds the compact context tool, browser OAuth for Codex, and provider stream self-repair. - New tool: compact context lets an agent compact itself and give it a continuing instruction. - Codex provider can now use browser OAuth. The config UI doesn’t support it; you have to use the CLI. - polytoken exec should now be smarter about tasks that use subagents. - Some additional heuristics to restart/repair malfunctioning provider streams. - Prompt history now sucks less. - You can now rewind to a facet switch/plan handoff. - Additional bug fixes and polish, and no, mystery Sentry requester, I’m not going to summarize all of these because they are my weird little housekeeping gremlins that run tirelessly day and night to fix bugs. Mostly. Sometimes they’re your feedback. The gremlins get those too. Don’t get them wet. Polytoken 0.8.3 adds OpenRouter as an experimental provider, new catalog models, and a batch of fixes. - OpenRouter is now supported as a provider. Experimental. - claude-fable-5-1 and gpt-6-astra are now in the catalog. gpt-6-astra may not work for you yet. That’s sad for you. - Providers retry failed queries in more circumstances. - polytoken new now accepts --model and --facet . - /feedback submissions can now attach clipboard screenshots. - You can now disable prompt caching keys for poorly designed providers. - Removed a couple of superfluous status line chips. - A bunch of smaller bug fixes and cleanups: - VFS URIs accept trailing slashes. - Jobs viewer navigation stays on visible cards. - Skill template interactivity now carries into subagents. - Configurator fixes to model selector visibility and saving. - Providers that strip reasoning now produce one summary warning per turn instead of many. - Hashline reads and edits accept more cursor shapes. - Give-up notices survive restarts. - Bounded waits when a provider stalls inside serialized request offloading. Polytoken 0.8.2 adds Fireworks AI as an experimental provider, starts a jobs-viewer makeover, and keeps thinking traces around as cards. - Fireworks AI added as a provider. Experimental. - The jobs viewer has gotten the start of a makeover. More to come. - Thinking traces stick around as cards, if tui.show all thinking blocks is enabled. - Added language servers for C . - Streamlining for hashline tools. Experimental. - GLM-5.3 and GLM-5.3-Flash added to providers that now support it. Try them, they’re good. - Additional bug fixes and pig lipsticking. Polytoken 0.8.1 adds the rust-glancer language server, lets any model serve any default tier, and tunes provider retries. - Added rust-glancer as an LSP. If installed, it supersedes rust-analyzer . - Any model can be assigned to any default tier. - Provider errors now retry more often, with longer backoffs. - Additional bug fixes and polish. Polytoken 0.8.0 adds tool flows, a new way to run many tool calls in one step, and lets you rename sessions mid-turn. - New tool flow feature lets the model run a collection of tool calls inside a single tool call. Tool flows are approved once, up front, and respects permissions deny rules and other guardrails. Tool flows can run asynchronously, like shell exec . tool flow is experimental. If it causes problems, add "tool flow" to daemon.disabled tools in your config to remove it. - You can now change the session title while a turn is running. - Fixed some bugs around command substitution in the config UI. - Fixed some additional bugs. These bugs were found via /feedback . /feedback good. Give /feedback . Polytoken 0.7.4 makes shell exec batchable, persists subagent lifecycle cards in session history, and refreshes provider catalogs. - shell exec is now batchable: the model can queue several shell commands in one message and they dispatch sequentially. - Subagent lifecycle cards are now first-class session history: they persist across resume, rewind, and compaction instead of disappearing. - The configurator’s custom-model form now exposes a vision toggle. - Z.ai catalog updates: glm-5.3-flash added and glm-4.7 retired; custom models on Z.ai are no longer blocked from effort-based reasoning. - glm-5.3-flash added to the OpenCode Go catalog. - Additional bug fixes and polish. Polytoken 0.7.3 adds Ollama Cloud as a built-in provider, fixes the TUI compaction stall, and trims retired models from provider catalogs. - Ollama https://ollama.com is now a built-in provider. Connect to Ollama Cloud with an API key from your subscription, set OLLAMA API KEY , and pick a model in the configurator. Ollama Cloud speaks the Anthropic Messages protocol, so streaming, tool use, image input, and reasoning behave the same as on other providers. - Eight launch models are available: deepseek-v4-flash:0731 , glm-5.1 , glm-5.2 , kimi-k2.7-code , kimi-k3 , nemotron-3-ultra , nemotron-3-super , and nemotron-3-nano:30b . - The TUI no longer stalls during compaction. Refreshes are coalesced and rendering is fairer, so long sessions stay responsive while compacting. - Subagent launched and returned cards now persist across history refreshes and session reattach. - Config-edit backups are only created when the content actually changed. - Symlinked directories outside the workspace root now get the same relative-path grant treatment as real ones. - Dropped models were retired from the zai, Qwen Cloud, and Codex catalogs. - Additional bug fixes and polish. Polytoken 0.7.2 adds a /lsp status command, skips invalid skill files with warnings, and hardens provider streams and retries. - New /lsp status command shows language-server health for your session. - Invalid skill files are skipped with a warning card in the TUI instead of failing the load. - Stalled provider streams now time out with clearer user feedback. - Retries on Codex and OpenAI are more targeted, and TLS record errors are retried as transport failures. - The goal driver waits for terminating background jobs before continuing. - ask user question prompts without context now start collapsed. - Additional bug fixes and polish. Polytoken 0.7.1 renders LaTeX math in the TUI, adds polytoken sessions --all , and restores Codex prompt-cache affinity. - LaTeX math now renders as readable Unicode in the TUI instead of raw markup. - polytoken sessions --all lists historical sessions, not just live ones. - Restored Codex prompt-cache affinity using stable session and thread headers, improving cache hit rates and turn latency. - Skills with unquoted metadata scalars no longer prevent startup or fail validation. - Additional bug fixes and polish. Polytoken 0.7.0 adds preliminary language server support with an lsp tool, and fixes external editor input handling. - Polytoken now has preliminary, experimental language server support. We expose an lsp tool to conversations that supports go-to-definition, references, renames, and LSP-specific diagnostics. You can disable it with daemon.lsp.enabled . polytoken lsp check allows you to see what’s supported. - Fixes to opening an external prompt editor with Ctrl+G . Should probably not steal inputs as much. 0.6.14 Section titled “0.6.14” 0614 Polytoken 0.6.14 adds image paste in the prompt, a session header for provider and MCP traffic, facet resume fixes, and better prompt-cache hit rates on OpenAI and Codex. - You can now paste images directly into the prompt with Ctrl-Alt-V . - Polytoken now sends an X-Polytoken-Session header on provider and MCP requests. Disable with daemon.send session header . - Fixed a bug where some sessions would resume with the default facet. - Improved prompt cache hit rates on OpenAI/Codex. - Edit tools no longer rewrite files when nothing changed. 0.6.13 Section titled “0.6.13” 0613 Polytoken 0.6.13 removes the edits array batch mode from file edit search replace , restoring the single-edit contract. - Breaking change: file edit search replace no longer accepts an edits array. Make several changes to one file as one edit per occurrence. A call that still carries a leftover edits key alongside old string and new string now performs the single edit instead of failing. 0.6.12 Section titled “0.6.12” 0612 Polytoken 0.6.12 adds effort reasoning for GLM 5.2 on Mistral, and fixes MCP OAuth credential rotation teardown and update channel fallback. - GLM 5.2 on Mistral now supports effort-based reasoning. - MCP server OAuth credential rotations now wait for in-flight tool calls to finish before tearing down the connection. - The updater now falls back to the default channel when your update config cannot be loaded. 0.6.11 Section titled “0.6.11” 0611 Polytoken 0.6.11 adds the Mistral provider, a first-start license agreement with third-party licenses, batch atomic edits, provider disabling, and fixes for shell cwd, MCP noise, and prompt history. - Added a click-through license. Sorry. - Added third-party licenses. Should have done it sooner. - Added Mistral as a provider. Supports both Studio and Vibe API keys. - file edit search replace now accepts an edits array for batch atomic edits. - Providers can now be disabled in your config without deleting them. - polytoken auth provider status now shows all configured providers. - File tools are a little more hardened against weird Unicode edge cases. - Shell commands now correctly run in the current working directory and not the project start directory. - MCP connections/reconnections should be quieter. - The plan facet no longer uses todos, because it can’t clear them after a handoff plan . - shell service only shows you a clickable link when the service is actually running. 0.6.10 Section titled “0.6.10” 0610 Polytoken 0.6.10 adds GLM-5.3 to the Z.ai Coding Plan catalog. - Added GLM-5.3 to the Z.ai Coding Plan catalog. Polytoken 0.6.9 adds an orchestrate facet, subagent state isolation, prompt rendering performance fixes, and OpenAI/Codex opaque reasoning fixes. - A new orchestrate facet joins plan and execute . It guides the model through deliberate subagent-heavy coordination: upfront research, dependency graph construction, wave-based parallel dispatch, and parent-owned synthesis. - shell service jobs now show a clickable link emoji in the sidebar. Clicking copies the URL and opens it in a browser when local . - Subagent todos, flags, and plans are now isolated per-subagent instead of leaking into parent state. - Prompt rendering no longer degrades quadratically in long sessions. - Post-compaction continuation prompts are smaller, leaving more context for the conversation. - Fixed spurious 404 errors when replaying conversations that contain opaque reasoning blocks on OpenAI and Codex providers. - pushd and subagent working directories now expand ~ to $HOME . - Added DeepSeek V4 Pro to the DeepInfra and OpenCode Go provider catalogs. - The terminal bell now waits for assistant work to finish and control to return to you instead of ringing at autonomous continuation boundaries. Polytoken 0.6.8 adds gpt-daybreak-blue-latest, improves the @-file picker in large directories, adds new-session from the resume screen, and refines prompt mode visuals. - Added gpt-daybreak-blue-latest as a Codex and OpenAI model. This model only works if you have enrolled in OpenAI’s Trusted Access for Cyber program. OpenAI strongly recommends using an autonomous permission system when using this model. Polytoken will remind you. - Fixes to the @ file picker in large directories. - Some visual improvements to shell mode formerly command mode and slash mode. - You can now start a new session from the resume/continue screen. Polytoken 0.6.7 adds experimental reasoning loop detection, agent tool performance improvements, subagent correctness fixes, and mode-aware prompt history. - Experimental fixes for unjamming GPT 5.6 opaque reasoning loops. - Performance fixes to agent tools. - Correctness fixes for subagents. - Fixes for multiline permissions popups. - Prompt history now supports command mode and slash mode in history. Polytoken 0.6.6 fixes an @-picker deadlock, DeepInfra catalog models, and plan facet handoff guidance. - Fixed a situation where @ could deadlock. - Fixed DeepInfra models in the catalog. - Minor change to plan facet to better guide towards handoff plan . Polytoken 0.6.5 adds a post clear hook, tool blocklisting, context-clear disabling, OTEL env var conformance, a list jobs tool, and fixes for Codex/OpenAI providers, skill discovery, and tool-group rendering. - A new post clear hook event fires after /clear , with support for injecting additional context into the fresh session. Also fires after session start with clear count at zero. - daemon.disabled tools allows you to block tools from ever being surfaced in any conversation. - daemon.disable context clear removes /clear from both the client and from the daemon API. - OTLP exporter settings now honor OTEL environment variables per the OpenTelemetry SDK spec. - A new list jobs tool lets the model check running background jobs. - After compaction, Polytoken reminds the model about jobs still running. - shell service added. Think of it as a tiny supervisord for your session. Anything running at session exit gets terminated. - job status now surfaces stdout/stderr log file paths for shell service jobs. - The config UI for system reminders actually works now. - Fixed an idle-timeout retry death spiral on Codex and OpenAI providers during high reasoning-effort conversations. - Polytoken now follows symlinked skill directories during skill discovery. - Fixed stale rendering and incomplete copy text when sessions are loaded from history with scaffolding messages between tool calls. Polytoken 0.6.4 adds shell command mode, shell completions, XDG path overrides, and fixes for session reattach, dynamic providers, and the autonomous classifier. - now opens a command mode to run a bash command through the daemon. directs the command so the agent can see the output. - polytoken completions <shell is now a thing. - POLYTOKEN CONFIG PATH , POLYTOKEN DATA PATH , and POLYTOKEN CACHE PATH now override XDG-based directory resolution. - Improved backoff for the autonomous classifier in cases of failing upstream requests. - Clicking inside the feedback editor now moves the cursor. - Fixed bugs around reattaching sessions with outstanding interrogatives plan handoffs, etc . - /clear now cancels all running subagents. - Other fixes to the autonomous classifier. - Bug fixes around startup with dynamic providers. Polytoken 0.6.3 adds QwenCloud and OpenCode Go providers, improves Anthropic prompt caching, and fixes a subcommand permissions matching bug. - Added QwenCloud provider. - Added OpenCode Go provider. - Improved prompt caching adherence for the Anthropic Messages API. - Fixed a permissions-matching bug on multiply nested subcommands in shell exec . Polytoken 0.6.2 adds a custom Responses API provider option, protects session environment variables, fixes Anthropic stream wedging, and improves TUI recovery and configurator behavior. - Added an option for a custom Responses API provider. - Fixed a case where Anthropic streams would get wedged due to empty thinking signatures. - Added a recovery path for SSE events in the TUI when they get a bit sad. - You can now set POLYTOKEN DISABLE FEEDBACK=1 to hide the /feedback command. - All environment variables captured at session start are now protected from unsetting in shell exec . - Held-key auto-repeat acts a little more gracefully. - When the autonomous classifier errors, you’ll now see a flash message saying so while still falling back to human asks . - File tools now follow directory symlinks, so long as the origin was within the project root. - flag important now works relative to the cwd stack, so you can flag files after pushd . - Custom Chat Completions providers no longer show ignored reasoning effort. Polytoken 0.6.1 fixes a bug in parsing inference streams for classifiers. - Fixed a bug in parsing inference streams for classifiers autonomous permissions and title generation . Polytoken 0.6.0 adds per-model request options across the system, a DeepInfra provider, and fixes for empty provider responses, session reattach reliability, and DeepSeek V4 Flash in subagents. - Models now support request options. They are supported across the system. A fully-qualified model string looks like codex/gpt-5.6-luna high service tier=priority . - /model-option or clicking on the model name in the sidebar lets you change these on-the-fly. - Added DeepInfra provider. - Some readability improvements when switching models. - You can now switch permissions modes from within a permissions request prompt. - Fixed bugs around empty provider responses to user prompts. - Reattaching to a session now retries through transient transport errors. - Scrolling works in conversations longer than 65,000 rows. Oops. - Fixes to DeepSeek V4 Flash multiple providers when used in subagents. V4 Flash still has issues when used as an autonomous classifier. We’re working on it. - Completed goals fully clear goals from agent context. - Fixed some annoying bugs in the config UI. 0.5.10 Section titled “0.5.10” 0510 Polytoken 0.5.10 improves compaction reliability, fixes the self-updater on Linux, improves Azure OpenAI compatibility, and resolves TUI freezing and startup performance issues. - Compaction now retries if too close to the end of the context window. - polytoken update now correctly re-launches on Linux. - OpenAI and Codex providers adjusted to support Azure more effectively. - When a subagent fails to launch, the model now receives guidance to help it retry. - Markdown rendering fixes: links inside emphasis display correctly, bare URLs in emphasis are clickable, and inline links no longer show a trailing URL suffix. - Fixes to TUI freezing during very long autonomous sessions. - Startup in very deep, very large directories home directories should be faster. - Jobs viewer subagent scrolling is now cached better. - Fixed a bug that bricked sessions on resume. - Fixed regressions in shell exec behavior and notifications in subagents. - DeepSeek provider/model fixes to support deepseek-v4-flash-0731 . - It’s now legal to set a mini model as a full model config.yaml only . Polytoken 0.5.9 adds shell tool backgrounding, fixes OpenTelemetry export, and resolves several session crashes. - shell monitor can now be backgrounded. - shell exec is more easily backgrounded by models. - Ctrl+B or /bg-shell will send the currently-running shell exec or shell monitor to background. - OTel now, uh, works. - Provider URLs with a trailing slash now work. - Interrogative prompts don’t get lost behind the jobs or todo viewer. - Removed legacy pedagogy system. Sad. It was fun. - Fixed a couple of session-disconnection crashes. Polytoken 0.5.8 fixes a bug that could drop environment variables after shell commands, adds conversation scrolling shortcuts, and improves paste handling across text inputs. - Environment variables such as PATH and HOME no longer disappear after running shell commands that produce a partial environment snapshot. - You can now scroll the conversation with Shift+PageUp, Shift+PageDown, Shift+Home, and Shift+End from anywhere in the interface. - Custom Anthropic-compatible providers can now use prompt caching through a new cache control strategy config option. - Large pastes into clarification prompts and the feedback form now collapse into a compact marker, matching the prompt input. - The interactive configurator now supports configuring OpenTelemetry trace export settings. Polytoken 0.5.7 adds OpenTelemetry trace export, improves cache-miss detection accuracy, and lets you rewind across compaction boundaries. - Polytoken can now export trace data to an OpenTelemetry collector using the OTLP protocol, following GenAI semantic conventions. Telemetry is opt-in and disabled by default. Configure it through a telemetry block in your global config. - The prompt-cache-miss warning is more accurate and no longer triggers false alerts. - You can now rewind to before a compaction or clear boundary, removing the boundary card and everything after it. - Selecting an MCP server for OAuth authentication now targets the correct server instead of always using the first. - Polytoken now warns you when stored OAuth credentials already exist for an MCP server before starting a new authorization. - The configurator exit dialog now offers a Save and quit option when you have unsaved changes. - The collapsed compaction card now shows the continuation prompt size. Polytoken 0.5.6 makes session titles more reliable, recovers stalled subagents, warns on prompt-cache misses, and fixes several terminal UI issues. - Session titles now fall back to your first message when the title model keeps failing. - Stalled subagent provider streams now recover automatically, including those that happen while your computer sleeps. - You now see a status-line warning when a large session loses its prompt cache. - The prompt you are typing is preserved when you switch facets. - Repeated MCP lifecycle notifications during a reconnect cycle now collapse into a single updating card. - The context meter no longer shows a stale value briefly after /clear . - polytoken validate subagent now checks the global configuration directory for definitions. Polytoken 0.5.5 fixes subagent reasoning-level support and improves compaction context scaling. - Subagent frontmatter now properly supports reasoning levels for selected models. - Agent compaction has improved, with size/detail scaling with context window size. Polytoken 0.5.4 adds Claude Opus 5, shows hook output during active sessions, and fixes terminal UI and clipboard issues. - Claude Opus 5 is available in the Anthropic model catalog with a 1 million token context window and effort-based thinking controls. - Hook output text is now visible during the running session when a hook allows, blocks, or denies an action, instead of only appearing after you reattach. - When a session attach conflict occurs, the error message now suggests polytoken new so you can start a fresh session when your current one is busy. - The active plan file path no longer resets incorrectly after a daemon restart or session resume. - Pressing Escape twice no longer cancels a pending compaction by accident. A confirmation prompt now requires a deliberate second press within a short window. - The compaction card count no longer overcounts when earlier compaction events were cancelled or failed. - Long text in the feedback editor now wraps instead of scrolling off-screen. The phantom cursor is fixed, and pasted text now goes to the feedback modal instead of leaking to the prompt. - The cursor is now visible when editing a handoff plan refusal. - Clipboard copy now works correctly inside tmux. Polytoken 0.5.3 improves reliability of compaction, tool calls, and provider error handling. - Tool arguments are now type-coerced for unambiguous mismatches before schema validation. - Compaction now retries with backoff instead of failing on the first error. - Provider errors that Polytoken does not recognize now show the original error text from the provider instead of a generic message, so you can diagnose the cause. - Added polytoken.compaction hint facet frontmatter for controlling compaction prompt guidance. - Fixed a bug where pressing Escape during a goal-driven session caused the goal to re-iterate instead of pausing. Polytoken 0.5.2 eliminates session-length-dependent latency in the TUI and daemon. - Eliminated session-length-dependent rendering latency. Long sessions no longer cause progressively slower prompt acceptance, scrollbars, or tool-call rendering. Polytoken 0.5.1 adds a built-in self-updater, deprecates the Homebrew tap, and fixes MCP and provider compatibility issues. - Added polytoken update for self-updating. Checks for new versions at launch once per 24 hours and prompts to install. Disable with updates.automatic check: false . - Deprecated the Homebrew tap. New installs should use curl -fsS https://get.polytoken.dev | bash ; the tap will be decommissioned in a future release. - Changed the installer to download the bare binary directly instead of a zip archive, eliminating the unzip dependency on minimal Linux systems. - Added MCP server keepalive ping to prevent timeout during long sessions. - Fixed Kimi K3 tool-call errors when thinking mode was not explicitly enabled. Polytoken 0.5.0 adds the Kimi Code and Moonshot providers, bearer-token authentication for the daemon, automatic crash reporting with a /feedback command, per-tier reasoning effort, and recent-message retention through compaction. It also introduces a message subagent tool, subagent resumption from a prior job, atomic batch edits, global skill discovery, and broad improvements to the terminal UI, permissions, and providers. New providers Section titled “New providers” new-providers - Kimi Code https://www.kimi.com/code is now available as a built-in provider. It uses the Anthropic Messages API, ships models including K3 with effort-based reasoning, and authenticates with an OAuth 2.0 device-code flow, so you can log in without an API key. - Moonshot AI https://platform.kimi.ai joins the built-in providers with OpenAI-compatible Chat Completions and ships models including K3 with reasoning effort controls. - Both providers support vision and appear in the interactive configurator. Security Section titled “Security” security - The daemon now requires a bearer token on its control API. The CLI sets this up automatically, so no configuration is required, and requests without a valid token receive HTTP 401. See Daemon authentication https://docs.polytoken.dev/reference/daemon-auth/ . Crash reporting and feedback Section titled “Crash reporting and feedback” crash-reporting-and-feedback - Polytoken now sends automatic crash reports when it exits unexpectedly. A report includes the crash type, a stack trace, the component name, the version, and a pseudonymous installation ID, but never your logs, session records, prompts, or files. The setting is on by default; turn it off with data.automatic crash reporting = false . - /feedback opens a form where you can message the team and optionally attach session logs to help diagnose a problem. Models and reasoning Section titled “Models and reasoning” models-and-reasoning - Each default model tier :full , :mini , :nano can carry its own reasoning effort. Add the effort level in parentheses after the model name in the defaults block, for example full: "claude-sonnet-4 high " . The configurator offers an effort selector when you assign a default model, warns when a tier requests an effort the model does not support, and falls back to the model’s default. - OpenAI reasoning models now show readable thinking summaries when the Responses API provides them. - GPT-5.6 Sol, Terra, and Luna join the built-in catalog for the OpenAI and Codex providers, all supporting low, medium, high, xhigh, and max reasoning effort. Sol and Terra are full-class models. Luna qualifies as a default for both the full and mini classes, and the configurator offers a class chooser when you assign a multi-class model as a default. Session goals Section titled “Session goals” session-goals - When a saved-session goal is active, ask user question no longer interrupts you with an operator prompt. Polytoken directs the model to pick an answer, seek a second opinion, or block the goal when it cannot proceed. Pausing the goal restores ordinary question prompts. Compaction Section titled “Compaction” compaction - Compaction now preserves a budget of recent messages verbatim, sized to each model’s context window, so the model retains the immediate conversation alongside the summary. - Compaction summaries focus on recent completed tasks and tell the model how to retrieve older history when needed. - Bash hook handlers now receive POLYTOKEN FACET NAME , POLYTOKEN MODEL NAME , and POLYTOKEN GOAL ACTIVE alongside the existing POLYTOKEN variables. POLYTOKEN FACET NAME is the active facet, POLYTOKEN MODEL NAME is the active model’s config-key name, and POLYTOKEN GOAL ACTIVE is true or false depending on whether a saved-session goal is running. Skills Section titled “Skills” skills - Polytoken now discovers skills in ~/.agents/skills/ automatically. - MCP connections begin retrying sooner and keep trying longer, so transient outages recover more reliably. - polytoken models , polytoken sessions , and polytoken doctor accept --format human , --format json , --format json-pretty , and --format yaml . - polytoken openapi and polytoken event-schema accept --format json and --format yaml , while keeping pretty JSON as the default. - polytoken new --prompt <text submits the given text as the session’s first prompt as soon as the TUI attaches. With --no-attach , it submits the prompt and exits without starting the TUI. - The message subagent tool lets a parent session send a text message to a running subagent. The message arrives in the subagent’s context without interrupting its current step. - Use the subagent tool’s new resume from parameter to continue a completed subagent’s work in a new job. - The hashline edit format now supports atomic batch edits: multiple edits in a single call apply together or not at all. - shell exec now returns full execution context when a command fails, including persisted paths for the captured output. Permissions Section titled “Permissions” permissions - In autonomous mode, tool calls that match an explicit allow rule now run immediately. Filesystem grants still require approval, and Polytoken still evaluates calls that match no explicit rule. - Polytoken now recognizes literal relative directories in cd commands and can suggest appropriate filesystem permission rules for them. - Permission rules now handle paths that safely resolve internal .. segments. Paths that escape through a leading .. , such as ../secrets , remain blocked. - Plan handoff from the plan facet to execute now works in autonomous mode. File references Section titled “File references” file-references - The @ file picker now accepts ./ to browse from the current working directory. For example, type @./ to browse the project root or the directory selected with pushd . Polytoken continues to hide files blocked by .agentblock . Themes Section titled “Themes” themes - Polytoken now includes Sean Edwards’s rose-pine.yaml theme. Terminal UI Section titled “Terminal UI” terminal-ui - Press Ctrl+Shift+B or use the corner button to collapse the context area in ask user question and handoff plan prompts. - Provider error cards now show structured detail when expanded, with plain-language context and actionable guidance for each error type. - The context meter now stays stable during tool loops and reports the correct value after rewind, clear, or compaction. - Diagnostic output no longer corrupts the TUI or configurator. - Press Ctrl+Y in the Todo Viewer to copy the selected todo to your clipboard. - Ctrl-C during TUI startup now cleans up the daemon instead of leaving orphaned processes behind. - Modal text dialogs in the configurator now show Enter as the accept key instead of Ctrl+S. Providers Section titled “Providers” providers - Neuralwatt models now expose only the reasoning controls they support, with sensible defaults for models that offer optional reasoning. - Provider interruptions no longer silently discard partial responses; Polytoken preserves partial progress while recovering. - Custom models entries can now override selected settings on discovered models. Polytoken warns about overrides it cannot apply instead of rejecting the configuration. - Anthropic models now handle thinking text streams more reliably. - Polytoken now retries transient OpenAI server errors instead of ending a turn. - Kimi Code sessions now recover more reliably from temporary authentication failures and expired tokens. Performance Section titled “Performance” performance - Streaming renders faster in the TUI during long sessions. - Concurrent glob and grep calls are faster. - Long sessions stay responsive during file and compute-heavy work. Reliability and bug fixes Section titled “Reliability and bug fixes” reliability-and-bug-fixes - Session history no longer loses messages when you interrupt and resume a session. - Cancelling a response no longer makes later requests unnecessarily large or degrades response quality. Polytoken still preserves completed text and tool calls. - Polytoken now rejects empty and whitespace-only prompts instead of submitting them. - The TUI no longer crashes on ConPTY, WSL2, and Podman PTY chains; it redraws gracefully when a full clear fails. - Repeated invalid tool calls now end the session instead of looping indefinitely. - Sessions now recover when a provider rejects an oversized request. Session title generation is also more reliable across providers. - Permission checks no longer crash when you use reasoning-capable models. - Interrupting config edit no longer corrupts your configuration. - MCP servers now report expired authentication instead of disconnecting without an explanation. - You can now propose a new saved-session goal after the previous goal finishes. - User-level subagents now load from the global configuration directory. - Bare shell-interpreter commands such as bash , sh , zsh , or fish no longer trigger unsafe persistent permission suggestions. - Rewinding to an earlier prompt now restores @ -reference shorthand instead of raw XML tags in the prompt textarea. - Subagent definitions can now reference dynamically discovered models, such as Umans models, without being falsely rejected. Polytoken 0.4.2 fixes a bug where cancelling a provider stream could permanently bloat every subsequent request. - When you cancel a provider stream mid-response, Polytoken now drops provider-private reasoning blocks from the salvaged history instead of retaining them. Previously, a single cancelled stream could inject hundreds of reasoning items into durable history, causing every future request in the session to grow larger and degrade model quality. Text content and completed tool calls are still salvaged as before. Polytoken 0.4.1 improves model selection, permissions, file search, provider recovery, and terminal performance for longer sessions. Models Section titled “Models” models - GPT-5.6 Sol, Terra, and Luna are now in the built-in model catalog for both the OpenAI and Codex providers. All three models support low, medium, high, xhigh, and max reasoning effort. Sol and Terra are full-class models, while Luna is a mini-class model. Performance Section titled “Performance” performance-1 - TUI rendering during streaming is significantly faster for long sessions. Polytoken avoids recomputing settled conversation items, caches layout heights, and limits redraw work to visible rows. - Concurrent glob and grep calls are now faster. - Polytoken improves agent-loop performance for long sessions. Providers Section titled “Providers” providers-1 - Neuralwatt models now receive the correct reasoning configuration based on their capabilities. Models with reasoning effort support receive the full effort-level set, models with reasoning support reason by default with a disable option, and non-reasoning models receive no reasoning. - Mid-stream transport timeouts now enter idle-stall recovery, preserving partial-turn progress instead of ending the turn silently. Permissions Section titled “Permissions” permissions-1 - Literal relative cd targets in shell commands now create deterministic chdir requirements in the permission system, so Polytoken can offer filesystem candidate rules instead of treating the target as opaque. - Permission rules for filesystem paths containing resolvable .. components now canonicalize and apply correctly. Paths with unresolved leading .. components, such as ../secrets , remain blocked. - Plan handoff from the plan facet to execute now works in autonomous mode. - In autonomous mode, explicitly allowed tool calls now skip an unnecessary classifier pass. Polytoken still requires approval for filesystem grants and applies the classifier when no explicit allow rule matches. - Permission candidate scoping on macOS now recognizes canonical forms of /tmp and /var/tmp , so symlinked temporary paths produce consistent candidates. File search Section titled “File search” file-search - Recursive file searches now skip platform no-go directories and paths Polytoken cannot access, while still honoring directories you explicitly target. Agent behavior Section titled “Agent behavior” agent-behavior - The system prompt now encourages the model to give each subagent a descriptive name, making concurrent jobs easier to track in the Jobs viewer. Terminal UI and sessions Section titled “Terminal UI and sessions” terminal-ui-and-sessions - Switching between the built-in plan and execute facets no longer changes the model selected with /model . - The plan facet no longer asks whether you want a plan of plans. - The running timer card no longer disappears intermittently mid-turn. - A modal error dialog now captures keyboard input before a pending permission prompt. Pressing Enter or a digit while the error dialog is open no longer answers the concealed prompt. - shell exec now provides full execution context when a command fails, including persisted paths for captured output. Polytoken 0.4.0 adds the Neuralwatt and Kagi providers, Claude Sonnet 5, custom TUI themes, session goals that drive continuation, .agentblock permanent path blocking, a bypass+ permission mode, per-model compaction thresholds, and configurable keyboard shortcuts. The release also brings major improvements to planning, the Jobs viewer, file references, the @ picker, provider stability, and dozens of terminal UI fixes. New providers Section titled “New providers” new-providers-1 - Neuralwatt https://neuralwatt.com is now available as a dynamic provider. Polytoken discovers Neuralwatt models at runtime and integrates them into the model picker. - Kagi https://kagi.com is now available as a web search provider. Models Section titled “Models” models-1 - Claude Sonnet 5 is now in the built-in model catalog with effort levels low, medium, high, xhigh, max , a 1M context window, 128k max output, vision, and adaptive thinking. Themes Section titled “Themes” themes-1 - Custom TUI themes are now available. Polytoken loads theme files from tui.theme-file , validates theme tokens, and lets you preview theme choices through /theme before applying them to the session. - Theme files now support HSV colors and reusable palette entries, making custom themes easier to write and validate. Session goals Section titled “Session goals” session-goals-1 - The /goal slash command lets you set a goal for the current session. Polytoken persists the goal in the session record and a markdown file, so the goal survives restarts, reloads, and compaction. When a goal is active, Polytoken uses it to drive automatic continuation, keeping long-running work on track. - /goal set and /goal resume now start a turn, and resuming a paused goal shows a goal-paused card. Skills and customization Section titled “Skills and customization” skills-and-customization - Polytoken now discovers skills from .agents/skills before .polytoken/skills , so project-local skills written for common agent ecosystems are available automatically. .polytoken/skills still wins when two discovered skills use the same name. - New polytoken validate {skill,facet,subagent} commands check definition files for structural errors before starting the daemon. Each accepts a definition name resolved through standard discovery tiers or a direct file path. - The configurator’s recommended budget provider now points to Neuralwatt. Planning and facets Section titled “Planning and facets” planning-and-facets - Plan mode can now use read-only shell exec , so planning turns can inspect command output without switching into execute mode. - Facet switches now honor polytoken.model , so a facet can reliably select its configured model when you move into that facet. - Custom default facets that pin a model now apply that model when the daemon starts, matching the model selection behavior used when switching facets during a session. - Plan handoff refusal now includes a feedback field, so you can explain what needs to change before accepting a plan. - Plan files now use numbered names, old plan flags are cleaned up more consistently, and Polytoken reminds the model about open todos when re-entering planning. Permissions Section titled “Permissions” permissions-2 - A new bypass+ permission mode skips built-in protections and opaque-shell gates like bypass , but still honors explicit deny rules. - Permission rules can now include a custom message that Polytoken injects as a system reminder when the rule matches. - Polytoken now denies agentic edits to lock files by default for safety. - MCP synthetic tool approvals now persist as name-level rules, so approving a tool by its mcp server tool name sticks across sessions. - Press Ctrl+Shift+P mid-turn to open the permission typeahead and switch permission modes without interrupting the current turn. - Breaking change: Permission rule args now use a typed enum with strict validation. Unknown args keys produce hard load errors instead of being silently ignored. Cross-variant hybrid rules are rejected at load time. Correctly formed rules continue to work as before. - Fixed a regression where opaque shell command permission persistence options were not offered correctly. - Permission decisions that cannot be persisted no longer trigger persistence prompts. - Read-only tool denials from the autonomous classifier now escalate to a human prompt instead of blocking outright. File references and pathing Section titled “File references and pathing” file-references-and-pathing - The @ file picker now browses directories outside the project. Type @.. to see the project parent’s children, @~ for your home directory, and @/ for the filesystem root Unix only . Drill into subdirectories by typing path components. External file references resolve and inject content at submit time, just like project-relative references. - Reference resolution is now cwd-aware, so @ references resolve correctly after pushd changes the working directory. - A new tui.show-ignored-files-in-picker option shows all directories in the @ picker, including dotfiles and .polytoken . - .agentblock files block agent access to matched paths unconditionally across all tools and permission modes. Place one at the project root or in any subdirectory using gitignore-style patterns. Blocked paths cannot be read, edited, searched, or referenced by shell commands. - .agentblock discovery now protects against filesystem cycles. Configuration Section titled “Configuration” configuration - Breaking change: Compaction thresholds are now per-model instead of a single global setting. Each model derives its own threshold from its context window, and you can override it individually. - Breaking change: Keyboard shortcuts are now configurable through a unified tui.keyboard.commands map. - max output tokens per turn is now configurable, with defaults based on the selected model tier. - A new daemon.follow symlinks for configs option default true controls whether symlinked config and context files are followed during discovery. When set to false , symlinked files are skipped with a warning in the TUI and daemon log. Tool file access is unaffected. - The configurator prunes old config backups automatically, keeping the newest 10. - Config editing now creates backups with 0o600 permissions and auto-restores the backup on max-attempts exhaustion. Skeleton edits new files create no backup. - Config rollback now tolerates unrecognized providers and models instead of failing when a previous configuration references entries the current build does not know about. - Project-layer config validation now merges the real global config instead of a stub. - Polytoken caps Tokio worker threads at 3 by default to reduce resource usage. Providers Section titled “Providers” providers-2 - Provider usage throttling controls request pacing and concurrency for providers that support it. - MCP OAuth tokens are now refreshed proactively before expiry. - Dynamic provider models appear in @model: typeahead. - Umans flash models are classified as mini or nano defaults. - OpenAI and Codex Responses requests can now use prompt cache key , improving compatibility with provider-side prompt caching. - polytoken doctor now resolves dynamic provider catalogs before reporting provider readiness. - Polytoken resolves dynamic provider catalogs before model lookup, fixing model selection failures with dynamically registered providers. - The false-positive dangling-provider warning no longer fires for dynamic providers. - The polytoken models command now lists models from dynamic providers such as Umans. Dynamic models display a dynamic marker alongside existing default and small markers. When a dynamic catalog cannot be resolved, the command lists static and catalog models and prints a warning instead of failing. - Fixed a 400 error from Anthropic when using Sonnet 5 or newer models with deferred tools. Polytoken now places cache-control breakpoints on the nearest non-deferred tool instead of the last tool in the list. - Fixed OpenAI Responses provider misbehavior. - Fixed an issue where OpenAI Responses API tool calls could lose their arguments when output item.done omitted the arguments field. The parser now accumulates delta fragments as a fallback instead of producing an empty string. - MCP stdio servers accept a pass env field listing specific environment variable names to copy from the daemon process into the stdio child. Explicit env overrides still take precedence on conflicts. - MCP stdio servers now pass HOME , PATH , LANG , and TMPDIR by default unless the server config overrides the environment list. Agent loop and tools Section titled “Agent loop and tools” agent-loop-and-tools - Breaking change: The /prompt endpoint now auto-queues a submitted prompt when a turn is already in progress, instead of rejecting it. Clients that relied on a rejection response when the turn slot is held need to handle the queued behavior. - The ask user question tool result now includes the original context explainer text alongside the answers. This preserves the reasoning through compaction, so the model retains full context even after the original tool call is summarized. - web fetch now windows large pages instead of rejecting them outright, so the model can inspect long pages in chunks. - Tool inputs are now stripped of nullish values before dispatch, preventing malformed arguments from reaching providers. - The /history response now includes emitted at timestamps on each item. - Polytoken detects when the working directory is deleted and surfaces a system reminder. - Empty provider responses after tool results now trigger a re-entry instead of silently ending the turn. - Provider stream idle stalls now retry with exponential backoff, up to 6 consecutive stalls before failing. - Subagents now receive their working directory as an absolute path, so paths resolve correctly within the subagent context. - During automatic prompt draining, Polytoken now adds a reminder so the model does not mistake background notifications for new user input. - After compaction, Polytoken injects a CWD reorientation reminder so the model knows relative paths resolve against the current working directory, not the project root. - Fixed tool-call argument truncation failures. Context and agent grounding Section titled “Context and agent grounding” context-and-agent-grounding - Polytoken now loads AGENTS.md files from ancestor directories between the global location and the project root. This means shared project-level instructions are picked up when working in a subdirectory. Files are deduplicated by content hash and ancestor @-links are confined to the project root. Rewind Section titled “Rewind” rewind - Rewind now restores runtime context correctly when returning to an earlier point in the session. - The working directory stack restores correctly after rewinding past pushd / popd operations. Terminal UI Section titled “Terminal UI” terminal-ui-1 - The Todo and Jobs viewers now show scrollbars and support scroll interaction. - Subagent cards in the Jobs viewer now support keyboard navigation, expansion, and narration, so you can inspect subagent progress without leaving the viewer. - The Jobs viewer now shows the thinking level for each subagent. - Plan scrolling is faster after caching interrogative preview rows. - File edit cards can auto-expand with a new expand file edit cards config flag. - The ask user question custom answer area is now three lines tall and clears on Esc or Ctrl+C. - @// typeahead now runs asynchronously, so typing stays responsive when searching large directories. - The sidebar facet label and mini toggle now use the active facet’s accent color. - Polytoken now clears the prompt input lock on attach, preventing a stuck prompt after reconnect. - Diff colorization no longer disappears when scrolling with a card open. - Diff highlighting now survives history reloads, including attach, reattach, compaction, and rewind. - The flagged files pane now shows paths inside the session directory with the $SESSION DIR shorthand while preserving the real path for copying. - Esc now force-releases the prompt instead of leaving it in an ambiguous state. - /new is now an alias for /clear . - The session picker now supports K to kill a selected session. - Esc Esc opens a quick rewind picker for returning to an earlier prompt. - Invalid slash command arguments are rejected in the TUI instead of being sent to the model as ordinary prompt text. - Single Enter now accepts the active typeahead selection and submits the prompt. - The custom-text exit shortcut in ask user question changed from Up to Ctrl+Up to avoid conflicts with cursor navigation. - Polytoken now writes crash logs for TUI panics and lease-lost exits, making failed terminal sessions easier to diagnose. - Fixed several ask user question interaction issues, including checkbox toggling, custom-text navigation with Up, and text exclusion behavior. - Fixed missing and duplicate timer cards in the conversation. - Fixed an input handling issue in the session picker drive loop. Reliability and performance Section titled “Reliability and performance” reliability-and-performance - macOS filesystem scans now avoid repeatedly probing denied directories, and glob searches use a bounded worker count to reduce filesystem overhead. - Repository-status reminders no longer reappear after pushd and popd in a way that can corrupt history. - Polytoken now stops a turn when a provider repeatedly returns the same synthesized tool call identifiers, preventing a self-sustaining retry loop. - Fixed a CPU spin on macOS caused by terminal device handling. - Fixed rendering of trailing setext underlines in streaming markdown. - Resolved several CPU hotspots in the agent loop. - Improved daemon startup speed on large projects. - Queued prompts now start correctly after manual compaction finishes and releases the turn slot. - The configurator now exits with guidance when it closes without saving. Polytoken 0.3.3 fixes MCP tool availability in subagents and two terminal UI issues. - MCP tools now more reliably provision within subagents in edge cases. - Fixed a cursor flash that could appear during rapid file-diff rendering in the terminal UI. - Cleaned up the subagent spawn card display by removing a redundant type label and adding a “launched” suffix for clearer job tracking. Polytoken 0.3.2 adds a disabled state and /mcp command for MCP servers, custom provider HTTP headers, a provider stream idle timeout, and fixes several terminal UI and permissions issues. - MCP servers that cannot connect now stop retrying and enter a disabled state instead of looping indefinitely. A new /mcp slash command lets you enable or disable individual servers from the TUI, and servers can be configured to start disabled. - Fixed several bugs in MCP server reconnection so recovered tools reach the model after an auto-reconnect cycle, and silently degraded connections are detected before a tool call fails. Providers Section titled “Providers” providers-3 - Custom providers can now send additional HTTP headers on every request. This is useful for proxies, internal gateways, and other Anthropic-compatible endpoints that require backend-specific headers. - If a provider stream stalls mid-response, Polytoken now salvages any partial content received so far and continues the turn instead of hanging indefinitely. Permissions Section titled “Permissions” permissions-3 - Read-only tools like file reads and searches are never hard-denied by the autonomous classifier. In autonomous mode, questionable reads now escalate to a confirmation prompt instead of being blocked outright. Terminal UI Section titled “Terminal UI” terminal-ui-2 - Added mouse drag selection in interrogative preview areas. - Word deletion shortcuts Ctrl-Backspace, Ctrl-Delete now work in all text fields. - Fixed full-width character cursor drift and text highlighting so East Asian and other wide characters render and select more reliably. - Expanded configurator and viewer dialogs to use much more of the available screen space. - Clarified the configurator wording around when Polytoken switches into mini mode. - Interrogative popups now dismiss cleanly with a single Esc. - Added a 256-color fallback when truecolor is unavailable. - File edit tools now return a contextual diff in their results, so you can review the surrounding change without digging for it. - Fixed rewind so returning to an earlier user message does not leave that prompt duplicated in your history. - Fixed a high-CPU lockup that could occur when certain providers returned tool-use responses in an unexpected order. - Improved permission handling for shell-style patterns so Polytoken accepts more literal test commands and avoids overmatching some glob-like inputs. - Compaction no longer includes completed todos in the context summary, saving tokens when your todo list grows long. - Subagents can now be given a custom name when spawned, making job handles more readable. Polytoken 0.3.1 is a patch release. - Rumors of a bug in plan-reviewer rendering are grossly exaggerated. - You saw nothing. Polytoken 0.3.0 adds the Umans provider with dynamic model discovery, adds a TLS trust store option and a polytoken doctor readiness check, and brings a broad set of terminal UI, planning, and stability improvements. Umans https://umans.ai is now available as a built-in provider. Polytoken discovers Umans models at runtime from the provider’s live API and caches the last known good catalog for fallback when the network is unavailable. Discovered models appear as read-only entries in polytoken config ui and can be selected as full or mini defaults. New commands Section titled “New commands” new-commands - polytoken doctor checks startup readiness and reports what stands between you and a working session. - polytoken vfs ls and polytoken vfs cat inspect the resources Polytoken ships, such as facets and skills, through the internal polytoken:// virtual filesystem. Configuration Section titled “Configuration” configuration-1 - models and defaults are now user/global-only settings. The project config layer <project /.polytoken can no longer declare them. - polytoken config ui is the canonical command name for the interactive configurator. polytoken config configure remains as a backward-compatible alias. - A new daemon.tls.trust store option lets you point Polytoken at a custom certificate authority for daemon connections that use TLS. - Custom OpenAI-compatible providers now default to the chat completions protocol, so they work without an explicit protocol hint in most cases. - Static API keys are trimmed of leading and trailing whitespace when read, so a stray space no longer breaks authentication. - /reload returns a typed error when a configured dynamic default model becomes unavailable. - max tool batch size is now derived from the active model’s context window. Terminal UI Section titled “Terminal UI” terminal-ui-3 - A slow refresh mode makes Polytoken usable on e-ink displays. - You can configure the prompt cut shortcut and a local-paste shortcut. - The next queued prompt now appears above the prompt box, so you can see what is coming next. - Clicking outside the Jobs and todo viewers dismisses them. - The configurator window opens larger, and its validation errors wrap with hanging indents for easier reading. - First-run provider guidance is clearer. - Polytoken now infers a session title from your conversation, and you can override it. Prompt, planning, and model behavior Section titled “Prompt, planning, and model behavior” prompt-planning-and-model-behavior - A single Ctrl+C cancels an active assistant turn. - Before a plan handoff, Polytoken nudges you to close out open todos. - @skill: and @model: shorthand is now resolved at submit time, and @ -file reference completion is sourced from the daemon. - Queued prompts are appended to your local prompt history. - Templates can read the active plan through a scoped active plan context, and a new source control template variable exposes repository state. - Subagents accept an optional start directory, and pushd can move outside the project when permissions allow. - The terminal title is cleared when you quit or detach, so it no longer lingers as polytoken over SSH. - Polytoken recovers from a stuck busy state instead of locking a turn, with hardened daemon terminal handling. - Resuming a session no longer shows a stale in-flight lockout. - Repository reminders and status cache are constrained to your current directory stack. - Clicking the status bar dismisses flash messages. Polytoken 0.2.2 adds pushd/popd directory tools, command substitution in configuration, MCP OAuth credential isolation, and fixes several terminal UI and provider issues. New tools and configuration Section titled “New tools and configuration” new-tools-and-configuration - New pushd and popd tools give the model a per-session directory stack, so the model can work within a subdirectory without repeating full paths on every file call. - Configuration values now support command substitution with $ cmd and default values with ${VAR:-default} . Set config command substitution: true in your user config to enable it. Only the user or global config layer can turn this on. - MCP servers can use a per-server auth key name to isolate OAuth credentials when multiple servers share a provider. Terminal UI fixes Section titled “Terminal UI fixes” terminal-ui-fixes - Rewind now removes the target card and restores your prompt text to the textarea. - Typeahead and slash-command recognition are now case-insensitive. - You can switch between status bar typeaheads when one is already active. - The configurator shows protocol-aware provider URL help and a larger picker. Provider and shell fixes Section titled “Provider and shell fixes” provider-and-shell-fixes - A new /reset-shell command clears a stuck shell session; environment flowback from shell commands is now guarded. - Provider tool-use input is now coerced to a JSON object, fixing a class of provider errors. Tooling and docs Section titled “Tooling and docs” tooling-and-docs - polytoken print-tools exports built-in tool metadata. - The changelog now has an RSS feed at https://docs.polytoken.dev/rss.xml https://docs.polytoken.dev/rss.xml . Polytoken 0.2.1 improves terminal UI status, first-run setup, job navigation, multi-root file search, and permission suggestions. Terminal UI improvements Section titled “Terminal UI improvements” terminal-ui-improvements - The status bar now shows Git dirty state and ahead/behind indicators for the current branch. - A new first-run welcome dialog helps you choose initial configuration settings. - Polytoken fixed some bugs in the Jobs viewer. - The Jobs viewer now shows a scrollbar when the list is longer than the pane. - Polytoken fixed some bugs around canceling interrogatives. - Typeahead stays up when backspacing into a model reasoning suffix. - Double-clicking a flagged file in the sidebar now copies the file’s absolute path. Commands and tool behavior Section titled “Commands and tool behavior” commands-and-tool-behavior - /version shows the current Polytoken version, and GET /version exposes the same information over the local HTTP API. - grep and glob can now search more than one root in a single call and report missing paths separately from permission-denied paths. - transclude can now include project files by path. Permissions and customization Section titled “Permissions and customization” permissions-and-customization - Permission suggestions now handle filesystem paths and shell redirects more safely. - Rules can now match all tools from one MCP server with mcp <server . - Facets without a configured accent color now receive a generated accent color. - Projects can now customize plan handoffs to automate more of the transition from planning to execution. Polytoken 0.2.0 improves day-to-day prompting, planning, and subagent visibility. The release also adds project-level template variables and new CLI metadata export commands for users who build tooling around Polytoken. Planning and prompt behavior Section titled “Planning and prompt behavior” planning-and-prompt-behavior - Polytoken now gives more direct first-pass answers. The model should start with the answer or recommendation, include the essential rationale, name the next useful decision, and offer deeper detail only when useful. - Plan mode now starts from a more structured plan template. Plans include the goal, implementation shape, validation approach, acceptance criteria, and review expectations. - Projects can replace the shipped plan template by setting project vars.plan facet.plan spec override in .polytoken/project vars.yaml . - Templates can now read project variables through {{ project vars.foo.bar }} . Missing project variables resolve to None , so templates can provide fallback behavior. Prompt editor and picker improvements Section titled “Prompt editor and picker improvements” prompt-editor-and-picker-improvements - Clicking inside the prompt box now moves the cursor. - Text selection and copy now work inside the prompt box. - Double-clicking the prompt box or pressing Ctrl+Y copies the current prompt. - Ctrl+Y also copies handoff plans and ask user question explainers when Polytoken shows those prompts. - Clicking the facet, model, or permissions area in the status bar now opens the matching picker. - The mouse wheel now scrolls @ file/reference pickers and / command pickers. - Clicking a picker item now highlights the item. Clicking the highlighted item selects it. TUI display fixes Section titled “TUI display fixes” tui-display-fixes - Polytoken now updates the display more cleanly after a plan handoff, so the execute turn no longer drops the first conversation cards. - Turn counters render more reliably and avoid multiple on-screen counters. - The terminal bell now rings more predictably for turn completion and questions. - A new configuration option disables the animated terminal title bar. Jobs, compaction, and subagents Section titled “Jobs, compaction, and subagents” jobs-compaction-and-subagents - /jobs now shows cleaner job status details and better subagent call history. - Compaction cards now show more useful detail about what Polytoken compacted. - In auto-drain mode, Polytoken now reports subagent failures reliably when the failure arrives. - Polytoken now backs off and retries more gracefully when a provider returns intermittent rate-limit responses. CLI metadata export Section titled “CLI metadata export” cli-metadata-export - polytoken print-commands exports CLI command metadata. - polytoken schemas exports structured schemas, including configuration and template frontmatter schemas.