Skip to content
UHPUHPDeveloper Guide
Independent developer guide. Not affiliated with HarnessRouter or the official UHP project.

UHP and HarnessRouter release tracker

Track protocol versions separately from HarnessRouter software releases and conformance-package revisions.

Page reviewed 22 Nov 2026 Source and review policy

Protocol: 2026-10-19HarnessRouter: v0.41.24
CoordinateCurrent state
UHP protocol2026-10-19 Draft
HarnessRouter stablev0.41.24
Stable release SHA52c929f1cf3097d34a733fd3f419f454d0e7973f
Conformance package2026.10.19.post11
Executable checks111
Released backends30

Separate an upgrade decision from release discovery

Section titled “Separate an upgrade decision from release discovery”

When an upgrade spans several releases, use the dated update analysis to reconstruct which change introduced a behavior, then return to the release and implementation evidence for your installed version. A later suite or support matrix does not replace the denominator of an earlier measurement.

The stable release endpoint, rechecked 22 November, identifies v0.41.24 / 93994cfcea54. Development main was separately observed at the same commit the tag was cut from. v0.41.24 is the stable/Latest release; the v0.37.0-rc.1 through rc.4 and the v0.37.4-rc.1 candidates have not superseded it, and the v0.36.0-rc.1 through rc.4 candidates observed on 23 October have since been superseded by the published v0.36.0 line.

Change familyBefore upgrading, verify
New backend/modelRequired provider route and scenarios; upstream install/runtime policy
Runtime PluginWorkspace inclusion, member authority and connection prerequisites
Environment protocol supportAdvertised optional capability and metadata.environment request shape
Relay/artifact security fixTargeted failure path plus affected authentication/delivery behavior
Conformance package changeInstalled suite revision and a newly generated scoped report

Pin the source/tag or container digest you intend to deploy, preserve its configuration/data recovery path, then validate the workflows affected by those changes. A protocol version can remain unchanged while implementation behavior changes materially. Conversely, optional Environments do not become mandatory merely because a server updates.

Treat release-note measurements as publisher evidence at the named build. Keep your deployment test receipt separate, including failures and untested paths. The conformance evidence remains revision-bounded; neither upgrading the executable package nor reading a new release note refreshes an older 74/74 report.

Published 25 Sep 2026 from 288a6dead48e114acd950f044f131d1212a733bf. PR #266 adds a Browser Use runtime service connected once per workspace and included per harness. See Runtime Plugins.

Published 27 Sep 2026 from aab7597493b16f5ff12084d1f1400d3e7f4e46d5. PR #264 adds CheetahClaws 3.5.88 as backend sixteen. See CheetahClaws.

PR #321 introduces UHP 2026-09-28 Environments. PR #331 patches task selection to metadata.environment, with schema, conformance and implementation moving together. Current package is 2026.10.19.post4 with 102 executable checks.

v0.26.18 — Sonnet 5.5 and Runtime Plugin availability

Section titled “v0.26.18 — Sonnet 5.5 and Runtime Plugin availability”

Published 30 Sep 2026 at 21:33:01 UTC from 4fee4bcc4a4764e418b49d12d6d530a7de4b004f. PR #350 adds Claude Sonnet 5.5 coverage and removes the separate per-organization Runtime Plugin availability gate; PR #351 records the measured Sonnet 5.5 matrix.

Published 1 Oct 2026 at 04:53:54 UTC from 31167726566186e4009790fe7e7ab7061ed99425. PR #343 adds Microsoft 365 through each organization’s Entra integration, with delegated calls bound to the signed-in workspace member.

v0.28.6 — public artifact sharing hardening

Section titled “v0.28.6 — public artifact sharing hardening”

Published 2 Oct 2026 at 05:42:44 UTC from d80c540dcfd7ee8d09915fbedc12c8202ab9d527. PR #368 adds sandboxed public-artifact response policy and durable share-revocation revalidation.

Published 2 Oct 2026 at 10:32:49 UTC from de674959316f6f2ef27be0f9407553f7552ace76. PR #369 adds GPT-6.1 Sol to the GPT-6 Sol-capable harness bases and records connection-specific measured coverage. It does not change the UHP protocol version.

Published 20 Nov 2026 at 23:17:30 UTC from 93994cfcea54. The 0.41 line is a memory line: v0.41.0 polished the memory surface so every workspace remembers, what an agent is reminded of reads as a past note, a memory’s cover is a picture of its schema graph, and every engine is visible; v0.41.1 and v0.41.11 name each remembered line after the person who said it. v0.41.2 to v0.41.7 simplify the memory pages, fix the runner’s file access, take the task prompt off the agent’s command line, and install each agent as its own unprivileged user. v0.41.8 to v0.41.21 are fixes and dependency updates. v0.41.22 and v0.41.24 close a Graphiti recall defect from both sides: a fact drawn from a stated record can outrank the record itself in a search, and because a drawn fact carries no contradicts reference, a recall with a small limit could return one side of a disagreement without the other, which the Memories chapter’s §6.3 requires — UHP conformance ME-12 failed on Graphiti once a recall ran after Graphiti had read the records. v0.41.24 also made a self-hosted sign-in ignore spaces around the username. v0.41.24 fixed an organization switch that a background session refresh could undo: the console re-reads the signed-in person and their organizations in the background, and a refresh that was already on its way could write the earlier session back, returning the stored token and organization to the one the user had left. A refresh now merges what it fetched only into the session it was made for and leaves a newer session alone, and the same rule was applied to the other session writers, so a pre-switch token is no longer stored over a switched session and a 401 for the old token no longer signs the newer session out. The release also dropped x-ai/grok-build-0.1 from OpenRouter’s model list, where the provider had removed it; the model stays in the catalogue through xAI and Vercel.

v0.40.0 — every new agent remembers, per person

Section titled “v0.40.0 — every new agent remembers, per person”

Published 9 Nov 2026 at 09:16:49 UTC from ad19e05ce30a. Memory is now chosen per agent: Mem0 for an agent that serves people, Graphiti for one that tracks changing things, Hindsight for one that learns from its own work, ReMe for notes people read and edit, and none for one-shot transforms. Each person’s memory is their own, keyed by the task’s user, and a task without a user runs with memory off and says so.

Published 8 Nov 2026 at 04:35:16 UTC from 52c929f1cf3097d34a733fd3f419f454d0e7973f. First-start installs are now pinned and stamped, and a connection can carry request headers. Every agent harness on a volume is installed once at the first start after the upgrade, because each is now pinned to an exact version with its release assets checked against digests committed in the image (PR #646, #623): opencode 1.18.23, Qwen Code 0.26.3, Pi 1.5.0, and exact versions for Hermes’s and the other Python harnesses’ top-level packages. A stamp replaces the per-start version checks, and a CI source check fails on any unpinned install, so the list of exceptions can only shrink. The first start after upgrading takes 8 to 14 minutes — 13 minutes 40 seconds on the amd64 hosts measured — because all 30 harnesses are reinstalled; later starts read the stamp and install nothing. If the network fails during that start, the harnesses it could not reinstall are reported as not installed and are retried at the next start, and no harness is left running bytes that were not checked. The installer’s processes now see only the path, the locale, proxy and CA settings and your registry settings, not the instance’s internal key, the console password or the provider keys passed with -e; installs still run as root, so this stops a package that reads or reports its environment rather than one that sets out to read the instance’s secrets. A connection can also carry request headers, which the broker sends on every model call after the key (PR #636): the values are held like the key, are never handed to the agent, are masked when read and are re-typed when the address changes, and names the gateway sets itself — via, forwarded, proxy-*, x-forwarded-*, Cloudflare’s own cf-* — are refused, as are AWS request-signing names (x-amz-*). With HR_STEP_CHECKPOINT=1 the workspace is also saved after each finished tool step, so a sandbox lost mid-task keeps its steps (PR #634, #505); it is off by default because a self-hosted workspace is on the volume and already outlives a restart. HERMES_MCP_PIN is no longer read — use HR_HERMES_MCP_VERSION with an exact version — and HR_OPENCODE_VERSION no longer accepts latest. On a deployment shared by several organizations the setup builder now holds an organization’s own connection address to the same rule the model relay uses since v0.38.4 (PR #649); a single-operator install is unchanged.

v0.38.4 — a shared deployment holds an organization’s own connection address to the MCP rule

Section titled “v0.38.4 — a shared deployment holds an organization’s own connection address to the MCP rule”

Published 7 Nov 2026 at 23:23:15 UTC from 51c45bd62731. On a deployment shared by several organizations, the model broker calls each organization’s own connection, and that call now goes only to an address that resolves outside the deployment’s own networks — the rule an MCP server’s address already follows. An address that cannot be used as written is refused before any lookup, which covers anything that is not http or https, has no host, contains a user name or password, or contains a ? or #. The name is resolved once and every address it answers with is classified, so the socket opens only to one of those addresses: the address that was checked is the address that is connected to. A refused call answers 403 with the reason, on the first attempt and on a retry. These are unchanged: the operator’s own connections, the platform’s connections, and an install where one operator runs everything, which is the default.

v0.38.3 — the MCP probe reads a whole event, and Pi and Oh My Pi can use a custom Responses connection

Section titled “v0.38.3 — the MCP probe reads a whole event, and Pi and Oh My Pi can use a custom Responses connection”

Published 7 Nov 2026 at 22:44:38 UTC from 73dae19715ef. Test connection on an MCP server could find no tools. Some servers split one answer over several data: lines of a single event, for example a pretty-printed tools/list; the probe read each line on its own, so it reported the server as connected, with no name and no tools. It now joins an event’s lines before reading them, as the event-stream format specifies (PR #645), reads at most 4 MiB of an answer and spends at most 30 seconds on it, and says which of the two limits it reached. A custom connection in the OpenAI Responses format can now drive Pi or Oh My Pi (PR #642): the format you choose decides which API they speak, whatever the model is called. This release also carries the two runner changes previously tracked as development evidence on main: a turn’s status and cancel answer only the session that started it (PR #643), and a retried turn’s idempotency key now matches only within that same session (PR #644).

Published 4 Nov 2026 at 22:53:50 UTC from 8e1cf1b598a7ea864fda3a5de064b11289b6ac84. A licensing correction; nothing changes in how HarnessRouter works. The image shipped an LGPL library with no licence text beside it: Next.js lists sharp as optional for its image optimizer, the build copied @img/sharp-libvips-* (LGPL-3.0-or-later) into the image, and nothing called it because the console has never used that optimizer. NOTICE also said everything bundled was under a permissive licence, which was not true — it listed only direct dependencies — and the image did not carry LICENSE or NOTICE, though NOTICE says it travels with a copy. Now sharp and its native libraries are no longer installed or shipped, NOTICE is corrected and names certifi (MPL-2.0) and Pyphen (offered under GPL, LGPL or MPL, used under MPL), and the image carries /app/LICENSE, /app/NOTICE and /app/THIRD-PARTY-console.txt with the licence text of every console package. scripts/check-licenses.py runs in the tests and inside every release’s image, on both architectures, and fails on anything that is not under a permissive licence and is not named with a reason.

v0.38.1 — a security fix and a task that says why it cannot start

Section titled “v0.38.1 — a security fix and a task that says why it cannot start”

Published 4 Nov 2026 at 17:29:58 UTC from 682d843d0149. Reads and writes for a session no longer carry group root. When the server reads or writes a session’s files it takes on the session’s user and group, but through v0.38.0 it also kept root’s supplementary group, so an agent could use a link to reach what gid 0 can reach: reading a file only group root can read, such as one with mode 0640 owned by root, by swapping a file for a link during a download, and overwriting a file group root can write, or creating a file in a directory it can write, through a link at one of the session’s config paths. On the default image this reaches very little — four empty lock files under /var, and creating files in /usr/share/man/man1, man5, man7, man8 and /var/cache/debconf/tmp.ci; no file under /data, and no existing file could be overwritten. A volume you mount can hold more: if you mount files or directories group root can read or write, and an agent you do not fully trust has run on this instance, treat what they hold as exposed, check them, and rotate any credential among them. Files only their root owner can read or write were not reached, and each release now fails its check if the server holds a supplementary group. A task that cannot start says why: a link at one of those config paths used to make a follow-up wait about five and a half minutes and then report that no connection serves the model, which was not the reason; it now fails in under a second and says a file in the workspace cannot be written, and once the link is gone the same session carries on. Any other failure to start now says the task could not start and no longer blames the connections.

v0.38.0 — a minor release: MCP credentials stay out of the session, stored credentials are encrypted at rest, and a task can ask the person

Section titled “v0.38.0 — a minor release: MCP credentials stay out of the session, stored credentials are encrypted at rest, and a task can ask the person”

Published 4 Nov 2026 at 07:28:26 UTC from 80d5d369a1dd. MCP credentials stay out of the session. A remote MCP server’s address and credentials now stay in the server; an agent’s session gets a local address holding one token that stops working when the turn ends. Before this, most bases wrote the bearer token and the address into a config file in the session, where the agent could read it, and some of those files travel in the checkpoint. Rotate MCP tokens if an agent you do not fully trust has run on this instance — older sessions, checkpoints and their git history keep the copies they already hold, and this release does not remove them. Stored credentials are encrypted at rest on every install. An install with no HR_SECRET_KEY now gets one at its first start, kept in /data/secret-key (root’s, mode 600), and records stored in plain text are encrypted at that start; a token or header written on a harness moves to the same store. This protects a copy of the secret store, not a copy of the whole volume, because the key sits beside the data — to keep them apart, set HR_SECRET_KEY to the value in that file and delete it, and do not set a new value over a generated key, because it cannot open what is already stored; the log says how many records it cannot open. Back up /data/secret-key with the volume: rolling back to an older image needs its value in HR_SECRET_KEY. A task can ask you a question. A task can now end by asking — a choice, a confirmation, free text or a secret — the turn ends incomplete with reason input_required, and the answer continues the session. The console has an answer form, and a secret typed there is not saved or shown in the conversation; it works on the Claude Code base for now. This can change what your client sees: a Claude Code task that used to finish with a question in its text can now end incomplete. Every release is now checked for this: each release runs one task per base with a planted value in every credential, and fails if one reaches a session, the volume or a response (scripts/secret-canary-gate.py).

v0.37.18 — current stable at the 4 November publication

Section titled “v0.37.18 — current stable at the 4 November publication”

Published 4 Nov 2026 at 01:41:17 UTC from 433acfe0047e19f6d4610cda065d9287b9620876. Claude Code’s default model is now claude-sonnet-5.5. A task on the Claude Code base that names no model runs claude-sonnet-5.5 instead of claude-sonnet-4.6; claude-sonnet-4.6 is still offered, and a harness with its own saved default keeps it. Anyone upgrading from before 0.37.17 should read that release’s notes first — it closes critical security issues.

v0.37.17 — critical security issues closed

Section titled “v0.37.17 — critical security issues closed”

Published 3 Nov 2026 at 23:16:12 UTC from 433acfe0047e19f6d4610cda065d9287b9620876. An API key could read any file as root. A marker file in a session’s folder could point the file routes at any folder on the machine, and a caller’s own PUT could set it, with no agent involved — reaching the instance’s internal key HR_SECRET_KEY, the stored provider keys and the databases. A follow-up turn on Codex through its app server could also write through a link an agent left in its folder. Only the value the runner writes itself now counts, and these reads and writes run as the session. Upgrading does not undo a read that already happened, and the logs cannot show one, so the release advises: change HR_SECRET_KEY and the console password, replace the stored provider keys and any passed in the environment, reissue the API keys, then restart. File routes now refuse dot paths. Every route on a session’s files — read, write, move, mkdir, delete, file ids, the listing, the archive and share links — refuses a path with any segment starting with a dot, and refuses a link that lands on one. That covers the runner’s state, git’s folder, every CLI’s folder and any credential file, including one an older version left behind, without depending on a list of names; before this, the listing, the archive and share links also served a dot file inside a folder such as sub/.env, so anyone holding a share link could read one. The change was tested through the real routes with about 12,000 requests, with 0.37.16 as a control that failed. What changes for you: an app can no longer read or write .env or .gitignore in a session through the API, and such a file is not listed or downloadable — though the agent itself still works with them, and an agent can still copy a file it can read into an ordinary file. Root no longer runs what the product user can change. The gateway and the console run as agent, but that user owned files root runs, so a bug in either could become root at the next restart; /app is now root’s, root has its own HOME, and on the data volume root keeps its folders where agent cannot move or replace them — and the release gate now checks this on a running container. Also in this release: two organizations connecting the same plugin in a workspace called default no longer share one record; a MySQL-compatible database that refuses a read-only session now gets no query rather than the query running anyway; signed links on an instance with no internal key use a random key instead of a constant from the source; an unexpected error in a turn tells the caller only that the turn stopped, because its text could name internal addresses; the conformance workflow runs third-party code with no write access and with only secrets named UHP_*; and every start checks that the installed agent tools are root’s and written by nobody else. The Bedrock and Vertex relay paths from 0.37.15 were run with real credentials: the agent’s session held no AWS credential, the relay’s route stopped working when the turn ended, and AWS refused any other model.

v0.37.16 — several security issues closed

Section titled “v0.37.16 — several security issues closed”

Published 3 Nov 2026 at 11:35:34 UTC from 433acfe0047e19f6d4610cda065d9287b9620876. The console’s image optimizer was reachable with no sign-in through published Next.js remote-code-execution bugs (GHSA-2464-fhr4-w9xw, GHSA-2xp9-vwfh-vxw4); Next.js and its image library are updated, and since the console does not use the optimizer it is now off. Connecting a database: connections to MySQL and MariaDB are hardened against a hostile database server. Building an environment: a build’s own code could plant a link that made the runner write or read a file as root; system packages now install before any of the project’s code runs, leftover build processes are ended before the runner’s last steps, and the runner refuses to follow links there. Read-only database queries: the check could be fooled into passing a statement that was not read-only; it now reads a statement the way the database does and refuses anything it cannot read unambiguously — a backslash before a quote, dollar quoting and # comments on MySQL are now refused. Files an agent wrote: opening an agent’s HTML or SVG through the console ran it with the console’s origin and the session, and every such file is now sandboxed. A relay route now ends with its turn: a route held a connection’s key until the runner restarted, so an agent process left running could keep spending it, and a process left behind now gets a refusal. Workspaces: a task or a harness can name only an environment its own workspace can reach. Browser plugin: a host can no longer answer a public address once and a private one afterwards. On a cloud VM an agent’s shell can reach the provider’s metadata address, which with an instance role means the role’s credentials — nothing inside the container can stop that, and the self-hosting guide now carries the one host command that blocks it. The release also fixed a regression it had introduced: 0.37.15 sent Gemini’s key on every call both as x-goog-api-key and as a bearer token, and Google reads a bearer as an OAuth token, so every Gemini turn on a Google API key failed with a 401 — the key now goes only as x-goog-api-key. Google API keys no longer offer gemini-3.5-flash: Google answers that id with gemini-3.6-flash, the Gemini CLI failed such a turn as a substitution, and other bases ran 3.6 under the 3.5 name; a provider that serves 3.5 as itself still offers it. A refusal from AWS or Google now says why on the turn’s error and in the log, as it does for every other provider.

v0.37.15 — current stable at the 3 November publication

Section titled “v0.37.15 — current stable at the 3 November publication”

Published 3 Nov 2026 at 03:16:52 UTC from d11a20845b6ae48f83c955d733562b6b37b0c1bd. Security: an agent turn can no longer reach your provider keys or the instance’s secrets. An agent turn runs with real shell access, and everything it can read is also readable by anything that reaches the agent, such as a prompt injection; this release closes every path from a turn to a credential. Provider keys — every key, including the image-generation credential and Bedrock and Vertex credentials — now stays in the runner’s relay, and a turn gets the relay’s address and a placeholder on every base and every kind of connection: Bedrock requests are signed by the relay, Vertex access tokens are minted by it, and the relay accepts only a path and never follows a redirect. A test builds every base’s turn for every connection shape with a canary key and fails if the key appears anywhere the turn can read. The instance’s own secrets no longer travel with git and the other processes the runner starts for a session, and hooks the agent leaves in its repository no longer run. Files outside the session are protected by writing into the session’s folder with the session’s own permissions, so a link the agent planted cannot redirect a write or a cleanup elsewhere. Other sessions are protected by sending the jobs that need one privately instead of on a command line other sessions could read. Saving an integration no longer lets a stored key follow a changed address or provider. And there is no default password: the published default console password is no longer accepted, an instance without a password of its own makes one at first start (read it with docker exec <container> cat /data/selfhost-initial-password), and the compose file now publishes the console on this machine only. Advisories GHSA-cww9-m4gr-cmw5, GHSA-m479-gmx7-8r8h (reported by Omer Bar-Ness, MergeSafe) and GHSA-r37g-f5m3-523c (reported by AkshatJoshi2000); if untrusted people can start turns on your instance, or your agents read untrusted content, rotate your provider keys after upgrading. The starter kits are rebuilt with every critical and high advisory closed — Sheets and Harness Bot read and write Excel files with xlsx 0.20.3 pinned by checksum, Slides and Video use DOMPurify 3.4.17, and every kit builds with vite 6.4.5 — and a deck’s PowerPoint export now includes its diagrams (as an image) and its code blocks (as editable text), naming anything with nothing to export. PR #615, HarnessRouter/starter-kit#54 and #55.

v0.37.14 — current stable at the 2 November publication

Section titled “v0.37.14 — current stable at the 2 November publication”

Published 1 Nov 2026 at 10:24:02 UTC from f6592b27e0d46478f895100a20827d60b538acf4. A file the agent deleted is no longer counted as one it made. Sometimes an agent deletes a file on a later turn after already handing it back, and that turn still listed the deleted file as produced: nothing could be downloaded for it, but it was counted, and the session’s changed files (GET /v1/sessions/{id}/files?changed=true) reported produced: 2 beside a single file. A deletion is now recognised however it was made, so a turn lists and counts only the files that exist. PR #619.

v0.37.13 — current stable at the 31 October publication

Section titled “v0.37.13 — current stable at the 31 October publication”

Published 31 Oct 2026 at 04:38:37 UTC from 558a8bbbd548c57e6b5afb21541993a164a94203. The starter kit formerly called My Workplace is now Harness Bot. It lives at /kits/harness-bot, every teammate’s face is the Harness Bot mascot animated in one of six colours — drawn in a circle so they read on any background, with a still frame when reduced motion is on — and everything launched as My Workplace keeps working: the recruiter, the teammates, their faces, direct messages and groups carry over, and a face from before maps to the nearest of the six so each teammate keeps one look. Bookmarks to /kits/workplace/... open the same page under /kits/harness-bot/..., and pressing Launch on the kit again moves the recruiter to the new name — a name the owner gave the recruiter themselves is kept. A renamed kit keeps what was launched under its old name. kit.json can now say what a kit was called before, and the server reads it: aliases lists the kit’s old ids, so a setup launched under one of them belongs to the kit, the record and catalog report the current id, launching under either id reuses it and the next launch stores the current id; harness.replaces_packages lists the old names of the kit’s package, so the current package takes such a package’s place on the next launch instead of being installed beside it; and harness.replaces_names lists the setup’s old default names, so a setup still called exactly one of them takes the kit’s current name. The console sends an old id’s address on to the current one, read from the same aliases, and kit maintainers can rename a kit with a few lines in its own kit.json. This is a product-side change: no protocol, conformance, backend or capability coordinate moves with it. PR #616 and HarnessRouter/starter-kit#52.

v0.37.12 — current stable at the 30 October publication

Section titled “v0.37.12 — current stable at the 30 October publication”

Published 29 Oct 2026 at 18:50:37 UTC from 5520b956ffc15373a89ddfe00d944c271661f21a. A refused turn says why, when the agent could not read the refusal. A Qwen Code turn refused by vLLM 0.10.0 or older failed with only API Error: 400 status code (no body): Qwen Code reads a refusal’s reason only under an error key, and those vLLM versions put message at the top level, so the one thing a person could fix — a context window too small for the request — was never shown, while the runner’s relay had read vLLM’s sentence all along. A failed turn now gives the provider’s own words whenever the reason it would otherwise show lacks them, for every agent harness whose calls go through the relay: The model provider refused the request (400: This model's maximum context length is 8192 tokens. However, you requested ...). An agent that already says the reason keeps its own sentence, and only the turn’s last call counts — a refusal the relay repaired, or one before a call that went unanswered, is never given as a later failure’s reason. PR #614 and #612.

v0.37.11 — a retried harness create no longer makes a second harness

Section titled “v0.37.11 — a retried harness create no longer makes a second harness”

Published 29 Oct 2026 at 17:15:36 UTC from 8d5c031c320afd0c7e6fd527fbbcf3705e7b0687. POST /v1/harnesses honours an Idempotency-Key, as a task already did. When a create reached the server but its answer was lost — a timeout, a dropped connection — the client could not tell whether the harness existed, and its retry made a second one. With a key, a repeat answers with the harness the first request created; a repeat that is still writing waits for that harness rather than answering not found; the same key with a different request is refused 409 idempotency_key_reused; and once the harness a key created is deleted, the key names none and a repeat creates one. A key belongs to the workspace the harness is made in and is kept for 24 hours, and the server reports the new harness_idempotency capability. Deploy to Cloud sends a key on each create, so a deploy whose answer was lost no longer leaves a second copy in Cloud — the known issue since 0.36.3 — and it sends one only to a Cloud that reports the capability. The protocol text is in Harnesses §5.1 of UHP 2026-10-19, an addition within that version, and conformance suite 2026.10.19.post4 checks it as F-10. PR #613 and #582.

v0.37.10 — current stable at the 29 October publication

Section titled “v0.37.10 — current stable at the 29 October publication”

Published 29 Oct 2026 at 00:38:16 UTC from 3fa39ad1947f9a3d1526e7d1814399498f7b1be0. The console’s setup screens now share one bar, as on the hosted service, with Tasks · Settings at the bar’s right end and the bar’s action reading “Connect your app”; the tabs move to their own row at 860 px and under, and the bar’s actions become icons at 480 px and under. Two smaller fixes come with it: a setup created through the API with its base named by its backend ("base": "claude") shows that base’s name and mark, and a Start demo picked before the list of agent harnesses has loaded takes its own harness and model once the list arrives. DeepSeek Harness model calls go through the runner’s relay. The DeepSeek Harness ran a relay of its own inside its driver, which held the provider key and lacked most of what the runner’s relay does; its model calls now go through the runner’s relay like every other agent harness’s, and the key never enters the driver’s process. That brings the dsh base the runner relay’s repairs and its answer when a provider fails: a refused, dropped or timed-out connection now ends the turn with an error instead of a dead connection, its turns record the model that served them and their usage, and a provider that takes no reasoning_effort gets the request again without it. Measured on the release candidate, Anthropic, Google, OpenAI and Azure each pass the five matrix scenarios, and a deck conversation that changes model family on every turn passes 5 of 5. PR #597, #382 and #611.

v0.37.9 — a produced file shows as soon as its turn is written, and Qwen Code works with a small window

Section titled “v0.37.9 — a produced file shows as soon as its turn is written, and Qwen Code works with a small window”

Published 28 Oct 2026 at 23:44:34 UTC from 164fe1262d4bdbf27656341f93d19351ad56b51f. A produced file shows as soon as its turn is written. A turn’s file card sometimes did not appear until a reload, though the file was in the turn’s record; after a turn ends the console reads the record again to pick up what the stream did not carry, and it waited for a field this edition’s server does not send, so that second read never added anything. It now treats the record as final once it says when the turn ended, which is the same write that stores the files, and a record that does not hold the new turn yet is no longer read as its end. Qwen Code works with a model whose window is smaller than the cap Qwen Code picks. Qwen Code puts its own max_tokens on every request, chosen from the model’s name — 64000 for most names, 16384 for gpt-6.1-sol, 32000 for a name it does not know — so a self-hosted model served with a smaller window refused every turn (vLLM: max_tokens=64000 cannot be greater than max_model_len=8192) and no setting could change it. HarnessRouter now drops that field on Qwen Code’s requests, as it already does for Kimi Code, OpenHands and Vibe, so the provider’s own default applies; Qwen Code’s own prompt still needs about 16,600 tokens, so a model needs a window well above that. And MiniMax M3.1 Flash is offered on MiniMax Code through OpenRouter, Vercel and TokenRouter, having passed the five matrix scenarios and a deck conversation that switches to it from minimax-m3. PR #608, #609, #610 and #578.

v0.37.8 — current stable at the 28 October publication

Section titled “v0.37.8 — current stable at the 28 October publication”

Published 27 Oct 2026 at 20:24:45 UTC from b2091142cdca2f1d0cbaffc3e0e42c983dc88b24. Two harness and model pairs are no longer offered. On the hosted service, fast-agent with Gemini 4 Argon on Google’s own API failed in deck conversations that switched models partway — Argon made one tool call and then gave an empty answer, or said what it would do and made no tool call — and Docker Agent with Grok 4.20, asked to restyle a deck another model had made, said the deck was restyled and made no tool call. Both pairs are no longer offered, here and on the hosted service; fast-agent keeps the rest of the Gemini family and Docker Agent keeps the other Grok models, and the support matrix and its notes are updated to match. A Gemini thought signature is also known by the signature itself: HarnessRouter tracked which model made each Gemini tool call’s signature by the call’s id alone, so a harness that replayed a call under an id of its own while keeping Google’s signature would have slipped past the check. It now recognises the signature itself and keeps only a digest of it.

v0.37.7 — text-only models are known, a Gemini signature goes home, and DeepSeek Harness gets its skills

Section titled “v0.37.7 — text-only models are known, a Gemini signature goes home, and DeepSeek Harness gets its skills”

Published 27 Oct 2026 at 08:54:40 UTC from 20d73fe727c82d13eff8f49b938acbb4d7ca294f. Nemotron 3 Super is treated as a text-only model: it refuses images on both providers that serve it — Vercel answers “This model doesn’t support the image content block” and OpenRouter answers “No endpoints found that support image input” — so a conversation that holds images now leaves them out and the turn goes through. OpenRouter reports a text-only model with a different error from the other providers, and that is recognised too, so a text-only model HarnessRouter has not been told about recovers on OpenRouter as well. A Gemini thought signature only goes back to the model that made it: when a conversation switches between Gemini models, the first model’s earlier calls are sent with Google’s documented placeholder instead of its signature, and a model’s own calls keep their signatures. DeepSeek Harness can use your skills: HarnessRouter now puts each setup’s and plugin’s skills where DeepSeek Harness looks, fresh on every turn, and where a setup’s own skill and a plugin’s share a name, the setup’s wins. DeepSeek Harness no longer offers a web search that cannot work — its built-in web search calls DeepSeek’s own service directly, outside the chosen model connection, and every search failed with an authentication error; it is switched off and no longer listed among the harness’s tools, while web fetch still works and a search plugin in the setup is how the agent searches. And a refused request reaches the agent harness as the provider’s own error: since 0.37.5 HarnessRouter could fail while reading a refusal if one of the setup’s tools had a parameter named type, so the harness saw “Connection error” instead of the provider’s message and HarnessRouter’s own refusal fixes never ran.

v0.37.6 — fast-agent no longer gives up on a slow model before HarnessRouter does

Section titled “v0.37.6 — fast-agent no longer gives up on a slow model before HarnessRouter does”

Published 27 Oct 2026 at 06:43:39 UTC from fca0987ca181fa660da3bff0bbe90a7064e7b383. fast-agent stopped waiting for a model’s answer to start after 150 seconds, which is shorter than HarnessRouter waits on the provider, so a Gemini 4 Argon turn with a long conversation full of images failed with fast-agent’s own “OpenAI stream did not start within 150.0 seconds” while the model was still working. fast-agent now waits a little longer than HarnessRouter does: a slow model gets its full time, and if it never answers, the error comes from HarnessRouter and says what happened.

v0.37.5 — image-reading models see images on fast-agent, and a switch no longer breaks the conversation

Section titled “v0.37.5 — image-reading models see images on fast-agent, and a switch no longer breaks the conversation”

Published 27 Oct 2026 at 06:10:47 UTC from d125bb1fe988c36538d6515714bb23650e383ecf. fast-agent lets Claude Opus 5.5, Sonnet 5.5 and other image-reading models see images. fast-agent decides from its own model table whether a model reads images — whether to offer its attach_media tool, and whether to keep an image in a message — and a model its table does not know counts as text only. That covered 65 of the model ids HarnessRouter sends, most Claude models among them, so those models on fast-agent were never offered attach_media and never saw an image, and a deck they made was worse for it. The runner now tells fast-agent what each turn’s model reads, wherever fast-agent’s own table says otherwise. A conversation with images no longer breaks a switch to a text-only model or one with an image limit: each call now carries only the images the model takes, the most recent ones, and every image left out is replaced by a line that says so — a text-only model gets none, and Mistral models get their eight. When a provider refuses a request’s images and states its limit, HarnessRouter keeps that limit for the model and sends the call again within it.

v0.37.4 — the console gains the team’s letter, and Deep Agents Code defaults to gpt-6.1-sol

Section titled “v0.37.4 — the console gains the team’s letter, and Deep Agents Code defaults to gpt-6.1-sol”

Published 27 Oct 2026 at 05:35:18 UTC from 10fe99bb35ea2a0960716ea73fd746aa110e46ac. The console’s rail now has a Letter button above the account button, and the account menu opens with How it works — the same explanation the hosted console shows, minus the P.S. about the hosted service’s free requests, because a self-hosted instance does not have them; on a phone the letter is under More. Deep Agents Code defaults to gpt-6.1-sol, as Codex, Hermes, Pi, OpenCode and the other Responses API bases already do, having run all five matrix scenarios with it on OpenAI, Azure OpenAI, TokenRouter, OpenRouter, Vercel and the HarnessRouter API. Mistral Vibe, Docker Agent and fast-agent keep gpt-5.4: they speak chat/completions, and OpenAI, Azure OpenAI and TokenRouter refuse gpt-6.1-sol there when the request carries function tools, whether the request leaves the reasoning effort out or sets it to low. Through OpenRouter and Vercel, which reach the model over the Responses API, those bases do run gpt-6.1-sol, and the model picker still offers it there.

v0.37.3 — current stable at the 27 October publication

Section titled “v0.37.3 — current stable at the 27 October publication”

Published 26 Oct 2026 at 23:29:42 UTC from ef9e12e29b61c7bb913a767a7bf1bee96e851ef8. fast-agent can use MCP servers whose address does not end in /mcp. fast-agent rewrites the address of an MCP server on its own, so a server at /v1/mcp/plugs got its requests at /v1/mcp/plugs/mcp, was answered “Not Found”, and the turn failed at startup with “Core not initialized”; the runner now connects fast-agent to those servers itself, at the address as given, while a server whose address ends in /mcp or /sse still goes through fast-agent’s own client. The release records that 0.37.0 had read this as fast-agent refusing local addresses, and the hosted service’s browser check showed the path was the cause. docker-agent no longer runs part of the officecli skill as a command: docker-agent runs a skill’s inline shell-escape sequence as a command when it reads the skill, and one line of the built-in officecli skill contained it, so each time docker-agent read that skill, part of its text ran as a command, failed, and left an error message in the skill the model was reading. The line is reworded in the skills repository and this image carries the new text. PR #596.

v0.37.2 — nemotron-3.5-lightning is no longer offered on three of the new bases

Section titled “v0.37.2 — nemotron-3.5-lightning is no longer offered on three of the new bases”

Published 26 Oct 2026 at 17:38:34 UTC from d64959932c27b298cb9529c861ba7a705a3e5296. On 0.37.1, a custom harness on Deep Agents Code asked to build a Markdown file saved it as tmp/hello_world.md, and the workspace’s instructions say that folder is never shown, so the console had no file to show. The same harness and the same two turns were then run in four fresh conversations on each of the four newest bases, all through OpenRouter: Deep Agents Code saved the file in tmp/ again and once a .pptx nobody asked for; fast-agent saved it in tmp/ and ran about 150 shell commands for the same file; docker-agent saved the .md and a .pptx to /tmp/content-studio-output/, outside the workspace; vibe saved the file at the workspace root every time and the console showed it. A model that fails once on a base, through its own fault, is not offered on that base, so nemotron-3.5-lightning is no longer offered on Deep Agents Code, fast-agent or docker-agent. Every other model on these bases is unchanged, and the measurements are recorded in the upstream support-matrix notes. PR #592.

v0.37.1 — fast-agent and Deep Agents Code show each step as it happens

Section titled “v0.37.1 — fast-agent and Deep Agents Code show each step as it happens”

Published 26 Oct 2026 at 10:59:27 UTC from b4d53a10d0882d6fe7edfce0d59a03ce2e61a2d6. On 0.37.0 both new bases showed “Thinking” for a whole turn and then every step at once; each tool call, its output and the answer now appear the moment the agent reports them. fast-agent now runs over its own Agent Client Protocol server (fast-agent-acp, the same 0.10.47 release), and its shell commands run through HarnessRouter, so each one shows its command, its output and its exit code. The release records the measured latency for one turn running two commands of four seconds each: 17.9 s against 16.6 s for the whole turn on 0.37.0, with the first command’s 7.5 s visible rather than hidden behind the turn. A fix comes with it — a fast-agent conversation that holds an image no longer carries it into later models that take text only, which had failed every subsequent turn with “No endpoints found that support image input” — and Deep Agents Code’s shell tool, named execute, now shows its command instead of nothing. PR #591.

Published 26 Oct 2026 at 05:37:39 UTC from 5abab5b002bcac7b0fccd4bf83d47a8a7f2f146a. HarnessRouter now runs four more agent harnesses, each on its latest release: Mistral Vibe 3.0.2 (Apache-2.0, the bundled Rust CLI from mistralai/mistral-vibe), Docker Agent 1.152.0 (Apache-2.0, its binary pinned by checksum), fast-agent 0.10.47 (Apache-2.0) and deepagents-code 0.2.1 (MIT), the last driven over the Agent Client Protocol so a conversation continues across turns. Each is installed on first start from dependency trees pinned by hash and checked as they install; deepagents-code needs Python 3.14 and gets its own pinned CPython 3.14.9, used for nothing else. Every base was measured end to end on every provider before release — a first turn, a follow-up, a model switch, a file, and a conversation continued after its sandbox restarted — and a base offers only the models that passed all of it, so the release carries a per-base model matrix rather than a uniform catalogue. The released backend set moves from 26 to 30 with this release, and the runner’s BACKENDS registry now holds exactly those 30 entries: Gajae-Code, which the previous stable implemented but did not offer, is no longer in the registry. PR #541.

v0.36.4 — current stable at the 26 October publication

Section titled “v0.36.4 — current stable at the 26 October publication”

Published 25 Oct 2026 at 02:42:30 UTC from edca78bffa40418a00f3c7bc524b2ba9890caecc. Two fixes. A setup’s open task stays open when the window narrows: on a wide window the newest task opens beside the list by itself, but narrowing the window to a tablet’s or a phone’s width switched the page to the list while the address still named the task, and the same happened to a task opened from its link after visiting Settings; the task now stays shown with its header and back button, while a page opened on a phone still starts on the list. And codex’s own warnings are no longer listed as tool calls: with codex in exec mode (HARNESS_CODEX_APPSERVER=0) a warning from codex itself, such as “Model metadata for openai/gpt-6.1-sol not found”, appeared in a task’s tool calls as a call named error; it now goes to the runner log, and the default app-server mode was not affected. PR #584 and PR #586.

v0.36.3 — Deploy to Cloud works for a second destination, and a tool output that could not be read is marked incomplete

Section titled “v0.36.3 — Deploy to Cloud works for a second destination, and a tool output that could not be read is marked incomplete”

Published 24 Oct 2026 at 06:35:09 UTC from e2e8ce41b64a910a2b4616fa4744166470bd70c0. Four fixes. Deploy to Cloud works for a setup’s second destination: since 0.9.0, deploying a setup that already had a copy in one HarnessRouter Cloud workspace to another workspace or organization always failed with “cloud answered 404: No harness with that id.”; a copy is now created the way the protocol creates a harness (POST /v1/harnesses) and replaced with PUT, and a copy deleted on Cloud is created again by the next deploy. A codex command’s output is stored whole: output a command printed before a pause was missing from the task’s record though the model saw it — an upstream codex bug (openai/codex#28207) — so the output now comes from codex’s own record of what the model was given, including codex’s own cut of a very long output. A shared link answers 503 when its Task cannot be read, instead of 500. And Deploy to Cloud stores the destination names the key is bound to, instead of the instance filling in “Default Workspace” itself. One known item is recorded with the release: if Cloud creates a setup’s copy but its answer never reaches the instance, the next deploy creates a second copy, which can be seen and deleted in Cloud. PR #577, PR #580 and PR #581.

One protocol-surface change is worth naming. A function_call_output item now has status: "incomplete" when the tool’s output could not be read in full, instead of "completed". The release states that this stays within UHP 2026-10-19: the schema leaves an output item’s status an open string, and "incomplete" is the value a response already uses for work that did not finish. The Tasks §3.1 example still shows "completed", so a client that assumes every tool output is completed is reading a field the specification does not fix.

v0.36.2 — current stable at the 24 October publication

Section titled “v0.36.2 — current stable at the 24 October publication”

Published 23 Oct 2026 at 21:38:52 UTC from 66eb6e55f956c9d5fcfee55d504bf39954a0be57. Three fixes. A shared link survives a brief storage hiccup: if the share index cannot be read for a moment, the link answers “could not be loaded just now” with Try again, instead of saying the link was revoked or never existed and then saying so for ten seconds after the store recovered; a link that is really gone still says it is unavailable. DeepSeek Harness tool calls work with every provider’s stream format: through a provider that writes stream events without a space after data: — which the format allows — and sends an empty name on each continuation of a tool call, every tool call on the turn failed as an unknown tool. And the workspace menu fits the window: with many workspaces the list scrolls inside the menu, and Workspace settings and New workspace stay in reach. PR #572, PR #573 and PR #575.

Published 23 Oct 2026 at 19:57:09 UTC from 7fa821e0948283be801583ce41bd7630fce9f40b. A setup on a self-hosted instance can now go to HarnessRouter Cloud from its page bar, beside Set up the app. The first time, it opens HarnessRouter Cloud’s sign-in in a new tab; sign in or sign up, and the deploy goes on by itself — no key to copy or paste, and the key stays on the instance’s server. The button always says where the setup stands: Deploy to Cloud when it is not deployed, Update Cloud when it has changed since its deploy, and In Cloud when the copy is there. The instance stays the source, and every deploy replaces the same copy. Two fixes come with it: a key from HarnessRouter Cloud’s sign-in now deploys to that organization’s Default Workspace instead of being refused, and a setup deployed to Cloud no longer reads “changed since” immediately after its deploy. For API clients, POST /v1/cloud-upload/connect and GET /v1/cloud-upload/connect/{code} carry the sign-in for a deploy, and a deploy’s answer and GET /v1/harnesses/{id}/upload gain open_url. PR #574.

v0.36.0 — the redesigned console, and the optional input_history capability implemented

Section titled “v0.36.0 — the redesigned console, and the optional input_history capability implemented”

Published 23 Oct 2026 at 09:27:53 UTC from 08cc4ac3136fff8700f1bbd493ebee3ab84ab964. The redesigned console reaches self-hosted instances, with the setup flow of HarnessRouter Cloud: describe what an agent should do for your users, and the instance builds it. A builder reads your workspace and writes the setup — a name and standing instructions, the harness and model when left on Auto, the skills, plugins, MCP servers and environment the job needs, and one example request. The builder runs on your own model keys; with no key connected, New setup says to connect one first. The sections are now Start (example jobs and guides), Studio (your setups, their tasks and settings), Shop (starter kits), Add-ons (plugins, skills, tools and environments), API Keys, Bring your own key, Profile, and Workspaces. Old addresses redirect to their new places, query strings included.

The optional input_history capability is now implemented and advertised. POST /v1/responses can bring a task’s earlier turns: with no previous_response_id and no session named, the user and assistant messages before the last user message become the task’s first turn, marked imported, with no model and no cost. GET /v1/bases and GET /v1/uhp report capabilities.input_history: true. This is a change within protocol version 2026-10-19; the UHP version, the schema and the conformance denominator do not change. The builder’s own routes, POST /v1/harnesses/builds and GET /v1/harnesses/builds/{key}, are for the console signed in as a person, and an API key cannot call them.

v0.35.17 — current stable at the 23 October publication

Section titled “v0.35.17 — current stable at the 23 October publication”

Published 22 Oct 2026 at 09:58:40 UTC from 7f036edf545b5c55582c31f5750b70f24fa079c9. A failed Agent Zero turn shows its exception and stack trace in the trace, like every other harness. Agent Zero handles its own exceptions, so its failure never reached the output the trace’s diagnostic is read from, and the stack frames were dropped. A failed turn’s trace now carries the error with its frames: the message first, then the innermost frames, capped at 4,000 characters and with secrets redacted. The turn’s one-line error is unchanged. PR #546 and PR #571.

v0.35.16 — nothing is answered from a read that failed

Section titled “v0.35.16 — nothing is answered from a read that failed”

Published 22 Oct 2026 at 08:00:22 UTC from 00054dcc95947c5ca5b3b9f0cad2e62f46ac7a38. Seven fixes, all of one shape: a read that could not be made whole is reported instead of being answered from what happened to be there. A video’s canvas is never wiped by a read that failed. An export never delivers a silent film when a sound on the timeline can no longer be read. Pressing a kit’s Launch twice never makes two copies. Workspaces are never renamed or overwritten by a read that failed. The storage total is never a false zero. A package that declares plugs now gets them. And a download of files that cannot be made whole answers file_not_found, the protocol’s code for it, in place of a generic not_found. The conformance suite moves with it: R-07 now deletes a shared session at the path Sessions §6 names, falling back to the older one, so a server that serves only the newer path is measured; and an install of the suite missing its schema validator no longer reports R-01 to R-06 as ERROR, because the validator’s absence is its own condition, ValidatorUnavailable. The routes that could not answer whole now answer 503 where they answered a 200 with something missing, a 404 or a 400; those are codes the protocol already defines, so this is a fix within protocol version 2026-10-19. PR #561 to #569.

v0.35.15 — a file, a transcript and a response are never answered short

Section titled “v0.35.15 — a file, a transcript and a response are never answered short”

Published 22 Oct 2026 at 07:08:17 UTC from 3cfdeb97e0ede49fed8567b8e702b4a1777a0fba. Nine fixes, mostly about answers that were served incomplete. A file with an accented or quoted name is listed under its real name, so résumé.md is no longer listed under an escaped name no file has. An app cannot overwrite a harness’s own files, such as nanobot’s AGENTS.md or opensquilla’s SOUL.md. A turn’s file count counts what it produced, not the files the caller attached. A download of a turn’s changed files is refused with the names of the files that could not be read. A transcript is never shown with a hole in it. A response that could not be read answers 503 to try again instead of a permanent response_not_found; a response that does not exist is still a 404. A database whose schema was cut short by the row cap says truncated. A browser the person has taken over is not stopped for being idle. And an apt version with an epoch is a valid package spec. The routes that could not be read whole now answer 503 or a 404 naming what is missing, instead of a short 200; both are codes the protocol already defines, so this is a fix within protocol version 2026-10-19. PR #547 to #557.

v0.35.14 — gpt-6.1-sol becomes the default where it passes, and a failed task names the harness and the tool

Section titled “v0.35.14 — gpt-6.1-sol becomes the default where it passes, and a failed task names the harness and the tool”

Published 22 Oct 2026 at 05:35:08 UTC from 4d18f215c3882ae48b0017ee3a88a77149dd7e00. gpt-6.1-sol is the default model where gpt-5.4 was, on every harness where it passed every check on the connections people use (OpenAI, Azure and TokenRouter): codex, hermes, pi, opencode, kilo, omp, goose, aider, agentzero and mimo. A setup with no model of its own runs gpt-6.1-sol after upgrading; to keep gpt-5.4, set it as the setup’s model. cline, cheetahclaws, jcode, nanobot, zeroclaw and openhands keep gpt-5.4, because gpt-6.1-sol refuses the request those harnesses send through a connection that passes chat completions straight to the provider; opensquilla keeps gpt-5.4 because gpt-6.1-sol is not offered there. A failed task says which harness and which tool failed: the error’s detail names the base that ran the turn and the tool whose result came back as an error, and the trace’s last event keeps the harness’s own exception and stack trace, capped and with secrets redacted. SQL connections respect a statement’s own row limit in every form Postgres accepts — LIMIT ALL, FETCH FIRST n ROWS ONLY (or FETCH NEXT, WITH TIES) and an OFFSET in ROWS — where a second LIMIT was previously appended and Postgres refused the query. A pip wildcard pin such as numpy==2.* is a range, so an environment’s package check no longer calls it unpublished. Aider with gpt-5.4-mini writes the file it was asked for: a reply that is only a word of the request now gets the block to write, with the path in it — measured the same day on a copy of a real instance, after three exact-reply turns, 0 of 6 files written on 0.35.13 against 5 of 6 on this release, with a fresh session writing 6 of 6 on both. API keys can be renamed (PATCH /v1/orgs/{org}/keys/{kid}), where the console’s Rename had sent the request since the keys page shipped and every rename answered 405. Saving model connections with plain model ids works, where a plain id no longer fails the whole save with a 500. Two known items are recorded with the release: choosing gpt-6.1-sol on cline, cheetahclaws, jcode, nanobot or zeroclaw fails on a connection that reaches the provider directly, and through TokenRouter a codex task on gpt-6.1-sol cannot continue on its later turns, because TokenRouter serves the model from more than one account. PR #504, #513, #542 to #545 and #558 to #560.

v0.35.13 — current stable at the 22 October publication

Section titled “v0.35.13 — current stable at the 22 October publication”

Published 21 Oct 2026 at 08:10:14 UTC from 3668370c10255be7085bae207ace12a94422614f. A conversation is never answered or drawn as empty because the server could not read it. GET /v1/sessions/{id}/turns and a shared link’s turns now answer 503 with type: server_error when the conversation cannot be read, instead of 200 with the unreadable turns left out — so a finished task can no longer look shorter than it was, or new and empty. The console says “This task could not be read just now.” with a Try again action, and a shared link says “This conversation could not be loaded just now.”; Send stays off until a read succeeds. A task’s latest turn always shows in its conversation even when the session’s list of turns did not record it. A turn’s record and its entry in that list are each written twice before the server gives up, and a write failure is logged but never fails the turn. /readyz gains turns_reads, counting failed conversation reads and writes and turns shown from the session. A codex follow-up on a session from before 2026-09-05 now applies the model switch check to the models its earlier turns ran. The publisher records a known item on this release and on 0.35.12: Aider with gpt-5.4-mini answered a request to create a file without creating it, on the release day, having passed the same check the day before.

v0.35.12 — a task waiting past the hard limit is failed, and never starts

Section titled “v0.35.12 — a task waiting past the hard limit is failed, and never starts”

Published 21 Oct 2026 at 04:54:48 UTC from 8d2c01170641. A task still waiting when the hard time limit passes — queued for a free slot, or restoring a workspace that hangs — now ends failed with "reason": "timeout" in its response record as soon as it is settled, so a client polling GET /v1/responses/{id} sees it end. Before, the session said failed while the response record stayed in_progress. Once settled, such a task never starts its agent; before, it ran anyway and finished as completed after being settled as timed out. PR #539.

v0.35.11 — interrupted work reports itself, and survives

Section titled “v0.35.11 — interrupted work reports itself, and survives”

Published 21 Oct 2026 at 03:12:59 UTC from a69531d1ae85. A task cut short by a server restart, or by the gateway shutting down, now ends failed with an error that says so — {"type": "harness_error", "code": "harness_error", "detail": {"reason": "interrupted"}} — and a plain sentence; a task past the hard time limit says "reason": "timeout". Before, such a task could end failed with no error, or completed with no output. On self-hosted installs, the files an interrupted task wrote survive a recreated container, because the checkpoint markers now live on the data volume (HR_WS_MARKER_DIR, which the entrypoint sets to /data/ws-markers). A task restoring a large workspace keeps its heartbeat, so it is never taken for a dead one while it starts, and the hard time limit is now the runner’s limit plus a 30-minute margin (HARNESS_CAP_MARGIN_S). grok streams its answer word by word when partial messages are on (HARNESS_STREAM_PARTIAL=1 or metadata.stream_partial). A caller’s W3C traceparent and tracestate now reach the agent’s environment (TRACEPARENT), every model call carries a child traceparent, and a malformed value is dropped rather than refused. The conformance suite at this tag is 2026.10.19.post3, with T-12, which measures the step budget.

v0.35.10 — current stable at the 20 October publication

Section titled “v0.35.10 — current stable at the 20 October publication”

Published 20 Oct 2026 at 00:48:17 UTC from caa3c0a109fffd84ef245dcdad23f2cdfeff7153. An aider turn whose whole reply is an edit no longer answers with just the file’s name — it says what aider did, for example “Edited hello.txt.” And a running session with no heartbeat on record is aged from when it was created or when its turn started, never from 1970, so the gateway’s cleanup sweep can no longer mark a task that just started as failed. PR #529 and PR #531.

v0.35.9 — aider’s automatic file-add chatter leaves the reply

Section titled “v0.35.9 — aider’s automatic file-add chatter leaves the reply”

Published 19 Oct 2026 at 13:37:19 UTC from 4995106ea67d. PR #528: a follow-up edit on a file not yet in aider’s chat no longer shows aider’s request to add the file, or an empty message, in the reply. The reply is what aider did.

v0.35.8 — the models view’s fallback is deprecated

Section titled “v0.35.8 — the models view’s fallback is deprecated”

Published 19 Oct 2026 at 11:55:19 UTC from eaa37b8b640c. PR #524 deprecates the harness models view’s fallback field, following UHP 2026-10-19 Harnesses §3: it only ever repeats default, nothing changes on the wire, and it is removed no earlier than the second version after 2026-10-19. Read default and each model’s available instead. PR #526 has Hermes check its work once, before it answers, and keep the check out of its reply; PR #525 demonstrates idempotent retries in the quickstarts. The conformance suite at this tag is 2026.10.19.post2 with no check or verdict changes.

v0.35.7 — background answers use UHP’s status vocabulary

Section titled “v0.35.7 — background answers use UHP’s status vocabulary”

Published 19 Oct 2026 at 07:01:21 UTC from 7f3cf64e7e08. This is the release that carries most of the remaining 2026-10-19 additions into the product:

  • A new task may bring its earlier turns (PR #520, UHP 2026-10-19 Tasks §2.1). With input_history, the messages before the last user message are the conversation so far, kept as the session’s first turns marked imported.
  • Each turn says when it began and ended (PR #521, UHP 2026-10-19 Sessions §3). A session’s turns carry created_at and completed_at in Unix seconds, with completed_at null while the turn runs or when its end is not known.
  • Background responses report in_progress (PR #512). A client that checked for running should check for in_progress; UHP’s five-status vocabulary names in_progress as the only non-terminal state, and an implementation answering a poll with its own internal name leaves a strict client nothing to wait for.
  • An idempotent replay whose original response is gone now answers 404 response_not_found instead of 200 with "status": "unknown".
  • A produced file larger than the attachment cap is named in the answer and stays in the task’s files; the cap is now 64 MiB (HARNESS_RESP_MAX_FILE_BYTES, was 25 MiB).
  • A cancel answered cancelled now holds, including one sent the moment a background task starts.

The conformance suite at this tag is 2026.10.19.post1, which adds T-11 (a background task read mid-turn reports in_progress).

Published 19 Oct 2026 at 03:48:00 UTC from 28cac79a2dcac6e8c0f4a7f458417d1d6b014652. Matches UHP 2026-10-19 Errors §3.1: a model its base cannot serve is 422 model_unavailable; the harness-write path keeps 422 model_not_available for a default_model the base does not run. An aider turn that only edited files gets one nudge to give the words back when the request asked for them, and otherwise says what the turn changed. The released backend count remains 26.

v0.35.5 — UHP 2026-10-19: harness retirement and the two error moments

Section titled “v0.35.5 — UHP 2026-10-19: harness retirement and the two error moments”

Published 19 Oct 2026 at 00:41:10 UTC. PR #515 implements the harness_retirement capability: a server that reports it may retire a harness’s sessions and responses with it once a stated restore window has passed, and a restore inside the window brings them back. PR #516 separates model_not_available (a harness write named a default_model its base does not run) from model_unavailable (the model the request named cannot be served on this harness), and adds the F-09 Full check that exercises the new retirement behavior.

v0.35.4 — environment build that stopped is a failed build

Section titled “v0.35.4 — environment build that stopped is a failed build”

Published 19 Oct 2026 at 00:07:17 UTC. PR #514 makes an Environment build that stopped without recording its end read as failed, so a client can tell “stopped” from “completed” without inspecting the build record.

v0.35.3 — README names all 26 harness bases; aider nudges

Section titled “v0.35.3 — README names all 26 harness bases; aider nudges”

Published 18 Oct 2026 at 23:49:32 UTC. PR #510 updates the README and the implementations page so the released harness bases are named in the source itself. PR #499 gives an aider turn one nudge when the request named a file to write and the turn did not, and PR #502 (stacked on #499) makes aider’s no-files messages say that a new file was written directly.

v0.35.2 — npm dist-tags, SQL driver cap, step budget

Section titled “v0.35.2 — npm dist-tags, SQL driver cap, step budget”

Published 18 Oct 2026 at 19:15:50 UTC. PR #496 resolves npm dist-tags to their versions in the Environment package check. PR #497 bounds the SQL fetch itself at the driver. PR #495 enforces the step budget for bases with no limit of their own.

Published 18 Oct 2026 at 06:26:23 UTC from 41a28230d52681157913c572fc5244bc5c00f7f0. Twelve fixes with no protocol revision: a request’s max_output_tokens is named in metadata.ignored_fields; the self-hosting guide no longer says max_cost_usd caps a turn; a route’s totals are folded and read under one lock; the three copies of the Gemini tool-schema repair are made to agree; a SQL result cut short by the row cap says so and a statement’s own LIMIT offset and count are respected; an empty media answer says what was missing; an aider answer line that starts with a shell’s name is kept; a small binary file is not taken for text; a cheetahclaws failure reason that names a path is kept; an npm range is not called an unpublished Environment; and a zeroclaw tool result without rawOutput is its content. UHP remains 2026-09-28 as CE advertises it, and the released backend count remains 26.

v0.35.0 — session workspace file operations

Section titled “v0.35.0 — session workspace file operations”

Published 18 Oct 2026 at 03:27:35 UTC. PR #478 adds delete, mkdir and move beside the session workspace’s existing GET/PUT file routes.

v0.34.2 — pinned console fonts and a native Qoder driver

Section titled “v0.34.2 — pinned console fonts and a native Qoder driver”

Published 16 Oct 2026 at 18:08:08 UTC. The console serves its fonts from pinned packages so the build never fetches them from Google, and the Qoder driver speaks Qoder’s own Python SDK while the relay reads a DELETE’s body.

v0.34.1 — the relay names the model that served the turn

Section titled “v0.34.1 — the relay names the model that served the turn”

Published 15 Oct 2026 at 22:11:26 UTC. Hermes and Codex on Azure name the model that served them and count cached input.

v0.34.0 — runner, MCP bridge and support-matrix corrections

Section titled “v0.34.0 — runner, MCP bridge and support-matrix corrections”

Published 15 Oct 2026 at 02:46:25 UTC. The runner answers Hermes’s local model-server probes itself, removing about 1.3 s from the start of every Hermes turn; the MCP bridge forwards only what a closed tool declares, so jcode’s browser calls are refused; Qoder’s efficient, kmodel and qmodel_latest are not offered because they build the file and never deliver it; nanobot’s shell receives the turn’s own settings; a served alias of the model asked for is not a substitution; and a saved goose provider is overridden on every turn.

Published 13 Oct 2026 at 16:19:07 UTC. MiMo Code (Xiaomi’s coding agent, an opencode fork), jcode, OpenSquilla and nanobot run in the sandbox like every other base, and Qoder Cloud Agent is the first remote runtime: its agent loop runs in Qoder’s cloud and the files it delivers land in the workspace. Each base offers only the models it was measured to run reliably. The released backend set moves from 21 to 26 with this release.

Published 8 Oct 2026 at 03:17:35 UTC. Deleting a task now also removes its responses, the files it produced and their previews, its changed-file list and its media-job records; the same rule covers API keys and credentials. The self-hosting guide documents what a delete removes and when. Back up the data volume before upgrading.

v0.31.0 — Microsoft Entra sign-in for Azure

Section titled “v0.31.0 — Microsoft Entra sign-in for Azure”

Published 6 Oct 2026 at 05:06:37 UTC. An Azure connection can sign in with Microsoft Entra instead of an API key, for an organization that issues no API keys for its Azure resources. The service signs in on every model call and again before the token expires; the agent never holds the secret or the token.

Published 5 Oct 2026 at 21:47:46 UTC. A task can send "reasoning": {"effort": "low"}, or a harness can set Thinking as the default for its tasks; the levels are none, minimal, low, medium, high and xhigh, and a task’s own level wins. GET /v1/bases lists the levels each model really has, and a model asked for a level it lacks gets the nearest one it has. At this release 19 of the 20 bases ran a model with levels.

v0.30.1 — faster task end and streamed answers

Section titled “v0.30.1 — faster task end and streamed answers”

Published 5 Oct 2026 at 22:50:44 UTC. The workspace save after every turn no longer carries the large files Codex and Oh My Pi each re-download, and a streamed answer is forwarded as it is written rather than in 1.2-second jumps.

Published 18 Oct 2026 at 06:26:23 UTC from caa3c0a109fffd84ef245dcdad23f2cdfeff7153. Twelve fixes with no protocol revision: a request’s max_output_tokens is named in metadata.ignored_fields; the self-hosting guide no longer says max_cost_usd caps a turn; a route’s totals are folded and read under one lock; the three copies of the Gemini tool-schema repair are made to agree; a SQL result cut short by the row cap says so and a statement’s own LIMIT offset and count are respected; an empty media answer says what was missing; an aider answer line that starts with a shell’s name is kept; a small binary file is not taken for text; a cheetahclaws failure reason that names a path is kept; an npm range is not called an unpublished Environment; and a zeroclaw tool result without rawOutput is its content. UHP remains 2026-09-28 as CE advertises it, and the released backend count remains 26.

Published 18 Oct 2026 at 03:27:35 UTC. PR #478 adds delete, mkdir and move beside the session workspace’s existing GET/PUT file routes.

Published 16 Oct 2026 at 18:08:08 UTC. The console serves its fonts from pinned packages so the build never fetches them from Google, and the Qoder driver speaks Qoder’s own Python SDK while the relay reads a DELETE’s body.

Published 15 Oct 2026 at 22:11:26 UTC. Hermes and Codex on Azure name the model that served them and count cached input.

Published 15 Oct 2026 at 02:46:25 UTC. The runner answers Hermes’s local model-server probes itself, removing about 1.3 s from the start of every Hermes turn; the MCP bridge forwards only what a closed tool declares, so jcode’s browser calls are refused; Qoder’s efficient, kmodel and qmodel_latest are not offered because they build the file and never deliver it; nanobot’s shell receives the turn’s own settings; a served alias of the model asked for is not a substitution; and a saved goose provider is overridden on every turn.

Published 13 Oct 2026 at 16:19:07 UTC. MiMo Code (Xiaomi’s coding agent, an opencode fork), jcode, OpenSquilla and nanobot run in the sandbox like every other base, and Qoder Cloud Agent is the first remote runtime: its agent loop runs in Qoder’s cloud and the files it delivers land in the workspace. Each base offers only the models it was measured to run reliably. The released backend set moves from 21 to 26 with this release.

Published 8 Oct 2026 at 03:17:35 UTC. Deleting a task now also removes its responses, the files it produced and their previews, its changed-file list and its media-job records; the same rule covers API keys and credentials. The self-hosting guide documents what a delete removes and when. Back up the data volume before upgrading.

Published 6 Oct 2026 at 05:06:37 UTC. An Azure connection can sign in with Microsoft Entra instead of an API key, for an organization that issues no API keys for its Azure resources. The service signs in on every model call and again before the token expires; the agent never holds the secret or the token.

Published 5 Oct 2026 at 21:47:46 UTC. A task can send "reasoning": {"effort": "low"}, or a harness can set Thinking as the default for its tasks; the levels are none, minimal, low, medium, high and xhigh, and a task’s own level wins. GET /v1/bases lists the levels each model really has, and a model asked for a level it lacks gets the nearest one it has. At this release 19 of the 20 bases ran a model with levels.

Published 5 Oct 2026 at 22:50:44 UTC. The workspace save after every turn no longer carries the large files Codex and Oh My Pi each re-download, and a streamed answer is forwarded as it is written rather than in 1.2-second jumps.

Published 4 Oct 2026 at 15:17:35 UTC from d257219281ef3b5a113fe8d8f2550ae0d12be320. PR #381 fixes provider-stream forwarding that could hold short responses behind a 4 KB read, and makes relay liveness depend on real provider events so comment-style keep-alive traffic cannot suppress HR_RELAY_UPSTREAM_TIMEOUT_S (600 seconds by default). The DeepSeek Harness relay takes the same rules. The release also fixes publication semantics so release candidates are marked prerelease and no longer take GitHub’s Latest badge. UHP remains 2026-09-28, conformance remains 2026.9.28.post3 / 85 checks, and the released backend count remains 20.

Publisher validation drove the real relay handlers over real sockets and reports 75 harness/model pairs on nine bases with 359/375 scenario runs, plus 14/14 family tours on DeepSeek Harness and Grok Build and 25/25 released-image checks on four bases. These are upstream release measurements, not this site’s independent provider run; the release explicitly says the relay defects were not caught on a live provider.

v0.29.0 — previous stable / backend expansion

Section titled “v0.29.0 — previous stable / backend expansion”

Published 4 Oct 2026 at 00:33:05 UTC from 2826cf633b0588fa494da20fdf2828deacfb7e79. PR #373 adds Kilo CLI 7.8.1, MiniMax Code 0.5.4, Grok Build 1.0.41 and Agent Zero v2.13 as four new harness bases, taking the released backend set from 16 to 20. The same merge adds relay resilience and secret-handling hardening. UHP remains 2026-09-28 and conformance remains 2026.9.28.post3 / 85 checks.

Historical conformance results remain tied to the source revision that produced them. A later check-count increase does not rewrite an earlier denominator.