Repository navigation
AgentRun v0.1.0-beta.39
Pre-releaseAgentRun 0.1.0-beta.39
AgentRun is an open-source platform that supercharges your agents. It adds self-improvement, durable execution and a
control plane on top of your favorite agent harness. Your agents become smarter, more fault tolerant and use fewer
tokens with every run. This release is about an inbox assistant that runs on your own machine, and about starting with
no model key. The server reads and sends mail through one account of yours, an email trigger starts a run on each new
message, and Settings → Email sets the account up. npx @parcha/agentrun now ends its first screen with Continue with
ChatGPT: Codex and pi run on your ChatGPT plan with no key saved. A run's cap is never passed (one worst-case mode, and
cap_mode is gone), a free model costs a known $0 and stops at a request limit, an OpenRouter run books what OpenRouter
charged, and a run that parks for you keeps its desktop, phone or sandbox for 10 minutes. Mode 2 now starts on Agent37,
Modal and E2B from AgentRun's public box image. Existing Claude Code and Codex setups connect without copying their
configuration, and the Learning tab replays a held workflow and grades a run. There is no fresh store this time: the
SQLite store stays at version 11 and Postgres migrates itself. The launcher is @parcha/agentrun 0.4.7, the Python
client parcha-agentrun is 0.3.7, and the server image is ghcr.io/parcha-ai/agentrun-server:v0.1.0-beta.39
(linux/amd64 and linux/arm64).
No fresh store: beta.39 opens the data beta.38 left. The SQLite schema is still version 11, and Postgres goes from
11 to 14 (migrations 0012, 0013 and 0014). Keep the same data folder and the same start command. On the Docker image
that is--data /data/b37 --listen 0.0.0.0 --port 7340 --drain-timeout 6sas before, and on Render the Docker Command
you set for beta.37 stays. Change only the image toghcr.io/parcha-ai/agentrun-server:v0.1.0-beta.39. The console
keeps its login, and your agents, MCP servers, keys and sign-ins are all still there.
- SQLite. The server adds three columns when it opens your folder, so there is nothing to run:
revisions.monid_pin
(an agent revision's monid pin, #2863),settings.connector_idwith an index on it (#2839), andruns.cost_upper_bound
(#2911; every run that ended before reads 0). Start the server first: a runner started on a folder the server has not
yet opened refuses with "Start the AgentRun server ... before opening a runner". Copy the data folder before you
upgrade. beta.38 still opens the folder afterwards (it checks the schema version only, 11, and its writes name
their columns, so the new ones take their defaults). To go back, restore the copy and set the image back to
v0.1.0-beta.38; the command stays. Two things beta.38 does not read: an agent revision saved on beta.39 with a
monid grant keeps its pin inrevisions.monid_pin, where beta.38 looks in the revision's body, so its runs are
offered no monid endpoint; and a trigger of a kind beta.39 added (composio,
delete such triggers before you go back.- Postgres. Migrations
0012_revision_monid_pin.sql,0013_native_command_owner.sqland
0014_run_cost_upper_bound.sqlrun when the server starts, and the server moves the schema to version 14. Going
back does not work without a backup: beta.38 refuses a schema at version 12 or later, withthe Postgres schema agentrun is schema version 14, set up by a later release than this one (this release knows version 11). Run that release or a later one.Take a backup of the store's schema before you upgrade (for examplepg_dump --schema=agentrun <url>, with your own schema name if you set one), and to go back restore it and run beta.38 on the
restored data. Start the server first: a beta.39 runner started on a schema the server has not yet migrated refuses
to start, and so does a beta.38 runner on a migrated one, so upgrade the server, then the runners.- Two things to do after the upgrade. An agent saved on beta.38 with a monid grant has its pin in its revision's
body; beta.39 reads it from the new column and does not copy it across, so that agent's runs are offered no monid
endpoint, and none runs unheld, until you save the agent again (#2863). A webhook trigger made the old way for
Composio's events (signature: "composio") stops firing: delete it and make a Composio trigger for the same event
from Triggers → New app trigger (#2961).- Nothing else under the store changed shape.
Highlights
- A local inbox.
- The server reads and sends mail through one account of yours, over IMAP and SMTP with an app password, sealed like
every other secret (GET,PUTandDELETE /v2/settings/email, presets for Gmail, Fastmail and iCloud or servers
of your own; Outlook is refused for now, as it needs a Microsoft sign-in).POST /v2/settings/email/testsigns in
to both servers and saves and sends nothing. - An email trigger (
{kind: "email", mailbox?, every_s?, accept?}) starts a run for each new message itsaccept
takes, from the moment the trigger is made (mail from before starts nothing, and that is said), and shows its last
check. It turns itself off (sign_in_refused) after the mail server refuses its sign-in three times in a row. An
edit of itsacceptorevery_swhile its start is being taken keeps that start (#2963). - Five tools, on every harness:
email_search,email_read,email_draft,email_sendandemail_file.
email_sendasks a person by default and its approval shows who it goes to and what it says;email_filefiles a
message under a Gmail label or moves it to a folder and never deletes one. A run an email started reads and files
only that email's thread, and only the messages its sender took part in, unless the agent'sscope: "mailbox"
setting widens it. Each call reads as one line on the timeline, with no message body in it (#2874). - Settings → Email sets the account up: Gmail, Fastmail, iCloud or your own servers, where each app password comes
from, a Test that signs in to both servers and sends nothing, Save (the password is never shown again) and Remove
(#2912). - A template's eval can run an inbox agent on a fake mailbox seeded from a fixture (
email: fixtures/mailbox.jsonin a
case), never real mail: each run gets a fresh account per layout (Gmail labels or folders), starts as the package's
email trigger would, and graders see each email call and the mailbox after the run, with amailboxgrader for
where a message ended up.eval-template --on-data DIRruns the eval on a stopped server's data folder, so a
package's cases are measured on the owner's ChatGPT plan with every fake still bound (#2923).
- The server reads and sends mail through one account of yours, over IMAP and SMTP with an app password, sealed like
- No model key needed to start. Continue with ChatGPT, on the console's first screen and in Settings → Model keys,
runs Codex and pi on your ChatGPT plan. The sign-in is the person's who runs the server, finishes by itself on the same
computer or by pasting the sign-in tab's address on a server elsewhere, renews once across the server and every runner,
and signing out revokes every token it leaves behind. A run on the plan says its tokens and "Using ChatGPT plan", never
$0.00 or unknown, and a capped run on a plan model is admitted.npx, Render, Docker and Compose all start with no
key and end the same way (#2862, #2865, #2866, #2872). - A cap is never passed. Every model request is held at its worst case, its whole input priced uncached, so a cache
miss can no longer take a run pastmax_cost_usd. Theexpectedmode is gone with thecap_modesetting. A request's
answer is clamped to what the cap has left, parallel steps near the cap wait for each other instead of parking, and a
capped OpenAI-model run counts its input with a local o200k_base count, with no wait on OpenAI's count request
(#2871, #2920, #2935, #2947, #2950, #2956). - Free models. A run whose every request went to a free model (an OpenRouter
:freemodel, say) costs a known $0
and says so; the model picker marks free models and can show only them; a run on one stops at 60 requests, and is
told three requests before. Two free OpenRouter models that pi's catalog lacks are listed at $0 (#2879, #2883, #2900,
#2917, #2929). - Mode 2 on every vendor. An Archil run folder starts on Agent37 and Modal in any account from AgentRun's public box
image, and on E2B from a public template. Hosted computers take networkopenand allowlists through an egress
proxy, so a hosted desktop's browser and a sandbox'spip,npmandgitreach the web. E2B's system paths are
hardened against the computer's user (#2832, #2861, #2888, #2933, #2889). - A park keeps the computer. A desktop, phone or sandbox run that parks for a person keeps its computer for 10
minutes; an answer inside that time resumes on the same machine. A run that does resume on a new computer is told so,
and its live view says what it waits for instead of "ended" (#2931, #2936, #2906, #2915, #2926). - One cost figure. An ended run's cost is read from its booking, and a request cut with no answer under a cap is
booked at what the cap held, shown as "at most $0.97". An OpenRouter run books what OpenRouter charged, and a
provider's error inside a 200 stream is booked by what it reported (#2911, #2901). - Connect the Claude Code and Codex you already run. A local setup connects without copying its configuration or
credentials, runs through an outbound worker on another machine if you choose, imports selected history with consent,
and can be observed read-only. The client wraps existing Claude Agent SDK queries and Codex SDK clients to supply
checked notes and skills (#2837, #2838, #2839, #2840, #2841, #2842, #2843, #2844). - Learning you can replay and grade. The Learning tab has a Replay button on a workflow revision that is not
serving, a run's page has a Grade control, a replay scores a wrong grade against its grade, and "7x cheaper" is said
only when the replay backs the accuracy (#2823, #2907, #2825, #2858, #2877, #2921). - Triggers from the console. Composio triggers are made from a saved Composio server with no curl and no secret to
copy, and are now the one way to take Composio's events; a webhook's or GitHub trigger's filter is set and edited in
the console (#2873, #2857, #2961). - One codemode on every run. A desktop or phone run scripts its computer through the host's
codemode, each call a
gated, recorded nested call;computer_scriptis gone (#2892, #2925, #2909, #2932).
Breaking
cap_modeis refused (#2871). A run start (POST /agents/{agent_id}/runs,POST /runs) that sendscap_modeis
refused 422 with{"field": "cap_mode", "reason": "not_allowed"}, and so is an agent definition (create or update) that
sends it (definition.cap_mode,not_allowed).--cap-modeis gone from the CLI, andcost.overshoot_bound_usdand the
budget pause'sovershoot_bound_usdare gone from the run's view and both generated clients. Every run now behaves as
strictdid. An inline agent's version that first ran with a cap mode keeps it in its stored body, so running it
again with the field dropped is 409external_body_conflict: publish a new version. A run whose requests read the
cache spends less than it is held at, so it pauses earlier than underexpected: raise the cap by the cached share if
a run that used to finish now pauses.- The budget pause's
spent_usdis what the run spent (#2920), never what its running steps hold. The newheld_usd
says what they held when the step was refused, andsuggested_cap_usdisspent_usd+held_usd+ 10 ×needs_usd. A
client that readspent_usdas everything the cap had taken addsheld_usd. For kernel hosts:OPENAI_INPUT_COUNTER
is gone (#2950); aProviderPolicythat named it ascountInputdrops it and sets
countInputLocally: responsesLocalInputTokens. The kernel also dropsCapMode,RecentRequest,CapOptions.mode,
Admission.boundUsd,RefusalFigures.boundUsd,DECISION_OUTPUT_TOKENSandModelRequest.inputTokensCounted(#2871). - Capabilities carry more (#2865, #2879): the harness mark is
harness/5, each harness entry gainsneedsand
free_models, and a client that validates againstharness/4refuses them. DELETE /settings/chatgptanswers 200{"revoked": boolean}instead of 204 (#2872).- A stopped turn says so (#2960).
GET /v2/runs/{run_id}and its item inGET /v2/runscarry
pending: {"kind": "stopped"}besideinput_requiredandidleafter a person's stop, wherependingwas null. A
person's stop reaches the webhook asrun.needs_inputwithpending.kindstopped, once, as documented, where it
went out asrun.endedwith statusinput_required;run.endedcomes when the run ends. A turn its harness ended on
its own reads the same status and state withpendingnull, and itsrun.needs_inputcarries nopending(it said
stoppedbefore). Both generated clients gain the kind. - The old manual Composio webhook is gone (#2961). A webhook trigger with
signature: "composio"is refused422
undersource.signature(on create and on change), and one made that way before stops firing: delete it and make a
Composio trigger (Triggers → New app trigger) for the same event, which makes and keeps its subscription itself.
features.triggers.webhook.signaturesis["standard", "github"].POST /v2/triggers/{id}/secrettakes no body: any
body,{}too, is refused422.GET /v2/capabilitiesgainsfeatures.triggers.public_url, whether the server has a
public address (never the address). computer_scriptis gone (#2925). Usecodemode, whose script has the computer ascomputer; each call is held to
the run's gate (a shell command tobash, a started process tobash_background, a file written towrite, a file
read toread). An agent definition that namedcomputer_scriptnamescodemode. pi's own computer code mode, with
itscomputerglobal, is gone too (#2892). Codex and an MCP client of a claim's computer server call the step tools
andcomputer_op.- A ticked template tool is granted, not allowed (#2834). Apply writes no permission for it: under approvals
ask, a
ticked write now asks before it runs, unless you allowed that tool in Settings → Tools, and a ticked read runs
unasked only where Allow reads is on for its server.neverstill runs every granted tool. An agent applied before
keeps the allows apply wrote: remove thepermission: allowfrom those tools to have its writes ask. - A learned workflow asks once per call (#2903). Its approval card lists each MCP action set to ask, and one allow
lets them run; each MCP node is a call nested in the workflow's call. Removed from@parcha/agentrun-serve's modules:
host-calls.js,learnedAskFits,LEARNED_ASK_CARD_BYTES,learnedMcpTools(uselearnedMcpCalls),DraftAsk,
RunPermissions.decide,SpecGate.askedForandLearnedWorkflowRun.decide. - Monid (#2863, #2831). The kernel's
isCatalogNameisisCatalogWire, matchingmonid__where it matched
monid_. An agent's monid pin lives on its revision (see the note at the top). A workflow's monid grant is refused
("a workflow's nodes don't bind monid endpoints here yet; grant it to an agentic agent"). Endpoints priced by usage
can be granted again, with a warning, and one call may pass its estimate and the cap; the run stops when it does. - Computer providers (#2832, #2861, #2888, #2802).
AGENT37_BOX_TEMPLATE,MODAL_DEFAULT_IMAGE,E2B_BASE_TEMPLATE
andE2B_DESKTOP_TEMPLATEare gone from@parcha/agentrun-computer(ask the provider'sprepareImage()), the
descriptors have nodefaultImage, andbox/build-agent37.mjsis gone. A vendor on the shared adapter states only its
own facts:HostDescriptoraddscapabilities;HostFacts.gpu,gpuReadBack,vendorSockets,specFieldsand
runtime()are gone andlabelRulesislabels: { count, bytes };MachineRequest.networkisegress: boolean,
workdiris gone andlifetimeMsis always given;Machine.gpuis gone (a label records a GPU computer's GPUs); E2B
and Daytona refusecpusandmemoryMbin their own words; every child host closes the session buses in/tmpto
the computer's user; a Daytona VM-class computer is held to a VM runtime. - Desktops reach the public web by default (#2933). Under the default
AGENTRUN_BOX_ALLOW, a desktop isopen:
public internet only, through its egress proxy. Setnoneor a list to narrow it. A Docker computer onopennow
reaches public addresses only (no LAN, private ranges or metadata) and its one way out is its own egress proxy
container;openis refused on a Docker host that remaps user ids unlessegress.hostAddressesis given. No network
setting reaches a host or LAN service, and a private CIDR is refused at check. - Run cost fields (#2911).
so_far_usdandunpricedare no longer absent wheneverusdis known: a client that
took their presence to mean an unknown cost checksusd === null. A workflow request stopped while it streamed now
leaves the cost unknown, as on an agent run, where it used to store the cap's estimate. - Learning reads (#2858, #2893, #2921).
GET /runs/{id}/learning'sgradedis the grade asPUT /runs/{id}/grade
answered it, withfeedback_status. A replay's read nests its counts asoutcomes: {answered, handed_back, failed, skipped_write}in place of the four top-level counts andalone, and a promotion record'sreplay.skipped_writeis
gone. The learning cost CSV gains acost_upper_boundcolumn aftercost_usd, which shifts a script that reads it by
position. - A run in its end-of-run learning while its runner is upgraded ends failed ("the end-of-run turn is open, but the
job's end was never kept before it"), because the older runner never kept the job's end. Let running learning settle
before upgrading a runner (#2864).
Behaviour changes
- E2B desktops in Mode 1 now start from AgentRun's box template (Debian 13) instead of E2B Desktop's (Ubuntu 22.04); time to ready about the same (7.2-7.7 s vs 6.6-12.0 s).
- A park keeps its computer 10 minutes (#2931, #2936). The timeline says "The desktop is kept for 10 minutes while you
decide."; after 10 minutes the kept computer is ended, also across a server restart, and a later resume gets a new
one. A stop or cancel lets it go at once. A sandbox run is kept the same way.GET /v2/runs/{run_id}/computeranswers
a parked run's kept desktop asnonewithsession: "closed", thenendedafter the time passes. - A resumed run on a new computer is told (#2906, #2915). The ready line says it is new, the model's next request
carries one note that starts "AgentRun server note:", and acomputer_actthat names a ref or state the new computer
never gave is refused. A run that made its computer before this upgrade has no record, so it is not told. - Near its cap, parallel steps wait (#2920): eight subagents started at once under a $0.50 cap, or a burst of 25 fetches
and searches under a $4 cap, now finish where they parked atspend_cap_reached. A step waits at most 10 minutes for the
others to settle. - A run on a free model stops at 60 requests (#2900): the next request is not sent and the run ends
failedwith
spend_cap_reached;AGENTRUN_FREE_MODEL_REQUESTSraises it. ChatGPT-plan requests do not count. - OpenRouter (#2901, #2950): a request is booked at the cost OpenRouter states, and its routers (
openrouter/auto,
openrouter/free) have no price here, so a run with a spend cap on one is refused before it starts. - A template model that this server can't run is refused (#2886):
model.preferthat has no key, catalog entry or
harness here is 409agent_not_runnableundermodel, where it used to take the tier's model silently. A package can
offer up to three model choices (model.choices) and the owner picks one at install. The bundled code-review package
has noprefer, runs on the server's fast model, offersparticle(particle-1) as a choice, and its per-run cap is
$2.00, up from $0.50 (#2904, #2898). - A capped run on a ChatGPT-plan model is admitted (#2871). Its model requests are held at $0; a paid tool beside them
still counts and stops the run at the cap. A 429 for the plan's usage limit ends the run with "Your ChatGPT plan's
limit is reached. Manage usage in your ChatGPT settings, then run it again." - web_search stops asking a provider that said no (#2919): out of quota holds for the run, rate-limited holds for the
wait the provider named (60 s when none, at most 15 minutes). - A desktop or phone run gets its kind's skill again (#2941), so a desktop run's first request is larger. Through a
gateway, the smallest cap for a pi desktop run on Sonnet through five steps is now $0.09, the same as straight to
Anthropic: each request after the first is bounded by the last one's bill and the bytes it adds (#2956). - A run's pull request checkout is removed before its end is committed (#2942).
- A server note counts as told only once a request carrying it is answered (#2927).
- A run's date is told in its time zone (#2957): the system prompt's first line reads
Today is Sunday 15 November 2026 (America/Los_Angeles).where it said the UTC date, a day ahead in an evening west of UTC. A run a cron schedule
started takes that schedule's zone; every other run takes the server's (TZ). A claude-agent session told a later
day in its turn gets that day in the same zone.
Known issues
Known issue, Mode 2 (the run folder on your Archil disk):
- Works: sandboxes on Agent37, Modal and E2B, and desktops on Agent37 and E2B, in any account. The first computer in an account starts from AgentRun's public box image: Agent37 registers it once ("pulling" for about 15 s), Modal pulls it once (about 35 s), E2B starts from a public template in about 5-10 s. Shell commands start in /workspace, the run folder.
- Not yet: Modal sandboxes with GPUs (the Archil mount isn't proven on Modal's GPU runtime), and Daytona (no Archil mount yet).
- What to do: for those two, use another provider for Mode 2, or set the computer to none.
Known issue, spend caps: a cap below what the model needs for one step and an answer wraps up early; saving such an agent now warns.
Dependencies
- The cap counts an OpenAI-model request's input with a vendored o200k_base vocabulary file (3.6 MB, MIT, in
@parcha/agentrun-kernel, checked against tiktoken's sha256 pins); no tokenizer package is a dependency (#2871, #2950). @parcha/agentrun-serveaddssmol-toml1.9.1 (BSD-3-Clause), to read a Codex configuration (#2837).@parcha/agentrun-clientdeclares@modelcontextprotocol/sdkas an optional peer and gains the subpaths/learning,/claude,/codexand/workflows(#2841, #2842, #2843).- Local email (#2874):
@parcha/agentrun-serveaddsimapflow,mailparserandnodemailer, loaded only when the account is tested, a trigger checks the mailbox or a run is given an email tool, never at startup. The graph grows from 307 to 343 package/version pairs (about 11.9 MB in node_modules) and gains two license expressions:MIT-0(nodemailer 10.5.0) and(MIT OR EUPL-1.1+)(@zone-eu/mailsplit 5.4.30, under mailparser, used under MIT); see docs/dependencies.md.
What you can do now
- Run Codex or pi with no key: Continue with ChatGPT on the first screen, then your first agent (#2862, #2865, #2866).
- Give an agent an inbox: save the account in Settings → Email, add an email trigger, and grant the email tools (#2874, #2912).
- Pick a model when you install a template that offers choices, and turn on an optional schedule a package ships behind a boolean parameter (#2886, #2891).
- Set up a Composio trigger from Triggers → New app trigger, and set a webhook's filter in the console (#2873, #2857).
- Replay a held workflow revision from its card, or
POST .../revisions/{digest}/replays, and grade a run from its page (#2907, #2823, #2825). - Read a free model's mark in the picker and filter to free models only (#2879).
- Let
web_fetchlist the URLs a JSON, text or XML page writes out (#2938). - Fake the web, monid or the clock in a template eval case (
web:,monid: unavailable,now:), and gradetool_usedwithat_least: n(#2894, #2905, #2896, #2897). - Build agents from a brief or the full template catalog through a remembered coding-agent dropdown (#2846).
- Connect a Claude Code or Codex setup on this server or another machine in one onboarding step, and inspect an agent's
configuration beside its runs and versions in Mission Control, with no model key needed (#2847, #2848). Claude Code
and Codex sessions running on their own show read-only in Connected agents, Runs and Mission Control with their model
and token use (#2850), and the Learning Dojo shows each agent's retained knowledge and reviews and each run's learning
outcome and grades (#2849). - Name a run's opening message to a subagent by its file,
inputs/message.md, saved byte for byte in the run's folder on
every harness; the bundled code-review package hands a pasted diff to its reviewers that way (#2951).
Fixed
- Runs and servers:
- A run that answered and lost its runner's claim during its end-of-run learning ends completed with its answer; learning spend that passes the cap drops the learning as
no_budgetand never fails the run (#2864). - A server restarted after a crash during a harness install no longer fails the install, and a local eval installs Claude Code and Codex once, sweeping what a killed eval left (#2914, #2924, #2928).
- A run whose runner was lost and whose next claim ended before its computer started no longer leaves the lost runner's computer running until its deadline (#2926).
- A hosted create the caller gave up on, then refused by the vendor, no longer deletes another instance's machine (#2913).
- A scrubbed URL keeps an apostrophe inside it, so no query value after one shows (#2934).
- A stop that reaches a run while it parks on a tool approval, and whose end is refused, waits out its retry backoff
instead of being tried again at once (#2958).
- A run that answered and lost its runner's claim during its end-of-run learning ends completed with its answer; learning spend that passes the cap drops the learning as
- Spend and cost:
- A capped run whose model's tokenizer is o200k_base, reached over Messages (pi on GPT-6.1 Sol through OpenRouter), bounds each request by a local count: a $1.50 run no longer stops at $0.10-0.27 (#2935).
- A capped run's wrap-up note is priced on the same input count its request is admitted by (#2947).
- A capped Claude Agent or pi run on a Claude model through a gateway (Sonnet 5.5 through OpenRouter) no longer stops at
its second step with most of its cap left: a request that carries the last one unchanged plus new messages is
bounded by what that one billed and the bytes it adds, not by its whole body's bytes (about four times its tokens).
An inbox agent on claude-agent at a $0.10 cap now runs (#2956). - A cap below what a model needs for one more step and an answer says so with both amounts at wrap-up, and an agent saved with such a cap saves with a warning on
definition.max_cost_usd(#2935). - pi never sends thinking off to a model that must reason, and sends no
noneeffort to gpt-6.1-sol on the ChatGPT plan (#2918, #2890). - The model picker's free mark uses the kernel's free-model rule, the one that decides what a run books at $0 (#2929).
- A free request a crash cut is a known $0 marked
free(#2900). - The analytics page names a long MCP tool by its catalog name and counts CLI runs' tools (#2887).
- Tools:
- An MCP tool whose
mcp__<server>__<tool>is past 52 characters can be granted (Composio's Slack thread tool); every harness calls it by its name cut to 52 characters with a hash (#2881). - A tool call that ran beside a call waiting for approval keeps its result on pi, Claude Code, Codex and OpenCode (#2860, #2853, #2867).
- A tool row on the timeline carries the call's catalog name, not the harness's wire name, for runs recorded before too (#2821).
- In a
codemodescripttools.missingreadsundefinedfor a tool the run lacks; calls read under their script's row withparent_tool_use_id(#2909, #2932, #2940). - claude-agent tells its model that Skill is its tool; two watch templates went from 2 of 6 and 0 of 6 skill loads to 6 of 6 and 4 of 6 (#2869).
- The destructive-call 422 reads as one sentence that says what to send (#2853).
- The eval link check ends a URL at
|(a Slack-style<https://x|text>citeshttps://x), counts the URLs a non-web
tool's result holds (a calendar event's link), and decodes a feed's entities; the link check, a page's links and the
timeline's scrub read URLs through one module (#2946). It also takes a host and the same host withwww.in front
as one (https://www.cpsc.gov/xmatches a call that fetchedhttps://cpsc.gov/x); no other subdomain is dropped
(#2959). - A Codex code-mode
execthat ends ok after a call parked on an approval and then ran no longer reads "cut short" on
the timeline, and an exec the run's end closes reads "interrupted" with no status, never failed (#2953). - keyless Exa's daily count running out turns web_search off for the rest of the run, not just busy for 60 s (#2948).
- A web_fetch or web_search answer keeps its
linksand still parses after the run's redactor cuts it: the redactor
cuts each string of its JSON, never the whole line, so the link check no longer reads a cited link as "no matching
call" (#2962).
- An MCP tool whose
- Computers:
- Learning: a run its learned workflow answered keeps path
workflowwhen the agent restates the JSON in prose (#2854); a replay scores a wrong grade against its grade (#2877); a rerun call keeps its earlier attempt's guard rows (#2939); a compiled workflow's answering step runs on the agent's own model (#2895); one replay scorer (#2833); a grade's words are told once (#2856); a replay needs 5 graded inputs (#2893).
Also in this release
- Computers: the computer protocol overview for vendors (#2835), PROVIDERS.md follows the vendor's create (#2916), the computer tests remove a live-Docker test's containers (#2922), the shared host-facts adapter (#2802).
- Learning: a workflow checks in code that its answer's quotes are in the pages it fetched (
quote_missing), and a compiled workflow that quotes fetched pages must use it as its quote check, never a judge (#2937);softEscalates(doc)in@parcha/agentrun-dsl(#2852), a held revision never re-promotes by tying its own wrong answers (#2877), the Learning tab's cost claims (#2921), the replay API and button (#2823, #2907), grade tidy (#2856, #2858). - Console: consistent action controls and clipboard paste into a secret field (#2845), the trigger filter editor (#2857), Composio triggers (#2873), and the console reads a monid grant's name with the server's own tool-name grammar instead of a copy (#2880).
- Templates:
trigger.enabledby a boolean parameter (#2891), model choices (#2886, #2898), a one-page tools guide for template authors (#2829), ticked grants (#2834), template tools in the console's handoff (#2846). - Evals: a case's
now:(#2896),web:(#2894),monid: unavailable(#2905), anllmgrader that sees each call's arguments, the first 8,000 characters of its result and every url a web call answered, and a link check against the run's web calls (#2870, #2899, #2930),tool_usedat_leastand a monid endpoint by name (#2897, #2859). - Docs: present AgentRun Cloud first in getting started (#2752), the deployment guide's one
docker runline and the ChatGPT sign-in (#2866). - Dependencies:
smol-toml; see Dependencies. - Tests and CI: a deterministic test of a workflow's lost-claim answer (#2855); the pi park-siblings cases measured from the model's ask (#2868, #2878); a ChatGPT sign-in test that survives another process holding port 1455 (#2875); a late unanswered claim's give-back test ordered by the runner's calls (#2876); two CI flakes fixed (#2882); CI shard timings refreshed and a seventh any shard (#2884, #2885, #2910); three stop-withdraw tests skipped for their 30 s wall-clock wait (#2908); the replaced-sandbox test waits on the runner's release (#2945); the lapsed-keep sandbox test holds the new sandbox while the run's row reads released (#2949); a live-Docker computer test's containers are removed (#2922); a two-runner lost-claim test that timed out under load is deleted (#2836); an email search that matches nothing answers exactly
{messages: []}(#2964).
Known limits
- The learning loop is further along but not finished: a learned workflow that answers wrong with confidence is not caught by the money rule.
deploy/compose.yaml(a server and runners) has no Docker socket, so its agents get file tools andweb_fetchbut no shell; for a Docker sandbox, use the rootcompose.yaml, run the server on the host, or use a Daytona or E2B computer.- The other limits listed for beta.32 still hold.