Repository navigation
Changes since v16.0.0.
Compatibility notes
Map overflow control (#1504)
Maps no longer expose the overflow group control or its ok|overflow labels. Update custom graphs that read this control before upgrading.
maxItems now limits active complete item bodies, rather than input length. Maps process the full input unless authored control flow stops them. For example, three input items with maxItems: 2 now run with up to two item bodies active; previously the map ran no items and reported overflow. Graphs that need an input-size limit must check it explicitly before entering the map.
Features
Process complete maps with bounded concurrency (#1504)
Process the complete ordered input of a map. maxItems bounds active
complete bodies; optional maxConcurrency bounds running leaf and
auxiliary work (default 64, maximum 65,536). Count visits and loop
entries independently of retained outputs, and retain capacity through
cleanup.
Settle joins from durable history through one reducer for dispatch and
replay. Preserve pending candidates and conditional writes, use the
current loop round's facts, and discard closed-group frontiers before
later work. Scoped static proofs preserve correlated outcomes and
implicit failure/presence observations while saturating counts at
authored thresholds.
Carry the same execution contract through templates, SDK, numeric
editing and queued/live history. Remove overflow routing and
workEnforcement; retain the structural-bounds format. Validate
selected or replacement profiles while preserving unrelated opaque
entries.
Support suspension, waits, and vector records (#1573)
Add checkpoint suspension, successor continuation, the built-in wait
worker, and ordered vector records across local/direct targets, CLI,
SDK, protocol, and UI.
Fix the coverage-job race by atomically publishing the Git cancellation
fixture’s complete PID. Preserve all cleanup assertions and earlier
repairs.
Carry agent conversations across checkpoints (#1574)
Checkpoint and suspended-run successors retain captured agent
conversations, with private checkpoint storage, safe fallback and
unchanged restart behavior.
Validation passed: 445 Rust test executions, 467 UI tests, 159 tooling
tests, 972 pinned-matrix tests, formatting, Clippy, Rustdoc, protocol
generation, CLI docs and strict MkDocs. Four accepted real Codex/Claude
recall checks and strict validation of 2,788 coverage profiles were
reused after verifying source and binary identity.
Pending: required merge-queue/full coverage checks, privileged root
checks, fresh Windows/macOS CI, external integration settlement and
AC-10 owner checks.
Co-authored-by: Zeroshot delivery@zeroshot.invalid
Ask for human input with request nodes (#1576)
Graphs can suspend at request nodes, accept validated answers in
successor runs, and route expired requests through timeout. Includes
inbox commands, target support, SDK/protocol integration, UI details and
documentation.
Adds eight regression tests covering local answer refusals, read-only
inbox behavior, hosted history restoration and target capability checks.
Support local worktree submissions (#1586)
Expose worktree and worktree_base on RunRequest and string-based
Client.run()/submit(), preserving native defaults and preflight
validation. Document base selection, retained worktrees, and cleanup.
Record when each run event was stored (#1597)
Every stored run event now records when the ledger stored it, so run
views can show how long steps took and when things happened. The time
sits beside the event's cursor (recordedAt, Unix milliseconds), never
inside the event itself, so replay and the reducer stay time-independent
and order is still defined by the cursor alone.
Add durable run steering (#1584)
Add durable run steering through the CLI, direct targets and Python SDK,
with conversation receipts and full vector readback. Preserve SQLite
decoder compatibility.
Show human requests in run history and the graph (#1598)
Run views now show when a run is waiting for a person, what it asked,
and how the question was resolved. They also read the work a continued
run carries over from the run it continued. Before this, a paused run
looked idle, and a continued run's carried-over work appeared as
unlabeled events.
Stacked on #1597 (ledger-recorded times); review that first.
Size. About 825 of the added lines are hand-written: 354 of app
code, about 250 of tests and 225 in the example generator. The other
~8,160 are ui/public/history-examples.json, written by
ui/qa/generate-history-examples.py. This PR marks that file as
generated in .gitattributes, so GitHub collapses it in the diff.
Add a breadcrumb, one inspector header and simpler canvas toolbars (#1599)
The canvas and inspector chrome, shared by the run view and the profile
editor:
- One inspector header with a breadcrumb. Both inspectors open with
where the node sits in its graph, each enclosing group a link up the
hierarchy, then its title and a close button. Before this there was no
way up from the run inspector, and the editor had a separate "In
‹group›" button that named groups differently from the canvas. - Simpler canvas toolbars. The run graph's "Run overview" and the
editor's "Run settings" buttons are gone: the workflow's own details
open from the inspector button or, while the inspector is open, from the
empty canvas or the breadcrumb's first link. "Add activity" leads the
editor's toolbar, both canvases gain "Expand all" beside "Collapse all",
and the status bar says "1 activity".
Stacked on #1598; review that first. This PR also absorbs #1600
(toolbars), which was closed to keep the stack shorter.
Slim the replay bar and share three layouts at every width (#1601)
Sizing and responsiveness for the workspace:
- A slim replay bar. The replay bar under a run is one quiet row
instead of two tall ones: small step controls, a thin track whose
replayed part is ink, the event being shown, then the live status and
"Show node". Replay is a tool to reach for, so it no longer competes
with the run for attention. - Three shared layouts. Every page changes layout at the same two
widths, 1150px and 720px, instead of a dozen page-specific breakpoints.
Medium puts the sidebar behind a menu button at the header's left edge,
the side it opens from; narrow wraps the profile header and slides the
inspector over the canvas, below its toolbar. The top bar keeps one
height and both sidebars share one width. Canvas toolbars keep one row
by following the canvas's own width, folding the editor's quieter tools
into a "More tools" menu.
Stacked on #1599; review that first. This PR also absorbs the
shared-layouts change to keep the stack shorter.
List and remove named targets (#1619)
- Add secret-free named-target listing and removal to the CLI and typed
Python SDK, with updated documentation. - Safely clean credentials and resume permissions across concurrent
operations and Linux store switches, preserving unrelated targets and
retry state.
Add authorized node steering networks (#1611)
Agent nodes can send peer guidance within the same graph through steer --network. Graphs declare every network member and explicitly authorize
publishers. Messages support once and standing, reach recipients
that start later, deduplicate overlapping networks, and exclude the
originating node instance.
CLI and local RPC calls enforce publisher mandates even after an agent
clears its execution labels or targets another local run. Native
execution keeps its existing OS permissions.
Add isolated read-only context repositories (#1657)
Add up to 16 immutable, same-forge context repositories to Docker-target
and Platform runs. Keep full-history checkouts read-only and isolated
outside workspace checkpoints and delivery. Preserve recovery pins and
omitted-empty compatibility, and reject environment overrides that
remove exact-path Git trust.
Reconciles
#1602 with
current main. Network regressions also confirm durable publications
after lost responses without retrying.
Fixes
Scroll only the inspector when a run is embedded (#1578)
When a run is embedded (for example on a hosted run page) and the window
is wide enough for the inspector to sit beside the graph, a long
inspector scrolled the whole frame, graph included, instead of just the
inspector panel.
Rediscover restic when resuming runs (#1581)
Rediscover Restic when restoring checkpoints and creating resumed
checkpoint stores, retaining the recorded executable as fallback.
Preserve storage compatibility, conversation restoration, and
executable-independent checkpoint pinning.
Bound local controller requests (#1580)
Bound local controller replies and subscription setup to ten seconds.
Lists share one deadline and retain recorded status for unresponsive
controllers. Force-stop timeouts identify the run, controller PID and
queued effect. Established quiet streams remain open. Keep coverage
profiles outside conversation checkpoint workspaces.
Match the workspace UI to the zeroshot.sh paper look (#1583)
Brings the workspace UI (profile editor and run viewer, standalone and
embedded) in line with zeroshot.sh's hand-marked paper look, and adds
ui/DESIGN.md as the UI's visual source of truth.
Handle unavailable branch rules on private Free repositories (#1589)
Allow PR and merge delivery when GitHub's branch-rules API returns its
known Free-plan entitlement denial. Recognize only a complete standalone
response before redaction, preserving other errors, pagination
enforcement and existing policy gates.
Include provider causes in retry logs (#1590)
Include native causes in Claude, Codex and Pi retry progress. Redact
before bounding lines to 1024 bytes, preserving failure classification,
continuation and final diagnostics. Add portable regressions and macOS
CI coverage.
Submit final answers through native structured output (#1609)
Direct Claude initial and correction answers through its native
StructuredOutput tool once. Preserve response contracts, validation,
sessions and other harness prompts.
Leave default stage assignment to the support service (#1615)
Stop adding triage automatically when an issue is opened or reopened.
The label workflow continues enforcing maintainer and service decisions,
preserving approved stages and manual holds without introducing a second
stage.
Harden suspend, requests and steering against end-to-end failures (#1610)
Fix the defects an end-to-end stress test found in suspend and resume,
request nodes, conversation checkpoints, steering and the run viewer: an
interrupted answer or resume no longer blocks the workspace,
resume and answer work right after a suspension, concurrent steers
no longer refuse each other, steers are no longer lost at Claude Code
subagent or oversized tool calls, Pi conversations are restored on a
target, and a steer keeps standing on Copilot, Pi and OpenCode. Two
limits found on the way are lifted: request nodes with different answer
shapes can share a graph, and a failed run restarts where it started, so
an answered question is not asked again.
Name files in connection and target store errors (#1618)
Name files in connection-store and target-registry errors. Connection
decode failures include safe parser locations; target failures retain
parser detail and identify invalid records or unsupported versions.
Preserve whole-store reads and credential redaction.
Propagate parallel branch terminals (#1622)
Propagate parallel-branch fail and succeed outcomes when they
precede or tie join satisfaction, preserving the selected result and
cancelling siblings through existing cleanup. Align workflow authoring,
history replay and documentation.
Stop repeated unchanged push rejections (#1592)
Stop delivery after the same branch, revision and ref-status reason are
rejected twice, preserving work for checkpoint retry.
Normalize HTTP fixture sockets and cover fragmented requests in macOS
CI.
Skip unreadable runs during submission scans (#1620)
Allow local submissions and resumes to skip unreadable run directories,
including directories whose child metadata remains accessible. Preserve
cleanup and duplicate/conflict checks for readable runs, and name failed
paths in reported scan I/O errors.
Strengthen default software-change reviews (#1631)
Strengthen the default software-change reviews while avoiding repair
loops caused solely by unavailable checks.
- Prepare documented prerequisites and shared candidate builds, and send
both reviewers freshreviewContexton every worker visit. - Check explicit requested behavior and trace concrete failure scenarios
through affected callers. Require source or execution evidence for
material defects or unmet requirements; keep unavailable checks,
unrelated failures, and optional improvements advisory. - Separate required fixes from advisory notes and leave shared runtime
setup guidance neutral about verdicts. - Preserve PR-thread response verification and ship delivery's CI,
conflict, and merge-policy checks.
This is a non-breaking update: caller input, delivery contracts, and
saved custom graphs are unchanged.
Preserve early CI failure context (#1632)
Preserve early CI failure context and final summaries through bounded
GitHub/GitLab feedback. Carry raw failure positions alongside redacted
text so credentials cannot hide failure markers during capture or
aggregation.
Retain complete redaction, cleanup exclusion, older GitHub CLI
compatibility and existing limits. Add portable regressions and update
repository guidance.
Fail unchecked worker errors (#1627)
- Fail final unchecked step/verifier errors with
unhandledafter crash
retries, preserving authored recovery guards. - Retain named diagnostics, failed nested-group history and parallel
cleanup. - Cover built-in delivery modes, documented review-loop behavior and
checkpoint recovery assertions.
Accept exited process groups during macOS cleanup (#1633)
On macOS, a command no longer fails with "process group termination
failed: Operation not permitted" when the last members of its process
group had already exited. The delivery command tests become reliable on
the macOS runners.
Send steers as user messages and keep lessons out of tool results (#1621)
A steer is the operator's own words, so at a tool call it now reaches
the running conversation as a user message through the harness's own
message channel. It is no longer placed inside a tool result (Pi,
OpenCode), which models are trained to distrust, or in hook context that
the harness does not keep (Copilot). Tool-call Active Memory lessons
leave the tool result too, and Copilot is given a lesson again in a
later turn.
Includes current main through #1633 (44571e74). The merge preserves
#1611's steering-network contracts.
| Harness | Steer at a tool call | Received when | Tool-call memory
lesson |
| --- | --- | --- | --- |
| Claude Code | user message on stream-json stdin | Claude Code replays
the message (--replay-user-messages) | hook context (unchanged) |
| Copilot | session.send with mode: "immediate" | user.message
session event | hook context, given again in a later execution |
| Pi | steering queue (pi.sendUserMessage, sent by the extension) | Pi
reports the user message in its event stream | session entry added at
the turn boundary |
| OpenCode | session message stored by the plugin (noReply) | the
plugin acknowledges after storing it | stored session message |
| Codex, Hermes | hook context (unchanged) | hook returned | hook
context (unchanged) |
The trigger is unchanged: a steer is picked up at the conversation's
next tool call or goes with its next prompt.
Decide conditions by search instead of enumerating every outcome combination (#1638)
Graph verification no longer refuses a valid graph because of how much
work its control analysis takes. The organisation's default graph
zeroshot-change, which main refuses with control analysis exceeds 65536 states or 8388608 count-cell work units at the graph root, is
accepted again. Saving it takes 11 ms; v16.0.0 took 0.24 s, and main
needs 1.8 s to refuse it. Adding a reviewer or a signal to a graph now
adds to the checking work instead of multiplying it.
The verifier decides each condition by search over the controls that
condition reads. The work meter shared by the whole verification is
gone. What remains is a backstop per check, reported at the choice, join
or guard that could not be decided.
Skip syncing temporary workspace copies (#1636)
Skip per-file and per-directory syncing of temporary workspace copies
before Restic reads them. Preserve complete capture/restore contents,
metadata/catalog durability, and Restic completion before catalog
publication. Add a portable cancellation regression selected by Windows
and macOS CI.
Refs #1634
Report submission and node input size limits (#1629)
Report oversized submissions before local run creation with byte counts
and stable submission.too_large diagnostics, preserved by Python
InvalidRequestError. Oversized node starts produce descriptive
runtime_failed results in foreground output, status and logs, with
active sibling cleanup.
Preserve the existing 2 MiB admitted-definition and 1 MiB node-start
limits, including bookkeeping, and valid mapped inputs above 1 MiB. Add
regression coverage for submission preflight boundaries.
Expose recorded execution timing in status and watch (#1641)
Expose optional ledger-recorded startedAt values for active executions
and recordedAt values on watch notifications, including hosted
observation and Python SDK support. Preserve legacy unknown times,
execution identity, cursor ordering, import semantics and read-only
inspection.
Record bounded native tool activity (#1623)
Record native tool calls and completions for Codex, Claude, Copilot and
Pi in existing logs and node transcripts. Preserve native
identity/status with credential redaction, bounded 8 KiB excerpts and
duplicate suppression.
Keep Claude tool completions separate from prompt replay
acknowledgements, and preserve Pi user-message acknowledgements
alongside tool activity. Document Claude’s 60-second post-result
acknowledgement deadline: an unconfirmed steer receives no receipt and
remains due for a later prompt.
Remove unadmitted local run directories (#1628)
Remove newly created local run directories when startup fails before
admission, using the existing lease and ledger checks. Preserve original
diagnostics, admitted runs, existing leftovers and resume behavior.
Fold independent parts of a map item separately (#1640)
One guard can now read many independent nodes of a map. The verifier
folds the parts of a map item that no rule ties together separately, so
such a guard costs about the sum of what it reads instead of the
product. Before, a guard that read nine reviewers of one map item was
refused with ceiling_exceeded; it now takes about 1,900 search steps
of the 1,048,576 a check may use.
This is the follow-up to #1638, which is merged; the branch is rebased
on main.
A second commit fixes a cost regression that #1638 put on main and
that the automated review of this PR found: listing the branch sets a
wide join can complete with was not charged. See "Join listing" below.
Report failures in logs and foreground output (#1642)
Report bounded Active Memory preparation, specialist and confirmed hook
failures through existing logs and foreground output while preserving
graph outcomes. JSON decoding diagnostics retain the database path,
category and location without exposing stored values. Finished runs
report an incomplete memory-log check with the full-log command; an
ongoing resumed run keeps its normal bounded read quiet.
Allow checkpoint discard for finished local runs (#1639)
Allow finished local runs to discard lineage checkpoints while
preserving the checkout, run history and delivery state. Expose Python
Run.discard_workspace() and preserve checkpointless resume and request
continuation, with actionable refusals for active runs and interrupted
restores.
Preserve outer-item dependencies in nested maps (#1651)
Preserve the actual map ownership of implicit control dependencies so
nested-map verification rejects incomplete choices and accepts reachable
branches. Add admission and reducer regressions covering mixed and
uniform outcomes, empty maps, multiple inner items, and shared controls.
Retain styled compiler errors in CI feedback (#1653)
Preserve early ANSI-colored compiler diagnostics and source locations in
GitHub and GitLab CI feedback. Keep original byte offsets, redaction,
excerpt bounds and final exit summaries.
Enforce the shared local preparation deadline (#1655)
Prevent local runs and checkpoint resumes from admitting after their
shared preparation deadline. Bound version checks and runtime
preparation, await helper cleanup, and check expiry before durable
admission while preserving checkpoint recovery and retry.
Limit Windows test and doctest compilation to one Cargo job to reduce
peak memory use.
Explain why local runs cannot be resumed (#1617)
Explain local resume refusals with specific run-state, recovery-record,
workspace and claim diagnostics. Missing-workspace guidance
distinguishes suspended runs and honors explicitly discarded
checkpoints. Preserve structured request-error classification and
read-only status/list behavior. Publish delivery conversation fixture
state atomically during recovery polling.
Retain tool activity when assistant steps reuse IDs (#1654)
Preserve later Pi tool calls and failed completions when assistant steps
reuse a native tool ID. Reset activity deduplication at each
assistant-message start while retaining duplicate suppression within the
step. Add fragmented-stream and durable CLI log/transcript regressions.
Make the Restic discovery fixture tolerate brief Windows access/sharing
failures when deleting the retired executable, with a bounded wait and
the final I/O error in failures. Restore still requires the old
executable to disappear first.
Compile Windows workspace test executables before starting the native
execution host and preserve compiler settings across WMI so that the
host reuses those artifacts. Keep its thirty-minute execution bound,
every test and all doctests. Include the full CLI response when a
positive network-publication test call fails.
Make baseline mismatch review guidance neutral (#1656)
Make delivery baseline-mismatch guidance neutral while preserving the
existing repair/review round. Add regression coverage for unchanged
targets with unpushed local work across four PR/merge modes.
Hold a started receipt write's reservation until its outcome is known (#1658)
Successor to #1606 (original work by Anton Emanuel; the original PR and
branch are preserved). It repairs the steering receipt publication race
on top of current main, keeping main's provider acknowledgements and
peer steering.
- A steering receipt append now runs in its own task. The callback stays
bounded by its deadline, but the deadline no longer abandons the write. - A started append keeps the conversation's reservation until the ledger
reports its outcome. This stops callbacks from re-offering the steer
while a commit may still land. - On commit, the receipt is applied and the reservation cleared under
one lock, so there is a single canonical receipt. On a confirmed failure
the reservation is released and the steer can be offered again. - Once is held per node instance and Standing per provider conversation.
Snapshots reference a pending receipt, and it counts only once
committed. AGENTS.mddocuments the write lifecycle.- Regression tests cover a held write past the callback deadline with a
late commit, confirmed failure and retry, cancellation, and network Once
and Standing behavior.
Recover summaries and document map compatibility (#1659)
Restore release-note generation for 16.1.0 by recovering the missing
summaries of two immutable README squash commits. Put a compatibility
notice before the feature list explaining the removed map overflow
control, the new meaning of maxItems, and the updates required for
affected custom graphs.
Documentation
Replace the tabloid opener with one plain argument (#1575)
The README opening is now one argument instead of a hook line plus two
loose questions.
- First paragraph says what goes wrong as an agent's context fills up,
even when it is told to check its work. - Second paragraph is the line "The agent that wrote the code shouldn't
be the one that says it works", linking the review-loop benchmark. - Third paragraph says zeroshot runs the team's process as a graph the
engine enforces, with who writes, who reviews, what must pass and where
failures loop back. - "Do you still look at the code?" is no longer in the README, and
zeroshot.sh still uses it. - The "Why not a skill?" and "Do loops work?" lines under the example
graph are removed, since the benchmark link carries that evidence.
Checks run locally are prettier --check README.md and node --test tests/tooling/docs-agent-contract.test.js (19 pass).
Co-authored-by: Eivind eivindcovibes.ai@Eivinds-MacBook-Pro-5.local
List plans in README features (#1577)
Adds a Plans bullet to the README Features list, linking the planning
guide: plan this with zeroshot turns a big change into small PRs, each
with the tests it must pass.
Co-authored-by: Eivind eivindcovibes.ai@Eivinds-MacBook-Pro-5.local
Say local runs use your agent subscription (#1593)
- Adds "Runs locally on top of your agent subscription." under the top
install command, so the first screen says you can try zeroshot without
an API bill. - Hosting now says local runs use your Claude, ChatGPT or Copilot
subscription, next to self-hosting and Platform with API keys.
Co-authored-by: Eivind eivindcovibes.ai@Eivinds-MacBook-Pro-5.local
Say it runs locally on your agent subscription (#1595)
Add the README tagline that local runs use your existing agent subscription.
Drop duplicate subscription line from tagline, say "your process" (#1596)
Remove the duplicate subscription sentence from the README tagline and describe the graph as your process for solo users too.
Maintenance
Freeze the stalled controller between ledger writes (#1630)
Make the test
unix_controller_stalled_requests_return_without_killing_the_controller
reliable. It failed in 4 of the 52 macOS merge-queue runs that contained
it and twice removed an accepted pull request from the queue. Only the
test changes; no product behaviour changes.
Update Prettier to 3.10.1 (#1608)
Update the repository's Prettier development dependency from 3.9.10 to 3.10.1, including its lockfile.
Run only the README check for README-only changes (#1594)
Run only root README formatting in local pre-commit and pre-push hooks when the shared CI classifier proves a README-only change. Keep the full tooling suite for mixed, empty or unavailable history; a missing nonzero remote commit always selects full checks, and only new branches may use the merge base with main. Direct main pushes and release tags remain blocked.
Align live recall assertions with provider homes (#1652)
Correct live recall assertions to validate worker session reuse and
verifier isolation independently of durable execution IDs. Add offline
regressions for sparse IDs, incorrect reuse, wrong scopes/nodes, and
missing callbacks while preserving production semantics.
Upgrade
npm install -g @the-open-engine-company/zeroshot@16.1.0Versions 10.4.0 through 10.8.0 cannot update themselves: zeroshot update fails with release archive contains unexpected entry restic (restic.exe on Windows). Upgrade those installations once with the npm command above.