Repository navigation
Changes since v15.0.0.
Breaking changes
Publish working branches and add opt-in worktrees (#1541)
- Publish local deliveries from the checked-out branch, with opt-in
worktrees and attached resume. - Document the breaking delivery behavior.
- Stabilize the OpenCode recall test by checking receipt bytes directly,
preserving payload assertions and production deadlines.
Features
Connect consented website journeys (#1501)
Connect documentation visits to the consented website journey using the
shared analytics entry served by zeroshot.sh. The entry only activates
on the zeroshot.sh origin and reuses the website consent banner and
PostHog session.
Fixes
Verify live manifests after Pages deployment (#1500)
A successful GitHub Pages deployment can leave the public site serving
an older artifact. The v15.0.0 release exhibited this: its published
branch and uploaded artifact contained the correct manifest, while the
public URL returned 404.
Verify the selected live manifest after the official deployment. Require
the validated docs version, product version and source commit, with
explicit JSON null for Current's product version. Bounded HTTP polling
runs under a ten-minute cap; stale or unavailable content fails
publication.
Report tool permission denials in run logs (#1503)
Claude Code reports, in each result record, the tool calls that its
permission settings or hooks
denied (permission_denials: [{tool_name, tool_use_id, tool_input}]).
Zeroshot ignored that field.
An unattended Claude node whose writes were refused looked exactly like
one that ran normally:
zeroshot logs had nothing about the refusals.
Now each Claude result with denied calls writes one error-level line to
the node's run logs and
retained output history. The line gives the denied tool names and
counts, for example:
Claude Code denied permission for 3 tool calls: Write (2), Bash (1). Denied calls do not change the node's outcome.
This change only adds diagnostics, as agreed for #1376. A task that
completes after some denied
calls (for example 3 of 10) still succeeds with its structured output
and usage unchanged. The
graph's verifiers, not denial counts or the permission mode, judge
whether the work was done.
Failures, retries, cancellation, usage accounting and permission
selection are unchanged.
Tolerate empty git file placeholders (#1490)
In combination with uv some project can include an empty .git file
inside .uvcache. While that is irrelevant to zeroshot it still
triggers an error in the checkpoints filesystem git discovery logic.
Discard every lineage catalog on local force-stop (#1524)
Force-stopping a local run deletes the checkpoint repository shared by
its whole recovery lineage. Until now it removed only the stopped run's
own catalog. The original run and earlier successors kept listing
checkpoints whose snapshots were gone, and resuming from one of them
failed with an I/O error.
Local force-stop cleanup now reuses discard_lineage, the same cleanup
successful runs use. It removes every catalog that names the shared
repository, then the repository, so no run in the stopped lineage lists
checkpoints anymore. Other lineages keep their catalogs and repository.
Force-stop stays destructive for the lineage. As before, a removal
failure keeps the repository and is reported as unavailable cleanup.
Part of #1368. This covers only the force-stop listing defect. Plain
resume preserving the workspace and the --discard-workspace-changes
consent for selected checkpoints are handled separately.
The report suggested keeping the other runs' checkpoints restorable, or
else dropping their listings. Following the agreed scope, force-stop
keeps deleting the lineage's checkpoints and drops the listings; no
checkpoint retention is added. Direct-target force-stop cleanup is
unchanged.
Accept invalid Unicode in provider output records (#1525)
Accept provider JSONL records containing invalid UTF-8 or lone surrogate
escapes, replacing invalid characters with U+FFFD through shared
decoding. Preserve malformed-record handling and redaction after
decoding. Add regressions beside Claude, Codex, and Pi.
Retry merge request creation during branch indexing (#1526)
Retry GitLab’s HTTP 400 missing-source-branch creation rejection within
the existing review-sync batch, rechecking the published branch and head
before each attempt. Preserve terminal handling for other validation
errors, at-most-once handling for uncertain responses, and the original
reason and branch on exhaustion.
Validation: 297 source-provider tests and 232 delivery tests passed; one
existing matrix test remained ignored. CLI build, version smoke check,
all-targets Clippy with warnings denied, and formatting passed.
Closes #1505
Co-authored-by: Zeroshot delivery@zeroshot.invalid
Show recorded failure causes in failed turns (#1528)
Failed ACP turns now show the existing bounded foreground failure
summary and logs command when recorded context is available. Messages
without context and the documented rawOutput shape remain unchanged.
Added regression coverage and updated guidance.
Closes #1510.
Preserve run cache totals when script steps execute (#1530)
- Document every
metadata.tokenUsagefield, harness-reported
summation,complete, and omitted cache counters in
docs/guides/observe-and-control.md. - Record explicit zeros for all four counters in
zeroshot/src/native_v2_candidate/script.rs. - Extend
script/tests.rsto verify zero counters and preserve agent
cache totals in a mixed run. Missing-counter aggregation remains
unchanged.
Fixes #1515.
Fail early on unwritable installs and show npm recovery commands (#1529)
- Check installation-directory writability before downloading assets for
a newer release, preserving the current-version no-op. - Add shared npm recovery commands to permission failures, archive
rejections, and release-lookup HTTP 403/429 failures. Known releases use
their exact version; failed lookups use@latest. - Document npm’s retained package version after in-place updates and the
archive-layout compatibility contract.
Closes #1509.
Explain recovery after local history rewrites (#1533)
When a worker rewrites published local history, delivery now names the
branch, published head and local head, then explains checkpoint listing
and resume recovery. The terminal refusal and existing remote-history
refusal remain unchanged.
Extend shared-transport regression coverage to verify recovery guidance
and preservation of HEAD, workspace state, repair files and remote refs.
Add one sentence to docs/guides/forge-delivery.md.
Settle completed turns before process exit (#1531)
Preserve completed Codex answers and usage across late exits, signals
and stalled teardown. After a 250 ms grace, use bounded release with
confirmed process-tree cleanup and log teardown problems once. Preserve
existing cancellation and failure behavior.
Include failed pipeline jobs and log excerpts (#1535)
Failed required GitLab pipelines report failed job names, stages and
reasons, configuration errors, and bounded, redacted log excerpts.
Optional-pipeline handling is preserved.
Merged main at 7c7ef768, preserving published commit 2e0c772a. Fixed
the API-path mismatch in GitLab api.rs and added a regression in
diagnostic_tests.rs, retaining merge-request creation retries.
Validation passed: 357 Rust regressions, 158 tooling tests, CLI
build/smoke check, workspace Clippy, formatting, lint and distribution
checks.
External platform matrix, coverage and release/image CI remain pending.
Fixes #1508.
Closes #1508
Co-authored-by: Zeroshot delivery@zeroshot.invalid
Record Zeroshot and Claude versions in run logs (#1534)
Record the Zeroshot version first in each run’s logs, including resumed
successors, and Claude’s accepted CLI version once per run. Release
Claude’s run-owned version bookkeeping when runs close while preserving
ACP sessions.
Distinguish failed runs and permanent refusals (#1532)
Return run.failed with details.runId for failed runs, and
request.invalid for unknown targets and registry resume refusals.
Preserve transport errors as target.unavailable. Update SDK mappings,
regression tests, documentation, and AGENTS.md.
Name the program in spawn failures (#1542)
Include the program in spawn failures as an escaped preview bounded to
256 bytes plus a truncation marker. Preserve OS details without adding
arguments or environment values.
Add regression coverage for script log replay, waiting for terminal
failure before controller exit and using the fixture’s existing CLI
timeout.
Name GitLab objects and published state in refusals (#1539)
Name projects, merge requests, branches and revisions in delivery
refusals. Distinguish closed and merged reviews, explain recovery,
report published state, and classify merge-only 401 refusals as missing
permission.
Preserve credential redaction and diagnostic bounds, with focused
regression coverage.
Keep the checkout on plain resume and require consent for checkpoint rollback (#1538)
- Plain local
zeroshot resume RUN_IDrestarts the graph on the
retained checkout as it is now. It no longer restores the latest
snapshot over work done after a failure. Commits, staged, unstaged,
untracked and ignored files are kept, and so are the attached branch,
HEADand index. - Restoring a selected checkpoint is now an explicit rollback:
--from-checkpoint IDrequires--discard-workspace-changes. Without
it, the command fails before credentials are resolved, a successor is
claimed or the workspace changes. The error explains what the restore
replaces and prints the command to opt in. - Interrupted restores can still be retried against their pinned
snapshot. While a restore is pending, any resume that would not complete
it is refused, and the error names the command that completes it. Older
pending records for the latest snapshot complete only with consent. - The flag applies to local runs only and conflicts with
--target.
Named-target and hosted recovery are unchanged. - The CLI reference, the observe-and-control and forge-delivery guides,
andAGENTS.mddescribe the new behavior. - The new
resume_workspacetests compare the checkout's Git status
without*.profrawfiles. Coverage builds write these LLVM profiles
into the checkout, because the CLI and controller processes run there
withLLVM_PROFILE_FILEcleared.
Bound reconnects when following target runs (#1544)
When run, watch, logs or attach follows a direct or hosted
target, announce the first connection loss and retry with capped backoff
while preserving the event cursor. After ten minutes without
reconnecting, exit with the transport error and a zeroshot watch
pickup command. Ctrl-C detaches immediately, and detaching or giving up
leaves the target run running.
Preserve typed Python SDK errors after connection notices and document
the policy in AGENTS.md.
Stop local controllers cleanly on SIGTERM (#1536)
On Linux and macOS, SIGTERM now stops a detached local controller
cooperatively instead of ending it mid-work. The controller cancels
owned work, confirms cleanup and records the existing force_stopped
terminal result before it exits, within 30 seconds of the first signal.
The workspace and checkpoints stay available for zeroshot resume. On
Windows, console close, logoff and shutdown events remain unhandled;
zeroshot force-stop is the cooperative stop there.
Repair authorized reviews and resolve verified threads (#1545)
Route authorized app reviews through repair and resolve only verified
conversations with durable receipts.
Fix native CI failures by reusing the reducer’s neutral contract alias
and restricting Unix-only test imports. Preserve existing assertions and
upstream changes.
Expose graph results while Active Memory finalizes runs (#1550)
Expose optional graphResult through status, list and watch while
Active Memory finalization continues. Foreground runs print it
immediately and retain existing completion timing and exit codes. Record
the graph-result log once, preserve force-stop outcomes, update Python
models, and regenerate protocol artifacts and CLI documentation.
Closes #1547.
Confirm fast-forward merges after target integration (#1551)
Fix the two remaining failures found during live verification of #1395.
Fast-forward merges with null merge/squash SHAs now complete only after
target equality or GitLab’s merge-base
API proves
the reviewed head is on the captured target history. Post-push metadata
synchronization now uses the existing bounded MR/ref rechecks before
rejecting a stale MR head. Published history remains intact.
Changed: AGENTS.md, docs/guides/forge-delivery.md, and
zeroshot/src/source_code/adapters/core/gitlab/{review.rs,tests.rs,effect_tests.rs,merge_completion_tests.rs}.
Normalize uncached input tokens across harnesses (#1549)
Define inputTokens as uncached input across harnesses. Codex runs now
report lower inputTokens, excluding cache reads and writes after
resumed-session deltas. Normalize Copilot usage, preserve unknown
counters and recorded runs, and document verified upstream definitions.
Fix the merge-queue cancellation test’s readiness race by atomically
publishing its completed PID file.
Check CLI versions before local execution (#1553)
Check adapter-owned CLI ranges before local run, checkpoint resume or
ACP nodes start, refusing unreadable or incompatible versions with
upgrade guidance. Minimums are Claude 2.1.310, Codex 0.155.0, Copilot
1.0.70, Pi 1.4.0 and Hermes 0.22.5. Hermes requires >=0.22.5, <0.22.6;
OpenCode retains its existing admission. Contained Claude disables its
auto-updater.
Version probes now observe controller shutdown and await confirmed
cleanup before reporting startup failure. A no-prompt v0.22.6 bootstrap
confirmed the cleanup incompatibility. The maintainer files the drafted
follow-up to restore compatibility starting with v0.22.6.
Preserve provider rejection and force-stop reasons (#1552)
Record native credential/model rejections and force stops as refusals
with specific reasons. Prevent rejection retries, preserve capsule
classification, skip postmortems that inherit rejected sources, and
update protocol, Python, logging, and documentation.
Retain documented crash fallbacks for message-only Codex errors, Pi
assistant failures, generic Copilot model-query errors, unsupported
OpenCode error shapes, and legacy Hermes results.
Preserve Copilot’s bounded, redacted provider explanation when cleanup
fails, while retaining CleanupUnconfirmed as the settlement error.
Fail clearly on Claude quota exhaustion (#1560)
- Report confirmed Claude quota exhaustion through existing refusal
plumbing, preserving reset details, affected-node context and
manual-resume guidance. - Require explicit native failure evidence. Malformed result subtypes
preserve existing failure handling, addressing the reviewer finding. - Preserve authored recovery, cleanup and usage accounting; add stub
regressions and recovery documentation.
Keep run inspection read-only (#1558)
Local run inspection is now read-only. zeroshot status, list,
logs, watch, attach and checkpoints read a local run's recorded
history without taking the controller lease, starting an observer,
rewriting controller.ready.json or recording runtime_lost, and they
work without write permission on the run directory.
For an unfinished run whose controller does not answer, status, list
and checkpoints exit 0, keep the recorded status and add an optional
controller entry beside it: {"state": "lost", "observedAt": 1794470707000}. lost means the original controller process is
confirmed absent and its original lease is released. Anything less
certain, such as an unreachable endpoint, a stalled controller, a reused
PID, a replaced lease or a permission failure, is unknown.
observedAt is the time of the check, not the time the controller
stopped. The entry is omitted when the controller answers and for
finished runs.
watch, logs and attach print what is recorded, then exit 1 with a
message that names zeroshot resume RUN_ID and zeroshot force-stop RUN_ID. Only those two commands confirm a loss under the original
controller lease and record runtime_lost; they refuse to take over a
controller that still holds its lease or whose loss cannot be confirmed.
Finished runs are read without a controller and keep their exit codes.
The Python SDK exposes the entry as RunStatus.controller and
RunSummary.controller.
Preflight forge access with caller network settings (#1559)
Local --pr and --ship runs, and their resumes, now check forge
access before any agent starts. One authenticated, read-only request
confirms that the forge is reachable, the token is accepted, the project
is readable and the server version is supported. A failed check refuses
the run and names the forge, the failing step and the reason. Before
this change an unreachable forge, a rejected token or an unreadable
project was found only in the deliver node, after the worker and both
reviewers had run.
Local delivery now also works for forges that are reachable only through
a proxy or that use a private certificate authority. The early check and
delivery honor the invoking shell's HTTPS_PROXY, HTTP_PROXY,
ALL_PROXY, NO_PROXY, their lower-case forms, SSL_CERT_FILE and
SSL_CERT_DIR, the same settings Git transport already used. No flag or
stored network configuration is added. GitLab accepts private CA
certificates from the named PEM file or certificate directories and
supports socks5:// and socks5h:// proxies. The GitHub CLI honors
those CA paths on Linux, and on macOS and Windows when it is built with
Go 1.27 or newer; an older build uses only the system certificate store
there. TLS verification stays on, an unreadable certificate file is
reported by name, and proxy credentials are kept out of output, logs,
status and the run ledger.
Show one title when the editor and run view are embedded (#1565)
When a host page embeds the profile editor or the run view, the title
now appears once. The host already shows the profile or run name and
status, and handles unsaved state and saving, so the embedded view no
longer repeats them. Standalone pages are unchanged.
Classify bare gh HTTP statuses for delivery retries (#1556)
Ship delivery now retries when GitHub briefly returns a server error
that gh reports without a structured message. Previously these
failures had no HTTP status, so a temporary outage was treated as
permanent and the run stopped as a delivery failure even after the
candidate had been accepted.
The GitHub CLI reports these errors in two forms that the GitHub adapter
did not recognise:
gh apiwrites a non-JSON (or message-less) error body to stdout and
onlygh: HTTP 503to stderr. Delivery usesgh apito read and update
pull requests, reviews, comments, checks and merge queues.gh pr merge, used for direct merges, reports a failed merge mutation
asnon-200 OK status code: 503 Service Unavailable body: "…".
The shared GitHub status parser now reads both forms, so these failures
go through the existing status policy:
- 429 and 5xx are retried within delivery's existing retry, deadline and
cancellation bounds. - 401 remains an authentication refusal.
- 403 without a recognised rate-limit message, and other 4xx refusals,
stay permanent. - 404 uses the existing not-found handling.
No new retry layer or configuration is added. This follows the reported
proposal to rely on the existing transient policy, with no deviation.
Responses where gh prints no HTTP status at all remain statusless and
are not presumed transient.
Say why a native forge credential lookup failed (#1568)
When Zeroshot asks the local gh or glab CLI for a forge credential
and that lookup fails, the error now ends with the step that failed, for
example native forge credential lookup failed: command timed out after 10 seconds. Before, every such failure read native forge credential lookup failed with no reason. Command output is never included.
Preserve bridge startup diagnostics (#1569)
Report Hermes bridge startup failures with node identity and preserve
their cause after confirmed process cleanup. Release inactive processes
without parsing leftover help as cleanup frames; activated bridges still
require native cleanup acknowledgement.
Activate bridge through symlinked provider homes (#1572)
Normalize the Hermes bridge marker before its exact path comparison so
sessions start through symlinked provider homes. Add startup regressions
for real paths, direct and ancestor symlinks, and inactive invalid
markers.
Return repairable reviewer setup gaps to workers (#1571)
Require workers to prepare shared dependencies, tools and candidate
binaries and report reviewer setup. Return installable setup gaps as
actionable rejection feedback while preserving review protections. Keep
the prompt snapshot within source-line limits without changing its
bytes.
Documentation
Join the review sentence (#1499)
The README now reads "Independent, adversarial review fixes that,
because a reviewer starts with fresh context and looks for reasons the
work should fail." instead of splitting it into two sentences.
Prettier and the docs-agent contract test pass locally.
Co-authored-by: Eivind eivindcovibes.ai@Eivinds-MacBook-Pro-5.local
Put getting started before detailed features (#1502)
The README now introduces the workflow, shows how to start a first task,
and then lists detailed features. Move Get started before Features and
place the plugin setup paragraph there alongside CLI installation. Put
the first-run example before optional custom graph design.
Rename the diagram section to “How it works” and the feature subsection
to “Graphs and agents” to remove the repeated graph headings. Keep the
existing diagram and caption without adding an explanatory paragraph.
Say which checks are deterministic, with an example delivery check (#1540)
Simulated CTO visits to zeroshot.sh kept asking which gates are
deterministic and which are model judgement, and one asked for a check
script they could adapt (tests plus a smoke boot) with a sample verdict.
- How execution works gets a short "Which checks are deterministic"
section. The engine enforces order, routes and retry limits. Agent steps
and verifiers are model judgement. Astepbound to a
script returns your program's
result (local only for now). Outside the graph, delivery waits on CI and
required checks, and local runs can add a delivery check command. - Delivery validation gets an example. It's a shell check that runs
the tests on the exact candidate in a scratch copy atpublishand
merge, with install steps, the request it receives, the four verdicts,
and the rules the executable must meet. Those rules are taken from
native_v2_delivery/validation.rsandexecution/platform: absolute
private path outside the checkout, no arguments, aLANG=C-only
environment on Unix, exit 0, one JSON object of at most 1 MiB, and a
reason of at most 8 KiB.
Open in the site's voice and add two questions (#1546)
The README opening now reads like zeroshot.sh instead of an essay.
- The three paragraphs about reliability ceilings and renewed scrutiny
are replaced by the site's lines ("Do you still look at the code? We
don't. We built zeroshot so we wouldn't have to." and why one agent gets
sloppy), plus one plain sentence on drawing your team's process as a
graph, with per-step agents and models and fresh-context review kept. - Under the example graph, two short answers from the site: why a graph
and not a skill, and whether loops work, linking the review-loop study
(a small model reaching a big model's score at 58% of the cost).
No other sections change.
Co-authored-by: Eivind eivindcovibes.ai@Mac.Home
Say zeroshot sends no telemetry and list its network access (#1543)
Simulated CTO visits asked whether the local CLI phones home to
app.zeroshot.sh and whether that can be turned off. Nothing public
answered it.
- README (Hosting) now says: zeroshot sends no telemetry. The CLI
contacts GitHub only to install or update, and otherwise only the model
providers, Git hosts and targets you set up. - Installation reference gets a "Network access" section with a
table of every request the CLI makes. Those are the npm install
download,zeroshot update(anonymous, never in the background), a
run's runtime-plan connections, delivery and review, and--target/
target login. It also notes that zeroshot starts Pi with
PI_TELEMETRY=0and Copilot with session telemetry off, while Claude
Code and Codex follow their own settings.
Sources on main:
- No telemetry, analytics or crash-reporting code in
zeroshot/,ui/,
npm/orsdks/. native_v2_cli/update.rsruns only for theupdatecommand
(execution/submission.rs).npm/zeroshot/lib/install.jsis the postinstall download.- The
reqwestcall sites are the target authority, hosting
connections, decision bindings and the GitHub/GitLab adapters. - The UI source references no external hosts.
Upgrade
npm install -g @the-open-engine-company/zeroshot@16.0.0Versions 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.