Releases
v4.70.0
Compare
Sorry, something went wrong.
No results found
v4.70.0 (2026-11-10)
Features
reverse-engineering skill mirroring REA (4.64.0) (#926 )
evidence-driven model routing, automatic probes and new-model adoption (4.61.0) (#918 )
every Jev gate is observable, and integer-only legs that cannot act are retired
panes compact themselves on a Compact verdict
zirv brand guide, approved by the operator
a brand-guide skill and an in-repo brand/ convention
an elevate pass and a life layer for interfaces and films
rewrite frontend and motion-graphics skills around craft
built-in skills carry embedded reference files
discovery-only model candidates for the benchmark (#891 )
read persisted records that lack tier fields (#891 )
benchmark every discovered model, no tiers or hardcoded ids
add zirv benchmark to find the best stack on this machine
make the shell edit guard an opt-in [edit_guard] toggle
lean zirv:worker plugin agent for orchestrator worker dispatches
lean headless launches drop tools a headless seat never uses
headless launches are lean by default
headless cache TTL defaults to 5m
headless effort defaults to low (4.54.0)
add goal-driven FAFO experimentation skill (#880 )
clicking a Claude subagent opens that subagent
auto-start a workflow for programming and investigation sessions
a workflow awaiting approval shows as a gate wait on the seat
AskUserQuestion latches a Question wait until the next tool call
status shows whether the supervisor is on with its harness and model
keep worker env scrubbed across a pane handover
withhold secret-shaped env from dashboard worker panes
withhold secret-shaped env from delegated workers by default
agent tree, orchestrator dashboard, supervisor, model discovery and approvals inbox (4.42.0)
add built-in motion-graphics skill
report optional motion-graphics prerequisites
classify Hyperframes render as a heavy operation
add Video artifact kind for motion-graphics renders
pre-registered larger handoff-thin confirmation (validate/holdout reps 6)
confirmatory handoff-thin cases and trimmed candidates
keep cached Jev answers for a week by default
wire the remaining six jev probe sites (missing-tests, stop-verify, review-disposition, review-dedup, artifact-substance, gate-reclass)
confirmatory approve-escalate cases and trimmed candidates
conservative-direction approve-escalate floor candidates
add crash, judge, approve, intake-plan, inject-screen to jev probe; probe-only floor override
determinism corpora and campaigns for the remaining twelve Jev sites
size determinism criteria to detect small stability wins; targeted handoff-thin and intake candidates
record the screen stage's quality point estimate
finer intake margin candidates around the 0.2 default
show d_quality in the candidate summary and results.tsv
measure intake stability over every acted decision field
round-2 handoff-thin determinism corpus and campaign
add zirv ctx jev probe floor-determinism measurement verb
promote on a quality objective
jev determinism probe backend and corpora
pair every code review round with a simplify pass
make the JEV sidebar errors line clickable
make report.rs cover the issues' definition of done
stage class filter, split discipline, and a min_pairs plan check
implement the warm cache mode's uncounted warm-up trial
implement the optional proposer, testable spawn path
cohort stratification by corpus task class (#804 )
proxy:decider receipt proof for the Jev intake path (#803 )
bounded, resumable autoresearch campaign runner (#802 )
add pure statistical promotion gate for autoresearch (#801 )
OutcomeRow schema v2 -- direct rows, route metadata
campaign attribution, spend reconciler, per-site Jev floors
rot track, rollover distance, Jev section and clock-driven motion
inline intake plan card replaces the stderr proxy prompts
wrapped-session dashboard refresh -- 28-col sessions, pane header, limits, per-session workflow (#807 )
[jev] compaction_select keep list for compaction focus (#789 )
[jev] launch_effort gate for the headless effort lever
force single seat for headless launches
scope guard fires on shell edits, owes tests up front, lists stated details
scope-creep guard -- a checkpoint before the first edit, a backstop at Stop
operator-only headless cost levers -- prompt-cache TTL, effort by intake class, lean launch
Jev advisory gates for approve, classify, handoff, screening, inject and stop-verify
usage rollup in jev status (#758 )
zirv-jev-full condition for wrapped-vs-vanilla (#758 )
record outcomes and propose routing recalibration (#757 )
relevance-ranked core memory selection (#760 )
server-side context editing for native Anthropic sessions (#756 )
measure and repo-filter the skill index (#755 )
sync --check CI gate (#754 )
warn on duplicate remember writes (#773 )
intake classification and discipline for wrapped sessions (#753 )
closed-set admin dispatch on the UserPromptSubmit hook
re-project dispatch-tier advisory onto the metadata-only contract
re-project memory rerank to the metadata-only contract
screen memory harvest before the generative helper (#742 )
add harvest_screen and admin_dispatch operator gates
add private Jev token-saving decisions
reuse a warm worktree pool for --worktree-reuse (#718 )
honest Claude Code egress allowlist reporting (#727 )
typed per-worker Conditions with reason codes (#723 )
resume workspace setup and gate workers on goal bootstrap
add declarative worker workspace binding (#729 )
ctx watch, ctx reconcile, worker post-mortem, MCP self tool, agent --manifest (google/ax batch 1) (#730 )
portable cross-model skill library with integration-aware activation (#539 ) (#708 )
mask sensitive data before model egress (#706 )
choose the default harness by what is installed, and say so (#690 )
route a seat's model tier, and say "configuration error" once
reuse a passing check whose inputs did not change (#699 )
shift the review rubric and test-first loop into implement (#699 )
cover the script runner and workflow
make the implement-vs-validate split measurable (#699 phase 0)
finish the first-run wizard and classify the new commands
add zirv tour and zirv ctx jev status
make Jev decisions deterministic: margin gate, frozen intake state, pinned model, decision cache
gated Jev advisory for review triage, gate reclassification and artifact substance
Jev supervisor pre-filters and dispatch model tiering, off by default
proxy clarification and domain tags, leaner intake state, Jev-ranked memory retrieval and harvest gate
jev::advise as the single gated advisory entry point
shared Jev decision client and operator-only [jev] config table
harness proxy -- intake decision via TypeSafe Jev with helper and deterministic fallback
surface blocked commands in the delegation receipt
extend the shipped allow list and escape_allow for the worker capability set
blocked-command observability (Change 5a/b/d, reduced 5c)
Fixes
an empty PATH entry must not re-open the unfiltered presence search
only list harnesses whose real CLI can run here
shell escapes, command args and IDE context are injected user text, not the task
before-done rulings follow the subagent graph, only a done-vs-not-done ask lifts them, and they name the work journal
injected text, dead sessions and stray files no longer move the task or the done ruling
an approve_allow lowering counts as acted for approve_allow
Jev counter rows are written atomically and only real effects count as acted
retire the integer-only TypeSafe intake leg
handoffs, done consults, team plan and the proxy helper keep the task
the Jev relay pins its own model and bounds each request frame
the Jev relay forwards only zirv's compiled questions
pane compaction never types over a draft or blocks the dashboard
one failed compaction ends automatic compaction for that session
a failed compaction resumes the same session and suggests a restart
zirv never restarts a session because of rot; it only suggests
rot band and recommendation follow the verdict
fold context-fill pressure into the rot score; restart only after a compaction
learn the Claude context window from evidence and resolve it once
opaque floor under GIF compositions; unattended builds never wait on brand approval
brand guide fingerprint tracks length and degrades; stubs stay untrusted
correct the life-layer recipes and capture script
GIF renders carry no grain; motion type chosen with the typography reference
owner ids over MCP, frame-relative motion type, consistent timing and target rules
scope the chat view's agent strip to the open session
parse relabelled choice replies, show the unparsed reply, drop the consult cap (#877 )
orchestrator guard ignores heredocs in $(...) and follows cd (#887 , #883 )
route artifact-publishing workers to general-purpose
launch_effort never lowers effort; reuse guard skips conventions (#874 )
harness notifications are never treated as the user's prompt
a session-tagged intake row is never cut by the launch window
the Jev box shows only the intake that launched this seat
the subagent drive only selects a row that can only be that subagent
zirv's own helper sessions never auto-start a workflow
auto-start detect recognises bug reports and investigations
no USD anywhere in the dashboard
a subagent never receives its lead's scope checkpoint (#849 )
starting from a task ignores untracked noise
an active workflow is bound to its session, not the repo
a Codex approval wait names the command it asks to run
hold only the modes that show the operator a dialog
finished subagents of a live session show as idle
drop USD spend from the orchestrator dashboard
scope the Jev box to this dashboard's sessions
every pending operator wait shows in NEEDS YOU
a binding supervisor done ruling also blocks a Codex Stop
launch settings wire the Stop and tool-failure hooks while the supervisor is on
a repo-forbidden ctx.toml no longer silences supervisor rulings
re-check the routed adapter's read-only floor at launch, scope Anthropic cloud names to their provider
scrub worker env on refused config, scope cloud credentials to the selected provider
refuse read-only workers on harnesses with no read-only floor
a repo-forbidden ctx.toml no longer silences the Bash safety hook
zirv's own commands run without approval prompts (#845 )
stalled-compaction mail names what it measures (#829 )
inbox returns root-level mail addressed to the session from any cwd (#829 )
count intake rows toward the runtime session a chat attaches to
attribute exec rows to the live session, and match intake failure rows
never re-send a relayed request directly, and keep tests off the relay
count every live Jev request once, under its session
pad the launch banner box on both sides and start it at column 0
stop the compaction checkpoint claiming selection is preserved
restore the hook and safety CLI help that #821 shortened
fold in the back-squat test findings
missing Chrome is not a blocker; browser ensure fetches it
gate doctor on named checks, snapshot transitions with --at
native video artifact kind, anchored render pattern, skip Hyperframes skill installs
pin every hyperframes call, drop unverified ElevenLabs path
extract node_version_meets_floor to satisfy clippy
narration is Kokoro-only, and clean up setup.rs formatting
resolve the objective judge's state dir through the injected env
kill Windows process trees natively instead of via taskkill
bound the stderr-pump join in drain_stderr_after_exit
drain stdio MCP stderr deterministically instead of racing a sleep
compact the launch posture announcement; seat the orchestrator on Opus 5.5
compare the acted clarify decision, not its raw probability
drop the probe's no-op --json flag and harden the probe trial
harden decision_trial exit codes, rep stability and state isolation
align jev-case labels with the probe's production action names
keep execute-plan out of the simplify step; scope gate wording
replace the broken delegation-ledger JEV session closure
scope the JEV sidebar total to this dashboard's own session
keep the intake cursor inside the box at an exact-width wrap
wrap the intake prompt/clarify boxes and restore empty-Enter
clean report paths, reproduction --dir, and per-stage noise
classify the research/spend verbs and document the new proposer call site
accept run.py's cache-token field spelling on a Receipt
a resumed headless session no longer appends a second Direct outcome row
case-fold patch-scope matching and refuse traversal/absolute paths
reject a proposed candidate id that collides or is malformed
an orphaned retryable crash retry no longer vanishes or gets treated as resolved
re-hash evaluator drift before every dispatch, not just at stage boundaries
explicit child env + jev_ran from the production receipt (P3, P4)
timeout paths never drop or misreport an agent receipt (P1, P2)
integrate against the real promote gate, fix path/budget/dispatch bugs
SpendReport Money fields always present, never omitted
compaction_select costs nothing when off and never stalls the pump
scope-guard shell checkpoint skips read-only commands, staged new files, sub-second rewrites
headless clarify line stops telling unattended runs to ask the user
read-only approve skip excludes git remote/reflog writes and --output
the Jev approve gate makes no call under dontAsk
headless effort is sticky per conversation, and compaction launches get the levers
review findings on the headless levers and the missing-tests gate key
PR #788 review findings -- headless allow, usage cache, compaction deadline, missing-tests gate key
never launch a real browser from the unit-test suite
address release review findings (#754 #756 #757 #759 #773 )
native-parity row for workflow calibrate (#757 )
classify prompt.intake_discipline as narrow-only (#753 )
re-project handoff and harvest gates onto metadata-only facts (#759 )
review findings for the intake floor and staged rollover sockets
probe Windows Program Files/LocalAppData for Chrome and Edge (#678 )
stage rollover successor socket under the seat short (#681 )
floor intake complexity by request size; v7 stated-detail check
match punctuated remember markers; bump 4.25.0
launch headless browsers with a disposable profile; stop probe argv races
apply the proxy's decided seat model to the native pane (#703 )
discover launchable browsers, including macOS bundles (#678 , #676 )
record native-subagent review runs toward the gate (#685 )
record measured native Jev context savings
keep the proxy workflow live for the launch that actually runs, close credential asks properly
give headless Worker/Single seats a skill pointer instead of the full index
gate the frontier proxy seat on complexity/risk, name the started workflow
keep supervisor notices out of JSON receipts
keep review dispatch private and preserve evidence
share goal budgets and keep workflow reviews inline
classify the worktree pool keys as narrow-only (#718 )
scope WebFetch/WebSearch to the egress allowlist in the launch argv (#727 )
claim a reusable worktree under a lock and age it from idled_at (#718 )
resume a ghost-parked seat without opening a handover (#721 )
validate and report the egress allowlist honestly (#727 )
drop unwired condition reasons and test the condition surfaces (#723 )
recognize a ghost-parked seat on mail/nudge addressing (#721 )
migrate workspace setup progress with repository state
remove unsupported GPT-5.4 mini (#732 )
deterministic workflow triggering; steer wrapped sessions and subagents to skills (#714 )
resolve thin-margin complexity/risk answers upward in the proxy intake (#709 )
stop repeated wrapped-session permission prompts (#712 )
preserve source sessions when automatic rollover fails (#711 )
address review findings on the three terminal and intake fixes
floor complexity for an explicit parallel/multi-agent delegation request
keep a drag-selection highlight across ticks, and give the proxy intake prompt a real cursor
re-prompt instead of skipping the proxy on a blank first line
stop overriding the operator's own claude permission mode
run the proxy intake on the native runtime too
fail fast when the harness is not installed (#690 )
pin the hook-install fallback test to an injected machine
select and scroll together, and copy only the pane (#697 )
collapse the first-run harness storm and stop claiming success without one
name the harness, the config layer, and the package manager
accept --version, mark the machine configured, repair a partial .zirv
correct package metadata and make release templating robust
dispatch tier override carries the whole tool input in updatedInput
keep the deterministic frontend trigger byte-identical when the Jev gate is off
approvals never become findings, no auto-dismiss of Critical/Major, wider test-runner safety net
drop the removed harnesses field from the typesafe wrapper test
memory reranking never truncates or drops candidates when Jev is off or silent
keep Question::noul allowed as dead code until its intake caller lands
append Jev decisions atomically and re-point the parity matrix at the moved tests
proxy derives execution from complexity and launches a single seat
narrow launchctl SSH agent access
silence safe command permission prompts
preserve pipe structure across a keyword-stripped segment head
keyword-aware segment candidates and docker/kubectl exec decoding
Performance
skip the attention lock round trip when no Approval is pending
session supervisors relay Jev calls over one keep-alive connection
the Jev approve gate skips read-only local commands
benchmark-to-fix loop rounds 1-5 and long-session rot fixes (4.30.0)
skip the adapter readiness probe unless help renders; widen recursive-delete of own temp scratch
Documentation
scope Windows dev-machine notes to Windows (#928 )
record the operator's kill of the Jev artifact text tier
Jev cannot judge artifact substance from text (Stage 1 r2)
the logo is logo.png, used as-is
revise the brand guide to the approved site redesign
FAFO round 2 -- lean headless tools and zirv:worker
FAFO part 3 -- macOS validation and headless levers
XL benchmark noise floor
FAFO record on worker-prompt speed/cost/quality levers
README covers the dashboard follow-ups
describe the relay's no-resend fallback
record the workflow gate, probe and platform gotchas
keep the why, drop the history in comments
record the native worker, terminal, transcript and path gotchas
keep the why, drop the history in comments of native_account..worktree, rollover and session
record the catalogue, mail and memory bank gotchas
keep the why, drop the history in comments of agent_manifest..model_window, api and mcp
record the wrap and exec pty and harness gotchas
keep the why, drop the history in comments
keep the why, drop the history in comments
record the config, agent and chat harness gotchas
keep the why, drop the history in comments
record the prompt, compile, policy and permit gotchas
keep the why, drop the history in comments
record the dash terminal and platform gotchas
keep the why, drop the history in comments
record the safety and hook bypass and harness gotchas
keep the why, drop the history in comments
point the critical-file list at the split modules
record why native MCP keeps bounded transport
catalogue row for motion-graphics; heavy permit applies only to zirv-run commands
document the motion-graphics skill
research note matches Kokoro-only narration
narration is Kokoro-only in the motion-graphics intent
motion-graphics plan
motion-graphics intent; research notes live in .zirv/work
move motion-graphics note to its workflow folder, add audio
motion graphics from wrapped agents
say what each tree-kill snapshot actually adopts
final Jev determinism reports and findings
record the Jev determinism noise finding
Jev determinism campaign results and findings
state when the screen d_quality point matches the gate's own comparison
list the research and spend verbs under Features
document the built runner and add a sample report
drop leftover internal-process wording from plan.rs
document [stages.*] classes, split discipline, and the min_pairs check
document zirv workflow research next to workflow spend
spend reconciler is zirv workflow spend
autoresearch design for #799 -#805
README tagline and measured-cost section from the round-7 benchmark
compaction_select references its own issue #798
round 7 full-grid results; chain wall excludes stagger pacing
round 6 and r6-t24 results and findings
round 5, 5b and 5b-t24 results and findings
drop the 4.24.0 handover now that verification is done
split the 4.24.0 handover under the entry size cap
handover for the unfinished 4.24.0 release verification
drop the last pointer to removed result files
drop the committed result reports and CSVs
summarise the benchmark as a headline claim
rerun against superpowers vanilla with zirv 4.21.0 and large tasks
plan Jev savings implementation and validation
separate review convergence facts from reuse eligibility
preserve accepted intent artifact
specify optional Jev savings and fallback contracts
define optional Jev savings scope
measure what wrapping a session costs (zirv vs vanilla claude)
note that extra_allow can re-widen the egress allowlist (#727 )
document tour and ctx jev status where the gates require it
spec for built-in safe-command policy and blocked-command observability
disclaimer, ctx module boundaries, latency targets, generated changelog (#430 , #431 , #432 , #433 )
Chores
advise_gated counts reached and eligible through pass()
share the resource directory table between disk and embedded bundles
one const for the lean headless tool list; unwedge the routing test
bump version to 4.43.0
simplify the dashboard follow-ups
bump version to 4.42.1
share the worker scrub-env decision-log write
bump version to 4.41.10
bump version to 4.41.9
bump version to 4.41.8
remove the never-set native pane selection fields
bump version to 4.41.7
bump version to 4.41.6
bump version to 4.41.5
bump version to 4.41.4
anchor the unwrap check on the test module
split wrap.rs by responsibility
split hook.rs by hook event and gate
split exec.rs by responsibility
split safety.rs by responsibility
bump version to 4.41.2
share the SSE frame reader and error builders in transport.rs
rename optimize.rs to surface_collect.rs
drop the duplicate truncate_bytes, rename compaction's truncate_chars
remove unwired native pane and footer-hint scaffolding
share the create-only private writers with spawnreq and permit
co-locate tests with the code the splits moved
merge rollover.rs and rollover_runtime.rs into rollover/
split runtime/tools/mod.rs by responsibility
split ctx/config.rs by responsibility
split dash/mod.rs by responsibility
split adapters/mod.rs by responsibility
split workflow/engine.rs into a directory module
split ctx/sessions.rs into a directory module
split runtime/native.rs into a directory module
split ctx/agent.rs into a directory module
generic config narrowing and shared tool validators
share duplicated helpers across modules
drop stale dead_code allows and dead items
drop dead dash code and share one now_ms helper
trim unused deps and redundant CI work
bump version to 4.41.0
bump version to 4.40.0
accept either restart outcome in the goal whole-run deadline test
pin the Windows state dir inside the redirected test home
make dash pane/spawn timing tests deterministic, ignore live Jev battery
cover gate-reclass per-item floors on an answered probe rep
drop campaigns for the two text-bearing gates sites Jev never answers
size determinism campaign ceilings and caps from observed per-trial cost
extract shared jev question/action fns for crash, judge, approve, intake-plan, inject-screen
extract hook/review/engine Jev site questions, floors and action rules
bump version to 4.38.0
size the intake campaign's ceiling and caps for six candidates
cover the jev-determinism campaign manifests in the plan sweep
size the handoff-thin round-2 per-trial ceiling to fit its budget
extract shared question/floor/action fns for probe reuse
keep the Jev description fixture under the inline argv budget
bump version to 4.36.0
drop redundant ubuntu legs and a wasted CD release build
inline unit tests for schedule and reconcile
stop the proposer tests shelling out to python
split run.rs's three natural seams into sibling files
jev::floor unit tests, jev.floors env/repo-forbidden tests
regenerate CLAUDE.md from canonical context (#754 )
bump 4.27.0
bump 4.26.0
provider live-evidence manifest tooling and the frontend acceptance scenario (#592 , #610 ) (#675 )
release 4.24.0
remove load-sensitive timing and repo-root leaks (#669 , #713 , #450 )
bump to 4.22.0
bump to 4.21.0, document the proxy seat tiers, skill pointer and temp-dir deletes
bump version to 4.20.0
use the shared lock guard for the worktree store (#718 , #728 )
share one advisory-lock guard across ctx state stores (#728 )
prove the nudge path resumes a due ghost park in place (#721 )
isolate Claude score fixture from operator adapter settings
bump version to 4.11.0
bump version to 4.8.0
drop the dead-code allow on Question::noul now that intake calls it
cover the typesafe wrapper's delegation to jev::ask
4.4.0
Other
4.70.0: T3 wave 2: checkpoints, PR watch, eval corpus, judge failure logging, MCP hints (#962 )
4.67.1: mask the whole span when an entropy match overlaps an encoded secret (#961 )
4.69.0: fix stale supervisor rulings, ruling leaks, verify on nested repos, memory init, paste-blocked (#960 )
4.68.0: CI-parity verify discovery from GitLab CI and GitHub Actions (#958 )
T3 wave 1: prompt-cache TTL, encoded secret masking, side-call usage gate (#959 )
4.66.0: typed --network grant for workers; zirv ctx run in codex sandboxes (#932 , #933 ) (#934 )
4.65.1: stop background spawns opening terminal windows on Windows (#930 )
4.65.0: stop the harmless safety prompts from #927 (#929 )
4.63.0: fewer irrelevant permission prompts, fix #921 -#924 (#925 )
4.62.0: fix 13 open bug reports (#920 )
Release 4.53.0: validated cost, speed and quality fixes; NEEDS YOU keeps subagent dialogs; prompt de-dup (#882 )
Release 4.52.0: auto-started workflows carry step context, abandoned ones close, hook classification bounded (#878 ) (#879 )
Release 4.51.0: cost, speed and quality fixes from the 4.50.0 analysis (#876 )
Release 4.49.0: edit guard and Edit-tool worker rule cut scripted-splice repair loops; benchmark compliance columns (#873 )
Release 4.48.0: MCP tools visible again, leaner headless context, classifier and benchmark harness fixes (#872 )
Release 4.47.0: supervisor asks rule again, Jev in workers, valid Jev benchmarks, proxy validation gate, Windows fixes, subagent log (#871 )
Release 4.46.0: supervisor card, Codex idle, NEEDS YOU confirmed dialogs, supervisor budget, inbox socket (#867 )
Release 4.45.0: /settings session scope, proxy overrides and stats, resumed-seat transcripts, Windows CI (#861 )
Issue sweep: safety, dashboard, Jev, supervisor and /settings (4.44.0) (#858 )
bench(autoresearch): retune the 4 real campaigns so plan says valid
bench(wrapped-vs-vanilla): give every split a long_session task
bench(autoresearch): docs, README, and a real do_one_run integration test
bench(autoresearch): campaign manifests for #803 /#804 /#805
bench(autoresearch): intake decision-case corpus and decision_trial.py (#803 )
bench(autoresearch): run.py trial mode, corpus.toml, t25 long-session chain (#801 #802 #804 #805 )
bench: run.py comment for the round-7 gate keys
bench(jev-3way): give zirv-nojev the same headless proxy intake as zirv-jev-full
bench: t24b_long_haul states what steps 9 and 14 test
bench: a usage limit ends a zirv run like a vanilla one
bench: zirv-jev-full turns on the #781 -#786 Jev gates; zirv conditions get the headless levers
bench: wrapped-vs-vanilla harness, tasks t16-t24 and 2026-09-24/25 fix-loop results
You can’t perform that action at this time.