Repository navigation
Releases
v0.15.2: a journal path is always the literal file it names, a dsh 0.2 transcript that ran background jobs reads again, and a volta shim under a private HOME no longer recurses
v0.15.2 is a patch on v0.15.1. It carries the two fixes
v0.15.1 held for it, and the changes merged beside them. It includes
everything merged since
v0.15.1.
It needs no migration. Read
Upgrade and compatibility if you pass
--db :memory: or an empty --db, open a journal through a file: URI,
or match brokkr --help by its text.
What changed
- Every journal path opens the literal file it names
(#954,
#1161).- Before: the bundled SQLite is built with URI filenames on. A
--dbsuch asfile:other.db?mode=rwcopened, or created, the file
the URI named, not the path Brokkr had checked and shown. - Now: both of the store's opens make the path absolute before
SQLite sees it, so afile:path is an ordinary file name. Backup and
restore read the same home. - Refusals name the path you gave. A path holding a nul byte is
refused where you wrote it. - SQLite's two special names are no exception.
--db :memory:is
the literal file./:memory:, and an empty path is refused. This is
decision 0100's addendum, accepted by the operator.
- Before: the bundled SQLite is built with URI filenames on. A
- A dsh 0.2 transcript that ran background jobs reads again
(#1176,
#1186).- Before: a session.v4 seat log from dsh 0.2.1-rc.2 that used the
background-job tools carriesjob/registeredandjob/settledrows.
Version four did not name them, sobrokkr transcriptrefused the
whole log asunsupported-format. - Now: those two kinds, and
working-directory/change, are quiet
bookkeeping under version four. A job's output still reaches the
transcript throughtool/result. - A catalogue fixture: the 34 record kinds seen across 589 dsh
0.2.1-rc.2 transcripts are a test fixture, and every kind in it must
be placed. Decision 0086 ruling 4 (proposed) names the three kinds.
- Before: a session.v4 seat log from dsh 0.2.1-rc.2 that used the
- A volta shim under a private
HOMEno longer recurses
(#1136).- Before: the unboxed dispatch and the harness probe clear the
environment, keep the engine'sPATHand set a privateHOME. A
volta shim then looked for its home under that emptyHOME, fell
back tonodeonPATH, which was the shim again, and recursed until
fork failed. - Now: one resolver tells the child the volta home its
PATH
reaches: the engine's ownVOLTA_HOME, else~/.volta, but only
whenPATHlists itsbin. The workspace box is unchanged. - Decision 0111 records the rule.
- Before: the unboxed dispatch and the harness probe clear the
brokkr --helpcites no decision, ruling or issue
(#1170,
#1173).
Every help line that cited one says, in plain words, what the citation
meant. A test walks the whole command tree and refuses a citation. No
flag, verb or default changed.brokkr watch's plain frame paints its graph as the console does
(#1168,
#1174). The
frame printed without a terminal, or with--no-view, now classes each
node by the console's function and reads the same colours and glyphs.
One run no longer reads in two colour schemes.
Fixed
- The store's one append door inserts through the connection's
statement cache, and the test fixture planter seals through it rather
than a copy. A 20,000-event plant takes about 11 s, where the uncached
insert took about 36 s
(#1156,
#1179). - Six more refusals are typed, with their text unchanged byte for
byte: the transcript module's refusals, the dsh home's "no harness
home", the secret store's write path, Muninn's unusable report, the
composite pnpm reader's flow map and packages key, and the bundle's
pinned-script admission. Their tests now match variants, not
substrings
(#353;
#1182,
#1183,
#1184,
#1188,
#1189,
#1197). - The capability broker's deadline proofs wait on the broker's own
startup deadline, a hang is killed with its whole tree at the guard,
and the teardown proofs are bounded
(#1008,
#1172). - The cool-down test fixtures plant their waits an hour past the
clock, so a lock stall at 128 test threads no longer loses a test's
own row
(#857,
#1175). - The shard runner's
--helptest reads the help at a width it fixes
itself, so a narrow terminal no longer splits a path in it
(#1181,
#1185). - The flake's digests catch up with v0.15.1
(#1171).
Upgrade and compatibility
--db.- A
file:URI is now a file name:--db file:other.db?mode=rwc
opens a file of that literal name, notother.db. --db :memory:is the file./:memory:, not a private in-memory
database.- An empty
--dbis refused. - An ordinary path opens the file it opened before.
- A
brokkr --help's text changed where it cited a decision, ruling or
issue. Anything matching those lines sees the plain words instead.- The public API: unchanged.
- Contracts: unchanged. The engine version reads 0.15.2.
Known issues
- Three flaky tests are open, all seen only under heavy parallel load:
hands::session::tests::the_listing_removes_nothing_and_counts_what_the_reaping_would
(#1157);engine::broker::admission::tests::recovery::a_moved_head_end_keeps_its_attempt_reserved
(#1164);- the
hands::box tests besideprobe::at 16 threads. A test's
stand-in bubblewrap leads the sharedPATHand reaches its
neighbours. The cause is in the test support, not in the box
(#1187).
- Not in this release:
brokkr graph's states and rules as aligned
tables
(#1167) is
still in review.
v0.15.1: the shipped roster moves to Sol 6.2, Astra 6.2, Terra 6.1 and Fable 5.2, dsh 0.2.2-rc.5 is a measured hands release, and a damaged journal is one refusal that names the file
v0.15.1 is a patch on
v0.15.0 “Gullinbursti”. It carries the fixes that were in
flight when 0.15.0 shipped, and the changes merged beside them. It
includes everything merged since
v0.15.0.
It needs no migration. Read
Upgrade and compatibility if you keep your
own adapter files, pin a model id in a recipe, or match a journal read's
error by its text.
What changed
-
The shipped model roster moves to the latest models
(#1155,
#1165).Alias Was Now solgpt-6.1-sol gpt-6.2-sol astragpt-6-astra gpt-6.2-astra terragpt-5.6-terra gpt-6.1-terra lunagpt-6-luna gpt-6-luna fable,fable-talliedclaude-fable-5-1 claude-fable-5-2 - Where: the
modelsmaps ofadapters/codex.json,claude.json
andlanetally.json, and the copiesbrokkr initwrites. The shipped
recipes' inline pins move with them (preflight, review-first, standby
and wager-harness). - What stays: aliases, judges, efforts, flags, hands and tool
permissions. - Decisions 0031, 0041 and 0045 each record the new mapping in a
dated addendum; their alias rulings stand.
- Where: the
-
dsh 0.2.2-rc.5 is a measured hands release with its own composed set
(#1158,
#1163).- Before, Brokkr refused rc.5 for every hands seat. npm's
latest
dsh is 0.2.2-rc.5, andRelease::namedadmitted only 0.2.1-rc.2,
0.2.1-rc.3 and 0.2.2-rc.1. - Its own release. rc.5 is matched whole. Its judge admits
0.2.2-rc.1's set plus rc.5's measured delta: theagent-teamhost
row, thetool-agent-teamrow, anddeepseek-accountunder its new
package. - Fail closed on the new surface. The hands overlay writes
tool-agent-teamdisabled on rc.5, and a composed tree that enables
it is refused. - The measurement is in
the dsh hands record. 0.2.1
and 0.2.2-rc.1 are unchanged.
- Before, Brokkr refused rc.5 for every hands seat. npm's
-
A damaged journal is one refusal that names the file and the way back
(#1000,
#1143).- The sentence. A journal Brokkr cannot read prints as one
sentence: the journal's path, the cause, that nothing was read from
it, and where to restore it from. Before, every read surfaced bare
SQLite or serde words with no path. - Where.
runs,inspect,export,seats,watchandtui
print it and exit non-zero.watchused to read damage as transient,
and now ends at once. - The fleet listing quarantines a run whose envelope is not JSON,
and lists the healthy runs beside it. - Brokkr reports damage and never repairs it (decision 0001). The
journal guide's new section, "A damaged journal", is the one the
sentence names.
- The sentence. A journal Brokkr cannot read prints as one
-
One home for the payload shapes the sweep reads
(#430,
#1160).
brokkr-coregives an effect's id, and an operator command's id and
word (accepted_command,operator_command,succeeded_effect).
Eligibility's own decodes go.
Fixed
-
The sweep keeps an undecodable operator command's rows for a reason
of its own, "an operator command after its end does not read"
(#1160). -
brokkr dispatch'sstarted:line names the driver's log, and the
standing loop's sleep wakes on the signal rather than on a 50 ms beat
(decision 0110's lows,
#1152). -
Every match over
EventTypein the view names each arm, so a new
event type does not compile until its readout is decided
(#950,
#1153). -
The required RustSec audit is the pinned
cargo audit's own exit
code, in CI and in the release admission. It no longer fails a clean
audit on a check-run API error
(#1113,
#1150). -
The driver's reaper tests are bounded throughout.
- The single-orphan test sees its orphan adopted before it sees none
left. - A rerun past its deadline is killed, reaped and drained in time.
- A launched box releases its hold with its own wait.
- The single-orphan test sees its orphan adopted before it sees none
-
The reaped-pid test fixture is bounded by one patience, and a reap
first seen after the deadline never counts as within it
(#1160). -
The flake's digests catch up with v0.15.0
(#1151). -
The shard runner's
--helpnames the pinned nightly's coverage
layout besidedeps
(#1129,
#1162).
Upgrade and compatibility
- Adapter files you keep yourself are not touched. Only the shipped
adapters and whatbrokkr initwrites move to the new model ids. A
realm or recipe pinninggpt-6.1-sol,gpt-6-astra,gpt-5.6-terraor
claude-fable-5-1by its concrete id keeps it. Such a gate pin must
still map to a judge alias in its driver's adapter, orcompile
refuses it, as before. - dsh. A hands seat on dsh 0.2.2-rc.5 is now admitted. 0.2.2-rc.2,
rc.3 and rc.4 stay unmeasured and refused. You need not move from
0.2.1-rc.2 or 0.2.2-rc.1. - A damaged journal's error text changed. Anything matching the old
bare SQLite or serde words for a journal read now sees the one sentence
above, and a non-zero exit fromwatchwhere it used to keep polling. - The public API gains
brokkr_core::accepted_command,
operator_commandandsucceeded_effect. Nothing is removed. - Contracts: unchanged. The engine version reads 0.15.1.
Known issues
- Two flaky tests are open, both seen only under heavy parallel load:
- Not in this release. These are held for 0.15.2:
v0.15.0 “Gullinbursti”: `brokkr dispatch` stands and starts queued runs every interval until told to stop, a driver reaps the orphans its harness leaves, and a dry sweep names what a pass would keep
Gullinbursti is the golden-bristled boar that Brokkr forged. It runs
over sky and sea, by day and by night, and its bristles light the dark.
This release is about the work that keeps going while nobody watches:
- The dispatcher stands.
brokkr dispatchrepeats its pass every
--everyseconds and starts what admission admits, until a signal
tells it to stop. Decision 0110 is built, and with it decisions 0095
and 0068. - A driver reaps what its harness leaves. A seat whose tools leave
orphaned processes no longer piles them up as zombies under its driver
for the rest of the attempt. - A dry sweep tells the whole truth.
brokkr sweep --dry-runnames
the trees a pass would keep, with the pass's own reason.
It includes the changes merged since
v0.14.1.
It needs no migration. Read
Upgrade and compatibility if a script
relied on brokkr dispatch without --once being refused, or if you
build against Brokkr's crates.
What changed
brokkr dispatchstands (decision 0110, slice 6c;
#1140).- The loop. Without
--once,brokkr dispatchruns one pass, prints
the pass's report, reports the drivers that have exited since, and
sleeps for--everyseconds less the pass's own duration. Each pass
reconciles first.--everyis 5 to 3600 seconds, and 60 unless set. - A stop. The first SIGINT or SIGTERM wakes the sleep at once. The
entry in flight still gets itsspawnrow and its driver. Nothing is
claimed after the stop, and the dispatcher exits 0. A second signal
ends it at once, with 128 plus the signal's number. The next
dispatcher's reconciliation then settles anything left in flight. - Drivers are never signalled. Each runs in a session of its own and
is reaped between passes. A driver that dies while its run reads
runninggets a line with the command that resumes the run as it was
queued:brokkr dispatch drive --entry=<entry> --claim=<claim> --db=<journal>. The drivers still running when the loop ends are
named, and left running. - Its endings. A pass's ending ends the loop and exits non-zero, in
the same words--onceprints: a fence-out, a replaced binary, a
store error, or a tick whose report cannot be written. - The guide. Running the dispatcher
covers what a tick does, and running the dispatcher under a systemd
user unit (KillMode=process) or a launchd agent. It says what each
ending leaves and how to recover it, including after the host goes
down. - Built. Decisions
0110,
0095 and
0068 now read built.
- The loop. Without
- A driver reaps the orphans its harness leaves
(#1133;
#1148).brokkr driveris made its harness tree's subreaper, so
the orphans of a seat's tools are adopted by it. It now reaps them.- The cause. One claude seat's driver held 156,752 zombies after a
timeout-killedcargo testin its harness, until its attempt ended. - The reaper. While a driver serves on Linux, a reaper reaps each
zombie it adopted, per pid, after two sightings of one identity.
Reaping is best effort, and never ends an attempt or changes its
result. - The door. Every child the driver spawns goes through one door,
which registers it under the reaper's own lock as it is spawned. A
harness that exits at once is never reaped out from under its handle,
and its exit status is the one the driver reports. - The release. A child's hold is released by its own handle's wait.
A pidfd the kernel will not wait on is closed, and a reaper the host
refuses leaves nothing held. - The engine's own reap is unchanged. On macOS there is no subreaper,
and nothing changes.
- The cause. One claude seat's driver held 156,752 zombies after a
brokkr sweep --dry-runnames the hands trees a pass keeps (5c2's
follow-ups F3 to F5;
#1142).- The kept half. A dry run now lists the lockless hands trees a pass
would keep, with the pass's own words. The listing and the reaping
share one reading of the same candidates, so the two cannot disagree,
and the listing removes nothing. - One home for a refusal's words. A tree walk's refusal is worded by
TreeError::causealone. - Typed payload reads. The sweep reads
effect/succeeded,
operator/commandedandoperator/acceptedthrough typed decodes. A
payload that does not decode keeps the run's rows, with the reason
reported; it never panics.
- The kept half. A dry run now lists the lockless hands trees a pass
Fixed
- The coverage shard runner finds the pinned nightly's test binaries
(#1147, closes
#1129). The
nightly puts them inbuild/<package>/<hash>/out/, so the per-binary
bound never started a clock. It now judges a binary by its file name. - The cool-down readout is one narrow scan
(#1138, closes
#1112). The
readout behind the fleet listing and the console read the table squared,
one correlated statement per row. It now reads one narrow scan in one
snapshot. - The park-cause design test names its cause when it fails
(#1146, closes
#1132). A
failure printed a bareunwraponNone; it now prints the journal the
test read. It is tests only. - brokkr-store's slowest tests plant their journals in one transaction
(#1141, closes
#1130). The
ordinal test took nearly an hour on a SATA disk, one synced commit per
event. It is tests only.
Upgrade and compatibility
brokkr dispatchwithout--oncestands where 0.14.1 refused it.
A script that ran the bare verb to check that dispatching is off now
starts a dispatcher.--everybeside--onceis refused.- The crates' API moves.
- brokkr-runtime:
dispatch::stand,Stand,Stopped,Tick,
Exited,ReapedandEnding::Unreportedare new.Seamsandpass
take a third parameter,claims, read before each claim.
admission::Room::claimis no longer public. - brokkr-protocol:
hands::list_reapable_underreturns aListingof
the reapable trees and the kept ones, andTreeError::causeis public.
- brokkr-runtime:
- Contracts and storage. The engine, event schema, database schema,
queue storage and driver protocol versions are unchanged. - Engine version. Bundle digests cover the engine version, so a run
started under 0.14.x resumes only under 0.14.x's binary.
Known issues
- brokkr-runtime's library tests can fail one engine test per run at
high test-thread counts, for example 128 on a 128-core host. A running
driver ends with the wrong effect. Main fails the same way, so this is
interference between tests, not slowness
(#857). Run
them on about 24 CPUs. brokkr doctor's version probe has no time bound, and its tests run
the host's tools under a temporary HOME. On a host with Volta-installed
Node tools, a shim there can call itself without end
(#1137).- The dispatcher's
started:line names no driver log path, and the
standing loop reads its stop every 50 ms rather than waking on the
signal. Both are carried from 6c's review. - The known issues listed in v0.14.1 that
this release does not address still apply.
v0.14.1
One native binary per platform — no Python, no Node, no services.
The binary is brokkr, and now the only one: the forge shim
that rode along for one release is gone.
Linux also ships .deb and .rpm built from those same
binaries; every package manager channel serves these files.
Verify downloads against SHA256SUMS, then verify signed GitHub
build provenance with gh attestation verify <asset> -R feedback-loop-ai/brokkr.
See ARCHITECTURE.md and docs/decisions/ for what this engine is and why.
What's Changed
- store: every window statement leaves out the whole read's runs, and the parks statement's SQLite share comes under the cap (decision 0089 slice 4, U5b1, #753) by @valentinyanakiev in #1104
- packaging: flake.nix digests for v0.14.0 by @brokkr-release-bot[bot] in #1106
- dispatch: one pass starts what admission admits and hands each run to a spawn seam (decision 0110 slice 6a, #430) by @valentinyanakiev in #1107
- engine: a refused reseal leaves no successor standing, and settle's last clauses are bound (#1099, #440) by @valentinyanakiev in #1108
- readouts: a cooling run's waits are painted once, from core's written end, and a fold-refused run paints none (decision 0103 slice 4f, #430) by @valentinyanakiev in #1111
- engine: the admission lock's refusal names its errno or the clause it failed (#1082) by @valentinyanakiev in #1114
- analytics: a hearth's row read is one snapshot, so the row path holds no set of run ids (decision 0089 slice 4, U5b2a, #753) by @valentinyanakiev in #1115
- admission: a capacity reason's words move to their own module (decision 0110 slice 6d, #430) by @valentinyanakiev in #1117
- engine: the undo's docs state each arm, every arm is pinned, and #1099's last LOWs close (#1099, #440) by @valentinyanakiev in #1118
- park: a park names its cause, typed and closed; retryable reads only what stands (decision 0106 slice 0106a, #1035) by @valentinyanakiev in #1116
- tests: the cool-down probe and deadline-words tests do not depend on the host's load, and a failure says what each member did (#1110) by @valentinyanakiev in #1121
- dispatch:
brokkr dispatch --oncehands each run it starts to a driver of its own, withdrive,set-boundandstatus(decision 0110 slice 6b, #430) by @valentinyanakiev in #1122 - adapters: a seat's git environment is one setting under a name no credential scrub drops (#1109) by @valentinyanakiev in #1123
- decisions: 0109, accepted — a gate's re-visit reads what changed, unless a declared trigger asks for a whole read (#1088, #1094) by @valentinyanakiev in #1120
- tests: the cool-down sequence tests print the witness on every assertion they fail through (#1110) by @valentinyanakiev in #1125
- tests: heap_costs' compare holds to the byte on a slow host (#1119) by @valentinyanakiev in #1131
- probe: claude 2.1.326's stream, transcript and MCP log decode whole by @valentinyanakiev in #1128
- probe: a CLI that keeps its login in a file is given that file,
--credential-fileby @valentinyanakiev in #1134 - tests: the admission and readiness tests hold on a loaded 8-thread host (#1124) by @valentinyanakiev in #1135
- release: v0.14.1 — a park names its cause,
brokkr dispatch --oncestarts queued runs under a run ceiling, Claude Code 2.1.326 decodes whole, and the harness probe binds a login file by @valentinyanakiev in #1139
Full Changelog: v0.14.0...v0.14.1
v0.14.0 “Megingjörð”: a run is held by its driver and repaired by a record, a quarantined run releases its worktree, and a provider's rate limit is a cool-down the run waits out
Megingjörð is Thor's belt of strength, worn with Járngreipr, the iron
gloves 0.13 was named for. This release is about what holds a run
together under strain:
- A driver holds its run. A run is held by its driver's lock, and
nothing but an operator's annotation is written after it ends. - A broken run is mended by a record. It is repaired by a record and
never by an edit, and a quarantined run's worktree is released once
every writer is proven ended. - A provider's rate limit is a cool-down. The run waits it out and
probes the provider, instead of parking.
It also adds:
- a deadline floor for shared local hardware, and a sweep of the seat
trees Brokkr made; - retirement of a historical evidence root;
- typed run endings in every listing;
brokkr analyticsread from the journal's own rows.
It includes the changes merged since
v0.13.1.
This release breaks scripts that stop or conclude a run without a
cause, refuses a second driver on a run another process drives, migrates
the journal's queue storage to version 6 on its first writing open, and
moves the MSRV to Rust 1.89. Read
Upgrade and compatibility before you install
it over a working setup.
What changed
- A run is held by its driver, and a broken run is repaired by a record
(decision 0108; #440).- Nothing but an annotation follows a terminal event
(#1078).- The store refuses any other append to a run that has ended, inside
the append's own transaction, asStoreError::AfterTerminal. It
writes no event, no cool-down row and no projection. - A late driver ends typed,
EngineError::ConcludedBeneath, and
writes nothing. stop,retryand a member retry on a concluded run print why and
journal nothing.classifyandsupersedestill land.
- The store refuses any other append to a run that has ended, inside
- A run is held by its driver's lock
(#1080).- Every driving entry (run, rerun, a queue claim, dispatch and
resume) takes an exclusive, non-blockingflockon
<journal>.runs/<run-id>.lockbefore its first append. - A second driver refuses at once: "run '…' is driven by pid P on this
host since T (engine V)". brokkr concludetries the lock first, and refuses a run its driver
holds before writing anything.operator stopandretrytake no lock and stay lawful on a live
run.
- Every driving entry (run, rerun, a queue claim, dispatch and
brokkr operator repair --run <id> --reason <text>
(#1085).
It names the late tail after a run's terminal event. The fold then
reads that tail as late: kept and shown, never applied. The run reads
as its terminal status, with "a late tail of N events, repaired by
". Nothing is edited or deleted, and the hash chain is
untouched.- The record: an
operator/commandedannotation whose args are the
newoperator-repair.v1. - The reserved words (
repair,settle,classify,
supersede) are refused at the Looper bridge before anything is
journaled. The bridge answers with a rejected receipt and keeps
applying the commands behind it, so the channel never wedges.
- The record: an
brokkr operator settle --run <id> --reason <text>: a
quarantined run releases its worktree
(#1100).- The release is proven, never assumed. Each driver records its
process group, and the release asks every recorded group. It refuses
with a typed reason if a group is live or a cleanup is unresolved. - The settle record: an
operator/commandedannotation under the
newoperator-settle.v1, read by type and never by its word.
- The release is proven, never assumed. Each driver records its
- No broker write makes a reservation that is gone.
- Six paths reseal: every path that reads a reservation as
standing and writes it again (a run's later attempt, adoption,
recovery, the operator's release, an MCP server's extension, and the
crashed probe's release) now reseals it. - One atomic exchange:
renameat2(RENAME_EXCHANGE)on Linux,
renameatx_np(RENAME_SWAP)on macOS. A record removed in the window
stays removed. - One path creates: only a run's first admission, which read no
record.
- Six paths reseal: every path that reads a reservation as
- Nothing but an annotation follows a terminal event
- A provider's rate limit is a cool-down the run waits out (decision
0103; #430).- The host declares the bounds, in
forge.host/v2'scooldowns
block (#1007):default_secondsandmax_seconds, each 1–7200 s;- overrides per provider and per route.
- A single seat waits, probes, then retries, all inside decision
0006's attempt cap
(#1026):- a seat refused by a rate limit before its session records the
site's cool-down and waits in-process, honouring an operator stop; - it then probes the cooled route with one turn of the refused
candidate's own adapter, and retries only after the probe answers; - the wait is recorded under the new
provider-wait.v1.
- a seat refused by a rate limit before its session records the
- Panels and sequences wait out each refused member's or step's own
cool-down, and probe each one
(#1046,
#1049). - Cool-downs are rows in queue storage v4, written in the terminal
event's own transaction
(#1070,
#1074). - The engine waits the aggregate
(#1081):- that is the latest unexpired end for the provider and route across
every run; - a probe refused again extends the wait without spending an attempt;
- a refusal past the maximum parks, naming the cool-down.
- that is the latest unexpired end for the provider and route across
- Admission holds a queued entry while its provider's route is
cooling down (cooling_down), and the claim's fence rechecks it
(#1095). - A quota or usage limit is a rate limit with its own reset, for
Codex, dsh and Claude
(#1058).
- The host declares the bounds, in
- Shared local hardware sets a deadline floor, and the host sweeps the
seat trees Brokkr made (decision 0104;
#430).- The deadline floor. A shared-local route in the new
forge.host/v3may declaretimeout_floor_seconds, at most 7200. It
raises the deadline of an attempt on that route and never shortens an
authored one
(#1039). - Seat trees are recorded. Each run-local seat scratch tree is
recorded in queue storage v5'sseat_treesbefore it is made
(#1073). Each
complete removal goes in v6'stree_sweeps, made by one walk that
follows no link and crosses no filesystem
(#1079). brokkr sweep [--dry-run]removes the seat scratch trees the
journal proves free, under the dispatch lock
(#1096).- Every tree it keeps is named, with its reason.
- A dry run reads the same verdicts and writes nothing.
brokkr doctornames each directory under a recorded root that no
open row records, and removes nothing.
- The deadline floor. A shared-local route in the new
- An operator retires a historical evidence root, with a record
(decision 0105;
#1020).brokkr evidence listlists every historical root.brokkr evidence retire --root <dir> --reason <text>retires one,
under the admission lock, with exactly one of:--archived-to <dir>, which writes the newevidence-archive.v1
inventory last;--discarded, which removes the root by a walk that follows no
link.
- A historical root no longer keeps every unboxed writer on the host
refused forever
(#1047,
#1064,
#1069).
- A run says how it ended, in every listing (decision 0064, slices E3
to E6; #317).stopandconcludename the ending.brokkr operator stopand
brokkr concludetake--cause, one of the six operator-owned
endings, recorded underoperator-ending.v1
(#1006).- Every row carries it. A run's summary row carries its typed
ending, its citation and whether aclassifystood over an earlier
ending, on/api/runs,/api/view,inspect --jsonand
brokkr runs(#1014). brokkr runspaints it
(#1015):- each row by its ending's tone, with a one-character gutter for what
it waits on:!for a ruling,>for an action; - a heading of who must act, and a closing endings tal...
- each row by its ending's tone, with a one-character gutter for what
v0.13.1: a provider's pre-session refusal is typed, the operator names how a stopped run ended, and the dsh console reads its fleet a page at a time
v0.13.1 is a patch on
v0.13.0 “Járngreipr”. It carries the fixes that were in
flight when 0.13.0 shipped, and the changes merged beside them. It
includes everything merged since
v0.13.0.
It needs no migration. Read
Upgrade and compatibility if you run a
driver of your own or a recipe that resumes runs across the upgrade.
What changed
- A provider's pre-session refusal is typed (decision 0103, slice 4a;
#992).- The refusal. A refusal before a session starts is a closed class,
rate_limitorother. It carries the provider, the route, bounded
words, and optionalstatusandretry_after_seconds. - The contract. It reaches the engine on the driver's result under
the new extension schemadriver-refusal.v1, which is additive at
driver protocol v1. A present refusal is never read as absent. - The probe. A driver that advertises
capabilities.probecan be
sent a probe, which stops at the model's first answer. - dsh's startup refusal is kept on every path. A stopped,
interrupted, incomplete or unreadable capture of dsh's startup error
output no longer loses the refusal: an incomplete or unreadable capture
refuses the start by name. - Nothing waits on the refusal yet. #430's later slices turn a provider's
rate limit into a cool-down instead of a park.
- The refusal. A refusal before a session starts is a closed class,
- The dsh console reads each hearth's fleet a page at a time (#907;
#990,
#995).- Paged fleet reads. A hearth's fleet is read 100 runs at a time,
with anN+count where more exist, and "show more" extends one hearth. - One run on its own.
brokkr runs --json --run <id>prints that
run's row, at view version 28. - One gate publishes world data, and it holds the console's only
writer. Everything the console paints is fenced by the journal and the
map it read.
- Paged fleet reads. A hearth's fleet is read 100 runs at a time,
- The operator names how a stopped run ended (decision 0064, slices E1
and E2; #988,
#1002).- The vocabulary. A run's ending is one closed word that brokkr-core
derives from the ruling that ended it. brokkr operator --run <id> classify --cause <cause>. Names how a
stopped run ended:superseded(with--by-run),infrastructure,
delivered-elsewhere(with--by-ref),withdrawn,exhaustedor
operator-stopped.- The contract. The annotation is recorded under the new payload
schemaoperator-ending.v1. A completed run is never classified, and
the fold reads none of it, so a classified run folds byte for byte as
before.
- The vocabulary. A run's ending is one closed word that brokkr-core
Fixed
- Flaky tests that could drop a pull request from the merge queue:
- a heap peak is the measuring thread's alone, so another thread's
allocation cannot move a same-peak equality, on Linux and on macOS
(#979); - a test that re-runs its own binary captures the child's output, so the
child's test summary cannot interleave with the parent's
(#996); - a protocol test whose verdict depended on another test's stray
process now runs in an engine of its own
(#1001).
- a heap peak is the measuring thread's alone, so another thread's
- macOS CI time: the large run-page tests build their journals once per
binary (#985), and
the macOS tests are spread over four machines
(#983,
#999). - The flake's digests match v0.13.0
(#984).
Upgrade and compatibility
- Drivers of your own.
driver-refusal.v1is an extension: a driver
that sends no refusal is read as before. A driver that advertises
capabilities.probemust stop at its first model answer. - Engine version. Bundle digests cover the engine version, so a run
started under 0.13.0 resumes only under 0.13.0's binary.
Known issues
- A masked refusal can grow past its 300-character bound when a bound
secret's name is longer than its value. The attempt then fails closed,
as an unreadable refusal, and no secret is disclosed. The fix is in
progress: #991. - A paged run read over a run with about 66,000 checkpoints takes
about 0.3–0.6 s on a release build, where #753 aims at milliseconds.
About 94 ms of it is the chain verification at opening:
#994. - Every known issue listed in v0.13.0 still
applies.
v0.13.0 “Járngreipr”: MCP tools a seat grips through Brokkr's own broker, a journal read a page at a time, and one folder for Brokkr in every repository
Gloves are not done when they fit the carving. They are done when the
hand that will use them can approach, greet, grip and release.
Járngreipr are Thor's iron gloves, which, the Edda says, he cannot do
without when he lays hold of the hammer's haft. v0.12.0 “Gleipnir” bound
capabilities to the realm. In this release a seat actually grips a granted
tool: on Linux, a realm can grant a Claude seat an MCP server, and the seat
calls it only through Brokkr's own broker. The release also reads the
journal a page at a time, keeps an adopter's Brokkr in one .brokkr/
folder, and adds a run queue, journal backups and boxed hands for dsh. It
includes the changes merged since
v0.12.0.
This release refuses an adopter repository still in the 0.12 layout,
stops seats on Linux from starting your own MCP servers, and parks a
timed-out attempt instead of retrying it. Read
Upgrade and compatibility before you install
it over a working setup.
What changed
-
MCP tools, through Brokkr's own broker (decision 0065, slice two).
- Granting a server. Write a tool dialect of kind
mcp(a stdio
server's argv, its pinned version and its tools), and grant it in the
realm map'scapabilities. For each capability a seat holds, the
engine adds acap-<capability>server beside the box'sworkspace
hands. The harness starts it asbrokkr broker serve, Brokkr's broker
(decision 0077).brokkr doctorshows each seat's whole capability
plan. - Each server in its own box. The broker runs the real server in a
box of its own: its program read-only, a private home and temp
directory, no workspace, a cleared environment, and only the network
its egress class allows. - The broker is the gate. It lists only the granted tools, refuses
any other call by name, and masks every bound secret in a response.
Each call is in the attempt's ledger before it is forwarded or
refused, and lands in the journal as aseat-record.v6checkpoint when
the attempt settles. - Kept responses. A dialect may keep its responses, and the realm may
veto that withretain: false(forge.realms/v8).
brokkr inspect --run <run> --capability-call <id>opens a kept
response, verified against its digest. - No ambient MCP on Linux. Every model seat is launched with only the
servers Brokkr names: none, the box's hands, or its brokers. Your own
MCP configuration no longer reaches a seat. A harness not measured to
shut it out is refused at compile, and the final launch check refuses
a server set that differs from the sealed one. - A real seat, end to end. A real Claude seat holding a grant
completed a run, and its call's response read back through
brokkr inspect. On the way, the broker learned to answer any MCP
initializewith its own version (2025-06-18), a brokered Claude
launch allows each broker asmcp__cap-<capability>, and the broker
closes its ledger on SIGINT or SIGTERM. - Host checks.
brokkr driversetsno_new_privsbefore a brokered
harness, and refuses where it cannot. MCP servers refuse to start where
/proc/sys/dev/tty/legacy_tiocstiis absent or not0. - One run per worktree, for every run. Every dispatched attempt
reserves its sites, whether or not it brokers anything. A second run
whose writers would reach a worktree that a standing attempt writes is
refused until the first is proven settled.brokkr operator release
journals your attestation that a run's ended attempts left no process,
and frees the worktree. Where HOME does not resolve to a protected
coordination root, an attempt that protects no evidence proceeds
unreserved, and one that does refuses. - Harder boxes. Every workspace box hides the secrets store and opens
in a user namespace that can make no further one. - Limits. Linux only: on macOS every
mcpgrant still refuses at
compile. Claude only: a brokered grant on a Codex or dsh seat refuses
at compile. The seat needs workspace hands undernamespace, with the
network off. A server may bind a secret on a Claude work seat: Claude
Code 2.1.316's boxed replacement session is measured
(#958). URL
connections, secrets in the argv and nonempty restrictions refuse. - Documented. The guides, the status page and the security model say
what this slice built, and what it did not.
#956 - Audited. Before this release, every slice-two unit was checked
against its removals, its proofs and the frozen contract bytes. The
audit changed no production code; it corrected one evidence citation
in Codex's shipped adapter declaration. The proofs it found missing
are repaired by slice two's remaining units, after this release.
#963
#524,
#552,
#640,
#745,
#804,
#827,
#834,
#853,
#875,
#924,
#933,
#940,
#952,
#955 - Granting a server. Write a tool dialect of kind
-
dsh seats get boxed hands (decision 0098). A seat whose office
declares boxed hands runs dsh 0.2.1 with its ten effectful tool rows
disabled, so Brokkr'sworkspaceserver is its only effectful tool.- The driver checks dsh's composed configuration before the session and
refuses an unexpected row by name. A hand that did not activate fails
the seat closed. - On Linux the seat runs on an engine-only dsh home, measured to shut out
ambient MCP. - An office may declare
access: read-onlyorwrites; one that
declares neither reads aswrites. - Brokkr reads dsh 0.2's session transcripts, plain and compressed.
- The driver checks dsh's composed configuration before the session and
-
A journal read a page at a time (decision 0089). On a 1.1 GB journal,
0.12.0's console held 2.6 GB andbrokkr runs --jsontook 6.7 s. A page
now costs the same at any journal size.- Projections. The writer keeps a summary row per run, a seat row per
participant and an index row per event, in each append's transaction.
brokkr doctorchecks them against a fresh fold. - The fleet. The console (
brokkr tui),brokkr runsand
/api/runsread a page of summary rows at a time;brokkr runstakes
--limitand--after. The console refreshes only what landed since,
off the input loop (decision 0088). - One run. A run is opened by verifying its chain one event at a
time, and a corrupt run is refused as before.inspect,seats,
watch,transcript,costs,compareand/api/viewread through
the pages and print the same bytes. The run view's first screen costs
15 ms and 250 KB at any journal size. - The console's run screens read pages too. A movement key pressed
while a page loads is held, in order, and applied when that page
arrives, so the cursor moves exactly as it would on a whole read. A
held key belongs to the view it was pressed on, and is dropped if that
view leaves the frame (a narrower or shorter terminal, or another run,
seat or tab). The queue holds 64 runs of keys; an overflow drops the
oldest and says so.
#964
- Projections. The writer keeps a summary row per run, a seat row per
-
One folder for Brokkr in every repository (decision 0083).
brokkr init <dir>writes.brokkr/and nothing else, with a
.gitignorethat keepsstate/out of git.--cli <claude|codex|dsh>
picks the agent CLI instead of PATH's order.- Every verb finds the folder from anywhere in the repository, and keeps
the journal, secrets and results under.brokkr/state/. The folder is
the default bundle, sobrokkr run --feature "…"needs no--bundle. brokkr doctor's first line names the folder. In a dsh folder it says
whether dsh reaches a key for each route the folder pins.- The guides place a personal realms map outside the repository, at
$XDG_CONFIG_HOME/brokkr/realms/<name>.json, selected with--realms
(decision 0101).
swarmail engine a9893b4-p4
Not a brokkr release. This is the exact engine binary that feedback-loop-ai/swarmail pins in .brokkr/recipes/self/tooling.json (formerly bundles/self/tooling.json), published so that CI and the verify seat can fetch and check one artifact by digest. It replaces swarmail-engine-a9893b4-p3: same base and first three patches, plus patch 0004, which raises the hands tool's per-command timeout maximum from 10 to 60 minutes so a seat can run a cold Rust build and full test suite in one command.
- Base:
a9893b4dfb67417edc99d91f941a9bac3856f138(this tag's target) - Patches, applied in order: the four
000*.patchassets, identical to swarmail'stooling-patches/; 0001–0003 are the previous release's (upstream form: brokkr PR #899, still open) - Build:
cargo +1.98.0 build --release --locked --offline --bin brokkr(rustc 1.98.0, 88d9e12ae 2026-08-18) - Binary sha256:
29dc3309925a2556b9713b7a9b5fb8089b8e4610e86215fc89554e75634e7cad - Patched diff sha256: recorded in
build-receipt.json(the digest ofgit diff --binaryof the patched tree against the base)
SHA256SUMS covers every asset, and build-receipt.json is the operator's build receipt. x86_64 Linux only.
Swarmail pinned engine: 0.12.0 at a9893b4 + 3 patches
Not a brokkr release. This is the exact engine binary that feedback-loop-ai/swarmail pins in .brokkr/recipes/self/tooling.json (formerly bundles/self/tooling.json), published so that CI and the verify seat can fetch and check one artifact by digest. It replaces swarmail-engine-731189f-p3 because this base carries brokkr's accepted decision 0096 (attempt correlation, 9fa4f620).
- Base:
a9893b4dfb67417edc99d91f941a9bac3856f138(this tag's target) - Patches, applied in order: the three
000*.patchassets, identical to swarmail'stooling-patches/and to the previous release's (their upstream form is brokkr PR #899, still open) - Build:
cargo +1.98.0 build --release --locked --offline --bin brokkr(rustc 1.98.0, 88d9e12ae 2026-08-18) - Binary sha256:
4873928c94f2677ef552c9e3ed04785a732851b89e8a413c13ce1434d6066be6 - Patched diff sha256: recorded in
build-receipt.json(the digest ofgit diff --binaryof the patched tree against the base)
SHA256SUMS covers every asset, and build-receipt.json is the operator's build receipt. x86_64 Linux only.
Swarmail pinned engine: 0.12.0 at 731189f + 3 patches
Not a brokkr release. This is the exact engine binary that feedback-loop-ai/swarmail pins in bundles/self/tooling.json, published so that CI and the verify seat can fetch and check one artifact by digest (swarmail #206).
- Base:
731189f9097308c01ae4f2b4c9caea5f41224844(this tag's target) - Patches, applied in order: the three
000*.patchassets, identical tobundles/self/tooling-patches/in swarmail - Build:
cargo +1.98.0 build --release --locked --offline --bin brokkr(rustc 1.98.0, 88d9e12ae 2026-08-18) - Binary sha256:
ab93555806ee4ac402457e40c18250f0eead8e6a0588eb8c508682fe0c6a885f - Source composite sha256:
c972dde7785915743fb8d7365ad5bb5efecafd29be78520911f0cb4d74a75e5b
SHA256SUMS covers every asset, and build-receipt.json is the operator's build receipt. x86_64 Linux only.