What changes in Cardumen, release by release.

Auto

Auto picks the model by cost, efficiency and quota

#13167d893b5e

Every model you can launch gets three 0–100 scores: price (using the reference price even on a subscription), efficiency for the task type from the Artificial Analysis indexes, and available quota. The mode —balanced, savings or quality— sets how much each one weighs. A model beaten by a newer one from its own family that is more capable and no more expensive stops competing. The catalog refreshes once a day; launching makes no network call for it.

Auto

Jev classifies the task and chooses among the top candidates

#1315cbe1cef8

With the Jev engine on (the default), one short call classifies the task and chooses among the three best-scored candidates. It uses Cardumen's key, not yours, and every decision is recorded with its cost. If it does not answer in time, the launch continues with the locally computed order.

settings

The model ranking is visible in Settings

#13200eaf7828

In Settings → "Agents & Auto", below the model the next launch will use, a table lists every candidate with its cost, efficiency and quota scores, the weights in use, and the ones that do not compete, with the reason.

settings

"Agents & Auto" panel with a single switch

#1296ad9bcad5

Auto has one switch and one model blocklist, in a single panel. Settings that were repeated across screens are gone.

settings

"Account & usage" and "License & updates" panels

#1297ad699e68

Account, usage, license and updates move to two panels of their own, in the new Settings design.

settings

"Automation" panel with a single set of limits

#1298ea4a126d

Autonomous work and its limits —concurrent sessions, reviews, budget— are configured in a single panel.

sessions

Every session launches as a minion or an orchestrator

#13188da0cbb4

When creating a worktree or a terminal you can set the role with --role; without it, a brief with a single deliverable is marked as a minion. Auto uses the role to choose weights: a small model is enough for a minion.

sessions

One task per session, with a notice when it needs escalation

#131730931b6e

A brief with several deliverables warns before launching, and a session that cannot finish its task says so with needs_parent or blocked instead of improvising. The rule ships in the minion-task skill.

sessions

Automatic handover when a session runs out of context or stalls

#12856c578e87

If a session runs out of context or stops making progress, Cardumen hands the work over to another one that continues from where it stopped.

worktrees

New worktrees start with dependencies and a build cache

#1280d2885e4f

A freshly created worktree already has its dependencies installed and the build cache ready, including when created from the CLI.

sessions

Launching classifies the brief and applies its workflow

#12641758dd08

On launch, Cardumen classifies the brief by what it asks and by the repository, and the session starts with the workflow and skills for that kind of work.

CLI

CLI decisions use Cardumen's key

#1327b7f99f8f

The CLI decision engine no longer reads your OPENROUTER_API_KEY: it calls the Cardumen API with your license, under the same monthly budget and usage record.

agents

OpenCode receives the brief, or the launch says so

#1342f9df9aaa

With OpenCode 1.19 the brief could fail to arrive and the session stayed on the start screen while the command answered "ok". It now waits long enough and, if the brief does not show on screen, the launch fails with agent_launch_not_live.

Auto

A balance failure is confirmed with you before sidelining the agent

#1338b875f4c0

When a session ends with a balance or quota error, the notice asks you to confirm it. With cardumen agent balance --agent <id> --exhausted Auto leaves that agent out for 6 hours; with --available it keeps it and does not ask again in that window.

agents

Codex no longer offers efforts the model rejects

#13352975391c

Auto could launch Codex with the minimal effort, which the provider rejects, and the session died on its first turn. Each model's effort list now matches the one Codex declares.

Auto

Jev decisions arrive in time

#1333231381b9

The first decision after an idle period took too long and was lost. The service now stays warm and reuses its database connection: a full decision takes around one second.

Auto

The cost of every Jev decision is recorded

#13360346f596

Usage was stored with zero cost and the monthly cap was never consumed. The real cost of each decision is now recorded.

settings

The Model row lists each agent's real models

#13115bb02b52

In "Agents & Auto", the model picker shows the models each agent actually offers.

Arrecife

Launching ties the session to its card, or says why it cannot

#127768589907

A session launched with an Arrecife card stays linked to it, including when the claim is in the agent owner's name; if linking fails, the reason appears in the result.

worktrees

Verify's internal copies no longer show up as projects

#1307548a350d

The working copies verify creates are no longer listed as worktrees, and orphaned ones are removed at startup.

sessions

Sessions without orphan processes and with truthful end notices

#1273cb5f012a

Closing a session also closes its processes, closed terminals do not come back, and the end-of-turn notice says what happened: a child process ending no longer counts as the session ending.

sessions

A brief sent to a terminal is delivered, or you are told

#12655752c84b

Text sent to a new terminal reaches the agent or the command reports that it did not; terminal send no longer answers "not delivered" when it was delivered.

agents

OpenCode no longer updates itself mid-session

#1284026f2668

OpenCode sessions start without self-update, a provider 429 does not stop the session, and the watchdog no longer leaves it without commands.

Auto

The model you choose by hand reaches the agent

#12627f7b6562

A model given explicitly at launch is honored, and the launch receipt reflects the model actually used.

Auto

Auto stops discarding models it can use

#125411a37c8d

The ranking flagged valid models as insufficient and did not rank those with unknown quota. Both cases now compete normally.

CLI

Guards stop blocking legitimate work

#126782a2477e

The deletion and environment guards resolve the command's repository and no longer stop safe commands.

verification

The full verify fits its budget

#1289b402d0de

verify --full finishes within its time budget, reports every failure reason in a single attempt and does not lose an interrupted run.

desktop

Search in the OpenCode model picker

#1246c9a4ac3d

In Settings → Agents, the OpenCode default-model picker filters as you type by name, id or provider («go glm» leaves only the opencode-go GLM models) and hides empty groups.

verification

«verify --full» runs the desktop tests related to the diff

9b5c3dd5

«cardumen verify --full» could go green without running the desktop tests. It now uses the file list verify.sh computes against the base and runs the related tests; it never skips the desktop silently.

verification

«verify --full» is no longer rejected before it starts

9e5d37c1

The variable that marks full mode was not on the allowed-variables list, so the command was rejected in under a second. It is now accepted, with the values 0 and 1 only.

agents

The settings skill includes the idle notice

6ee8cd7b

The settings map agents consult was regenerated to include the new session idle-notice setting.

agents

Agent sessions stop stalling silently

#12508aea34ed

Hooks no longer cut the agent's turn (they become advisories unless the repository policy requires blocking). The orchestrator is notified when a session ends its turn or sits idle for 5 minutes. Launches check that the first message arrived and resend it if not. OpenCode sessions on models that degrade restart past 180,000 context tokens. Closing a worktree ends the processes Cardumen started inside it. Terminals keep their agent and model after an app restart.

Auto

Auto honors your excluded-models list

#1245b6dcca58

With «Auto active» off, or when falling back to the default agent, Auto could launch a model you had excluded. The exclusion list now always applies: to the pick, to the fallback and to shadow mode.

license

Updates no longer fail on heavy-usage accounts

#12446c73c740

The usage sum took longer than the server limit on accounts with a long history, and the update check returned an error. A new index brings it down to tenths of a second.

Auto

The 1.10.0 changelog covers the calibrated Auto ranking

#1243e6ae2870

The public 1.10.0 entry includes the Artificial Analysis ranking calibration that lets Auto clear the quality floor.

Auto

Auto panel in Settings → Agents

#1226a9ded50b

Shows what Auto would pick now and why (each subscription's quota, model rankings and your settings), how it works, how often it would have chosen differently in shadow, and the «Auto active» switch to turn it on right there.

Auto

Auto can apply its model and effort choice

#12171b93c561

With «Auto active» turned on in Settings → Auto, each launch uses the agent, model and effort the router picks; anything you request explicitly always wins and is recorded. It ships turned off.

Auto

The router decides with real data

#12149f7d51a0

Auto resolves in the main process with each subscription's quota, model rankings, your mode, budget and exclusion settings, and the card's real task category.

Auto

Every launch records what Auto would pick

#1215c60a47a7

Routing receipts store the router's decision next to what was launched, and cardumen route explain --shadow shows the differences so you can measure before turning it on.

desktop

A notice when an agent is waiting for you

#119858d63ab1

When a session ends its turn with a question, its launcher gets «waiting for reply»; notices no longer repeat and say whether verify is queued or running.

verify

Verify fails in seconds, not minutes

#1178081ff38b

Static and generated-artifact checks run before the suites. cardumen verify --fast returns that result in about 2 s without taking the verify slot, and a skills-only change no longer runs the full desktop suite.

verify

PR verify queues itself on push

#1177456a2f7f

Cardumen queues the verify on push, waits until GitHub has the commit, defers dirty working trees and publishes the result saved per SHA. It can be turned off per project in Settings → Repository.

desktop

PR auto-correction that keeps you informed

#103718496d3d

When a PR has failing checks or no verify, Cardumen queues verification or launches a managed corrector, notifies when it starts and finishes, and waits if a verify is already running.

graph

Compact graph built on demand

#107898b82c06

graph.db is compacted, graphs of inactive repos are pruned, it is built only when needed and stale cache results are rejected.

Auto

Auto picks models from a measured ranking

#1122abfa05fd

The Auto router ranks routes with Artificial Analysis signals and efficiency measured on real sessions, with a global ranking for paid plans.

Auto

Models excluded from automatic selection

#11240837e5f0

Settings → Auto lets you exclude specific models so the router never picks them.

desktop

Automatic handoff when quota runs out

#1071cad6c88d

When an agent runs out of quota, the router picks the handoff using remaining quota compared across providers, and OpenCode resumes the same session.

desktop

A routing receipt for every session

#10759c20cd4b

Each session shows which agent, model and effort were chosen and the alternatives, and lets you override the choice and give feedback.

mcp

Secrets requested outside the agent chat

#11502f6f2dec

secrets_request asks for the value through a link (URL-mode elicitation), so the secret never passes through the agent conversation.

security

Protected merges need human approval

#9901725f4c4

Merging into a protected branch requires a person to approve it from the app; an agent cannot approve itself.

telemetry

Anonymous statistics, off by default

#109895d619c0

Anonymous statistical reporting exists but ships turned off; it only runs if the user opts in.

Auto

The Artificial Analysis ranking finally counts

#1242925ec305

Artificial Analysis scores are compared against the best model of each index before applying the quality floor; before, no model cleared it and Auto always fell back to the default agent. The receipt shows the normalized and the original score.

Auto

With Auto you can launch any agent

#1232a658e419

When the default agent is Auto, asking for a specific agent, model or effort is no longer blocked in the app, the CLI or the hooks. Only a fixed agent in Settings restricts it.

Auto

The CLI decides with the same data as the app

#1233a5cfbcbe

CLI launches ask the main process for the decision: same quota, same base ranking and same settings. If something is missing, the receipt says what and for which agent.

Auto

Quota from every subscription and detection of all 9 agents

#1239ab0e50d3

The router refreshes quota before deciding and reads it from every provider, not only Claude and Codex. The app finds Codex and OpenCode even when opened from the Dock.

license

A device error no longer blocks the CLI

#1237001b3ffb

The app only marks the license as revoked when the server confirms it, like the CLI; a device mismatch or an invalid response no longer stops agents.

Auto

Auto explains its choice and works from the CLI

#122555d1f490

«Auto will use» no longer comes from a fixed list but from the router, with agent, model, effort and reason. cardumen worktree create goes through the same router, records its shadow decision and can launch with Auto as the default agent.

verify

Automatic verify no longer gets lost

#1201fdaa4717

Each queued commit is verified in a disposable checkout even if the agent switches branches; it also fires when the push comes before the PR, after a rebase, and without duplicate turns.

worktrees

Creating a worktree from the CLI no longer launches an agent by itself

#11907126d55b

The runtime only launches the default agent when asked with --prompt or --work; --agent still launches the chosen one.

graph

The graph can no longer freeze the machine

#120609d0a1fe

Only one graph build runs per machine at a time, and queries answer from the cache while it rebuilds in the background.

desktop

Effort selector for Claude Haiku

#11959ba44c21

Haiku 5.5 reasons like the other Claude models and now shows its effort levels.

hooks

Agents no longer ask permission to push their own branch

#119222763c39

They only ask for force pushes, protected or deploy branches and irreversible actions; pushing the working branch and opening the PR need no confirmation.

hooks

Hooks no longer stall the session

#117470cdccaa

Every hook has a hard 4 s deadline and falls back to a neutral answer. SessionStart moves network, online license checks and coordination to the background: from 6–10 s to a measured 0.57 s.

runtime

The daemon no longer dies with queued work

#1162af1d9efd

The coordination daemon keeps pending work, and the graph is sealed by content instead of size, which avoids needless rebuilds.

worktrees

Creating a worktree no longer duplicates or launches agents

#11543703aef6

worktree create does not duplicate an existing worktree and only launches an agent when asked with --agent, --prompt or --work, which checks the card again.

hooks

More precise safety guards

#108259e99af5

The session guard blocks environment stripping and secret dumps, and catastrophic-deny no longer blocks searches that merely mention destructive SQL.

hooks

Far less context in every session

#11637d170819

Low-value hooks are off by default. Session start context drops from 11.4 KB to 1.6 KB and per-prompt context from 2.9 KB to 0.5 KB.

agents

Compact agent instructions

#11655aff6605

The repo AGENTS.md goes from 37.8 KB to 1.9 KB, below the size at which Codex silently truncated it, and the cardumen init template shrinks too.

verify

A new commit cancels the previous verify

#11765fa22b94

When a newer HEAD arrives, the running verify of the previous one stops instead of finishing for nothing, and the queue history is kept.

verify

A flaky test no longer sinks the verify

#1123b0d769b5

Going over the wall-clock budget is a warning, not a failure. A failing Vitest file is retried once in isolation and, if it passes, is recorded as flaky in the receipt.

interface

English interface by default

#1058481903fc

The app starts in English and migrates users who had the system language set, so nobody switches language by accident.

desktop

The installer retries its own identity if ps is slow

45a06637

On macOS, ps with 1 s failed under verify --full and the hook lock was thrown away. It retries up to 3 times with 5 s. No invented identity: a live lock is not stolen.

quality

The QA report wraps any bare URL when it is saved

11d5442c

write_qa_report covers criteria, notes, and console errors, not only two producers. A qa run or an older binary no longer breaks markdownlint MD034 in docs/evidence/.

verify

Queued verify turns are no longer reaped for waiting

5a4384f9

A queued verify spends no CPU; the 10 min stall threshold no longer deletes it. Only pid and heartbeat count. After 10 min behind another, the turn stays and does not force a full rerun.

verify

Local verify fits in 10 min, with shared Rust cache and explicit CI

#97585f881c2

sccache is shared across worktrees; what does not fit in 10 min (queue included) is delegated on the receipt, never green. --full remains the full net. A scripts-only package.json no longer scales to workspace.

graph

Graph ensure rejects $HOME and does not duplicate the build per root

#97007447382

cardumen graph ensure on HOME hit 8 GB and swap. It now requires a git worktree, takes that root's lock, and exits if an ensure is already live. Desktop does not launch a detached ensure on HOME.

desktop

The launcher notice rides the same session-ended event

#973eaf46adf

Closing the child fires the notice and the notification together. It no longer depends on another pane's status that sometimes never arrives.

hooks

Process ownership no longer depends on the PTY sid

#9718cdaf2fc

Each OpenCode bash opens a new Unix session and the guard killed what the session itself spawned. cwd/worktree or the same OS session now wins. A heredoc that quotes kill does not count as a command.

quality

The QA generator wraps URLs so markdownlint passes

726ec7eb

landed_url and the authenticated-screen criterion come out in backticks. A test fails if the generated report has a bare URL. Reports from #964 and #965 are covered at the source.

quality

Evidence QA reports no longer leave bare URLs

55bbe3b1

The two reports that had broken --full on main are corrected and the generator stops emitting MD034. You no longer patch docs/evidence/ by hand on every UI PR.

Arrecife

Agent identities are isolated per installation

#96614bb9ccf

Each desktop profile stores its own keys. Renew rotates only the requested agent. QA or a second userData does not stomp the keys of the app you actually use.

verify

DATO_INERTE no longer blocks code that production actually uses

#9699526d524

Methods via this., parameters, obj.prop, and i18n copy no longer count as a dead binding. The waiver, if needed, is per name (allow-inert-binding:), not global.

desktop

Auto picks the agent from detected hosts

#965197dab8d

The Auto pill resolves an installed, enabled host. A disabled agent is skipped; a launch with no host stays put instead of inventing Codex.

desktop

Settings splits the Arrecife session from agent keys

#964599bea3f

Renew my session and Rotate this agent's key are different actions. You no longer rotate the human seat when you only wanted the minion's key.

desktop

Queued child-session notices drain on their own

#967ea0fe2f3

The queue retries with backoff (2 s → 10 s) even if the launcher is not settled yet. A done without a final answer is still delivered, as (no final answer), instead of sleeping forever.

remote hosts

Remote-host registration keeps the diagnostic

#963c94591cc

Account discovery and connect share the same runtime; registration does not wait on the relay being ready. You see why pairing failed, not a mute error.

verify

The impact map ties hook identity to its related tests

59b29c16

Touching managed-hook-owner-identity.ts no longer fires the full desktop suite. It selects lock, Codex backfill, daemon, and the child-process boundary.

research

The graph vs rg and repo-map bench is on record, with tokens, time, and cost

#9772c24095a

Ten tasks measure the graph against rg + repo-map: tokens, wall, RAM, and cost. Evidence lives in benchmarks/grafo-vs-repo-map/ so the debate is not from memory.

verify

Verify --full caps vitest workers when it runs beside Rust

1b25305f

In --full desktop and nextest run in parallel; vitest no longer opens one worker per core. It exports VITEST_MAX_WORKERS to half unless you already set it. On an 18-core Mac the pool no longer dies on timeout.

desktop

Launcher notices survive a stale handle

#9530a46468a

If the launcher handle no longer answers, the child-session notice is still stored. A missing probe does not swallow the notification.

desktop

Saved agent args are defaults, not a hard override

#952a162378c

What stayed persisted in Settings fills the launch; an explicit arg from this session wins. It no longer rewrites the command you just asked for.

Cardumen

The merge guard resolves the repository selector

#951b6298aff

The guard maps the selector to the PR's real repo. A merge with -R owner/repo anchors receipt and base there, not on the session cwd.

desktop

Opt-in automatic correction of PR conflicts and checks

#960b4f850e7

In Settings you can ask an agent to resolve conflicts or red checks on the open PR. If no session is live, it launches the default agent; at the attempt cap it notifies the launcher and stops.

verify

Temporary base worktrees stay protected

#959d24aa0bd

Verify no longer lets another process delete or stomp the base worktree while it compares the delta. The receipt stays anchored to that base even if the original checkout moves.

desktop

Child-session launcher notices are no longer duplicated

#95813375080

Each close sends one exact, unique notice. A stale handle or a second Stop does not fill the history with the same child session.

Arrecife

Agent MCP wiring uses a key per identity

#957efb1fe9d

Each agent carries its CARDUMEN_ARRECIFE_<id>_KEY. MCP no longer shares the human token or another agent's key on the same host.

Arrecife

Agent identities are persisted and can be checked

#9553a5c7f3b

Settings stores each identity's state and lets you check it against Arrecife. You see whether the key is still live before launch, not after MCP already refused.

verify

Verify --scope pr scopes tests and reuses the base cache

#961d4fa6003

Desktop tests for a PR are chosen by impact and the comparable base is cached. A --scope pr no longer repeats the whole suite when the base did not change.

Arrecife

Owner WIP shows agent keys

#947e925d1e4

The owner's WIP list includes the agent keys that hold the work. You stop matching the board and the terminal by hand to see who has it.

Cardumen

Notifies the launcher when a child session ends

#941480490cb

When a session launched from another one ends, the launcher gets a notice (with history and a Settings policy). You no longer watch the child terminal blind.

desktop

One interface language, without mixed fallback

#9387fa7d900

The desktop uses one whole language (English by default). You no longer see a screen half Spanish and half English because one key was missing.

desktop

PR conflicts reach the live agent session

#9401af41251

If the PR collides, the agent that is working gets it in the terminal. It no longer finds out only after the merge is already red.

desktop

Arrecife authorization progress is visible

#937206f68a4

Connecting Arrecife shows steps, not a mute spinner. You see whether it is on identity, token, or node, and where it stuck.

verify

Verify can target one module in the monorepo

61841318

In a tree with several packages, verify runs the module you asked for instead of the whole net. The receipt names that cut.

desktop

Arrecife agent identities can be delegated

#9227de537ee

Each agent starts with its native Arrecife identity, not the human's. MCP and WIP see the right actor.

verify

Verify --full measures against origin/main even if the host exports the file list

#950bbde0526

An empty CARDUMEN_VERIFY_FILES list is no longer read as a clean branch. --full still measures the delta against origin/main and does not fire the whole desktop+workspace.

verify

The semaphore test stops the process group, not an extra binary

2226750b

The holder runs in its own group and is cut with libc. CI does not depend on a tool the runner does not ship, and no orphan yes remains.

verify

The semaphore test no longer leaks a yes that hangs verify --full

2ae3c1d9

Stopping the holder also kills the child yes. The stderr pipe closes and --full does not sit 3600 s waiting for EOF.

hooks

The delivery-receipt check no longer duplicates the None branch

f0d43ec3

Clippy -D warnings no longer fails verify --full on if_same_then_else. The two None conditions join with ||; the receipt rule is unchanged.

desktop

The Arrecife runtime no longer imports Electron

14cc55ad

Agent identities leave the runtime-electron ratchet. IPC opens links; the pure runtime does not pull in electron.

skills

The Settings skill includes launcher notices

b2594197

After the launcher notice, the generated skill lists launcherNoticeHistory and the PR-conflict policy. Cmd+J and the sidebar can find them.

desktop

Codex wires Arrecife MCP again

#9487e180305

Codex launch puts Arrecife MCP where the host reads it. It no longer starts without the Arrecife board or tools.

desktop

Desktop timing tests are no longer flaky

#946817fb9b6

Clock-sensitive cases use fixed instants. A desktop verify no longer goes red because of a setTimeout.

hooks

The guard detects protected git and in-place edits

#945f2e0e5e0

A hook sees whether the command touches a protected repo or rewrites the tree in place. The agent gets the block before it harms main.

hooks

The merge guard gets an explicit repo

#943cb6be088

It no longer infers the repo from the session cwd. It anchors receipt and base on the PR repo being merged.

Arrecife

Arrecife MCP no longer sends the retired profile query

#942014168f2

Arrecife ignores ?profile= since profiles were retired. Desktop and probe URLs stop adding it; the tool list follows the caller's authorization.

skills

Skill coverage matches the canonical roots

#93644f13bd6

Agents read skills from the owning roots. An unowned provider copy is not published as the canonical catalog.

desktop

The retired Meta model provider is gone

#9351488f23e

Settings and the session catalog no longer list Meta. An old default does not try a model the host does not serve.

Cardumen

Retired Arrecife MCP hosts are migrated

#93345216458

On open, old host configs move to the current format. An agent does not start pointing at an endpoint that no longer exists.

desktop

Quota launch options use the Cardumen org

#93497c7e753

The quota picker no longer mixes in an Arrecife org. You launch with the Cardumen seat cap.

desktop

Worktree create degrades if Arrecife cannot claim

#9314adcf650

If the claim fails, the worktree is still created and the error lands on the receipt. The tree is not thrown away for an Arrecife 403.

hooks

The guard tells shell reads apart from disk changes

#93096a10807

A cat or a git log does not trip the same block as a redirect to a file. The inner loop lets inspection through and cuts mutation.

Arrecife

Agent MCP defaults to the all profile

#9293d12c6ee

If a profile is still needed, the default is all. A new agent does not start with an empty tool cut.

desktop

Linux systemd can hand off the update

#924163dc235

Serve under systemd no longer cuts the update halfway. After install, the unit gets the handoff and the new binary is in service.

core

Agent browser operations go through the embedded browser

#92760426322

Browser tools do not open a loose Chrome. They go to Cardumen's embedded browser, with the same session and the same isolation.

Arrecife

The agent MCP probe fails closed on a fake identity

#9284f8595ad

PASS requires a real agent_ id, not the words kind/actorId from the prompt. A lying green probe no longer slips through.

Arrecife

Agent MCP host formats are aligned

#926fff13408

Claude, Codex, Cursor, and OpenCode get the same shape. A probe does not clobber foreign config entries while migrating.

desktop

Arrecife agent home and environment stay separate

#925ef0a2070

Keys live under desktop state, not a HOME mixed with the human's. The process environment does not inherit secrets from another seat.

desktop

Foreign Arrecife MCP entries are kept

#923e998a8d1

When rewriting the agent MCP, entries that are not Cardumen's stay. A server you added by hand does not vanish on the next launch.

desktop

The floating window can leave the app and come back

#918c4d23324

Pop out opens a native window of its own with the same tabs and PTYs. Coming back hooks them into Cardumen chrome again.

telemetry

Copilot cost is read from session-store.db

#91969b19b53

Cardumen reads assistant_usage_events read-only and queues usage so the server can estimate cost from the catalog. Copilot no longer goes missing from spend.

deploy

Release checks Docker and cargo-xwin before it starts

#92162079d32

If the target includes linux, docker info must answer; if it includes windows, cargo-xwin must exist. 1.9.19 does not fail again 33 min into Linux.

desktop

There is a QA channel; pnpm run dev refuses inside agent sessions

#92094b5291e

dev:qa uses its own userData and bundle id, without the CLI front door. An agent cannot raise the dev profile and stomp your app.

graph

The task map covers briefs without a file or symbol

#9172a2ed0db

graph query search looks up a substring in graph.db and the map uses the brief's words. A product brief no longer comes out empty.

graph

Graph A/B measures Claude from its transcript and checks the map

#9161ea0c4e1

With Claude it counts tool_use from the transcript, not a weak proxy. A/B fails if the map was not actually injected.

SQL

Repricing leaves the 2026-09-26 price migration

#91474576599

The migration only inserts rates. Historical repricing goes in a separate job, so deploy does not die on statement_timeout like 1.9.19.

verify

Verify --full scopes Rust to touched crates; --all keeps the workspace

86e84af8

The long pole of --full is no longer cargo test --workspace. It measures crates plus direct dependents against origin/main. --all remains the full net.

verify

Verify --full runs only impacted desktop tests; --all keeps the suite

a03b084e

The ~64k desktop tests were 85% of --full wall clock. It now selects by diff from origin/main. --all keeps the whole suite.

evidence

Live Arrecife and Cardumen API integration probes are on record

55a9a5f3

The integration probe against the local stack is documented with HTTP 200 on both boundaries. Anyone touching identities has the evidence.

memory

The 2026-09-29 owner decisions are on record

#93927edb50e

The orchestrator reports ready PRs and the owner merges; integration is proven on the local stack; UI in one language, English by default. It is written down.

version

The CLI shows the channel version, not Cargo's

6e0a8b39

cardumen --version uses CARDUMEN_DISPLAY_VERSION and, when set, CARDUMEN_BUILD_LABEL. It tells develop apart from a real release.

daemon

The packaged daemon on macOS waits longer to boot

c51741a7

The macOS boot timeout rises so native modules can load. Packaged-daemon verify no longer fails on a cold start.

Settings

Settings can turn on result-spec mode

6670a821

There is an enabled flag in settings for result-spec mode. You turn it on from Settings, not by editing JSON by hand.

gates

One door: notices, deadlines, and one kill-switch

#8330aa66efd

Gates, githooks, and the inner loop go through one emitter with a notice and ops-log. Merge and deadlines stop being one-off rules.

Cardumen

Inner gates ease off for the single door

#8312984e984

The first third of one door eases result-spec, Stop, and the inner loop into notice mode. Hard blocking stays on the shared door.

CLI

task check-verify warns before you run verify

#813273e297b

cardumen task check-verify is the preflight: you see whether the contract and scope close before spending a semaphore turn.

org

Desktop sees the org workspace root the CLI declares

#80555484e3b

The workspace-root fact comes from the CLI and Settings reads it from the active worktree. They no longer diverge.

CLI

The CLI diagnoses incomplete organization roots

#8032dd65677

Cardumen names what is missing in the org profile (root, plan, seat) instead of a generic error. Setup aims at the real gap.

desktop

Launch declares which agent account was selected

#800f426d036

The account preference rides into the receipt and the agent process. It no longer starts with a different identity than the one you saw in Settings.

verify

Verify credits the session delta against the base

#775f792239c

The receipt compares what this session broke against origin/develop. An inherited red does not fail the work you actually fixed.

CLI

Organization profile status lives in one place

#799d350d650

CLI and desktop read the same org diagnostic. You no longer get a green pane and a red CLI for the same seat.

CLI

A direct UI QA runner from the CLI

#79082915828

Cardumen launches visual QA without a side harness. The report keeps locale and comes out clean to attach to the card.

C.R.E.A.

C.R.E.A. records spec amendments on the linked card

#791603c6050

If the contract changes and passes validation, it lands on the Arrecife card. The successor sees the live spec, not the one from start.

skills

The canonical skills catalog is reconciled across agents

#785faa7f8ff

One catalog feeds Claude, Codex, Cursor, and OpenCode. A skill no longer exists on one host and is missing on another.

landing

The landing replaces intro and solution visuals

#724b135c891

Intro and the solution block use new pieces. Plans, checkout, and analytics stay where they were.

Cardumen

Alerts when a license heartbeat goes dead

#7378e7204fe

If the node stops beating, there is an alert. You do not discover a dead seat only after a launch already failed.

API

Usage reports the identity crosswalk explicitly

#735520394fe

The usage API names how person, seat, and provider cross. It no longer infers an id from another field.

quality

There are quality rollups per tenant

#694fef7acc3

Quality aggregates per organization, not only per repo. The board sees the whole tenant.

Cardumen

Sessions get domain-scoped Arrecife skills

#6937d1f080c

On start, the card's domain skills come in, not the whole catalog. Less noise, same contract.

brain

SessionStart serves cached org memory

#691892b5e01

Cached organization memory is injected when the session opens. It does not wait for a fresh pull when the cache is valid.

usage

Usage quotas per provider and per team

#6817bf149ab

There is a cap per provider and per team; if the quota cannot be read, launch fails closed. You pick the quota at start.

Grow 360

There are organization team rollups

#6889ee20c4c

Grow 360 aggregates per team, not only per person. The project rollup reaches people who are actually on those teams.

API

The API lists tenant brain items

#6890fe0cbe9

There is a brain-item list scoped to the tenant. One customer does not see another's brain.

usage

Impact metrics are declared as evidence

#6800caa2270

Usage declares the impact metric with evidence, not a loose number. You can audit where it came from.

CLI

A launch flag to turn computer-use off

#690b63f17a4

The launch flag leaves the session without computer-use. Opt-in per session, not a silent default.

landing

New public home without breaking plans or checkout

#6494eaf6dda

Cristobal's home lands on develop. Pricing, checkout, and analytics stay; only the public face changes.

graph

Update acceptance report for epic with new metrics and task-map failure

2640b4c2

Update acceptance report for epic with new metrics and task-map failure. The case is covered in the binary; it does not depend on a clean worktree.

CI parity

Declarar perl en BINARY_ALLOWLIST

#911b69533c8

Declarar perl en BINARY_ALLOWLIST. If it came back every session, it is now closed for good.

desktop

Marcar el agente que espera un comando que ya no existe

#9109af3692e

Marcar el agente que espera un comando que ya no existe. The symptom no longer shows up on the happy path.

verify

La línea base no espera el cerrojo de caché de su padre

#90979fcaae6

La línea base no espera el cerrojo de caché de su padre. A new agent no longer trips on this edge.

graph

Mapa de tarea con fichero:línea y firma, una sola vez

#905c0d0323d

Mapa de tarea con fichero:línea y firma, una sola vez. The next run does not inherit the failure.

core

Toda espera de cerrojo de fichero con tope y error claro

#9069db0c18a

Toda espera de cerrojo de fichero con tope y error claro. The receipt and the UI match this fix.

graph

Fijar que el grafo rematerializado desde graph.db equivale al build completo

#90723eec2f4

Fijar que el grafo rematerializado desde graph.db equivale al build completo. Anyone hitting this no longer needs a local patch.

desktop

OpenCode sin --variant no soportado y sin encargo tecleado en la shell

#902161792ff

OpenCode sin --variant no soportado y sin encargo tecleado en la shell. After install, the flow continues without a bypass or an odd flag.

graph

El brain rollup lee el resumen desde graph.db, no graph.json

#903fcd9f361

El brain rollup lee el resumen desde graph.db, no graph.json. The case is covered in the binary; it does not depend on a clean worktree.

graph

Expose graph_query MCP to every launched host

#90475464c61

Expose graph_query MCP to every launched host. If it came back every session, it is now closed for good.

graph

Automatizar el A/B del grafo con agente real

#9012bb943b6

Automatizar el A/B del grafo con agente real. The symptom no longer shows up on the happy path.

desktop

Pulir cabecera y orden del mapa de tarea

#8998b981af5

Pulir cabecera y orden del mapa de tarea. A new agent no longer trips on this edge.

graph

Add epic acceptance check

#898b068a63a

Add epic acceptance check. The next run does not inherit the failure.

graph

Isolate shared graph from worktree refresh

#89523f510f5

Isolate shared graph from worktree refresh. The receipt and the UI match this fix.

core

Safely prune legacy worktree states

#897bcf536ee

Safely prune legacy worktree states. Anyone hitting this no longer needs a local patch.

graph

Prioritize code edges in neighborhoods

#896ecc7ac49

Prioritize code edges in neighborhoods. After install, the flow continues without a bypass or an odd flag.

desktop

Resolve task maps from repo graph

#894f4c5b47d

Resolve task maps from repo graph. The case is covered in the binary; it does not depend on a clean worktree.

guards

Serialize global git test state

#89089596ef8

Serialize global git test state. If it came back every session, it is now closed for good.

verify

Make full scope run complete suites

#89270c62625

Make full scope run complete suites. The symptom no longer shows up on the happy path.

gates

Complete hard-set regression coverage

#8915253410c

Complete hard-set regression coverage. A new agent no longer trips on this edge.

mobile

Stabilize Android audio focus lifecycle

#8896b168102

Stabilize Android audio focus lifecycle. The next run does not inherit the failure.

desktop

Stabilize flaky updater tests

#885118038d5

Stabilize flaky updater tests. The receipt and the UI match this fix.

desktop

Keep RPC creates alive past socket idle

#884241b9857

Keep RPC creates alive past socket idle. Anyone hitting this no longer needs a local patch.

automations

Rebind local imports

#882aed6356e

Rebind local imports. After install, the flow continues without a bypass or an odd flag.

desktop

Record effective agent credential source

#881d2c88207

Record effective agent credential source. The case is covered in the binary; it does not depend on a clean worktree.

desktop

Preserve agent presence during slow startup

#88042656bc6

Preserve agent presence during slow startup. If it came back every session, it is now closed for good.

automations

Refresh runs after manual dispatch

#879053a42d2

Refresh runs after manual dispatch. The symptom no longer shows up on the happy path.

desktop

Toggle explorer folders instantly

#8774e3b6d0e

Toggle explorer folders instantly. A new agent no longer trips on this edge.

desktop

Allow multiline proxy bypass rules

#8786fdaf10b

Allow multiline proxy bypass rules. The next run does not inherit the failure.

sidebar

Stabilize worktree scroll, keyboard navigation, and pinned host labels

#8742574299a

Stabilize worktree scroll, keyboard navigation, and pinned host labels. The receipt and the UI match this fix.

desktop

Preserve markdown scroll after image layout

#87644ad0613

Preserve markdown scroll after image layout. Anyone hitting this no longer needs a local patch.

source control

Keep narrow sidebar content readable

#8759082ebd7

Keep narrow sidebar content readable. After install, the flow continues without a bypass or an odd flag.

desktop

Let terminal themes override imported colors

#873a16f205a

Let terminal themes override imported colors. The case is covered in the binary; it does not depend on a clean worktree.

editor

Restore readonly preview behavior

#872227c5639

Restore readonly preview behavior. If it came back every session, it is now closed for good.

Cardumen

Stabilize new workspace project selector and Create more

#8714d718200

Stabilize new workspace project selector and Create more. The symptom no longer shows up on the happy path.

desktop

Stabilize flaky scheduling tests

#8708aab0533

Stabilize flaky scheduling tests. A new agent no longer trips on this edge.

desktop

Avoid duplicate project groups during refresh

#8691b4644fc

Avoid duplicate project groups during refresh. The next run does not inherit the failure.

tabs

Cancel drag gestures on window blur

#868ed58d334

Cancel drag gestures on window blur. The receipt and the UI match this fix.

terminal

Preserve idle Linux IME direct commits

#86740d88a9f

Preserve idle Linux IME direct commits. Anyone hitting this no longer needs a local patch.

desktop

Coalesce advertised URL port scans

#8667cacf272

Coalesce advertised URL port scans. After install, the flow continues without a bypass or an odd flag.

verify

Pipe macOS ps CPU output

#86422d2c07a

Pipe macOS ps CPU output. The case is covered in the binary; it does not depend on a clean worktree.

desktop

Type partial ssh2 mock default

#8656c60db68

Type partial ssh2 mock default. If it came back every session, it is now closed for good.

desktop

Evict stale runtime skill discovery cache

#863035a5c20

Evict stale runtime skill discovery cache. The symptom no longer shows up on the happy path.

desktop

Skip empty hidden restore frames

#862b16934d0

Skip empty hidden restore frames. A new agent no longer trips on this edge.

desktop

Route runtime-owned SSH image pastes

#8616d5a7195

Route runtime-owned SSH image pastes. The next run does not inherit the failure.

desktop

Release browser results and fence guest registration

#8590836366c

Release browser results and fence guest registration. The receipt and the UI match this fix.

desktop

Release parked terminal scroll intents

#858c212eab2

Release parked terminal scroll intents. Anyone hitting this no longer needs a local patch.

desktop

Preserve ssh2 utils in host key wiring mock

#860ac08e2f1

Preserve ssh2 utils in host key wiring mock. After install, the flow continues without a bypass or an odd flag.

desktop

Prevent false remote terminal closures

#857408579d9

Prevent false remote terminal closures. The case is covered in the binary; it does not depend on a clean worktree.

desktop

Honor SSH stream metadata callback in provider mocks

#856ea06633f

Honor SSH stream metadata callback in provider mocks. If it came back every session, it is now closed for good.

desktop

Detach aborted auth filesystem waits

#854a14ffdc1

Detach aborted auth filesystem waits. The symptom no longer shows up on the happy path.

desktop

Regenerate bundled skill catalog

#85552a5e07f

Regenerate bundled skill catalog. A new agent no longer trips on this edge.

desktop

Unwrap Electron IPC handler errors

#8537918c193

Unwrap Electron IPC handler errors. The next run does not inherit the failure.

hooks

Mark Cardumen-launched Copilot sessions

#8523ae96c65

Mark Cardumen-launched Copilot sessions. The receipt and the UI match this fix.

desktop

Preserve crash breadcrumb lifecycle trail

#851759262ab

Preserve crash breadcrumb lifecycle trail. Anyone hitting this no longer needs a local patch.

daemon

Avoid invalid snapshot scans and PTY claims

#85014922b26

Avoid invalid snapshot scans and PTY claims. After install, the flow continues without a bypass or an odd flag.

browser

Bound CDP output backpressure

#846a590a1d1

Bound CDP output backpressure. The case is covered in the binary; it does not depend on a clean worktree.

desktop

Optimize terminal PTY hot path

#847fb73aac6

Optimize terminal PTY hot path. If it came back every session, it is now closed for good.

Grow 360

Route Carnada through Arrecife

#841fd3e4d86

Route Carnada through Arrecife. The symptom no longer shows up on the happy path.

desktop

Handle Unicode-normalized filesystem roots

#842b3823879

Handle Unicode-normalized filesystem roots. A new agent no longer trips on this edge.

native chat

Reject unlisted launch default models

#843069dac15

Reject unlisted launch default models. The next run does not inherit the failure.

desktop

Release consumed runtime and SSH queue entries

#84891ab3129

Release consumed runtime and SSH queue entries. The receipt and the UI match this fix.

mobile

Prevent markdown stalls and timestamp drift

#845bbfd3bea

Prevent markdown stalls and timestamp drift. Anyone hitting this no longer needs a local patch.

desktop

Bound worker archive lines linearly

#844f72c2f07

Bound worker archive lines linearly. After install, the flow continues without a bypass or an odd flag.

graph

Query SQLite projection directly

#839c3773783

Query SQLite projection directly. The case is covered in the binary; it does not depend on a clean worktree.

hooks

Allow policy guard when Cardumen is disabled

#8374652235b

Allow policy guard when Cardumen is disabled. If it came back every session, it is now closed for good.

desktop

Resolve macOS CLI shim bundle executable

#838bad74589

Resolve macOS CLI shim bundle executable. The symptom no longer shows up on the happy path.

hooks

Unify exit code handling for hooks

2354650c

Unify exit code handling for hooks. A new agent no longer trips on this edge.

hooks

Enhance global hook directory resolution and legacy guard handling

c5544d64

Enhance global hook directory resolution and legacy guard handling. The next run does not inherit the failure.

usage

Bigint telemetry accumulators and new official list prices

#834ea2a50f1

Usage accumulators move to bigint. GPT-6 and Claude 5 list prices land; the xAI price is current without dragging fork brand names.

hooks

Unify managed tool-call hooks

#832e56fc303

Unify managed tool-call hooks. Anyone hitting this no longer needs a local patch.

hooks

Unificar identidad de sesión Copilot en gate, task y verify

#8303f84e437

Unificar identidad de sesión Copilot en gate, task y verify. After install, the flow continues without a bypass or an odd flag.

backoffice

Resolve org external ids for scoped usage

#82901d8438b

Resolve org external ids for scoped usage. The case is covered in the binary; it does not depend on a clean worktree.

backoffice

Align Stripe fake with product retrieval

#828a82367a2

Align Stripe fake with product retrieval. If it came back every session, it is now closed for good.

desktop

Keep Vitest module cache out of tmpfs

#8244aed115c

Keep Vitest module cache out of tmpfs. The symptom no longer shows up on the happy path.

Cardumen

Route org profile tests through path helper

#825fbc23a7d

Route org profile tests through path helper. A new agent no longer trips on this edge.

minions

Inject card contracts and scope skills by role

#816185f054e

Inject card contracts and scope skills by role. The next run does not inherit the failure.

MCP

Export registry to all supported agents

#8141d88e0cd

Export registry to all supported agents. The receipt and the UI match this fix.

Cardumen

Retry macOS Electron dependency downloads

#819d93976eb

Retry macOS Electron dependency downloads. Anyone hitting this no longer needs a local patch.

desktop

Clean Vitest temporary user data

#811c7714547

Clean Vitest temporary user data. After install, the flow continues without a bypass or an odd flag.

guards

Adopt attribution from dead sessions

#817a6911b66

Adopt attribution from dead sessions. The case is covered in the binary; it does not depend on a clean worktree.

Cardumen

Validate result spec verification surfaces

#815a3e39f89

Validate result spec verification surfaces. If it came back every session, it is now closed for good.

CLI

Warn when registered desktop lags

#8121556dcdc

Warn when registered desktop lags. The symptom no longer shows up on the happy path.

Cardumen

Restore quick-checks on develop

#810c6bc4a3a

Restore quick-checks on develop. A new agent no longer trips on this edge.

Cardumen

Arreglos del modo fast (hook de Stop, gate de entrega, borrado de planes, perfil de org)

#820fcdb83b1

Arreglos del modo fast (hook de Stop, gate de entrega, borrado de planes, perfil de org). The next run does not inherit the failure.

install

Onboard organization licenses

#798679bd002

Onboard organization licenses. The receipt and the UI match this fix.

quality

Use canonical QA report path in tests

#796e30ca4d7

Use canonical QA report path in tests. Anyone hitting this no longer needs a local patch.

desktop

Keep QA Arrecife actions inside card

#7976968d8f5

Keep QA Arrecife actions inside card. After install, the flow continues without a bypass or an odd flag.

desktop

Create SSH project setup with worktree

#794deaf312e

Create SSH project setup with worktree. The case is covered in the binary; it does not depend on a clean worktree.

core

Remove needless borrows in skill installer

#792b6de99f5

Remove needless borrows in skill installer. If it came back every session, it is now closed for good.

landing

Restore quick-checks on develop

#786c40fad55

Restore quick-checks on develop. The symptom no longer shows up on the happy path.

C.R.E.A.

Offer linked card remedy and flag gate loops

#7823d556e4f

Offer linked card remedy and flag gate loops. A new agent no longer trips on this edge.

core

Resolve skill home through paths boundary

#7843be3003d

Resolve skill home through paths boundary. The next run does not inherit the failure.

desktop

Guide unregistered device setup

#783733b5766

Guide unregistered device setup. The receipt and the UI match this fix.

checkup

Inspect wired hook shims

#778c87f0d51

Inspect wired hook shims. Anyone hitting this no longer needs a local patch.

memory

Persist session notes outside worktree

#7771bec1be9

Persist session notes outside worktree. After install, the flow continues without a bypass or an odd flag.

llama.cpp

Normalize streamed tool arguments

#7740b9aeb6a

Normalize streamed tool arguments. The case is covered in the binary; it does not depend on a clean worktree.

C.R.E.A.

List accepted verify forms for unresolved targets

#769e29685b1

List accepted verify forms for unresolved targets. If it came back every session, it is now closed for good.

router

Preserve managed llama context when profile is missing

#76741873ee0

Preserve managed llama context when profile is missing. The symptom no longer shows up on the happy path.

desktop

Retry transient macOS release failures

#766bc586d9e

Retry transient macOS release failures. A new agent no longer trips on this edge.

C.R.E.A.

Coteja verificación con alcance declarado

#764f3cbc446

Coteja verificación con alcance declarado. The next run does not inherit the failure.

task

Preserve atomic quick validation test

#7614988f75e

Preserve atomic quick validation test. The receipt and the UI match this fix.

C.R.E.A.

Validate resultSpec verification commands

#75530c16d5f

Validate resultSpec verification commands. Anyone hitting this no longer needs a local patch.

quality

Require repository evidence before insufficient info

#757c4e2e0bc

Require repository evidence before insufficient info. After install, the flow continues without a bypass or an odd flag.

Cardumen

Honor configured default agent in worktree create

#7273891fbb3

Honor configured default agent in worktree create. The case is covered in the binary; it does not depend on a clean worktree.

CLI

Detect queued session input

#754e7d03ed6

Detect queued session input. If it came back every session, it is now closed for good.

deploy

Disable AWS CLI pager

#73816a73b7a

Disable AWS CLI pager. The symptom no longer shows up on the happy path.

Cardumen

Deduplicate cardumen desktop fixture path

#7514be98565

Deduplicate cardumen desktop fixture path. A new agent no longer trips on this edge.

Cardumen

Enforce pinned pnpm for workspace installs

#746c2c96472

Enforce pinned pnpm for workspace installs. The next run does not inherit the failure.

quality

Read v2 targets in status

#748a2821fa4

Read v2 targets in status. The receipt and the UI match this fix.

desktop

Catch undeclared test-source dependencies

#7471c07dc6c

Catch undeclared test-source dependencies. Anyone hitting this no longer needs a local patch.

C.R.E.A.

Cotejar resultSpec al cerrar

#74108b89701

Cotejar resultSpec al cerrar. After install, the flow continues without a bypass or an odd flag.

usage

Distinguish unreported Cursor usage

#74473c18828

Distinguish unreported Cursor usage. The case is covered in the binary; it does not depend on a clean worktree.

desktop

Answer Claude trust on worktree launch

#74327966767

Answer Claude trust on worktree launch. If it came back every session, it is now closed for good.

desktop

Identify blocked Codex startup dialogs

#732662038b6

Identify blocked Codex startup dialogs. The symptom no longer shows up on the happy path.

desktop

Confirm text-only terminal delivery

#7287c7a77c7

Confirm text-only terminal delivery. A new agent no longer trips on this edge.

desktop

Keep worktree removal working when the cleanup CLI is missing

#731e5a6c087

Keep worktree removal working when the cleanup CLI is missing. The next run does not inherit the failure.

Cardumen

Fail Cargo verification when no tests run

#7300c61f862

Fail Cargo verification when no tests run. The receipt and the UI match this fix.

desktop

Isolate relay harness from AWS deps

#7236708b3ff

Isolate relay harness from AWS deps. Anyone hitting this no longer needs a local patch.

runtime

Publish QA report atomically

#726fd703cc5

Publish QA report atomically. After install, the flow continues without a bypass or an odd flag.

Cardumen

Limpia estado huérfano de proyectos

#7215afa5ab9

Limpia estado huérfano de proyectos. The case is covered in the binary; it does not depend on a clean worktree.

desktop

Release reentrancy slots after exceptions

#720d123f551

Release reentrancy slots after exceptions. If it came back every session, it is now closed for good.

Cardumen

Unblock C.R.E.A. result spec gate

#71571ed6e4f

Unblock C.R.E.A. result spec gate. The symptom no longer shows up on the happy path.

Cardumen

Confirm brain rollups before dedupe

#717dd3e8bbf

Confirm brain rollups before dedupe. A new agent no longer trips on this edge.

git hook

Use canonical disabled marker path

#718981b4831

Use canonical disabled marker path. The next run does not inherit the failure.

desktop

Detect incomplete worktree setup

#70315e9fcae

Detect incomplete worktree setup. The receipt and the UI match this fix.

orchestrator

Scope scratchpad exception to host session

#7124dc8f522

Scope scratchpad exception to host session. Anyone hitting this no longer needs a local patch.

Cardumen

Classify baseline verify failures at stop

#714a27a443d

Classify baseline verify failures at stop. After install, the flow continues without a bypass or an odd flag.

verify

Release stalled semaphore turns

#7081eeb2ce2

Release stalled semaphore turns. The case is covered in the binary; it does not depend on a clean worktree.

verify

Stabilize Rust scope gate test

#711f7da2c43

Stabilize Rust scope gate test. If it came back every session, it is now closed for good.

Cardumen

Respect disabled switch in pre-push

#707215d21c6

Respect disabled switch in pre-push. The symptom no longer shows up on the happy path.

Cardumen

Share cardumen kill switch across worktrees

#70501793258

Share cardumen kill switch across worktrees. A new agent no longer trips on this edge.

CLI

Route Codex models through Codex auth

#701f07d26ec

Route Codex models through Codex auth. The next run does not inherit the failure.

desktop

Declare tinyglobby for census tests

#7002761c68d

Declare tinyglobby for census tests. The receipt and the UI match this fix.

Cardumen

Fold team project aliases to canonical keys

#682cbc5e0aa

Fold team project aliases to canonical keys. Anyone hitting this no longer needs a local patch.

desktop

Derive terminal dialog guard from screen

#69857a0b011

Derive terminal dialog guard from screen. After install, the flow continues without a bypass or an odd flag.

CLI

Recover stale graph probe after merges

#69944528bbc

Recover stale graph probe after merges. The case is covered in the binary; it does not depend on a clean worktree.

CLI

Remove needless borrows in worktree test

#697ae3b68f0

Remove needless borrows in worktree test. If it came back every session, it is now closed for good.

C.R.E.A.

Fail closed on cargo metadata output errors

#6954bfc2471

Fail closed on cargo metadata output errors. The symptom no longer shows up on the happy path.

Cardumen

Preserve prompt on worktree first run

#696f278f234

Preserve prompt on worktree first run. A new agent no longer trips on this edge.

desktop

Deliver stalled worker signals to coordinator

#6851ef9400b

Deliver stalled worker signals to coordinator. The next run does not inherit the failure.

autopilot

Persist crash status and report sessions

#68451cf9cba

Persist crash status and report sessions. The receipt and the UI match this fix.

desktop

Guard OpenCode sends from shell fallback

#67888a0f4c4

Guard OpenCode sends from shell fallback. Anyone hitting this no longer needs a local patch.

Cardumen

Reclaim idle verify semaphore turns

#674f7a7010e

Reclaim idle verify semaphore turns. After install, the flow continues without a bypass or an odd flag.

Cardumen

Make computer-use opt-in per session

#67745fb4631

Make computer-use opt-in per session. The case is covered in the binary; it does not depend on a clean worktree.

Cardumen

Supervise autopilot worker failures

#679db866f8c

Supervise autopilot worker failures. If it came back every session, it is now closed for good.

Cardumen

Clean temp files left by tests

#66937cf3a71

Clean temp files left by tests. The symptom no longer shows up on the happy path.

hooks

Read merge verify receipt from PR worktree

#676e93555eb

Read merge verify receipt from PR worktree. A new agent no longer trips on this edge.

hooks

Keep RESULT_SPEC remedies unblocked

#6738c8bca85

Keep RESULT_SPEC remedies unblocked. The next run does not inherit the failure.

desktop

Make folder repo upgrade polling deterministic

#672661ab2f3

Make folder repo upgrade polling deterministic. The receipt and the UI match this fix.

orchestrator

Respects the orchestrator scratchpad and heredocs

#66874e6d777

The orchestrator does not clobber the session scratchpad or heredoc blocks. Working material stays where the host left it.

verify

Cap session verification runtime

#671a6f9c4e6

Cap session verification runtime. After install, the flow continues without a bypass or an odd flag.

router

Keep execution refactors out of plan mode

#663d1e19028

Keep execution refactors out of plan mode. The case is covered in the binary; it does not depend on a clean worktree.

guards

Protect cross-session processes and repo state

#6662c0bd3d4

Protect cross-session processes and repo state. If it came back every session, it is now closed for good.

Cardumen

Align pnpm and bootstrap desktop tests in worktrees

#667ad407cce

Align pnpm and bootstrap desktop tests in worktrees. The symptom no longer shows up on the happy path.

desktop

Keep OpenCode prompts out of argv

#66503aab55d

Keep OpenCode prompts out of argv. A new agent no longer trips on this edge.

C.R.E.A.

Reconcile orphan task contracts

#664b39dd434

Reconcile orphan task contracts. The next run does not inherit the failure.

Cardumen

Allow focused local package test commands

#66287c7e19a

Allow focused local package test commands. The receipt and the UI match this fix.

worktrees

Ignore disposable build output on removal

#65718c3eb96

Ignore disposable build output on removal. Anyone hitting this no longer needs a local patch.

Cardumen

Seal baseline-only verify receipts

#661669d3823

Seal baseline-only verify receipts. After install, the flow continues without a bypass or an odd flag.

Cardumen

Isolate visual QA browser sessions

#660660f07a6

Isolate visual QA browser sessions. The case is covered in the binary; it does not depend on a clean worktree.

desktop

Rollback failed terminal creates

#65995701863

Rollback failed terminal creates. If it came back every session, it is now closed for good.

hooks

El guard de gh pr merge resuelve repo, base y recibo desde la PR

#65549f5e2ab

El guard de gh pr merge resuelve repo, base y recibo desde la PR. The symptom no longer shows up on the happy path.

OpenCode

Run hooks asynchronously and cap stop retries

#656193973ad

Run hooks asynchronously and cap stop retries. A new agent no longer trips on this edge.

desktop

La sección de MCPs abre aunque un servidor no tenga headers ni env

#65332a026e8

La sección de MCPs abre aunque un servidor no tenga headers ni env. The next run does not inherit the failure.

CLI

Isolate test git commands from inherited repositories

#65445566649

Isolate test git commands from inherited repositories. The receipt and the UI match this fix.

graph

The graph epic acceptance report is green again

#912938e918e

The task-map FAILs were pnpm/corepack, not the code. The report is 7/7 again and restores the agent A/B.

graph

Task-map A/B evidence and the graph day's notes are on record

#908191f19bc

The report includes measurement with and without the map, plus 2026-09-27 notes. Anyone picking up the graph does not start from zero.

telemetry

How to measure Copilot's local persistence is on record

#82687b3b2c5

The telemetry guide documents which local file counts and how not to mix it with inventory. Fewer false positives.

quality

Model-selector viewport evidence is on record

#8219224ba38

QA leaves selector shots at real widths. A Settings change can be checked without reopening the visual debate.

Cardumen

The OpenCode account-selection request is on record

#807fb3267e4

The note describes what OpenCode must declare when picking an account. Launch can require the flag without guessing the host.

quality

The org-root warning after the fix is captured

#8060e0c6316

QA documents the post-fix state of the workspace-root warning. A bug that is gone does not get reopened.

Cardumen

Organization-profile end-to-end evidence is on record

#80462021c65

The org-profile flow is captured end to end. A regress shows up in the evidence, not in production.

Cardumen

Org workspace-root warning evidence is on record

#802ac2d600e

Before the fix, the warning was photographed. It is the baseline so we do not return to the broken state.

desktop

The isolated desktop-dev QA recipe is on record

#789012c9e4a

How to run desktop QA without touching the real profile. Evidence included so the pass can be repeated.

Cardumen

Why the macOS runner keychain password changes is on record

#7883382c033

The note explains the runner signing symptom and what not to touch. Less time lost on the same CI error.

Cardumen

The Settings skill names the correct MCP list command

#683b0949a28

The documented verb matches cardumen mcp list. It no longer points at a subcommand that does not exist.

MCP

One MCP registry, user or project, for each agent

#64755dacc96

Settings and cardumen mcp keep one canonical registry (user or project) and export it to Claude, Codex, Cursor, and OpenCode grammar. stdio, HTTP, and SSE; credentials stay in the host store.

desktop

Settings discovers each agent's live session models

#646d40ce54a

Claude, Codex, OpenCode, and Pi/Antigravity no longer depend only on seeds. The session-defaults editor lists what the agent actually has now.

desktop

Desktop gates that broke develop pass again

#651c3b56b1d

After four merges, typecheck and cross-cutting tests are green: MCP search registered, update handoff typed, live Codex inventory as source. Develop is no longer red from PR collisions.

desktop

One Arrecife session shared across every surface

#64811899890

Connect, brain, artifacts, WIP, and worktree claims use the same session. You stop signing in twice for the same account; Settings shows one connection state.

updater

Linux systemd installs can update the desktop

#645c5f2fefa

Serve under systemd no longer blocks the update as unsupported-headless-serve. After install, it restarts the unit. cardumen version --check no longer says up to date when the desktop does not declare a version.

verify

Verify runs rust and desktop by impact, gates in parallel

#64453a247ea

The impact map feeds import-graph; rust only touches dependent crates; gates run in parallel. Less waiting without cutting coverage.

desktop

A clean quit no longer relaunches LaunchAgent

#643fac553a2

Serve stays headless and stops from the menu bar. A real quit unloads the agent; it no longer comes back on its own after you close the window.

verify

Verify --full aims at five minutes when warm

#642ea935e8f

Loud impact selector, rust by touched crates, no duplicated suites, coord timeout. The 5 min cap warns on the receipt; it only blocks when you ask.

CLI

mcp set-enabled takes true or false by hand

#652eac71b3d

cardumen mcp set-enabled docs false now parses the bool as a value, not a flag. Without true/false the command fails instead of turning the server on by accident.

evidence

Verify accepts a concrete commit as the base

6dacc324

CARDUMEN_VERIFY_BASE and BASE_SHA now check that the ref exists in the repo. If the environment names a SHA that is missing, the committed diff does not break: it uses a base that is there.

desktop

Bash 5.3 no longer marks an extra command in the prompt

#640c0d31e8d

On bash 5.1+ each PROMPT_COMMAND piece runs on its own and the DEBUG trap fired an extra OSC 133 C marker. The wrapper now brackets the prompt with sentinels and the PTY stops painting ghost commands.

task map

The task map reads the local graph before the CLI

#628868fd0ce

If cardumen graph query times out, the map now opens graph.json from the checkout. Empty sections are skipped; you see 'no matches' instead of four hollow blocks.

merge

Merge looks at the receipt in the PR worktree

#63807dae8be

The guard no longer credits a green verify from the session directory. It anchors the receipt on the PR worktree (or the command's cd). The remedy is to rerun verify there, not a dangerous bypass.

OpenCode

OpenCode attributes edited files to the session

#629e6b6a84b

Each edit lands in the session ledger (async hook, non-blocking). verify --scope session sees those files. If the host cannot attribute, Stop asks for --scope workspace instead of a waiver.

OpenCode

OpenCode starts with less noise and async hooks

#632e036cf42

Startup context does not name .claude/ paths on hosts that are not Claude. Session-start and prompt-submit run in the background. /tmp, /dev and worktree state are preauthorized; .cardumen-out/ does not trip C.R.E.A.

desktop

Desktop typecheck on develop passes again

#6392781c4ef

The session-card via union includes take, and the brain test no longer invents fields TypeScript rejected. Three errors that broke develop are closed.

graph

The graph is reused across worktrees via overlay

#63474a3b634

If the cached graph is ahead of the worktree HEAD, ensure accepts it as overlay instead of failing. Several worktree create in parallel reuse the base graph under lock, via temp+rename.

brain

Brain setup survives a missing notice channel

#6379d238645

BrainNodeSetup no longer blows up if readBrainSetupNudge is missing from preload. Optional-chain and catch: the wizard continues, and tests do not depend on a mock nobody had set.

brain

A rollup stays visible to whoever publishes it

#62424024a6c

On publish, the project joins the pusher's teams. The CLI lists those teams and warns if the seat cannot see the rollup. brain resolve tells 'not visible' apart from 'missing'.

OpenCode

OpenCode's shell exports the session id to the CLI

#6365bf72da6

export CARDUMEN_SESSION_ID prefixes every bash. task start/spec/verify stay bound to the session. Stop without attribution asks for --scope workspace and no longer offers a waiver.

minions

The Codex roster honors CODEX_HOME

#62382ce510f

codex_roster_path() no longer looks only at ~/.codex. If CODEX_HOME points at another tree, minions routes with that roster. An empty HOME no longer hides the Codex you did configure.

brain

Brain setup notifies on startup and surfaces errors

#622d0f1f646

If the seat profile is incomplete, the desktop shows the notice on open. The wizard includes connecting Arrecife; background node failures arrive over IPC instead of being swallowed.

brain

Brain shows the real blocker and a setup button

#62176971817

There is no generic no_org_plan anymore: missing profile, root, or plan. The 'Complete brain setup' button runs the missing steps. You stop guessing why pull does not start.

MCP

MCP export keeps your config and adds the local MCP

#62004b3c650

Claude, Cursor, and OpenCode merge only the MCP key; the rest of the JSON is kept. OpenCode uses ~/.config/opencode/opencode.json with its own grammar. Codex touches only the managed block. All four export cardumen mcp serve --stdio.

verify

Local verify runs cargo test the same as CI

234cdc88

verify.sh passes --all-targets and --quiet to cargo test, like the CI job. A test that only lives outside tests/ is no longer green locally and red on GitHub.

CLI

The binary version stays readable with ThinLTO

14e562f4

ThinLTO on macOS x86_64 ate the user-agent. There is a static marker and the deploy scripts extract the version again. An embedded changelog no longer counts as a fake version.

CLI

Context handoff now carries real evidence

e55979b7

When context use crosses 80%, session-context-handoff.json includes the contract, touched files, findings, and the next step. The successor does not start blind.

CLI

Session review flags a context-high handoff

777e6238

Session review shows context-high attention when the model nears the ceiling. It appears in the list; you do not guess why the session stopped.

CLI

Context handoff is persisted in the worktree

e97264bd

On a context cut, Cardumen leaves the handoff JSON in the worktree and builds the next session's opening prompt. Finished work is not lost in the cut.

verify

The verify queue can be queried

051dc789

cardumen verify queue-status (and --json) lists who waits, who runs, since when, and in which worktree. You no longer guess whether the semaphore is stuck.

CLI

Minion metrics per session, model, and effort

676d1813

The CLI aggregates terminals with a foreground agent by effective model and effort. A shell does not count as a minion even if the runtime stamps defaults.

CLI

Sessions autopilot: nudges without stepping on a live turn

32ff60c5

sessions autopilot looks every ~90 s and sends only if the terminal is still an agent, not Working or in background, and past cooldown. It uses cardumen terminal, not a side channel.

C.R.E.A.

Flags a deadlock when the same gate blocks in a streak

7902a542

If a gate (e.g. result-spec) cuts several times in a row, session review and the terminal show gate-deadlock with the exact remedy. It no longer looks like a one-off block.

CLI

Handoff tests no longer leave inert bindings

74846f6c

The super helpers that were only assigned in tests were reshaped so verify does not flag them as dead API. Handoff behavior is unchanged for whoever uses it.

CLI

The handoff artifact counts as a touched file

a28a25af

touched_files includes session-context-handoff.json. The next session sees the file in the inventory, not only the prompt.

CLI

Context exhaustion leaves the session incomplete

a6b2ed71

If the model runs out of context, close does not fake a done. The state stays incomplete and the handoff is how you continue, not a silent success.

verify

Verify isolates worktree and test environment

883b1911

Verify commands pin the worktree and do not inherit the host's global settings. A test or a launch does not pollute another checkout's queue or home.

verify

An explicit --scope session on HEAD is preserved

acba50aa

verify.sh no longer widens a hand-requested --scope session on current HEAD to workspace. The cut you asked for is the one that runs.

verify

Stub lint ignores function text inside fixtures

6bda7bc7

A string that looks like a function (in a test or an example) does not trip the detector. Only executable code in the hunk is scanned.

verify

Session evidence looks at live content

183dfe4e

verify --scope session hashes the files the session touched in the live tree, not a stale snapshot. A reverted file no longer counts as green evidence.

verify

The configured worktree is kept when a turn is acquired

4c3088de

When the semaphore is acquired, the worktree path you asked for is not replaced by the process cwd. Two checkouts no longer fight over the same turn.

verify

The verify adapter gets the real worktree

fb6e4ba1

The binary passes the concrete checkout to the script, not a guessed one. A verify launched from another directory stays bound to that tree.

verify

The semaphore stores the worktree path on the turn

68c4db87

turn-*.state records and the bash helper note the worktree. queue-status and a stuck turn name the right tree.

verify

A function declaration is not an inert binding

13c89adb

The orphan-symbol detector no longer treats fn foo as a dead assignment. Session review can name helpers without failing verify.

verify

An inert helper is gone and the probe fallback is stable

63ebe112

Session review no longer exposes a helper nobody called. The integration-probe fallback does not depend on that symbol.

CLI

Session review passes the clippy that used to fail verify

a0284439

Small borrow fixes in sessions_review so clippy -D warnings does not leave review red in CI.

CLI

Fewer inert-binding false positives in review and minions

dfbeff62

Minion aggregates and desktop review use the bindings. Verify stops flagging variables that are actually read in the output JSON.

verify

The scoped verifier stays clean for the detector

a41ff183

Receipt and internal lints are reshaped so the verify crate does not report itself. A small --scope can actually be green.

verify

Session scope is preserved when archiving

f67552c9

verify.sh does not drop --scope session when copying the tree into a probe archive. The cut reaches the child.

verify

Stub scan skips the verifier's own implementation

940f61e9

The detector excludes itself more precisely. A linter return no longer looks like a productive stub.

verify

Scoped delivery checks are ratcheted

16bfd7c0

A --scope session or delivery check does not require the full suite to go green. The receipt records what actually ran.

verify

A probe archive does not trip false stub hits

c2143763

The scan ignores archive copies used by the probe. A temp tree does not inherit a stub red the origin does not have.

verify

The integration probe uses the real cwd

bd860acb

integration-probe-gate.sh resolves the root from the directory it runs in, not from a baked-in path. In a linked worktree it aims at the right tree.

verify

Scoped checks stay green inside the archive

08b56693

When verify runs on an archive, out-of-scope steps are not marked red. The probe does not widen the receipt with failures it did not ask for.

CLI

Autopilot passes the clippy that kept it out

3adbaf1c

A targeted allow in sessions_autopilot so the crate builds with -D warnings. The pilot loop is unchanged.

CLI

The verify baseline stays green

06f12d76

Two detector annotations avoid an inherited red when touching verify. Coverage does not drop; it avoids a false that was already on HEAD.

verify

Accented prose is not treated as a binding

289bb575

Spanish words (función, sesión) in comments or strings no longer look like orphan identifiers. The detector wants a real symbol shape.

CLI

Autopilot stays verifier-friendly

276573d4

The pilot avoids patterns stub-detect and clippy flag. It still reads cardumen terminal and respects Working/background/cooldown.

CLI

Session review stays green for verify

7e686e61

A comparison in classify is adjusted so the crate does not fail the verify gate when review is touched. Attention rules are unchanged.

C.R.E.A.

Remedies stay unblocked and task quick is atomic

d9a27b08

The sandbox message names the host remedy (Claude, Codex, Cursor). task quick does start+spec; if spec fails, it undoes start. No half contract remains.

CLI

The background-task hold passes clippy

816ffd16

The text that names the background task is formatted without warnings. The hold still states handle, agent, and reason.

verify

Desktop test budget does not escalate to workspace

92340e6f

Touching only the desktop test ceiling no longer fires the workspace suite. Verify stays in the desktop scope. A ratchet does not cost the full net.

verify

An expected regression-probe receipt is accepted

cbc368aa

If the revert-the-diff probe is inconclusive or the expected outcome, the receipt does not go red. A fix that cannot be probed that way does not block the push.

verify

Verify receipts are scoped per worktree

f0d0e687

Pre-push reads the receipt of the tree being pushed, not a neighbor. Two worktrees of the same repo no longer share a green that is not theirs.

verify

Masked stub-marker literals are ignored

19fd0ce3

A string that quotes unknown or a stub marker, split or escaped, does not trip the detector. Only a constant return in new code does.

verify

Verify fixtures survive a rebase

6e53d496

Baseline and runtime tests do not depend on SHAs or sizes a rebase moves. A replay on develop does not go red because of the fixture.

verify

Environment isolation pitfalls are on record

bb2ac7bb

The auto-guard skill documents global settings, HOME, and worktree that leaked into verify. Anyone isolating a test has the list of what to pin.

verify

The concurrent verify receipt workflow is on record

0a37fedd

The coordination skill explains the queue, turns, and that a receipt is per worktree. Two agents do not clobber each other's green if they respect the semaphore.

skills

The scoped verify fallback is on record

87a76719

Terminal-runtime-safety documents when a --scope does not apply and what runs then. Fewer panic --full verifies.

skills

The inert-binding parser pitfall is on record

86b46527

The same skill warns: declarations, prose, and strings are not bindings. Anyone touching verify does not reopen that false positive.

skills

Worktree verify and autopilot safety are on record

b1e74dae

Do not point verify at a neighbor tree; the pilot does not send if Working or a background command is up. The skill makes that explicit.

orchestration

Autopilot is documented as the orchestrator default

fae3bef4

The orchestration guide describes interval, cooldown, and the conditions for nudging a session. The pilot is no longer a hidden verb.

skills

The delivery-receipt handoff is protected

267a13dd

Maintenance states: do not move or copy the receipt by hand between worktrees. The CLI owns it; a cp breaks the seal.

skills

Impact-aware verify budgets are captured

ff8981d4

Maintenance documents when a desktop or test-ceiling change does not escalate to workspace. The cheap ratchet is the happy path.

verify

Credits the delta when the base was already red

5cb5076a

The receipt stores the steps that already failed on origin/develop. If this session added no new failures, pre-push and merge treat the receipt as green. An inherited red no longer blocks the work you actually fixed.

verify

Also seals nested failures inside the receipt

d150e5aa

A script that fails inside (.cardumen/verify.sh and its steps) joins the baseline seal like a top-level step. The waiver no longer leaves a child red that pre-push counts again.

verify

Ignores stub-shaped fixtures that live in quotes

b1f89fa6

The productive-stub detector does not flag a function that only appears inside a string (a fixture, an example). The lint looks at real code, not text a test quotes.

desktop

Merge requires a verify receipt for this HEAD

f7e052ff

The guard asks cardumen verify receipt --json in the worktree; it does not invent the state path. Without a green receipt for the current SHA, the PR stays on hold. Same rule as pre-push.

verify

A baseline waiver is sealed into the receipt

82459ecc

If waiver consult classifies the failure as baseline, the receipt records class, evidence, fingerprint, and steps. Pre-push reads that seal; a spoken pardon no longer opens the gate.

verify

Stub lint looks only at the hunk you changed

b0e1d707

A stub that was already in the file does not fail verify because you edited another function. The detector uses the + lines of the diff. New hollow code is still caught.

skills

The catalog is linked on every detected agent

d648e16b

reconcile copies the canonical catalog to Claude, Codex, Cursor, and OpenCode when those hosts are in the repo. A symlink that already points right is kept; a real folder is respected. Settings shows coverage per agent.

CLI

Durable memory stays out of the delivery diff

54f7a1e5

cardumen memory add stores the learning in the home (~/.cardumen/memory/...), not in the checkout .nca/memory.json. Closing the task does not force personal notes into the PR.

desktop

A discovered unregistered machine can be paired

3aec1a25

In Settings → remote control, Connect on an unregistered device calls pairAccountDevice and signs it up. You no longer hunt another screen to register what the list already saw.

verify

The installed verifier boundary is on record

966adba3

The receipt-driven audit records which edge the bundled verifier covers and which it does not. Anyone designing a new gate does not rebuild what the binary already seals.

merge

The merge-guard receipt rule is on record

1f67e250

The same audit now states: without a green receipt for this HEAD, merge holds. Desktop and pre-push share one owner (the CLI), not two filesystem reads.

verify

The baseline-waiver receipt seal is on record

871c1f68

The audit describes the evidence, fingerprint, and steps the seal requires. A waiver without those fields does not count for a push.

skills

Durable memory ownership is on record

6e977614

The maintenance skill makes it plain: durable notes live in the Cardumen home, not in the product diff. Fewer PRs carrying an agent's memory.json.

MCP

MCP registry: export, doctor, and a Settings pane

1f726e29

cardumen mcp export copies the registry into each agent's grammar; credentials stay in the host store. cardumen mcp doctor checks transport and scope. Settings gets the registry pane (one registry, four agents).

coord

The orchestrator gets the session verdict

d09b2711

On stop or when approval is requested, the hook publishes the verdict (blocked, dead, command) on the coordination plane. The orchestrator no longer guesses: it waits for that event and acts on that session.

merge

The test-budget hold shows the real command

68128908

If the branch crosses the test ceiling, the hold message no longer escapes backticks. You read cardumen ratchet resolve and the origin (base already over, or this branch's tests).

router

Router scoring no longer nests two ifs

6217054e

Clippy flagged an if inside another in scoring.rs. It is now one combined condition; lane scoring is unchanged, it just stops fighting the linter.

desktop

Runtime registration no longer fails in silence

ddd69a57

If startup has no license, no network exposure, or the devices POST fails, the desktop stores the error code. It no longer looks registered when the signup never left the machine.

telemetry

Non-unix telemetry fallback is declared

dec857af

On Windows or other hosts without Unix sockets, local llama.cpp telemetry support documents the alternate path. It no longer looks like an empty return with no reason.

telemetry

The local model id is stored on the event

ecf01638

The identity of the model llama.cpp served (name and quantization) travels with the event. Local spend no longer shows up as an anonymous engine.

telemetry

Claude local model ids count as local

e4004850

An explicit list of Claude-on-local-engine ids stops them being priced as cloud. If the transcript names one of those ids, the event goes to the local provider.

telemetry

Claude local is also detected from the transcript

a21caf93

If the payload has no model id, the parser looks at the session transcript. A Claude served locally is no longer lost because the official field came empty.

telemetry

Early parser exits are commented on purpose

6c2a739c

Intentional telemetry parser returns (incomplete event, unsupported host) carry a note. The stub detector stops treating them as hollow code.

telemetry

The stub scan does not flag the telemetry parser

7d8f04a2

A branch that exits early without mutating state no longer trips the productive-stub detector. The parser can drop a malformed event without looking like a fake implementation.

telemetry

Local telemetry is anchored to the session

574a0f3c

Local-engine events carry session_id and the cwd work_kind. They stop floating as device-level spend; they join the same session as the agent that fired them.

telemetry

Claude local also anchors to the session

99d37bc7

The llamacpp provider lands in SQL and on the Claude-local event. Same session, same work_kind; historical repricing no longer treats it as cloud.

telemetry

Another stub-detector false positive, gone

0fca4a74

Parser exits are reordered so the detector does not confuse them with a stub. Event behavior is unchanged; verify stops failing a file that does real work.

telemetry

Local llama.cpp usage is credited as local

7f83d0aa

Codex or another CLI against the llama.cpp proxy emits events with provider cardumen-llamacpp. In pricing and SQL it counts as LOCAL, not as tokens from a paid provider.

merge

Merge holds if tests cross the ceiling

5d2b5f60

Before merge, the guard reads the package test budget (e.g. desktop). If HEAD has more tests than the max and this branch added some, the PR stays on hold until cardumen ratchet resolve.

desktop

An unregistered remote machine is explained

cb181794

In Settings → remote control, a device with no signup no longer looks like generic offline. It says “Not registered” and asks you to enable remote control on that machine so it can publish its runtime.

desktop

The runtime registers even if the relay is down

b739ccd6

Device signup on the account no longer waits for the relay. If network exposure fails, the machine still appears in the org (with no endpoints) instead of vanishing from the list.

merge

The hold names the background task that blocks

330cf77a

If a session or background job blocks the merge, the guard names it (handle, agent, reason). It is no longer an opaque “busy”; session review lists those tasks by name.

C.R.E.A.

The shell cannot create or mutate files without a contract

db9de025

C.R.E.A. looks at the effect of a Bash call: new or changed repo files. Without an active spec, the hook stops it. Hiding the edit in a quoted or nested command is no longer enough.

verify

The auto-guard skill explains verify baselines

762422dc

Anyone running verify on autopilot has the rule in writing: a failure already on HEAD is baseline; one born in this session is not forgiven. Fewer waivers by ear.

ops

The license postmortem cites the right measurements

60f8c98e

Latency and timeout references from 16 September point at the numbers that were actually measured. Anyone reading the postmortem no longer follows a figure that is gone.

ops

The license outage postmortem is on record

ee0e53b0

On 16 September the heartbeat talked to the wrong host and a 3 s timeout swallowed valid answers. The postmortem covers cause, symptom (a dead flag that would not heal), and the fix.

verify

Detects bindings assigned only in tests

#448cdeb0036

verify flags public symbols or new assignments nobody calls in production: if the only use is in tests, the change fails. Dead API cannot hide behind coverage.

CLI and desktop

Live tokens per second, visible and logged

#415adcceab4

llamacpp status --json reads the engine /metrics: generation and prefill t/s. If the engine is stopped or silent, fields are null with a reason, never 0. Every 5 minutes and on stop, a sample lands in ops-log.jsonl. The desktop engine card shows that pace.

desktop

Graph-gated worktree create is atomic

#4593d087079

If the graph gate fails mid-way, no half worktree or orphan directory remains. It is created whole or rolled back; the error names the step that stopped.

desktop

Detects incomplete brain profiles

#458495e65ff

An org seat with brain needs plan.json, license.jwt, arrecife-node.json, the key, and .last-validate. If any is missing, setup lists them instead of pretending the node is registered.

backoffice

License status, devices, and code resend

#457945a166f

The platform console reads license details and an org's devices (GET/DELETE). Resending a code no longer asks for invite_code: it uses the activation contract already in place.

terminal

Live agent stays behind a foreground shell

#44077e55f52

If a shell command runs in the agent pane, presence does not flip to shell while the agent process is still alive. agentPresence stays live and tracking is not lost.

verify

Detects migrations edited after they applied

#45685cd5e65

migrations.mjs --dry-run exits 3 if a registered migration no longer matches the file. verify treats that 3 as failure; missing network or pending ones only warn. The three-day silence is over.

Cardumen

Pre-push requires a verify receipt; CI only to main

#4548843c3ee

Without a valid cardumen verify receipt for this HEAD and worktree, the push is rejected and the message names the command. PR workflows no longer run toward develop: only toward main.

desktop

Queued terminal delivery is visible and aged

#4465458d837

A terminal send that stays queued no longer vanishes: the CLI and runtime show it is queued, for how long, and whether it confirmed or was lost. It no longer looks as if the text arrived when it did not.

Cardumen

Review and session runtime skills ship in the binary

#453597aa67e

review-placebo-catalog and session-runtime-lessons travel bundled. The router loads them in review and in session; they no longer depend on the destination repo copying them.

Cardumen

Result specs validate regardless of language

#45118bfdf36

task spec no longer demands Spanish or English keywords. It looks at the contract shape (goal, criteria, verify) and fails on the missing field, not on a dictionary.

CLI

Surfaces pushed branches that have no PR

#44770d68048

Session review and the terminal list flag a branch already pushed with no pull request, plus the command to open one. Work no longer sits on the remote with no review.

desktop

Waits until the paired workspace session is ready

#4425f13b33e

The paired harness does not talk to the session until the workspace answers. That avoids races where the CLI assumes a runtime that has not started yet.

desktop

Auto takes claims from the agent and the org

#438d1275454

In auto mode, claims come from the orchestrator agent and the organization, not from a generic seat. A node does not claim work from another org or another role.

Cardumen

Result-spec guard blocks name the reason

#437dc4567dd

When the hook stops an edit without a contract, the message says what is missing and how to declare the spec. It no longer looks like a silent failure; the rollback runbook sits beside it.

desktop

Relay GC sorts listings before the cap

#439121974e2

The remote relay-* listing is sorted, then the cap of 64 applies. Without the sort, the trim could drop new versions and keep old junk.

Cardumen

The result-spec contract resolves per session

#43349917f9e

The hook no longer looks for one global contract: it uses the active session. One agent does not inherit another's spec, and verify fixtures are commands that actually resolve.

desktop

Folder repo upgrades no longer overlap

#42322fc2c28

If a folder-repo upgrade tick is still running, the next one waits. Before, two ticks at once left the repo half-done or repeated the same fetch.

quality

Live-evidence verdicts stay conservative

#4340a2ec414

Hierarchical QA does not upgrade to PASS if live evidence is weak or partial. A conservative fail stays; it is not dressed up because a later step looked fine.

hooks

Task done blocks empty delegated work

#432ea0ae34a

task done reads the delegation log in project state. If a minion launched and left no result, close is rejected. A waiver also cannot cover a step that never ran.

verify

Detects stubs that look like productive code

#4200c365da4

verify hunts new functions that compile and return something, but do not do the work (unimplemented! or a dressed-up empty return). A stub no longer counts as an implementation.

desktop

Local harness waits until the relay is online

#4177f52ccd3

Local paired tests do not fire commands at the relay until the online condition holds. Fewer false reds from a slow start.

verify

No waiver for a step that never ran

#428f0852868

verify-waive refuses to forgive a mechanism that did not execute. The receipt records those steps as not run; you cannot mark green what was never seen.

verify

A fix is probed by reverting the diff

#4260ca0a1e8

On a change marked as a fix, verify reverts the diff and runs the probe. If the test still passes, the fix fixed nothing and the gate fails closed. A compile failure on reverted tests is inconclusive, not green.

verify

No waiver for a failure this session caused

#41899e510a0

If the command failed in this session and not on the base, it cannot be forgiven as baseline. Only a failure already present on HEAD counts as preexisting.

C.R.E.A.

The shell cannot mutate repo code without a contract

#427871bcbf4

C.R.E.A. watches the observed effect: if a shell command changes repository sources and no spec is active, the hook stops it. Disguising the edit in a bash -c is no longer enough.

distribution

The JSONL queue releases the drain lock on purpose

#425c141252e

When a drain ends, nca-dist releases the lock explicitly. The next batch does not sit waiting for a lock nobody was going to drop.

llama.cpp

Aptitude against the card, agent start, honest engine card

#41655bc0614

APTO/NO APTO is decided against GPU capacity, not free VRAM the loaded model already holds. up computes the agent profile (1 slot, high ctx) and states it. The card shows the served model, effective slots, and that model's VRAM vs the rest.

OpenCode

OpenCode 1.18 loads the plugin with {id, server}

#414e134fd37

The default export is no longer a function: it is an object with id and server(). On 1.18 a function left the instance null and crashed config/dispose. node --test gates that contract.

OpenCode

The plugin installs only from plugins/opencode

#41499ba1ef5

Team-setup no longer copies skills-memory.js into .claude/opencode. The source is plugins/opencode; the duplicate copy is gone so 1.18 does not load two different contracts.

license

License host is no longer the Arrecife account

bc1ce3af

The heartbeat uses nca_dist::api_base() (api.cardumen.ai), not Arrecife's apiBase and not the landing. A 404 against cardumen.ai looked like no network, so revocation never landed. One owner, one endpoint.

desktop

Hidden serve will not open a window without a display

b1d6c270

On Linux, with neither DISPLAY nor WAYLAND_DISPLAY, cardumen serve does not promote the desktop. SSH or a headless daemon stays as serve; the message asks you to start from a graphical session.

Cardumen

Pre-push evaluates the receipt in the repo you push

#47266e99346

The hook resolves the git root of the repository being pushed. A green receipt from another worktree or the parent no longer opens the gate. The block names cardumen verify in that tree.

desktop

Onboarding guides the company brain setup

#47174c3912f

If the plan includes the brain, first-run onboarding registers the node with Arrecife (connect the account, pick the org, complete artifacts) before the seat is ready. You no longer hunt for it in Settings.

desktop

A stale dead flag does not hide a live license

#4704dfe4784

If the server says the license is active, the desktop opens even when the local .license-flag says dead. A network error with a JWT still counts as grace. The activation step does not stay stuck on that local lie.

llama.cpp

Streamed tool calls match the declared schema

#468b55e7f1d

The llama.cpp proxy no longer forwards SSE as-is. It gathers the tool-call JSON and coerces "true" or "3" to the type the schema asked for. Codex stops rejecting arguments the model already meant.

license

A stale dead flag clears when the server says ok

#469c3a944f3

The hooks heartbeat asks the server again even when the local flag is dead. If /validate answers active, the flag becomes ok and .last-validate refreshes. A grace exhausted by a bad host no longer locks you out forever.

desktop

Settings → Models on Linux runs on llama.cpp CUDA

#4114e2c93a5

macOS stays on oMLX. Linux with an NVIDIA GPU lists, installs, stops, and launches against llamacpp. Without a GPU it is unsupported, with a reason. Claude Code and Codex buttons only appear on APTO models already downloaded. On Linux there is no load/unload: launch is what serves.

llama.cpp

A local model launches only if it is fit for agents

#406954a9dbf

pull rejects what is estimated under 40 t/s or does not fit. launch trusts only a stored certification: a real bench plus two tool-call probes (Claude Code and Codex). Without an APTO certificate it names the nearest fit model; no flag skips that.

router

Local only on verifier and summarizer; fallback is not double-counted

#403f723b0e3

The local engine does not compete on other lanes, even when it is free. If it does not answer in 800 ms, the request goes to the paid provider and tokens count there, not as LOCAL.

graph

A broad search is answered by the graph or ceded

#39834576165

Symbols and references come from graph.db. UI labels, JSON, Markdown, or Prisma do not live there: the gate cedes to rg instead of blocking blind. If the graph cannot open, the search proceeds with no invented guidance.

llama.cpp

You bring the model you want: Hugging Face or a local file

#401d3b239f3

cardumen llamacpp models search|pull resolves GGUF, license, and gate before download. It checks disk and VRAM; the download resumes from .part and verifies sha256. The HF token lives in secrets/hf.token (0600) and never reaches stdout.

MCP

The local MCP server exposes graph queries

#3936917ca8c

cardumen-local serves graph_query against the project's SQLite. If the graph is missing, the error is structured and tells you to build it; it does not invent an empty result.

llama.cpp

llama.cpp CUDA is a managed engine, sibling of oMLX

#390a17fe0bb

cardumen llamacpp installs, serves, stops, launches, and diagnoses on Linux+NVIDIA. User-space build, server on 127.0.0.1, idle auto-stop, 0600 key outside the repo. router probe sees it as AccountKind::LlamaCpp.

agents

The agent prompt arrives with graph context

#385abec9354

When an agent launches, the task-map injects the graph slice for the job. The agent does not start blind on a tree Cardumen already indexed.

quality

Each repo declares its Arrecife destination

#3848d61c6f3

QA no longer inherits another checkout's Arrecife project. The destination is per repo; a sibling worktree does not dump cards or evidence into the wrong project.

verify

`cardumen verify` selects impact and reuses the receipt

#377e39bd679

The selector marks related vs full from the impact map. A receipt for the same HEAD SHA is reused; another commit does not count as green. Stop and task done still require the current receipt.

remote

SSH, runtime, and account live in one host registry

#3671f6f1bec

Remote hosts stop being three lists. The unified registry discovers account runtimes and Settings navigates one pane. An SSH host and a paired runtime are named the same in the CLI and the desktop.

desktop

Account settings drive remote runtime control

#36812888beb

The account pane registers and syncs the relay device. Runtime lifecycle (pairing, token, presence) comes from that account, not a loose per-machine setting.

integration

Local boundary probes are mandatory

#369051afbcd

The integration gate runs the local Arrecife and Carnada probes. A legacy CLI is tolerated; a broken boundary is not marked green. Without a probe, there is no delivery.

desktop

Linux Models copy shows t/s, disk, hash, and cause

#41356232c3b

The APTO badge shows t/s to one decimal. Local storage sums size_bytes. The llama.cpp hash shortens to 7. NO APTO classifies cause (tool calls, t/s, VRAM) in the UI language. The GPU name wraps at 360 px.

desktop

`worktree rm` does not send back a card whose PR is already merged

#409862e3614

If the PR is in the retirement journal, the card stays where it is (delivered). If that cannot be known, it warns and still does not move back. Only not-merged returns to the previous column.

email

Code emails send you to download the app and activate

#4128e87b058

Purchase, seat, trial, resend, and activation say: download Cardumen, open it, and continue with your code. curl|install.sh, Homebrew, and cardumen activate leave those mails. The model is the installer (DMG / AppImage).

router

`assign` and `probe` see the local engine and local accounts

#4104f62dd2b

The catalog is no longer hosted providers only. It merges local accounts and the managed llama.cpp engine. probe says which lanes it routes. Dispatch synthesizes the llamacpp-managed account so the local model is not sent to a paid provider.

verify

Contract `cargo` commands run in the nested workspace

#408596be29d

cargo test -p from the repo root could not find Cargo.toml: the workspace lives in apps/cardumen-cli. It now resolves that manifest. Several ambiguous candidates fail asking for --manifest-path.

verify

A missing rustc-wrapper no longer kills `cargo` in verify

#4054c4f172a

If .cargo/config.toml or RUSTC_WRAPPER point at a binary that is not there, verify and RESULT_SPEC set RUSTC_WRAPPER='' and cargo still runs. A wrapper that exists is kept.

desktop

The runtime-imports guard matches globs and multiline imports

#4073af67212

/** left a literal * and out/main/codex/** missed the artifact. The parser no longer goes line by line (oxfmt splits the import). Every explicit out/main path must be an electron-vite entry.

verify

verify.sh tests pin `CARDUMEN_DEV=1` and do not inherit the host

#404b2ade78d

A live development environment was leaking into fixtures. The tests pin CARDUMEN_DEV=1 and isolate the merge-base so they do not read the operator's integration evidence.

core

The price table names Haiku 4.5, Opus 4.5–5, Sonnet 5, and Fable

#402006cafc1

Explicit rows sit before generic claude / claude-opus. Haiku 4.5 at 1/5, Opus 5 and 4.5–4.8 at 5/25, Sonnet 5 at 2/10, Fable/Mythos 5 at 10/50. Retired Opus fallback stays 15/75. Source: Anthropic pricing, 2026-09-15.

graph

Incremental graph refresh has a cost ceiling

#40004b38642

Ambiguous candidates, dependents, and summary scans are capped. An invalid projection sidecar is rebuilt. Refresh no longer grows with repo size without a bound.

core

Credential storage paths are injected, not guessed

#3999138de96

Tests and runtime receive the store paths. They are not hand-built against the developer's home. A fixture does not stomp the real ~/.cardumen.

telemetry

Each drain queue has its own lock

#397ff6f665b

Usage, memory, and work events no longer share one .drain-lock. A usage drain does not block memory. The lock lives next to that queue.

hooks

RESULT_SPEC only watches editing tools, and it is not overwritten

#396dae11f0a

The matcher is no longer .*: it runs on tools that edit. A user settings.json cannot cover the bundled hook. Reading or listing does not require a contract.

graph

The graph refreshes live on commit and session close

#395ef33414c

Incremental index is sqlite-first: node identities are kept, scans go to the target index, and an invalid sidecar is rebuilt. Closing the session does not leave the graph on the previous commit.

llama.cpp

llama.cpp platform tests do not depend on the host

#394189ccbd6

The suite does not require an NVIDIA GPU or nvidia-smi on the runner. Hardware is detected with doubles; CI stays green on machines without CUDA.

Cardumen

Native Read, Grep, and Bash also go through the graph

#39183b1080b

Graph-first PreToolUse covers native tools, not only the wrapper. It reads paths and definitions from graph.db. A covered symbol search is sent to the graph; a literal sweep is ceded.

setup

A bare `pnpm install` does not fabricate an npm init `package.json`

#3893138e020

If the worktree has no package.json, that setup is skipped with a notice. An untracked package.json whose scripts.test is the npm init placeholder does not count as a test step.

quality

The UI gate is certified on authenticated screens

#3923950a677

certify-ui --url --environment --role signs in with the sealed QA credential, captures in that session, and fails if it lands on login. setup --add-target declares an origin without a wizard. A hand-made report still does not certify.

verify

`detect_pm` recognizes `bun.lock` and `packageManager`

#388e944da32

A modern Bun module fell through to pnpm and aborted. Order is now packageManager, lockfiles (bun.lock / bun.lockb, pnpm, yarn, npm), then the default, always requiring the binary.

Cardumen

The graph exists before an agent launches in a worktree

#3863f7d4ff6

If the graph is cold, ensure runs in the background and create does not block. A ready graph is awaited. The agent does not start against a fake empty index.

integration

MCP probes accept streamable HTTP responses

#387cf1bb3a2

Arrecife is probed with JSON-RPC POST and accepts the streamable transport. A body that is not a single JSON object is no longer read as a contract failure.

state

A worktree's project state is the parent repo's

#3820afc65e0

~/.cardumen/projects/<slug>/ paths resolve against the main checkout, not the worktree folder. Memory, journal, and graph are not duplicated per worktree.

auto

An automation has a single live coordinator session

#379a19256fb

If the coordinator is busy, the cycle records busy and does not launch another. A coordinator stuck in a dialog or idle is retired before replacement. Three heartbeats do not open three sessions.

CLI

`terminal stop` does not relaunch the agent; concurrent create is guarded

#380e0049e23

On stop, the worktree is marked agentless and pending prompts and launches are purged. A terminal create --agent with a live agent is rejected unless --allow-multiple.

hooks

The orchestrator Stop does not relaunch verify, and SIGTERM kills workers

#3789894c4da

With CARDUMEN_VERIFY_OWNER=orchestrator Stop only requires a PR to develop and the LISTO line. cardumen verify kills the process group on SIGTERM/SIGINT so workers die in under 10 s.

CLI

`--json` emits a single object and the terminal handle is populated

#37530e571ed

stdout does not mix prose with JSON. A denied RPC stays a stable error object. The created terminal handle is no longer empty for whoever chains the next command.

quality

The UI QA report is published on the Arrecife card

#3765ad898f8

If a card is linked, the report uploads there. .cardumen/qa stays out of the team checkout: local evidence, not git noise.

hooks

A read-only subagent does not require a C.R.E.A. contract

#374c04e72fe

Explore, Plan, or a prompt without mutation verbs pass. Editing, commit, push, or opening a PR are blocked naming the tool and the word. The orchestrator can investigate without opening RESULT_SPEC.

verify

Session files outside the repo do not break verify

#371466e78d3

Transcripts and journals live in Cardumen home. Verify no longer treats that path as a dirty checkout or as product scope.

skills

The QA skill stops seeding `api.arrecife.ai`

#332780d9756

That host does not exist (NXDOMAIN). Skills and docs are locked to ARRECIFE_QA_MCP_DEFAULT_URL. The desktop stops copying the dead literal into every workspace.

desktop

CLI license runtime dependencies are unpacked

#36336672863

out/cli/** is in asarUnpack. The packaged binary no longer looks for modules inside the asar, and the CLI license starts on Linux and macOS.

local models

The CUDA 1/3 report says which load belongs on the GPU

#383127bf405

Measured on Ubuntu with an RTX 4060 Ti: verifier and summarizer tie Haiku on atomic tasks. Executor, planner, and reviewer stay paid. That table is what router 3/3 turns into policy.

CLI