dokploy-ops2.1.2
Changed
- Code comments and setup examples use a placeholder project name (
acme) instead of real ones.
Every release of every tool, newest first. 87 releases across 15 tools.
An agent can read the same list as JSON at
/updates.json, compare it with what is installed,
and tell you exactly what changed before it updates anything.
acme) instead of real ones.acme) instead of real ones.acme) instead of real ones.scripts/lanes/run-lane.sh is removed. It only forwarded to the fleet command since 2.0.0. Call scripts/fleet (or core/scripts/fleet) instead.fleet schedule install --every 2h: unattended PM runs from a launchd agent (macOS) or a crontab line (Linux), with status, pause, resume and remove. run-once takes a lock, skips while paused or while an attended PM's heartbeat is recent, stops after fleet.unattended.max_hours and calls fleet.alert_command on failure. See core/references/unattended.md.unattended-guard hook, on only in an unattended run. It always denies a push, PR, merge or release tag on a production branch and a production deploy command. It allows merging and client messages only when fleet.unattended.allow lists them.fleet schedule heartbeat, run by an attended PM each loop.core/references/pm-history.md: the incidents and measurements behind the PM's rules.missed. A fail the tester put on the ticket, the spec, the environment or a PM ruling is not scored either way.selection.demote_window is 20 (was 5).fleet apply at preflight and after each merge group, and applies it.preflight.md (25.5 to 17.8 KB) and spawn.md (27.3 to 19.1 KB) keep every rule, and their history moved to pm-history.md.spawn.md said both "do not log tokens" and "pass tokens" for a Claude spawn. It now says not to: tokens is the last call's size, and cost --backfill writes the real usage.fleet intake: every intake source's new items as JSON. A broken source shows ok: false and never fails the pass. See core/references/intake-sources.md.sentry block. It reads only the Sentry projects the block names, in its one environment. A new issue qualifies when it is first seen inside window_days or regressed, and passes min_users or min_events (any issue for always projects). Severity comes from users, urgent_paths and regressions. Evidence carries the link, counts, release, in-app frames and breadcrumbs, with query strings, emails, phone-like numbers and token=-style values removed.fleet.intake_sources, for any other feed. The command prints {"items": [...]}, and an optional resolve_command closes the item.fleet intake check at run start. It warns when a repo ships a Sentry SDK but has no sentry block, so the PM tells the owner.fleet intake mark and fleet intake resolve. A filed item is not offered again. A merged fix resolves the Sentry issue in the deployed release with a note naming the ticket, and an issue that comes back is offered again with regression_of set to the old ticket.$SENTRY_AUTH_TOKEN or a mode-600 token_file in the per-user override, never the repo.fleet intake check and the first pass. Each loop runs a pass before pre-pick triage. deploy-and-done resolves the source item. The planner and fleet-board have rules for intake items (external_source and external_id set, deduped on the board first).workitem list by external_source and external_id returns a validation error on a match instead of the row. The docs say how to read it.fleet.test_rounds_max failed test rounds (default 2, 0 turns it off) is refused until the owner's answer is recorded with fleet rounds allow <TICKET> --reason "...". One allowance covers one more round. fleet rounds lists the count per ticket. Switch file: rework-cap-off.fleet.trials, off by default: - case_list: before a build, a tester turns the ticket into a numbered CASE LIST: comment, with a state table for a screen and the grid each rule holds over. The engineer builds to it and reports pass or fail per case from a real run. A list over about 25 cases goes back to the planner to split. - closed_retest: a retest runs the failed cases and the list's regression, nothing new. A finding outside the list is graded s1, s2 or s3. Only s1 blocks (data loss, money, a wrong or duplicate send, security, a blocked main path). The rest become a follow-up ticket that lands before the release carrying the feature.group.must_run and group.labels in the config. fleet gate-stamp check fails a group PR when a required job was skipped or missing, or a label is absent. A group PR had gone green in 17 minutes with its full suites skipped.fleet slot verify-clean <S> and fleet slot release <S>. Engineers and testers paste the verify line as their teardown report. release never kills a process and refuses when a port or a database is in use.fleet queue add | list | next: a spawn the machine gate refused is parked with its brief and offered again when there is room. "Room for N heavy" now also counts sessions already running.fleet questions add | list | close: open owner decisions, the PR each one holds, and its age. Every status to the owner starts with them.log add: --outcome spawn for the row written at the start of a session, so the result row needs no --force. --cause pm_ruling. --in-list, --severity and --repeat on a failure.fleet status shows failed rounds per open PR, slots with leftovers, "required ran: x of y" on a group PR, queued spawns and open owner questions.fleet retro counts PM-ruling failures (pm_ruling_failures), splits failures into new and raised before, reports how long queued spawns waited, and splits tester pass to merge into waiting for a group, CI and waiting for the owner.core/pm/skills/ruling.md: a PM scope decision quotes the contract or spec line it rests on, or is marked a guess to verify, and lists the tests it retargets.fleet.session_soft_cap (default 300) and fleet.client_engines.fleet cost, fleet review and fleet retro do not count a spawn start row as a session.fleet held set|clear|list: the PM records each Todo ticket it skips for its blockers, from the relations it already read at pick time. No extra Plane call.fleet status prints a held work line when every blocker of a held ticket has an open PR (draft, approved or in a group PR), a slot is free, and the machine is ok or tight with room. It names the ticket, the blocker PR and the branch to stack on. No line while a blocker has no pushed branch, no slot is free, or the machine is closed or critical. The compact-resume hook shows it too.fleet retro counts the hours a ticket sat held with a slot free, from the sightings status writes: held_hours, a PM alert over half an hour. A run with nothing recorded says so instead of showing zero, since the outcome log has no field for it.STACKED ON block for the engineer brief: start from the blocker branch, rebase onto trunk after it merges, open the PR against trunk.--outcome decision. Waiting for the blocker to merge is not a reason.STACKED ON block.fleet.trials in the config, all off by default. Turn one on only between runs. fleet review shows a "trials on" row, so each trial is judged against the runs without it. - path_map: on a money, stock, state or permission ticket the engineer posts a PATH MAP: comment before code, with every entry point and writer the rule touches. The blind tester may read it and checks every row in round 1. - money_ceiling: the planner splits money, stock and state outcomes at 1000 lines. - own_commits_only: a branch carries only its own ticket's work, in rework too. One ticket's branch took in a group branch and three other tickets' fixes, and went through 9 test rounds. - group_cut_minutes: the PM cuts a merge group no later than this many minutes after the first approval starts waiting. Approved to group opened was a median 1.5h, and for 54% of it no other group was open.--defect on log add for a missed verify_fail: criterion, other_path, stale_state, concurrency, money_rule, out_of_scope or stacked. The tester names it in its report, and the review counts missed failures by kind.fleet status warns about each ticket PR out of draft, with gh pr ready --undo. A draft runs almost no CI; ticket PRs out of draft used about 3,600 runner minutes in 18 days.verify_fail row needs --reason, like blocked. 67 of 186 failures had none.fleet cost and fleet retro left lane spend out. Lane tokens were read, but Codex had no price, and lanes were never added to Spend, so a Codex-only run showed agents at $0. Lanes are now priced and counted in Spend, the per-ticket cost and the PM share.fleet cost dropped the PM's last day: a window ending 2026-11-14 17:00 left out all of the 14th, while fleet retro kept it. PM spend now follows the window to the 10 minute, like the context per call already did. Whole days remain only for sessions known from an old cache entry.fleet cost counted every PM session the repo ever had as the run's sessions (213 for a run that had 1). It now counts the sessions with calls in the window.fleet retro showed no lost time both when nothing was lost and when fleet guard never ran. The guard now writes a start event, and the retro says "not checked" when no guard covered the window.gpt-6.1-sol and gpt-5.6-sol (OpenAI standard tier, short context, read 2026-11-16). A price can carry its own source and a valid_until date. fleet cost names a price whose date has passed. The gpt-5.6-sol price is promotional until 2026-11-30.fleet review: finds the runs in the log, measures each with the same code as cost and retro, keeps one scorecard per finished run in {state_home}/reviews/<project>.jsonl, and prints the last runs side by side with what got better, what got worse, the toolkit versions released between runs, the session mix and the decisions logged. --md writes it for the owner's records. Config: fleet.review.--rebuild keeps nothing when any run failed.--outcome decision: a PM marker row for a routing choice or an owner's order, with a time and a reason. Never scored, never refused as a repeat.fleet log add refuses a blocked, lane_failed, deferred or decision row without --reason. A lane row takes the lane's last message when no reason is given, marked lane said:.expat formula. If that formula is missing, it says to run brew install expat.zernio, in front of every Zernio workspace. It exposes Zernio's tools once, adds an optional workspace argument to each, and finds the workspace from account_id when it is left out. accounts_list with no workspace lists them all.ZERNIO_KEY_<NAME> lines in the toolkit's credential store. There is no fixed count, and a name can be a letter or a word. A new workspace is one new line and a restart.scripts/setup.mjs adds a key line for the person to fill in, checks each key against Zernio without printing it, and records the working workspaces in state/setup.json.fleet retro read the PM context per call for the whole UTC day, whatever --since said. It now follows the window, to the 10 minute. The cost figures stay counted by whole UTC day, as before. A session that only an old cache entry remembers still counts by day.fleet findings reported an empty or missing review file as clean. It now says the review did not finish and exits 4, the same as a usage limit. A file with a marker but no final message does the same.gh pr checks --watch could end early. A workflow can leave a skipped duplicate under the same check name, and --watch reads it as done while the real check still runs. The merge queue now waits on the run for the head commit (gh run watch <id> --exit-status) and confirms the required check concluded success on that head. fleet status and fleet gate-stamp check no longer let a skipped duplicate hide a running or passed check of the same name.fleet-planner and fleet-scanner agents could not read ticket comments. They now have workitem_comment, and their roles say they only list and retrieve.engineering/references/brief-checklist.md item 21: a rule about whether something may happen, or who may see or do something, gets one function every caller goes through, with the callers listed by grep before coding. It is numbered last so the numbers in existing briefs stay the same.fleet guard: one background process per run that holds off system sleep, checks that an agent call still works, and alerts through fleet.alert_command and a desktop notification.fleet retro reports lost time (asleep, agents down, tester pass to merge) and a lost_hours finding.ci.group_is_the_gate and quality.pre_push_full config: the group PR's CI is the one full run, and the hook can check only what the branch changed.blocked_by needs a named code dependency. A real chain is built stacked on the previous branch and does not wait for its merge.engineering/references/brief-checklist.md: the defect classes that caused repeat rework, grouped (before code, UI, money, dates and concurrency, review before handoff, testers, PRs and CI). The PM copies the items that apply into each engineer brief under CHECKLIST:, and the engineer names in the handoff which ones it checked.fleet.max_heavy_sessions (default 3, read by the PM): at most that many engineers or testers with servers and a test DB at once, even when the machine gate shows room. The gate's room figure ignores swap that is already high./tmp/<identifier>-gate.lock with mkdir before a push and wait while it is held.[skip ci] commit last before a release PR.fleet.max_heavy_sessions, at least 1, in place of a fixed 2.fleet gate-stamp post looks for. Until the script changes, the PM proves the gate by the remote branch head matching the local HEAD after a push without --no-verify. Nobody writes the log line by hand.pgrep inside sh -c matches itself.fleet gate-stamp check read a stale re-run. gh pr checks keeps one row per check name. After a PR title edit it kept a re-run attempt of the old workflow run, which still carried the old title and failed, over the newer run that passed. The check said NOT READY while GitHub showed the PR as clean. It now also reads the PR's statusCheckRollup and, for each check name, uses the run with the highest workflow run id. An older passing run never hides a newer failure.fleet slot-template build|clone|status: one migrated, seeded Postgres template per schema state. build makes <prefix>_<key12> from origin/<trunk> in a throwaway worktree with the configured commands. The key hashes the migration and seed files and the build commands, not the commit, so a template is reused until they change. A lock stops two builds at once, and only keep templates stay. clone <slot> [--test] drops the slot's DB (and its test DB) and copies the template in seconds. It touches only slot DBs named in fleet.slots, never the shared row's, and exits 3 when no ready template matches so the session runs its recipe. Off unless fleet.slot_db_template.enabled.fleet gate-stamp post|check: post puts a local-gate commit status on a SHA, only when HEAD is that SHA, the tracked tree is clean and the gate log shows success for it. check --pr N exits 0 when the PR head has the stamp and every check outside ci.covered_by_local_gate passed or skipped. A skip after a covered job was cancelled does not count as passed. Off while that list is empty.fleet slot-template clone and skip migrate and seed when it works. Otherwise they run the recipe.DATABASE_URL and every test URL, check each names your slot DB, and export each in its own statement. On one project a combined export left the test URL on its old value, and the test harness migrated the owner's shared DB.fleet gate-stamp check passes, even while a covered job still runs. Never the release PR to production, and never when the head moved after the stamp. Trunk CI still runs everything and stays the deploy gate.fleet cache-server start|stop|status|env: a small Nx remote cache on 127.0.0.1 (Nx 20.8+ self-hosted cache API), one per project, so every engineer, tester and group worktree reuses the others' lint, typecheck, unit-test and build results. Off unless fleet.nx_cache_server.enabled. eval "$(fleet cache-server env --cwd <repo>)" starts it and exports the two Nx variables. Oldest entries go past max_gb (default 20). Checked with two worktrees: the second got [remote cache].--skipNxCache is for a result that looks wrong for the change, not for every green: cached lint, unit tests and build are real results.Failed to send request ... /v1/cache). The brief says to restart the server and run again, never to skip the hook.fleet retro [--since ISO | --hours N] [--cwd repo] [--json]: per ticket rounds, spend, tool calls, time and outcome, totals, the PM's share and context per call, rework and its causes, lanes and the Codex burn rate, then numbered findings, each with a cause, a fix and who owns it (pm, toolkit, owner). Exit 3 on an alert. A finding seen in the previous window is marked REPEAT. Thresholds and levels are set under fleet.retro in config.retro skill: one watcher runs the retro every 30 minutes, and again after each group merge and at the end of a run. The PM says each finding to the owner in one line, applies pm fixes at once and logs them as retro_fix, and proposes the rest with evidence.fleet findings <file>: the final findings block of a codex review or codex exec transcript, deduplicated, with one count line. Exit 3 on an open P1 or P2, 4 when the review did not finish.fleet watch --stop-prefix <prefix> stops every live watcher whose name starts with it.fleet status.fleet findings.CAP with what is done, what is open and the findings left. Testers stop at about 300 tool calls. The PM decides the next step.In Review · needs PM scope call. The PM decides before a tester starts.fleet log add refuses a row that repeats one already in the log (same agent id and outcome, or same project, ticket, role, executor and outcome within 10 minutes). --force logs a real second session.fleet watch <name> --until '<test>' and fleet watch <name> -- <command> run a background wait under a name. A second start under a name that is still running starts nothing and exits 75, with the holder's pid and output file. --show, --list and --stop read and end them. Engineers and testers wait only through it (standing brief), and the PM waits for CI with fleet watch <TICKET>-ci -- gh run watch <id>.until or while ... sleep loop running more than once from the repo. --kill keeps the newest copy.fleet.resources.swapout_mb_s, default 5, 25 and 100 MB/s) decide. Linux, Windows and a Mac that gives no pressure reading keep the old rule.fleet.tool_caches config: a short tmpdir for Jest, Nx and temp files, so a repo on an external disk stops filling the internal one. Lanes and gates get TMPDIR from it, fleet status names a Jest cache over its limit, and fleet-sweep.sh --caches (or the launchd job from fleet-cache-sweep.plist.template) clears old caches, temp folders and Docker's build cache. A Jest cache had reached 26 GB.class checked:. Testers name the class with each finding, and the retest tries each one..watchmanconfig ignore_dirs at the repo root. watchman then treats the repo root as the project for every worktree, so Jest inside a worktree found no tests and a pre-push hook passed green with three failing tests. fleet status and run-lane no longer write any watcher config. The warning now says to drop the watch and turn watchman off for Jest, and the PM docs forbid the ignore_dirs fix. If 2.4.0 wrote the file in a repo, delete <repo>/.watchmanconfig and the /.watchmanconfig line in .git/info/exclude.fleet resources (core/scripts/fleetres.py, standard library only). It measures memory the OS can still hand out, swap against RAM, load per core (CPU busy on Windows) and free disk on macOS, Linux (cgroup limits included) and Windows. Limits scale to the machine and can be set in fleet.resources. Levels: ok, tight, closed, critical (exit 0 to 3). It names the biggest processes, marks the fleet's own, and says what each usually is and whether its fix is safe (watchman, Nx daemons, headless browsers, VMs, browser tabs, indexers). It warns when a file watcher is rooted over the worktrees.max_heavy.fleet resources --wait: blocks until the gate opens, at most 4 minutes. Briefs use it before Jest, Playwright or a server, in place of uptime arithmetic.--gate heavy|light for scripts.--no-watchman.run-lane defers a lane while the gate won't admit it (exit 75, resources <level> in the error). --ignore-resources skips the check.resources-off in PLANE_FLEET_HOME.fleet status opens with a machine: line, the reasons, safe fixes for machine hogs, and watcher warnings.fleet-sweep.sh prints the machine gate instead of the macOS-only check. --min-free-mem-pct and --min-disk-gb still override it for one run.fleet resources --wait, and no session starts a file watcher in a worktree.engineering/engineer/skills/open-pr.md §3b).engineering/engineer/role.md §6).core/pm/skills/spawn.md).core/pm/skills/spawn-tester.md §4). On one project it found a P1 or P2 the tester's blind pass had missed on four PRs in a row.opus engineer rebuilds it against three to five written invariants, or the planner splits it (core/pm/skills/pick.md §2).core/pm/skills/spawn.md §3, step 5).core/pm/skills/merge-queue.md §5).core/pm/role.md §7).--maxWorkers=2 and waits out a high load before starting Jest, Playwright or a server (core/pm/skills/spawn-tester.md, engineering/references/standing-brief.md). Four heavy lanes plus two pre-push hooks had taken a 12-core Mac to a load of 108.core/planner/skills/write-ticket.md §3).core/pm/skills/pick.md §2, merge-queue.md §5).engineering/engineer/skills/investigate.md).~/.cache/plane-fleet/shots/<IDENTIFIER>-<n>.png and name that path in the verdict comment. They never open Plane in a browser, never guess its web address, and never open a tunnel or upload host to get a public URL (engineering/tester/skills/verdict.md, engineering/tester/role.md, core/references/non-claude-adapter.md).fleet.parallel_schema in the project config: how many schema tickets may build at once. Default 2.fleet antigravity ..., the agy CLI on a Google AI plan). Models: gemini-4-argon, gemini-3.8-flash, claude-opus-5-5, claude-sonnet-5-5, gpt-oss-120b, with --effort. Opt-in: no candidates list names it by default.agy -p /usage, cached 5 minutes. Its three pools (Gemini, Claude and GPT, Argon) are judged apart, so pick and the lane gate look at the pool of the model being picked. fleet budget check antigravity --model <m> does the same by hand.fleet antigravity-setup, run once: registers Plane with agy through fleet mcp-exec plane, which reads the server entry from ~/.claude.json at start so no key is copied, and links the lane guard plugin into agy.PreToolUse hook that acts only in fleet lanes. agy cannot set tools per run and its own denials end a headless run, so every lane runs with skip-permissions and the guard denies browser tools, other MCP servers and, in read-only lanes, writes, unlisted commands and Plane writes. A denial reaches the model and the lane still ends on its report. result.json says whether the guard ran and what it denied.plane-fleet → fleet. Reinstall it under the new name. A linked install relinks skills/plane-fleet and each agent card.core/ (PM, planner, board specialist, scanner, solo mode, the fleet command, shared references), engineering/ (engineer, tester, their references and scripts) and social/ (planned). Every role is a folder with role.md and skills/, one file per thing it does, read only when that step comes.team-pm, team-engineer, team-tester, team-ticket-writer and solo-ship are gone as separate skills. Start the PM with /plane-fleet run; the others are role files reached through fleet (table in SKILL.md). The ticket writer is now the planner's skills/write-ticket.md.scripts/lanes/run-lane.sh is now the fleet command in core/scripts/, with scripts/fleet as its short path. The old path still forwards to it for this version.agents/<section>/. fleet-plane is renamed fleet-board. A plugin install names them plane-fleet:<section>:<agent>.fleet-scanner, a read-only agent for counting, triage and log reading on behalf of the PM and planner.role.md.§3) stay, and references between files name the file.~/.config/plane-fleet (PLANE_FLEET_HOME), so the log, cost cache and overrides carry over. Hooks and cost still recognize sessions and log rows from before the rename.hooks/pm-recompact.py, a SessionStart hook on compact: in a PM session it adds the live run-lane.sh status and the PM's hard rules to the compacted context, so the PM continues from facts, not from the summary.claude --autocompact 300k. Its context grew about 1.3k tokens a call, to 650k by call 400, and calls over 400k were 48% of PM spend; the default window on a 1M model compacts near 967k.hooks/cache-keeper.py: in fleet agents, refuses a foreground Bash timeout over 270000 ms and a Monitor deadline over 270000 ms (or none), and tells the agent to run the job in the background and check it every 4 minutes. An agent's prompt cache lives 5 minutes; waits past it made the next call rewrite the whole context, 12% of agent spend over a month and 24% in its last week. Refusals are logged; cache-keeper-off turns it off. cost shows the refusals next to the rewrite cost.run-lane.sh status: budget, slots and who holds them, lanes running and just ended (a dead lane is flagged), open PRs with required checks, labels and mergeability, trunk CI, and worktrees with no open PR, in one call. The PM used about 2,000 separate calls a month for these.status and waits for CI with gh run watch in the background.cache-keeper.py in user settings like the pacer (README).run-lane.sh free-models refresh discovers the free models opencode serves today: no cost, tool calls, and at least min_context. It keeps each model's first and last day seen in {state_home}/free-models.json. run-lane.sh free-models lists them. New free models appear almost daily, some only for a short trial; 2 of the 6 listed on 2026-10-22 were gone three days later.free_model_policy in the registry: the roles that may use free models (default scan; a free tier may keep the prompt), the providers to ask, and how old a list may be.pick adds every live free model as a challenger for the allowed roles, and an experiment tries the newest untried one first. A model that is no longer served drops out. Every other role loses free models, whatever a list says.run-lane refuses a free model outside the allowed roles or outside a read-only sandbox.free_models list is only a fallback when no discovery is fresh.free-models refresh in preflight.run-lane.sh cost: what tickets really cost and how long they took, read from the Claude Code session files of the PM and its agents, the lanes' usage and GitHub merge times. Per merged ticket: cost, agent + PM share, first spawn to merge, hours agents worked; cost by role and by rework rounds; where the money goes (cache reads and writes, cache rewrites after a wait, first-call context, the PM's calls and context). --ticket prints one ticket's sessions in order, --json gives the data.cost --backfill [--dry-run] writes the real usage onto Claude spawn rows, so the scoreboard compares Claude and lanes on the same tokens.prices block in the executor registry: API list prices per model, used only by cost. A model without a price shows tokens only.log add --cause missed|ticket|spec_changed|environment, required on every verify_fail. Of 114 test failures logged before this, none had a cause.fleet.cost.claude_project_dirs in the project config, for a repo whose old sessions sit under another folder name.agent_id and duration, not its tokens, names every spawn <role> <TICKET>, and runs cost --backfill and cost in preflight.tokens a Claude spawn row carried is the size of the agent's last call, not its spend: measured 2026-10-25, a median engineer session processed 9.8M tokens and a tester 14.3M, 44 and 56 times the logged figure. The docs that built cost claims on it are corrected: the agent types' start is about 17k tokens of context against 72k for general-purpose (1.2.0 said 23k against 40k), and the fixed cost of a ticket is now given in measured dollars.cost keeps a summary of each session it reads, so a session still counts after Claude Code deletes its file (30 days by default).source_key. On one project that was 459 of 1,442 rows, so the scoreboard, reports and promotions worked on doubled samples, and seed would copy the rows in again. A legacy row's identity no longer includes the folder.agents/: fleet-planner plans right-sized tickets, fleet-plane files and reshapes them on Plane, fleet-engineer and fleet-tester run the existing roles. Each carries only the tools its role uses. Measured on one project: a spawn starts at about 23k tokens, against about 40k for general-purpose.engineering/references/ticket-rules.md §2: one outcome per ticket, end to end; never split by layer while under the ceiling; siblings on the same screen or component ship together.provider/model. Free models cost 0 in the cost units. Model families ignore the provider prefix.disable_plugins in the executor registry turns off Codex plugins in a lane. The browser and computer-use plugins bring cua_repl, which drives the owner's real Chrome and desktop.{state_home}/lanes/, so a status view can list live and finished lanes without knowing each lane's --out.fleet-engineer and fleet-tester by subagent_type, and sends new work through the planner and the Plane specialist instead of writing tickets itself.fleet-sweep.sh finds slot scratch dirs by fleet.slot_scratch_prefix (default /tmp/<identifier>-slot-), not one hard-coded project name.executor:model@effort is scored, explored and promoted on its own. pick returns it, log add --effort and lane results record it, and opencode lanes pass it as --variant.* was counted twice on the scoreboard.engineering/scripts/fleet-sweep.sh finds what finished or dead sessions left running (servers on free slots' ports, processes in removed worktrees, Nx daemons, old headless Chromium, wait loops that can never end) and stops it with --kill. It also prints a spawn gate on free memory and disk. The PM runs it before every spawn.run-lane.sh report --project <p> --since <iso>: tokens per Done ticket, first-pass approval, rework rounds and the share of Done work the client can see, for one run.run-lane.sh log add --from-result <result.json> takes a lane's tokens and duration from its own file. selftest checks the log round trip.engineering/references/ticket-rules.md.log add --set dropped its values on spawn rows, so the outcomes log had outcomes but no tokens or durations.watch-video to video-watch, to follow the toolkit's naming rule: our own tools are named medium first, then action. Remove the old copy, then install video-watch: /plugin install video-watch@itqan in Claude Code, codex plugin add video-watch@itqan in Codex, or ./scripts/install.sh video-watch for any other agent. Nothing else changed.eleven_v4, ElevenLabs' newest voice model. Measured at about 0.11 credits per character, a quarter of eleven_v3.eleven_v4 by default, as ElevenLabs does. Pass --model eleven_v3 for the old default.--seconds.--max or below the reserve, and never paid for twice..plane/config.json, with a private per-user override for client contacts, server names and timezone, and run-lane.sh config to print the merged result.result.json with status, report line, duration and token use, and never commit on their own.pacer-off and pace-off switches, and a lane that has no budget exits as deferred with a resume time.uvx./cloudflare-ops:cloudflare-ops, with the plugin name in front. This has been true since 1.2.0, and this release marks it as breaking. Use the new name, or ask in words, which works as before. A copy in a skills folder keeps the plain name /cloudflare-ops.agents/openai.yaml inside the skill folder. Other agents ignore that file./dokploy-ops:dokploy-ops, with the plugin name in front. This has been true since 1.2.0, and this release marks it as breaking. Use the new name, or ask in words, which works as before. A copy in a skills folder keeps the plain name /dokploy-ops.agents/openai.yaml inside the skill folder. Other agents ignore that file./hetzner-ops:hetzner-ops, with the plugin name in front. This has been true since 1.2.0, and this release marks it as breaking. Use the new name, or ask in words, which works as before. A copy in a skills folder keeps the plain name /hetzner-ops.agents/openai.yaml inside the skill folder. Other agents ignore that file./namecheap-ops:namecheap-ops, with the plugin name in front. This has been true since 1.2.0, and this release marks it as breaking. Use the new name, or ask in words, which works as before. A copy in a skills folder keeps the plain name /namecheap-ops.agents/openai.yaml inside the skill folder. Other agents ignore that file./toolkit-credentials:toolkit-credentials, with the plugin name in front. This has been true since 1.2.0, and this release marks it as breaking. Use the new name, or ask in words, which works as before. A copy in a skills folder keeps the plain name /toolkit-credentials.agents/openai.yaml inside the skill folder. Other agents ignore that file./toolkit-updates:toolkit-updates, with the plugin name in front. This has been true since 1.1.0, and this release marks it as breaking. Use the new name, or ask in words, which works as before. A copy in a skills folder keeps the plain name /toolkit-updates.agents/openai.yaml inside the skill folder. Other agents ignore that file./watch-video:watch-video, with the plugin name in front. This has been true since 1.3.0, and this release marks it as breaking. Use the new name, or ask in words, which works as before. A copy in a skills folder keeps the plain name /watch-video.agents/openai.yaml inside the skill folder. Other agents ignore that file.SKILL.md. It tells the agent how to check for a newer version, and to report what changed before it updates.SKILL.md. It tells the agent how to check for a newer version, and to report what changed before it updates.SKILL.md. It tells the agent how to check for a newer version, and to report what changed before it updates.SKILL.md. It tells the agent how to check for a newer version, and to report what changed before it updates.SKILL.md. It tells the agent how to check for a newer version, and to report what changed before it updates.skills/ folder. The older layout still works.--available lists toolkit skills that are not installed yet, and --json gives structured output.SKILL.md. It tells the agent how to check for a newer version, and to report what changed before it updates.Built at Itqan Lab, a design and technology studio.
إتقان — itqan, the Arabic word for mastery: doing a thing precisely, and completely.