clankerbar-cli

README

clankerbar

Drive a coding agent through your clankerbar backlog, unattended. clankerbar respawns fresh harness sessions that work the backlog one task at a time, and survives usage limits — pausing on the cap and resuming (including on an early reset) instead of dying until you come back.

It's a local, open-source client of the clankerbar control plane: the hosted plane holds the state (your backlog, over MCP), this runs on your machine and drives your coding agent. Because it holds your credentials and can edit your code and run shell commands, the source is open so you can audit exactly what it does before you trust it with an overnight run.

Status: pre-1.0, released. The loop runs and is what drives clankerbar's own backlog; the harness adapters and the limit/budget machinery are still being hardened. It is a supported surface rather than an internal tool - see Versioning for what the v0.x line does and does not promise.

How it works

Each iteration is a fresh harness session told to work one task, landing a PR at in_review. That is the prompt knob, and the wording is exact: the served protocol reads "work the next backlog item" as one task and "work the backlog" as drain the whole ready queue in this session. One task per iteration is the default because it is what keeps a session's context bounded no matter how long the queue is, and because the operator's controls — console pause, HALT, max_iterations, the budget breaker — are consulted between iterations, so a shorter iteration is a shorter wait for them to bite. Set prompt to the drain phrase if you want the old behaviour, and expect both properties to go with it.

When a session ends, the driver spawns a new one — so context never rots across a long run, and a session killed mid-task is fine: the driver hands the task straight back to the queue (see Resilience) and the next iteration picks it up. The backlog is the durable state; the loop is thin.

Phases: splitting one task across two sessions

A session's context grows monotonically and is re-read on every turn, so a long task gets expensive at the end: one measured task spent 66.2M tokens over 370 turns, and its last four deciles were 53% of that purely because each of those turns re-read a quarter of a million tokens. The served protocol already tells a session to shed its context at a safe checkpoint, and says in the same breath that the rule is a no-op if the harness cannot discard context. Inside a claude -p run there is no tool and no slash command the model can use to discard its own context, so for the agent that rule is dead letter.

The driver does have options, and they were evaluated rather than assumed away (claude 2.1.226): --autocompact <auto|tokens> compacts at a threshold, and --continue / --resume / --fork-session restore a session rather than shedding one. A phase boundary was chosen over compaction because it lands where everything load-bearing is already durable somewhere else — code pushed, bar and standing decisions on the plane — so the next phase re-reads them from source, where a summary that garbled the bar would not announce itself. It is also harness-agnostic, where --autocompact is Claude-only. The two compose, and trying autocompact inside a long phase costs nothing.

phases splits one task across several sessions, so the context resets at that checkpoint:

"phases": [{ "name": "implement" }, { "name": "review" }]

Phase 1 claims the task, implements it in a worktree cut from the repo's integration branch, self-verifies, commits, pushes and records the branch — then stops. When phase 1 instead finds an existing branch or worktree already carrying the task's work (a recovered claim, a prior session's WIP), its first step is to merge the integration branch into it - a merge, never a rebase, since rewriting pushed history breaks tip comparison - and re-validate the task against the merged tip before touching what it found: superseded means record the decision and park the task with an outcome saying so, or salvage only what still adds value - never verify-and-push stale work. Phase 2 starts on a fresh context, resumes the same run (the driver substitutes the task and run ids into its brief, so it calls heartbeat instead of claiming), re-reads the bar and the standing decisions from the plane, works in that same worktree — never on the integration branch itself — runs the adversarial review, fixes what it finds, pushes, opens a PR targeting the integration branch if no PR exists yet for the branch, and hands the task to in_review. The claim is held across the seam, so the task is never posted back to the queue mid-sequence. See docs/releases.md for what happens to that PR after in_review — merged by a clanker only once a human approves it, promoted to a release only by the operator.

The split works because each phase's prompt is a scope instruction — "your job this session is implementation only, then stop" — rather than a request that the model manage its own context. Naming a phase takes its built-in brief; set prompt on it to write your own, and max_turns to cap it as a backstop for a session that works past its brief (whatever it leaves uncommitted is then caught by the salvage). prompt and phases are mutually exclusive.

It is opt-in, and worth understanding before switching it on. A task that reaches its checkpoint and stops is only half finished until the next session runs, so this trades a driver that cannot lose a task for one that can leave it half-done if the sequence is interrupted. The budget breaker fires at the phase boundary too. Two other consequences worth knowing:

• max_iterations counts task sequences, not sessions. A two-phase config spawns roughly twice the sessions for the same limit.
• A multi-phase config requires a harness that observes the session's task claim, which today means claude or opencode. The handback across a seam, the salvage and the delivery check all depend on it, so a config naming codex alongside two or more phases is refused at validation rather than quietly stopping after phase 1 on every task. A single phase hands off to nobody, so it is allowed anywhere.
• max_turns is a claude-only backstop. opencode can now run phases, but Invocation.MaxTurns never reaches its CLI and it has no mid-stream token ceiling either, so a phased opencode config has no per-session cap at all — only the budget breaker at the phase boundary and max_wall_clock. Set those deliberately rather than assuming the max_turns you wrote is doing anything.
• A sequence that ends on the implement brief is refused, on any harness: that brief tells its session to stop at the checkpoint and leave in_review alone, so with no phase after it every task would stop half-finished, forever, with nothing in the logs reading as an error.

The saving is projected at 20-28% for a two-way cut — modelled off one real task's decile curve, not measured from a phased run, and stated that way deliberately until a phased run has been measured. Splitting thinner earns less each time while every extra boundary still pays a session's full startup cost.

Session-initiated handoff: a boundary the session chooses

Phases cut at two fixed, driver-chosen points. A session can also choose its own cut: ending its final message with a marker line followed by a prompt makes the driver respawn a fresh session on that prompt, prefixed by driver-authored framing that carries the same "resume this run, do not claim" contract a phase resume gets from its brief — same task, same run-continuity rules as a phase resume — instead of moving on. The built-in briefs teach the mechanism and its trigger: hand off at a genuine pivot (exploration finished and implementation about to start; one sub-goal landed and the next beginning), writing the successor's prompt the way an operator's continuation prompt reads — decisions made, state verified, exact next steps, nothing about the journey. Most tasks need zero handoffs.

A session-authored respawn prompt is self-directed prompt injection, so it is honoured only inside three guards: each respawn consumes an iteration of max_iterations (the one exception to "counts task sequences") and the budget breaker runs before the respawn ever spawns; an over-size prompt (over 4KB) is refused with a logged fallback to the normal path, never truncated; and a chain of consecutive handoffs is capped at 3 per phase, after which the sequence falls back to the standard brief; max_iterations and the budget still bound the total across every phase. Every respawn is named in the daemon log and tagged in its iteration log's filename (-h1, -h2, ...), so the chain reads the way phases do.

The independence the split also buys

Context economy is why phases were built, but it is not the only thing they are worth. Before phases, one session implemented a task, dispatched its own reviewer and applied the fixes: the thing being reviewed wrote the reviewer's brief. An author naming what to attack frames away from its own weak spots without meaning to — measured on one task, a hand-written brief named three things to attack, all three came back fine, and the real defect the review found was something the brief never pointed at.

A phase boundary fixes that structurally rather than by asking anyone to be fair-minded: the driver is the caller, so the implementing session cannot write the reviewing session's prompt. Phase N+1's brief is built from your config plus exactly two values off the held claim — the task id and the run id — and phase N's session output is not an input to it at all. A test pins that absence rather than the wording of the shipped brief, because the wording is the part that is allowed to change.

The second phase also puts the review behind the checkpoint, which is where a session's context peaks and so where a usage limit or a crash is most likely to land. Before the split, a death there destroyed the findings and the half-applied fixes together; now it costs the review, never the implementation.

Per-phase models: a tier map you own

model pins one model alias for the whole run. models and a phase's tier let different phases run on different models:

"models": { "strong": "opus", "standard": "sonnet", "cheap": "haiku" },
"phases": [
  { "name": "implement", "tier": "strong" },
  { "name": "review",    "tier": "strong" }
]

The buckets are yours, and the indirection is the point. clankerbar drives three harnesses and cannot know which models you have, nor which of them is cheaper — that needs a price table, which goes stale silently and in the direction that costs money. So it never learns what models exist; it learns only that you bucketed them, and a phase names a bucket, never an alias. ("tier": "opus" does not pin opus. It looks for a bucket called opus.)

Everything falls back, and nothing here can stop a run. A phase with no tier runs on model; a tier naming a bucket you never defined runs on model and says so in the iteration log; model itself empty means the harness picks, as it always has. A config written before any of this existed behaves exactly as it did. A mistyped bucket costs a log line at 3am, not a stopped drain.

Which way to tier is a question of blast radius, not difficulty. A phase that produces a durable artifact — code, a recorded decision, a bar someone will be measured against — is worth the strong bucket, because a mis-recorded decision is read for months by people who cannot see it was wrong. A phase that only compresses a large output into a small answer can take a cheap one, because the error is visible immediately to whoever asked the closed question. Nothing is tiered by default: an untouched config runs every phase on the run's model, and these two shipped phases both produce durable artifacts.

Per-phase harnesses: implement on one, review on another

A phase can name the harness it runs on, not only the model:

"harness": "claude",
"harnesses": {
  "opencode": {
    "config_dir": "~/.config/opencode",
    "mcp_config_path": "~/.config/clankerbar/opencode-mcp.json"
  }
},
"phases": [
  { "name": "implement", "harness": "opencode" },
  { "name": "review" }
]

That runs the implementation on a cheap provider-agnostic backend and keeps the adversarial review on the harness with the subagent machinery. Nothing in the seam resists it: each phase is already a fresh session seeded from the observed claim, so what crosses the boundary is a task id and a run id, and both live on the plane rather than in a session.

The harness-shaped fields stop being run-wide, and that is the whole of the configuration. config_dir, mcp_config_path, settings_path, model and models at the top level describe your harness and no other, because each is a dialect: config_dir is CLAUDE_CONFIG_DIR for one adapter and OPENCODE_CONFIG_DIR for another, mcp_config_path is Claude's .mcp.json for one and an opencode-schema config for another — and opencode does not ignore the difference, it refuses to start on mcpServers. So a phase on another harness inherits none of them and reads its own harnesses.<name> block instead. A phase naming a harness with no block is refused at validation, rather than spawning a session with no clankerbar tools that looks in the log like a model that declined the work.

Tiers are resolved per harness. A phase's tier names a bucket in the tier map of the harness that phase runs on, so one "tier": "strong" can mean opus here and something else there — the bucket name is your policy and travels, the alias inside it is a provider's and does not. A harness whose block names no models runs on that harness's own configured default, which is where opencode's model lives; a claude alias is never handed to another harness's --model.

Every phase's harness has to observe the session's task claim (claude or opencode), not only the first one's. The first phase has to observe a claim or there is nothing to hand on; a later phase is handed one - but seeding it is the adapter's job, and an adapter that does not track claims does not seed either. A non-tracking phase mid-sequence would end every drain early with a log line that reads like an ordinary finish, and one at the end would leave the driver handing the task back over its own work. Validation names the phase and the harness when it refuses.

A block that declares nothing is refused like a missing one, since an empty one spawns exactly the session the rule exists to prevent. And --harness / --model are refused outright on a config that selects harnesses per phase: both flags assert the run has one harness, and --harness would silently re-point which harness the run-wide fields belong to. Edit the harnesses block instead.

A single-project run is slug-checked too. With no projects block the poll's slug comes from the top-level mcp_config_path while each harness's sessions work whatever its file names, so two files naming different projects are refused the same way a mismatched projects entry is - the poll would gate on one project's counts while sessions claimed and worked another's.

Multi-project runs need one MCP file per project per harness. That file carries two facts at once — which project (its /mcp/<slug>) and which schema (the harness's) — so a projects entry declares them under mcp_config_paths:

"projects": [
  { "slug": "clankerbar", "workdir": "~/dev",
    "mcp_config_paths": { "opencode": "~/.config/clankerbar/opencode-clankerbar.json" } }
]

Omitting one is refused rather than resolved: the single top-level harnesses.<name>.mcp_config_path names one project, so falling back to it would poll one queue while sessions worked another. doctor reports each project-and-harness pair separately, along with each harness's binary, config dir and permission policy.

On a usage limit the loop doesn't die — it pauses and polls for the reset (catching Anthropic's semi-random early resets), then continues.

Install

Download a prebuilt binary. Each release ships cross-platform binaries on the releases page - macOS and Linux, amd64 and arm64, as clankerbar_<version>_<os>_<arch>.tar.gz. Extract it and put clankerbar on your PATH. No Go toolchain required.

# macOS arm64; swap the os/arch for yours.
VERSION=0.1.0 OSARCH=darwin_arm64
curl -fsSLO "https://github.com/lecstor/clankerbar-cli/releases/download/v${VERSION}/clankerbar_${VERSION}_${OSARCH}.tar.gz"
curl -fsSLO "https://github.com/lecstor/clankerbar-cli/releases/download/v${VERSION}/checksums.txt"
shasum -a 256 --ignore-missing -c checksums.txt   # Linux: sha256sum --ignore-missing -c checksums.txt
tar -xzf "clankerbar_${VERSION}_${OSARCH}.tar.gz"
./clankerbar version

Then verify where it came from, not just that it arrived intact. The checksum above catches a corrupt or truncated download, and that is all it catches: it is served from the same release page as the archive, so anyone able to replace one could replace both. Every release is also signed with build provenance - a Sigstore attestation held by GitHub, not by the release, recording that these exact bytes came out of this repo's release.yml at a specific commit:

gh attestation verify "clankerbar_${VERSION}_${OSARCH}.tar.gz" \
  --repo lecstor/clankerbar-cli \
  --signer-workflow lecstor/clankerbar-cli/.github/workflows/release.yml

--signer-workflow is the half that makes the sentence above true: --repo alone accepts an attestation signed by any workflow in this repo, while naming the workflow pins it to the one that publishes releases. That is worth the extra line - this binary holds your credentials and runs shell commands on your machine, and provenance is the part that would notice if the archive were not ours. Needs the GitHub CLI and a gh auth login; run gh attestation verify --help for the offline and JSON-output variants.

Downloading in a browser instead of with curl gets the archive quarantined by macOS, and the binaries are not notarized, so the first run is refused with "cannot be opened because the developer cannot be verified". Clear it with xattr -d com.apple.quarantine clankerbar, or use the curl route above, which never sets the attribute.

Or build from source (requires Go 1.26+):

go install github.com/lecstor/clankerbar-cli/cmd/clankerbar@latest
# or, from a checkout:
go build -o clankerbar ./cmd/clankerbar

A source build reports the version it was stamped with - go install leaves it at the 0.0.0-dev default, so clankerbar version naming a real vX.Y.Z means you are on a release archive.

Versioning

Releases are semver tags (vX.Y.Z) and the tool is pre-1.0, so expect the v0.x line while the surface settles. Concretely, while pre-1.0:

• A breaking change can land on a minor bump (v0.1 -> v0.2). That is what the v0.x line means; pin a version if you are automating against it.
• A breaking change is never silent. Permitted is not the same as unannounced: breaking commits carry a ! (feat(cli)!: ...) and the release config hoists them into a Breaking changes heading at the top of the notes, rather than leaving them as one line among thirty.
• Releases are cut when there is something to ship, not on a schedule. The commitment is that a tag produces working, downloadable binaries.

Those promises harden at v1.0.

Usage

clankerbar run --harness=claude
clankerbar run --harness=claude --model=opus --max-iterations=10
clankerbar run --config ./clankerbar.json     # or: -c ./clankerbar.json

Flags are GNU-style: --long options, -x shorts. --config (-c) and --help (-h) are the only short aliases; everything else is long-form only. clankerbar run --help and clankerbar doctor --help list them. A short flag's value is separate (-c ./x.json) or =-joined (-c=./x.json); the inline -c./x.json form is rejected, so a typo like -cofnig cannot quietly become --config=ofnig.

If you built from a checkout before v0.1.0, this changed under you. A single dash now introduces short flags, so the Go-stdlib spelling -harness claude no longer works - use --harness claude. Anything invoking clankerbar with single-dash long options (a cron wrapper, a launchd plist, a shell alias) needs updating. The old spelling is rejected with a message naming the double-dash form rather than being quietly reinterpreted. v0.1.0 is the first release, so it ships the GNU-style spelling from the start.

Preflight: clankerbar doctor

Most of what goes wrong in an unattended run only shows up as degraded behaviour, hours in — a rejected key silently drops the loop into blind mode, a missing binary kills the first session, an unreachable plane looks exactly like an empty queue. doctor turns all of that into one cheap answer before you start:

clankerbar doctor
clankerbar doctor --config ./clankerbar.json --harness codex   # or: -c ./clankerbar.json
clankerbar doctor && clankerbar run --harness=claude   # gate a cron wrapper

It prints one PASS / WARN / FAIL line per check, with a one-line remedy under anything that isn't a PASS, and exits non-zero if any check FAILs — so it composes with && as above.

PASS  config       loaded /Users/you/.config/clankerbar/config.json
                   harness: claude
                   workdir: /Users/you/dev
                   api key origin: https://clankerbar.com
                   backlog: https://clankerbar.com/api/projects/acme/backlog-summary
PASS  harness      /usr/local/bin/claude (2.1.0)
WARN  config_dir   not set — the session inherits the ambient environment
                -> set config_dir (or --config-dir) so a cron/launchd run loads the same skills, plugins and auth as your terminal
WARN  backlog      https://clankerbar.com/api/projects/acme/backlog-summary — 0 claimable, 2 open question(s) — nothing to claim; the loop will idle without spawning
                -> answer the open question(s) at clankerbar.com, or expect an idle run
WARN  state_dir    /Users/you/.local/state/clankerbar/loop/dev-9f70ef211d1e0549 has a leftover STOP marker
                -> delete it, or the loop stops immediately: rm /Users/you/.local/state/clankerbar/loop/dev-9f70ef211d1e0549/STOP
WARN  workdir[acme] /Users/you/dev has no agent-instructions file (AGENTS.md / CLAUDE.md)
                -> add one here naming each repo below and where its protocol lives — a session started in a multi-repo parent loads nothing from the repos under it
PASS  permissions  /Users/you/.config/clankerbar/headless.json parses
WARN  toolchains   no grant for: go (/Users/you/dev/acme-cli)
                -> allow the verbs each one needs in /Users/you/.config/clankerbar/headless.json (e.g. Bash(go build:*), Bash(go vet:*), Bash(go test:*)) — a headless session fails closed, so an ungranted tool is refused with no prompt and its task ships unverified
PASS  budget       no ceiling configured — the loop stops on a STOP/HALT marker or signal; a dry backlog idle-polls rather than exiting

The checks: config (discovered, parses, validates — plus the resolved harness, workdir, api key origin and derived backlog URLs), harness (binary on PATH and runnable, with its version), config_dir (resolves, exists, looks initialised for the chosen harness), backlog (creds present and the summary read succeeds — distinguishing no creds, a rejected key, a project_required key/route mismatch, and an unreachable endpoint — plus whether the queue is gated on your open questions, or paused from the console), state_dir (the driver's own directory: writable, no leftover HALT/STOP, and not sitting inside a configured workdir - a state dir under one a session is spawned in is writable by that session, which hands it the loop's own STOP/HALT switch; under a workdir nothing runs in yet it is the same trap for the next projects[] entry that inherits it), workdir (per project: it resolves, an .mcp.json reaches it, and it carries an agent-instructions file), permissions (harness-specific policy sanity), toolchains (the build tools the project's repos need are actually granted), power (whether the machine will stay awake long enough to do the work), and budget (ceilings parse and are sane).

A multi-project config gets one backlog check and one workdir check per project — one queue can be wired wrong while the others are fine. A project entry that omits workdir or mcp_config_path inherits the top-level one, and doctor resolves it exactly the way the loop does, so it never reports on a directory your sessions will not use.

The toolchains audit reads every settings file Claude merges — the --settings policy, the config dir's settings.json/settings.local.json, and each session workdir's .claude/settings.json/.claude/settings.local.json. A narrow deny (Bash(go run:*) alongside an allowed Bash(go test:*)) is reported as a hole in an otherwise-granted tool, not as a blocked one — only a bare Bash(go:*) deny means the toolchain is unusable.

Three of those exist because of failure modes that cost a real overnight run whole iterations:

• The queue said there was work and there wasn't. A ready task gated on an unanswered question still counts as claimable, so the loop spawns, the session correctly declines to pre-empt your decision, and you pay for that report ten times. Preflight now says so before the window opens.
• A session started in a multi-repo parent reads nothing. A harness loads its instruction file, skills and project settings from the session's cwd and upward — never from the repos below it. Spawn in ~/dev and every session begins by rediscovering the layout, with none of your conventions.
• An ungranted toolchain does not stop a run, it un-verifies one. A headless session fails closed, so go test that was never allowed is refused with no prompt reaching you — and the task ships written, pushed and never compiled.
• A sleeping machine does not pause a run, it freezes one. Timers use the monotonic clock, which does not advance while the machine is suspended, so a wait stops mid-flight and the loop goes silent exactly like a hang. One run lost 5h31m of a 10h window this way, waking only in 45-second Power Nap bursts.

The dead-phase rate: clankerbar dead-rate

The opencode implement phase has been observed dying silently — a session whose final step finishes with reason unknown and no branch recorded, so it produced nothing. That rate is the number that decides whether a fix to the harness or provider worked, and the driver now tallies it live: each phase session that got past its claim counts as a phase run, and one that then died producing nothing counts as dead, broken down by phase name and harness, and reported after every drain with its denominator:

dead-phase tally: implement/opencode: 6 dead of 23 (26.1%); review/claude: 0 dead of 4 (0.0%)

A session that never got past its claim — a correctly refused takeover — counts toward neither counter, because it satisfies both conjuncts of the dead signature while being a correct refusal, not a death.

The same predicate, applied to the existing iteration logs, is available as a subcommand — the retrospective view over the logs the driver has already written, per day, per phase and per harness:

clankerbar dead-rate
clankerbar dead-rate --error tool_count_limit   # logs whose APIError events match

The scan is verified against known-positive controls. --error tool_count_limit finds exactly the three 2026-08-19 logs that carry it as an APIError event, not the later logs where the same string appears as task-body text an agent merely read. Against the recorded 2026-08-20 hand count (6 dead of 16 implement sessions, taken mid-day) it reconciles session by session: three of the six are dead; the CLA-386 takeover held WIP inherited through its own claim result, so it ran but did not die; the two CLA-390 sessions never got past their claim and count toward neither counter; and two further deaths fall after the mid-day snapshot. Over the full day the scan reports 5 dead of 23.

Sleep, on laptops

clankerbar run now holds a no-idle-sleep assertion for its own lifetime (macOS; it dies with the process, so nothing outlives the run). Two caveats it cannot fix, both of which have cost a real run its night:

• A closed lid sleeps regardless. No assertion overrides clamshell sleep.
• Start the run on AC. Plugging in later does not wake a sleeping Mac, so a machine that idle-slept on battery stays down even once power arrives — its never-idle-sleep-on-AC setting never gets a chance to apply.

When a wait does get frozen, the loop now says so on the way out, rather than leaving an unexplained gap in the log:

wait of 30m0s took 2h28m11s of wall clock — the machine was suspended for ~1h58m11s;
timers are frozen while it sleeps, so an unattended run stalls silently

When the queue lies about having work

claimable > 0 means work is available, not that it can be done. A task gated on an unanswered question is claimable and unworkable: the loop spawns, the session correctly declines to pre-empt your decision, and you pay for that report every cycle. One run did it ten times.

So after three consecutive sessions that settle nothing — nothing reaching in_review or done — the loop backs that project off for 15 minutes, then 30, then an hour, capped at two. Other projects keep draining.

So does the queue simply going quiet. A poll that shows nothing to spawn for at all means the project is idle, not fruitless — there is nothing for the loop to spawn, so it has not failed at anything — and the count is forgotten along with any wait still running. Without that, parking the blocker would leave the count standing (parked is not progress a backlog can see), and the next task you file would serve out a two-hour wait it did nothing to earn.

Note "nothing to spawn for" rather than "nothing claimable": abandoned work counts as something to spawn for. A project can hold a branch whose session died, and the sweep that hands it back runs only when an agent asks the plane for its next task — so if the loop declined to spawn while the ready queue was empty, nothing would ever ask, and the branch would sit there for good. The gate is claimable + stale_claimable, both of which the queue line reports.

What bounds a recovery that never finishes is the plane, not this back-off. Taking a branch over renews its lease, so the task stops being offerable until the lease lapses again — and the poll in between reads as idle and forgets the strike. The count therefore does not climb across successive takeovers of one branch, and the local back-off will usually never reach its threshold. That is deliberate: the plane counts hand-offs per task, and once a branch has defeated its allowance it is parked and raised to you as a question instead of being offered again, which drops it out of stale_claimable for good. So the spend is bounded per branch by the plane, at a session or so per lease period, rather than by three strikes here. The local back-off still bites the case it was built for — a target that keeps being offered the same unfinishable work without the lease resetting.

Three rather than one, because a genuinely large task can span several sessions before anything reaches a reviewer, and backing off then would throttle exactly the deep work the loop exists to do.

Answering the question ends the wait, and so does any other progress: you merging a PR, another machine's loop settling something. Both are seen on the next poll, so "immediately" means within one idle_poll_interval (60s by default), not sooner. The answer is the one worth spelling out, because it is what the back-off message asks you to go and do: an answer moves neither in_review nor done, so a falling open-question count is the only trace of it the loop can see, and it watches both signals for that reason. Answer while a session happens to be running and the loop gives that project one immediate retry rather than the next rung of the ladder. That session still settled nothing, and its strike stands.

WARN vs FAIL is the "would this still make progress?" line: no creds and an unreachable plane WARN, because the loop drains blind and still gets work done; a rejected key and a key/route mismatch FAIL, because they never self-heal and every session the loop spawns carries the same broken credential.

Control an in-flight run with markers in the state dir, which lives OUTSIDE the workdir - $XDG_STATE_HOME/clankerbar/loop/<workdir-slug>, i.e. ~/.local/state/clankerbar/loop/<workdir-slug> by default. doctor prints the resolved path:

touch ~/.local/state/clankerbar/loop/dev-9f70ef211d1e0549/STOP   # stop gracefully (responsive even mid-wait)

An explicit state_dir wins, but pointing it back inside a workdir gives every session spawned there the ability to write these markers. doctor WARNs whenever the resolved state dir sits inside a configured workdir - including the default, if your workdir happens to contain ~/.local/state.

You can also pause a run remotely from the clankerbar web console (Settings → Pause) — per project. A paused project stops getting new sessions (other projects keep draining) and the loop idle-polls until you Resume — without exiting, and without killing an in-flight session (a pause is honoured between iterations, like STOP). Remote pause rides on the driver's cheap backlog read, so it needs a wired poller: a CLANKERBAR_API_KEY and a resolvable plane URL. Your account key (mint at clankerbar.com/account/api-keys) is the normal choice — the driver polls the project-in-path summary route whenever it knows the slug, which it derives from your .mcp.json's /mcp/<slug> URL or from the projects config (below). A project-scoped key works too (the CI-style setup). With no creds or no resolvable endpoint the loop drains blind and can't see the flag — the local STOP/HALT markers remain the fallback there. Two misconfigurations hard-stop (non-zero exit) instead of blind-draining doomed sessions: a revoked/wrong key (401/403), and an account key with no derivable project slug (400 project_required from the legacy slug-less route) — give the loop a slug via projects or an /mcp/<slug> .mcp.json.

Visibility

The driver logs milestones to your terminal as it goes (timestamped): spawning, queue state each poll, done + tokens/cost, usage-limit pauses, transient retries. The Claude harness runs with --output-format stream-json, so the agent's own progress — assistant text and → Tool markers — streams live too, and each attempt is captured to its own <state-dir>/iteration-<ts>.log.

Checking what a session says it delivered

The control plane holds the backlog and takes a clanker at its word. When a session records a branch (update_task(branch: …)) or declares a delivery merged (delivery.commit + integrationBranch), nothing on the plane can look at a repository to see whether either is true. One task read done for four days while about 900 lines of its work sat unpushed on one laptop, and its own PR merged a stale snapshot.

The driver is already local and already in the git tree, so it checks both claims after each session — no new credentials, no plane change:

• A recorded branch is really on the remote, and the local tip is an ancestor of the remote tip. A local branch ahead of its remote is the failure above.
• A declared delivery really landed: commit is an ancestor of the remote tip of integrationBranch — the same ancestor check the plane asks the clanker to attest to, run rather than trusted. The integration branch is read from the remote, never from a local main that is routinely tens of commits stale. Only a closure is checked this way: a hand-off to in_review carries the delivery it is proposing, and at that moment nothing has merged yet. And only a commit id counts — a revision expression like main or HEAD is trivially its own ancestor, so it is reported as unverifiable rather than checked.

DELIVERY UNVERIFIED — CLA-134: branch "clanker/x" is ahead of origin/clanker/x by
12 commits — local tip 4a91c0de, remote 0b3f21aa; that work is UNPUSHED and
exists only in /Users/you/dev/acme

It warns; it does not refuse. A failed check is logged loudly and names what is unpushed or unmerged, but the session's own report stands — the driver does not revert a status a session chose on the strength of a check that could be looking at the wrong tree. When the merge check ran, its verdict is written back to the task as delivery.mergeVerified, true or false, so the answer outlives the log.

It fails open. No git, no remote, a workdir that is not a repository, a remote tip that cannot be fetched — all report could not verify and the run carries on. Blocking a legitimate closure because the tree could not be found would be worse than the gap this closes. A check that could not run never writes an attestation: not knowing is not the same as knowing it is fine.

Finding the tree. The driver spawns sessions in a workdir, but the work happens in a per-task worktree it did not create — and the workdir is routinely a multi-repo parent (~/dev) that is not a repository at all. Linked worktrees share their repository's refs, so the specific directory does not matter: the driver looks for the repository whose refs carry the branch, searching the workdir and its first two directory levels (enough to reach both ~/dev/<repo> and ~/dev/<repo>-wt/<task>). Repositories are never descended into and bare ones are skipped, so the search costs a bounded handful of git rev-parse calls. If two repositories under the workdir carry the same branch — a review clone sitting beside the session's own tree — that is reported as unverifiable too: checking the wrong tree is worse than not checking.

This covers unattended runs only, by design: an interactive session bypasses the driver entirely.

The PR a delivery names

When a session's delivery names a pull request, the driver also checks it through the gh CLI - the one part of the check that reaches GitHub rather than your local tree. Two conditions, both required:

• The PR is MERGEABLE. A CONFLICTING PR is refused: a conflicted branch gets no pull_request event at all (GitHub cannot compute the merge ref), so it reports zero checks rather than a failing one, and quiet CI proves nothing.
• The check rollup actually ran and passed. An empty rollup is a refusal, not a pass - "no checks found" is the exact signature of the conflicted PR above, so silence never reads as green. Skipped/neutral checks (path-filtered jobs) pass but are named in the log.

GitHub computes mergeability lazily, so an UNKNOWN first read is polled within a bounded window (30s) and then refused, never assumed. Failing or unfinished checks refuse by name. If gh is not installed the check degrades to an explicit could-not-verify - never a pass - and clankerbar doctor's pr_gate line warns about it up front.

Repos with no CI at all: refusing every delivery there would wedge the driver shut, so a per-project opt-out exists - set "allow_unchecked_pr": true (top level, or on a projects[] entry) and an empty rollup is logged as a WARNING instead of a refusal. The mergeability condition is never relaxed: a conflicted PR refuses either way. doctor prints the resolved state per project, so a loose project is visible before it fires.

Rescuing work a killed session left behind

A usage limit kills a session mid-turn. There is no graceful shutdown, so the protocol's commit, push, record the branch before you let go never runs — there is no before. The uncommitted work survives in the worktree and nowhere else: invisible to the plane, invisible to another machine, invisible to the next session. The lease then expires with no branch recorded, the task correctly returns to ready, and the next clanker starts from nothing — redoing hours of work sitting intact on the same disk. One such task cost 112.7M tokens, a third of a day's spend, for work that was done twice.

So when a session ends still holding a task, the driver looks for the worktree it was working in and, if anything is uncommitted, commits it, pushes it, and records the branch on the task — turning a redo into a takeover:

salvaged CLA-314: committed the uncommitted work in /Users/you/dev/acme-wt/cc561415
as 4a91c0de and pushed it to origin/clanker/cc561415-a-session-killed-mid-task
recorded clanker/cc561415-a-session-killed-mid-task as the hand-off branch for
CLA-314 — the next clanker takes it over instead of starting again

On every abrupt ending, not only a usage limit. A crash, a killed process and a limit are different to the supervisor and identical to the worktree, and the limit is only detected by parsing a stream that may not have arrived whole. Running it when it was not needed costs nothing: a clean worktree commits nothing, pushes nothing, and records nothing — an empty hand-off would send the next clanker to fetch a branch with nothing on it, which is worse than recording none at all.

It commits work it cannot vouch for, and says so. The driver cannot tell a half-applied refactor from good work, and the failure it exists to prevent is caused by dropping work, never by keeping it. So the commit lands with a subject that reads WIP salvage: … (unreviewed, may not build) and a body saying in as many words that nothing here was reviewed, built or tested. It skips hooks and commit signing on purpose: a lint hook would reject exactly the half-finished tree this exists to save, and a signing passphrase has nobody to ask at 3am. One exception: a worktree holding an unresolved conflict is left exactly as it is, because committing it would record a state nobody chose — it would fabricate a resolution out of the conflict markers and push a file that reads as source with <<<<<<< in it. That covers both a git operation still in flight (a merge, rebase, cherry-pick or revert) and a conflict nothing is holding open: a git stash pop, git apply --3way or git checkout -m leaves unmerged entries with no state file at all, so the unmerged index is checked as well as the operation. Either way it is logged as STRANDED WORK LEFT AS IS with the path.

It cannot touch a tree that is not this task's. The worktree is identified by its checked-out branch carrying clanker/<first 8 characters of the task id> — the task-id half, not the title-derived slug, so a task retitled mid-flight is still matched. A detached HEAD is never matched, main and staging can never match, and two trees answering to the same task is reported rather than guessed at. Nothing is ever forced: no push --force, no checkout, no branch creation, no worktree removal. A push the remote rejects leaves the commit safely on this machine and records no branch — a recorded branch is a promise that another machine can fetch it.

Recording a branch is not settling a claim. The record carries no status, so it cannot move a task, clear a holder, or post ready over work already in review — which is why it is still safe on a session whose output could not be read whole, where handing the claim back is forbidden. Once a branch is recorded, the claim is deliberately not handed back either: the lease expires and the task stays a takeover, so the next clanker is told there is work to fetch.

Skills, plugins, and auth

A headless session loads project skills from the working directory (.claude/skills/) and user/plugin skills + auth from the config dir. Set config_dir (→ CLAUDE_CONFIG_DIR / CODEX_HOME) so an unattended run loads the same skills, plugins, and login as your interactive session — a bare cron environment would otherwise have none of them.

Config

Flags override the config file, which overrides defaults. Exactly one file is discovered on its own: ~/.config/clankerbar/config.json. Any other file has to be named - clankerbar run -c ./clankerbar.json - and an unnamed clankerbar.json sitting in the directory you launch from is refused, not read; where the key may go says why. (JSON today; TOML is the likely final format.)

{
  "harness": "claude",
  "backlog_url": "https://clankerbar.com",
  "model": "opus",
  "models": { "strong": "opus", "standard": "sonnet", "cheap": "haiku" },
  "prompt": "Work the next backlog item.",
  "max_turns": 400,
  "max_session_wall_clock": 0,
  "mcp_config_path": "./.mcp.json",
  "config_dir": "~/.claude",
  "idle_poll_interval": "60s",
  "poll_interval": "30m",
  "max_retries": 0,
  "retry_cap": "5m",
  "max_zero_spend_attempts": 3,
  "budget": {
    "max_tokens": 0,
    "max_cost_usd": 0,
    "max_wall_clock": "6h",
    "max_session_tokens": 0,
    "per_harness": {}
  }
}

max_turns bounds every session's turns when its phase sets none of its own: the phase's max_turns wins, then this, then a built-in default of 400 (CLA-343 — an unphased run used to have no cap anywhere in the chain, and one session ran 1093 turns / 285.9M tokens with nothing able to stop it). "0" no longer means uncapped; it means "defer upward". Claude only: the other harnesses take no turn flag, so for them the cap is inert.

budget.max_session_tokens is the per-session runaway detector: when a session's own cumulative usage crosses it, the harness kills the session mid-stream and the salvage commits whatever it left (CLA-343). Unlike the run-wide breaker, which only checks BETWEEN sessions, this one can stop a single huge session in flight — the 285.9M runaway was 3.8x its run's whole max_tokens ceiling. 0 defers: to 2x budget.max_tokens when that is set, else to a 150M floor (~2.3x the largest legitimate session measured, CLA-309). It is deliberately always on — it is a detector, not a budget. Claude only today: the other harnesses' streams are parsed but never killed mid-session, so for them the dial is inert until an adapter grows the kill.

max_session_wall_clock is the same backstop measured in TIME, for a harness whose CLI takes no turn flag: when a session outlives it the adapter ends the process and the driver treats that exactly like a hit turn cap — the session ends, nothing is retried, nothing fails, and the run carries on (CLA-368). A phase's own max_wall_clock wins over it, and 0 is off, which is the default: unlike the turn cap it ships no built-in number, because a duration that catches a runaway on one model/provider is a routine session on another. It is measured per SESSION, so it is not budget.max_wall_clock, which is the run-wide ceiling and counts the hours a run spends waiting out a usage limit. Since CLA-406 the cap bounds an opencode session's WHOLE life including any quiet-death resurrections: the budget is computed once at spawn and spent down across the original run plus up to 5 resume rounds, not refreshed per round. The inter-round overhead (25s backoff plus a 30s probe box per round) is charged against it like any other time, so a fully-capped session does not materially outlive the dial; a resurrection round is not started unless the remaining slice covers that overhead plus a minute, and a continuation is not started on a remaining slice under a minute.

opencode only today — it is the harness with no turn flag, so this is its backstop; under claude or codex the dial is inert and doctor says so. Two consequences of that pairing are worth knowing before you set it. The salvage that commits and pushes what a cut-off session left runs only for an adapter that observes the session's task claim, which opencode does not — so a capped session's uncommitted work stays in the worktree, and the driver's log says exactly that rather than claiming a salvage that never ran. And because phases is refused for opencode for that same reason, the per-phase max_wall_clock is reachable only in a single-phase config today; the run-wide dial is the one an unphased opencode run uses.

mcp_config_path defaults to <workdir>/.mcp.json when that file exists — Claude's headless mode does not auto-discover it, so under harness: "claude" the default is what gives spawned sessions their clankerbar tools. It always gives the driver a project slug to poll with, whichever harness is configured.

The other two harnesses do not read that file, and one of them chokes on it. The default is applied regardless of harness, so this matters the moment you switch:

harness	what it does with mcp_config_path
claude	passed as --mcp-config (with --strict-mcp-config) and read as Claude's .mcp.json.
codex	ignored. codex has no per-run MCP flag; its servers come from [mcp_servers] in config.toml under CODEX_HOME (which config_dir pins).
opencode	passed as OPENCODE_CONFIG, and must be an opencode config - servers under mcp, not Claude's mcpServers. Pointed at a Claude-shaped .mcp.json, opencode refuses to start and every session dies at spawn.

Under opencode, set mcp_config_path explicitly to an opencode config. Leaving it out does not opt out: an empty value re-runs the <workdir>/.mcp.json discovery, so a workdir that carries a Claude .mcp.json hands it over no matter what. Run clankerbar doctor - it prints the caveat instead of a verdict where the file is not read, and FAILs the workdir check when opencode is pointed at a Claude-shaped one.

Point it at your workdir, not at a checkout of this repo. The .mcp.json at the root here is the maintainers' own agent wiring and names the clankerbar project slug; running the loop from inside this checkout would have it poll a queue you cannot read, which it refuses rather than drains.

Where the account-scoped key is allowed to go

CLANKERBAR_API_KEY is an account-scoped bearer token — one key covers every project you belong to, not just the one being drained. So where it may be sent is pinned, not inferred, and the pin is your own backlog_url:

• backlog_url is the one trusted origin. It is the only origin the driver ever sends the key to. It defaults to https://clankerbar.com; to point at a self-hosted plane, name it there instead (e.g. "backlog_url": "https://plane.example.com"). It is no longer derived from .mcp.json — a checkout's .mcp.json contributes at most a project slug, and never redirects the key.
• Set it in a config file the sessions cannot rewrite. backlog_url is read from ~/.config/clankerbar/config.json - the one file discovered automatically - or from a file you name with --config. A clankerbar.json sitting unnamed in the directory you launch from is refused, not read, and the loop stops rather than falling back to your home config: that directory is, by default, the checkout the spawned sessions edit, so a session could rewrite the field that decides where tomorrow's key goes. Naming it (-c ./clankerbar.json) does lift the refusal - but it lifts only the implicit adoption and guards the file no further, so keep backlog_url itself in ~/.config/clankerbar/config.json.
• The scheme floor is TLS: https anywhere, or http only to loopback (a plane under local development).

Fixating on the origin is what makes the file-for-credential leak closeable. The MCP config is handed to the harness whole, and the harness's session carries the key in its environment, so a checkout's .mcp.json could name any host it liked; that is the exfil it refuses rather than silently polls.

clankerbar run logs where the key will go, once, at startup (with the usual clankerbar: <time> log prefix in front of it):

clankerbar: 12:54:24 sending the API key to https://clankerbar.com

and clankerbar doctor prints the same origin as a preflight line:

api key origin: https://clankerbar.com

That line is the thing to check in an overnight log: if it names an origin you did not choose, stop the run before it drains anything.

Refusals, each with its remedy. A config or MCP file that would move the account key off that origin is refused at startup and flagged by doctor — never silently dropped, because dropping the file would spawn a session with no clankerbar tools and read as "the backlog was empty":

Refusal	When	Remedy
Foreign origin	an MCP server that carries the key points at an origin other than backlog_url	set backlog_url to that origin if you mean it, or fix the file — a checkout's .mcp.json is not trusted to redirect an account-scoped key
Cleartext non-loopback	backlog_url (or a server URL) is plain http to a host that is not loopback	use https — the key would cross the wire in cleartext
Local server handed the key	a server entry that carries the key (clankerbar, or any entry whose headers/env reference the variable) declares no url — a local/stdio server	give the key only to a URL-based (HTTP) server on backlog_url — a spawned process may send it anywhere
Unparseable MCP config	an MCP config that exists (at mcp_config_path, or a per-harness/per-project equivalent) cannot be read or parsed, so its destination cannot be checked	make the file readable and valid JSON so the check can see where it points

Each of those errors is written to carry its own remedy, so they read as policy rather than as a bug.

Multi-project: one instance, many queues

One loop instance can drive several projects with a single account-scoped key — you never need an instance per project, or per-project keys. Declare the projects; the driver polls each queue and round-robins sessions across whichever have claimable work, each session spawning in its own project's workdir (whose .mcp.json names /mcp/<slug>, selecting that project's tools):

{
  "harness": "claude",
  "projects": [
    { "slug": "clankerbar", "workdir": "~/dev" },
    { "slug": "ezyapp", "workdir": "~/work/ezyapp" }
  ]
}

Per entry: slug (required, the <slug> in /mcp/<slug>), workdir, and optionally mcp_config_path (defaults to <workdir>/.mcp.json). Budgets, max_iterations, and the STOP/HALT markers stay instance-global — one operator, one spend pool; the console pause is per project. With no projects list, the top-level fields drive a single project exactly as before.

Resilience

A fresh session can die on a server blip (API 5xx / overloaded) or a network hiccup — nothing to do with the task. These are retried with exponential backoff (30s → 60s → … capped at retry_cap), re-running the same iteration; a fresh session reclaims any half-done task, so a retry costs minutes, not work. Detection is anchored (Claude's API Error: prefix, connection-error strings), so a task log that merely mentions an HTTP 500 isn't mistaken for a dead session, and under Claude a 400 still stops. max_retries: 0 (the default) means never give up — keep retrying at the ceiling until the API recovers, right for a daemon; set a positive number to bound it. A usage-limit pause and a transient retry both re-run the same iteration and neither advances the iteration count. STOP stays responsive during any wait.

One thing is bounded even under max_retries: 0: attempts that report no spend at all. A budget ceiling can only stop spend it is told about, and the figures it reads arrive in the harness's own usage events. A session killed before any of them arrives - the very thing a transient retry exists for - contributes zero, so with a token- or cost-only budget the ladder is: attempt, zero spend, back off at retry_cap, attempt, forever, with only max_wall_clock able to end it. So max_zero_spend_attempts (default 3) caps consecutive attempts within one retry ladder that ended without the harness reporting any usage, and the run stops with a distinct error naming the zero-spend loop - not a budget stop, not a wall-clock stop, because what it means is that sessions are dying early rather than working. An attempt that does report resets the count, including one that honestly reports zero, so an ordinary retry ladder is untouched and a session that legitimately did nothing is never counted against you. A usage-limit pause is counted neither way: a session the cap turned away often reports nothing at all, and waiting out a quota overnight must not read as sessions failing to start. The ladder is per phase, so a phased config allows the bound in each. Raise it if a harness of yours dies silently more often than that.

And one more thing is bounded: retries taken on a failure the harness could not name. opencode's error surface changes between versions, so beyond the anchored patterns it also guesses: an exit 1 from a session that had already reported usage is retried, on the reasoning that a session which authenticated and did paid work hit something mid-flight rather than a bad config. That guess is what carries a stream drop from a version the patterns haven't caught up with — but a deterministic fault satisfies it just as well, and every dial above is blind to that one. max_retries and budget are off unless you set them, and max_zero_spend_attempts cannot fire here at all, because the reported usage the guess requires is exactly what resets its counter. So a guessed retry is capped at 2 per phase, after which the run stops with its own distinct error quoting the diagnostic nothing recognised. Retries the harness did recognise are not counted against it — a provider flapping 5xx is still yours to bound with max_retries. The cap is deliberately not configurable: the dials you set are for failures the tool understands, and the honest number of guesses is small at any budget. If you hit it, the error text is the thing to report — it is a failure shape the adapter should learn a pattern for.

The quiet-death signature. A session that dies with a final step_finish reason of unknown and all-zero usage on its own last step is the quiet-death signature. The cause is an opencode client defect, not the provider - see docs/opencode-build.md. The marker itself is described in docs/opencode-tool-schema-limits.md, which is a separate failure with a separate cause.

Since v0.10.1 the adapter does not accept the signature as terminal: it backs off, resumes the SAME session in place (opencode run --session <id>), and asks the agent to prove it is intact by naming its task ref. A mechanical match continues the session where it left off (bounded at 5 resurrections per session, one probe per death); a failed probe falls through to the dead-phase path unchanged.

What gets read as a failure. A session's output is the whole event stream, and the events quote the backlog verbatim — the task the session claimed is sitting in the same bytes. So both classifications read only what the harness said: its stderr, its own non-event output, its typed error events, and Claude's terminal_reason. A task whose body happens to say "hit your", "usage limit" or "api error: 500" is narration, and narration is never a cap and never a blip.

Scoping to the harness is necessary and not sufficient, which is why the patterns are anchored on JSON fields and word pairs rather than single words: a harness's stderr still carries text it did not write — a provider SDK's error prose, a Node stack trace — and the MCP SDK names its own classes StdioClientTransport and SSEClientTransport, so a bare transport would read a dead MCP server (a config fault that must stop the run) as a network blip to retry.

Handing the task back. A claim carries a 30-minute lease that the session heartbeats. When a session ends mid-task, that lease is left ticking with nobody renewing it — and clankerbar charges the task a reclaim to sweep it up, of which there are only two before the task is parked for a human. So the driver watches the session's stream for what it claimed, and hands the task straight back (update_task(status: ready)) on every exit: usage limit, transient blip, outright failure, or a clean finish with the work unfinished. The usage-limit case is why this matters most — the pause that follows can be hours, and the release is ordered before it.

One case is deliberately left to expire: a task whose session pushed a branch. clankerbar computes its takeover hint only for an in-progress task with a dead lease, so handing that one back would discard the very flag telling the next clanker there is work to pick up. There it takes the reclaim and keeps the hand-off. Releasing is best-effort throughout — if the plane refuses or is unreachable, the loop logs it and carries on, since an expiring lease is merely the behaviour it already had.

Not getting locked out

No coding-agent harness lets a headless caller read its remaining subscription quota, so clankerbar cannot stop at a precise "80% of the window". Two reliable alternatives, in order of preference:

1. Separate credentials — run the loop under a different account / API key than your interactive agent use, so it physically cannot spend your daytime quota. This is the only can't-be-wrong option.

2. Budget circuit breaker — set budget.max_tokens / max_cost_usd / max_wall_clock; the loop self-accounts and stops early. Blunt but simple — tune it by watching a couple of runs.

   Prefer max_cost_usd - under claude or opencode. It comes straight from the harness's own reported cost, so it measures the thing you actually care about. max_wall_clock counts the hours a run spends waiting out a usage limit, in which almost nothing is billed - one run elapsed 10h23m against an 8h ceiling with 5h31m of that asleep. Keep it as an outer bound on how late a run may finish, not as your spend ceiling. doctor warns if it is your only dial.

   Under codex, max_cost_usd is inert - set max_tokens beside it. codex exec --json reports tokens, not money, so the adapter never populates a cost figure and no code path can reach that ceiling: it is not approximate, it is absent. A codex run with cost as its only dial therefore has no effective ceiling at all, whatever the config says. Give it max_tokens (or max_wall_clock) alongside; doctor WARNs on the cost-only case rather than reporting a budget that cannot fire.

   A mixed-harness run needs one block per harness - budget.per_harness.

   "per_harness": {
     "claude": { "max_tokens": 75000000 },
     "opencode": { "max_cost_usd": 2 }
   }
   

   The dials above are one ceiling over every session a run spends, and no single number means the same thing on two backends: 75M tokens is a sane week of Claude and roughly $2 on a DeepSeek-class backend, so the same max_tokens that paces a week ends a metered drain after a task or two; and a dollar figure is meaningless for a session billed to a subscription, which reports a price nobody is charged. So each harness can carry its own block, keyed by harness name, counted over only that harness's own sessions and measured in the unit that harness understands - tokens for claude, dollars for a metered backend. Any block that trips stops the whole run: these are circuit breakers, not per-harness quotas.

   A token ceiling in a block derives the per-session runaway detector (max_session_tokens) exactly as the run-wide max_tokens does, so moving a ceiling into a block does not quietly loosen it.

   The dials above keep working exactly as they always have, over the run's whole spend, and a config that sets no per_harness block behaves as before. Set both and both apply, whichever is reached first. Wall clock is deliberately not available per harness: it measures the run, not a harness. The same goes for the CLA-262 stop above - a session whose spend cannot be measured stops the run when that session's harness is under a spend ceiling, its own block's or the run-wide one.

   doctor reads these the way it reads every other ceiling: what can actually fire. A block naming a harness this config does not run, or a max_cost_usd on a harness that never reports cost, is annotated INERT, and a config whose only ceilings are unreachable is told it has none. A block keyed by a name no adapter answers to is refused at load, since nothing could ever be charged to it.

   max_tokens counts about ten times more than it used to. Token totals now include cache reads and writes, which input_tokens excludes and which dominate a long agentic session — one run reported 140,387 tokens against $147.98 of real spend, about $1.05 per thousand, which is no model's price. The old number was the wrong one, but a max_tokens you tuned against it will now trip roughly ten times sooner. Re-tune it, or move to max_cost_usd.

   max_wall_clock is measured on the monotonic clock, so time the machine spends suspended does not count against it. That is also the clock the limit-reset check uses, so the two cannot disagree about how much of a run is left.

   When a limit's stated reset falls beyond max_wall_clock, the loop stops right then and tells you when the quota returns, rather than sleeping through the window to run one session against the fresh quota and stop on the next check.

   Every ceiling is also checked inside an iteration, before each re-run of a session that paused on a usage limit or backed off on a transient blip. Those waits can repeat for hours, and a run that reached its ceiling during one stops there instead of spending another session first.

   The polls during a pause count too. Waiting out a usage limit means probing for an early reset every poll_interval, and a probe is a real harness session — cheap (a one-character prompt, with the harness locked down to no tools or read-only), but not free, and a week-long cap polled every 30 minutes is a few hundred of them. Their tokens and cost go into the same running total the breaker reads, so a pause whose probes alone cross a ceiling ends on that ceiling rather than polling on.

   A session whose spend cannot be measured stops the run, if you set a spend ceiling. Every figure the breaker reads is parsed out of the harness's event stream, and one enormous line — a tool_result carrying a large file read — can end that stream early. The session's own total arrives in its last event, so a stream cut short reports near-zero for a session that may have cost hundreds of dollars. The loop refuses to count that figure, says so loudly, and then stops if max_tokens or max_cost_usd is set, because the ceiling you asked for can no longer be honoured. Under max_wall_clock alone, or with no ceiling at all, it carries on: the clock does not depend on anything the child said. The same session's task is left to its lease rather than handed back — the settle that released it may be in the bytes that never arrived.

Harnesses

Harness	Status
Claude Code (claude)	primary
Codex (codex)	adapter present; parsing being hardened
OpenCode (opencode)	adapter present; provider-agnostic (model comes from opencode's own config)

New harnesses are a small adapter (internal/harness). Contributions welcome.

Permissions are set per harness, and the loop fails closed under all of them. claude gets its allowlist from the settings file (--settings), codex from config.toml, and opencode from OPENCODE_PERMISSION — a JSON policy baked into the spawned process env. The opencode policy is a path-scoped PermissionConfig: the run's workdir subtree (where the session's per-task worktrees live, e.g. ~/dev/clankerbar-cli-wt/<task>) is allowed for read/edit (the workdir root's own listing included), external_directory — opencode's gate for any path-taking tool on a path outside the project boundary — allows the same subtree while denying every other path (so bash cp ~/.ssh/... stays blocked, exactly as before), and bash stays allowed tool-level (its permission patterns match parsed commands, not paths). Everything else — reads or edits outside the workdir, lsp/task/skill, and the network tools (webfetch, websearch) in every shape — is denied by a * catch-all rather than asked, and denied tools are hidden from the session's catalog entirely. The read-only search tools are an exception (CLA-390): grep, glob and list are allowed so a session can search the code it is editing instead of burning turns on bash fallbacks. Their asks carry the search pattern (a regex, a glob, a directory) rather than a workdir-relative path, so the workdir boundary cannot live in the ask for them — it lives in external_directory, which each tool runs before searching and which stays scoped to the workdir subtree, so they reach exactly what read already can and nothing more. Three more exceptions keep the session alive inside that catch-all: *_* allows the MCP tools (tool names are <server>_<tool> in opencode — clankerbar_get_backlog_summary, context7_query-docs, chrome-devtools_click — and the plane itself is reached over MCP); read also allows mcp:*, because opencode's MCP resource tools (list_mcp_resources, list_mcp_resource_templates, read_mcp_resource) do not ask under their own names — they ask under read, with mcp:<server>:<uri> patterns that the path-scoped read rules never matched, so the catch-all denied every one of them and a session could call the plane's tools while being unable to read the protocol served over the same connection; and a probe/read-only run additionally denies edit and bash for zero writes. The policy is pinned by unit tests in opencode_test.go, including the JSON key order the catch-all depends on and a replica of opencode's rule evaluator over a table of effective decisions.

Verifying the policy against a live session. The unit tests replicate opencode's evaluator; to watch the real one decide, run a session with the policy in OPENCODE_PERMISSION and read the evaluated permission=... pattern=... action.action=... lines in ~/.local/share/opencode/log/opencode.log — one line per ask, naming the rule that won. That log is the cheapest live proof available, because it records the decision whether or not the session goes on to produce any output.

The workdir should be a multi-repo parent, not a git checkout. The read/edit rules are emitted to match the patterns opencode asks for a session whose project is not inside a git repo (worktree /; the multi-repo-parent case, ~/dev). Point the workdir at a git checkout and opencode asks with checkout-relative patterns the rules cannot express, locking the session's structured Read/Edit tools out with no error — the exact failure this change removes. A per-task worktree always lives under a parent workdir, so the parent is the correct setting.

Docs

• docs/releases.md - the release/publish pipeline: how a change gets from a task PR to a published binary.
• docs/large-tool-results.md - how the driver reads a tool_result that Claude Code persisted to a file instead of inlining.
• docs/opencode-tool-schema-limits.md - the opencode-go request-time tool-schema rejection family (unsupported_tool_schema / tool_count_limit): the incident, the real root cause behind the misleading error code, and what the CLI does about a recurrence.
• docs/opencode-build.md - which opencode build the fleet runs and why not opencode2, how to tell which one actually ran, and the upstream defect that makes a session exit 0 having done nothing. Start here for a session that dies with reason: "unknown" and zero usage.

License

MIT (see LICENSE).