feat(pool): no machine-wide width - 128 passes per available account

Jorijn: "there is no need to do a global cap anymore, the cap should be
128 per available account". `--concurrency` no longer defaults to 128:
the limits are each OAuth account's 128 (and the local login's), every
driver's counted, an API seat's own REVIEW_GATES_API_CONCURRENCY, and
free memory. `--concurrency N` still caps the machine when given; a value
below 1 is a usage error.

The dump's args.concurrency is null then. The board's driver settings
read `128 per account`, which every driver's `seats` shows as that
driver's width, and the `seats` header says a width is that driver's cap.
The Scheduler test runs 300 passes on two accounts: 256 at once, 128 each.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Jorijn van der Graaf 2026-10-31 17:33:11 +01:00
commit 61ee4602f4
Signed by: jorijnvdgraaf
GPG key ID: 2937E59CDCC1BCFB
12 changed files with 98 additions and 19 deletions

View file

@ -110,6 +110,7 @@ Conventions:
- **Seat placement** (SeatPool, Seats, Drivers):
- An account whose 5-hour limit Anthropic has removed reports the weekly window alone. Its burst window (`Seats::BurstWindow`) is the weekly one: it is ranked and admitted by the weekly window with the passes in flight projected on it (0.04 points per pass at xhigh, 0.125 otherwise), under the 95 % cutoff.
- Every OAuth account and the local login takes at most 128 passes at once, every driver's counted.
- There is no machine-wide width (2026-10-31, Jorijn: "there is no need to do a global cap anymore, the cap should be 128 per available account"). `--concurrency` caps only when given, and the dump's `args.concurrency` is then null. The board's driver settings read `128 per account`, which every driver's `seats` shows as that driver's width (Scheduler, Runs).
- A 'concurrents' rejection pauses the seat for 20 s or its retry-after, not until a window's reset.
- Board entries and the folded `foreign` record carry `cost7d`, and so does the dump's seat summary.
- The `seats` report and the log table say `no limit` in such an account's 5-hour column.

View file

@ -27,7 +27,7 @@ alias crafter-review='flatpak run net.catcrafts.CrafterReview'
What the machine provides:
- **The Claude Code CLI.** It is proprietary and not bundled. crafter-review looks for it in this order: `CLAUDE_BIN` or `REVIEW_GATES_CLAUDE_BIN`, the PATH, the editor extensions' bundled binary (newest first), then `~/.claude/local` and `~/.local/bin`. It runs inside the sandbox, so every pass is measured and every pass's process tree can be killed.
- **Seats.** The lens processes run on the seats in the repository's `.env` (`CLAUDE_CODE_OAUTH_TOKEN1..6`, `ANTHROPIC_API_KEY1..6`), else this machine's own login. `crafter-review seats` probes and reports them. It never prints a token. An account whose 5-hour limit Anthropic has removed shows `no limit` in that column and is placed by its weekly window; every account takes at most 128 passes at once.
- **Seats.** The lens processes run on the seats in the repository's `.env` (`CLAUDE_CODE_OAUTH_TOKEN1..6`, `ANTHROPIC_API_KEY1..6`), else this machine's own login. `crafter-review seats` probes and reports them. It never prints a token. An account whose 5-hour limit Anthropic has removed shows `no limit` in that column and is placed by its weekly window. Every account takes at most 128 passes at once, every driver's counted, and there is no machine-wide cap beyond that and free memory (`run --concurrency N` sets one).
- **Chromium, for a lens with the browser.** Run `crafter-review browser install` once. It installs the Chromium build the pinned `@playwright/mcp@0.0.83` drives into `~/.cache/ms-playwright`, where the MCP server looks. It runs Playwright's own installer, which also removes the browser builds no installed Playwright on the machine refers to.
- **A local Supabase stack, for a lens with the database** (`supabase start`, or the repository's own bring-up). The `sql` helper speaks the Postgres protocol itself: no `psql` and no driver to install. Every call is one transaction, rolled back.
- **gpg, if your commits are signed.** `freeze` reads each commit's signature through gnupg's keyboxd socket, which is all the sandbox is given. The agent, which signs, stays out of reach.

View file

@ -294,8 +294,9 @@ inventory unit, against a frozen commit range.)";
{"--units", true, "UNITS", "comma-separated unit ids (test only)"},
{"--warm", false, "", "test runs only: let one pass per prompt prefix go first, as a full run does"},
{"--concurrency", true, "CONCURRENCY",
"lens processes in flight at once across all seats (default 128, 2026-10-15; the pool still places each one by seat headroom, launches a pass only when it fits in free "
"memory at its environment's measured cost, and gives a pass with the browser a stack of its own)",
"a cap on the lens processes in flight on this machine, every driver's counted (default none: each OAuth account takes at most 128 - "
"REVIEW_GATES_OAUTH_CONCURRENCY - and an API seat REVIEW_GATES_API_CONCURRENCY; the pool still places each pass by seat headroom, launches it only when it "
"fits in free memory at its environment's measured cost, and gives a pass with the browser a stack of its own)",
Kind::Int},
{"--effort", true, "EFFORT", effort},
{"--reads", true, "READS", reads, Kind::Int},
@ -355,7 +356,10 @@ inventory unit, against a frozen commit range.)";
if (namesNothing(a.only)) return UsageError(UsageLine(Prog, specs), Prog, std::format("argument --only: names no lens file ({})", Dyn::Repr(*a.only)));
if (namesNothing(parsed.Value("units"))) return UsageError(UsageLine(Prog, specs), Prog, std::format("argument --units: names no unit id ({})", Dyn::Repr(*parsed.Value("units"))));
a.warm = parsed.Flag("warm");
if (const std::optional<std::string> n = parsed.Value("concurrency")) a.concurrency = *Dyn::ParseInt(*n);
if (const std::optional<std::string> n = parsed.Value("concurrency")) {
a.concurrency = *Dyn::ParseInt(*n);
if (*a.concurrency < 1) return UsageError(UsageLine(Prog, specs), Prog, "argument --concurrency: a number from 1 up (leave it out for no cap)");
}
a.effort = parsed.Value("effort").value_or("");
if (a.effort.empty()) a.effort = std::string(Plan::DefaultEffort());
a.noReuse = parsed.Flag("no-reuse");

View file

@ -784,7 +784,7 @@ namespace Crafter::Review::Pool {
Sleep(std::min(60.0, Number(pauseUntil) - now));
continue;
}
while (!queue.empty() && static_cast<std::int64_t>(running.size()) < o.concurrency && !paused) {
while (!queue.empty() && (!o.concurrency || static_cast<std::int64_t>(running.size()) < *o.concurrency) && !paused) {
std::vector<Json*> live;
for (Slot& slot : running) live.push_back(&slot.job);
// Without a board every queued pass waiting for a stack or
@ -820,8 +820,8 @@ namespace Crafter::Review::Pool {
if (o.board) {
// The board folded, the seats ranked, the admission
// checked against every driver's passes and the pick
// registered in one locked section; the pool width is
// the machine's cap.
// registered in one locked section; a pool width, when
// given, is the machine's cap.
const SeatPool::Board::Placed placed = o.board->Place(*o.seats, o.now(), o.seatConfig, o.apiConfig,
[&](const std::vector<Json*>& c, const SeatPool::Others& others) { return choose(c, &others); }, o.concurrency);
cands = placed.ranked;
@ -833,7 +833,7 @@ namespace Crafter::Review::Pool {
if (full) {
if (roomWait.is_null()) roomWait = pauseUntil;
RaisePause(o.now() + o.noRoomWait);
Log(std::format("pool: {} passes in flight on this machine, the cap is {} - re-checked in {:g} s, {} queued", full, o.concurrency, o.noRoomWait, queue.size()));
Log(std::format("pool: {} passes in flight on this machine, the cap is {} - re-checked in {:g} s, {} queued", full, o.concurrency.value_or(0), o.noRoomWait, queue.size()));
paused = true;
break;
}

View file

@ -184,7 +184,7 @@ namespace Crafter::Review::Run {
{"without", opt(a.without)},
{"units", units},
{"warm", a.warm},
{"concurrency", a.concurrency},
{"concurrency", a.concurrency ? Json(*a.concurrency) : Json()},
{"effort", a.effort},
{"reads", a.reads},
{"no_reuse", a.noReuse},
@ -655,7 +655,9 @@ namespace Crafter::Review::Run {
{"dump", Basename(dumpPath)},
{"scratch", scratch.string()},
{"effort", a.effort},
{"settings", {{"maxUtil", pc.at("maxUtil")}, {"apiConcurrency", apiCfg.at("concurrency")}, {"lensCostPct", pc.at("lensCostPct")}, {"concurrency", a.concurrency}}}},
{"settings", {{"maxUtil", pc.at("maxUtil")}, {"apiConcurrency", apiCfg.at("concurrency")}, {"lensCostPct", pc.at("lensCostPct")},
// Shown as the driver's width by every driver's `seats`.
{"concurrency", a.concurrency ? Json(*a.concurrency) : Json(std::format("{} per account", Dyn::Str(pc.at("oauthConcurrency"))))}}}},
Platform::Now());
}
const bool openrouter = std::ranges::any_of(seats, [](const Json& s) { return s.at("kind") == "openrouter"; });

View file

@ -1466,7 +1466,9 @@ namespace Crafter::Review::SeatPool {
for (auto it = all.begin(); it != all.end(); ++it) {
if (it.key() != me) drivers[it.key()] = it.value();
}
const std::string width = cap ? std::format("; {} in flight on this machine, the cap is {}", TotalInflight(data), *cap) : std::string();
const std::int64_t perAccount = Dyn::Has(config, "oauthConcurrency") ? Int(AsDouble(config.at("oauthConcurrency"))) : OauthConcurrency;
const std::string width = cap ? std::format("; {} in flight on this machine, the cap is {}", TotalInflight(data), *cap)
: std::format("; {} in flight on this machine, at most {} per account", TotalInflight(data), perAccount);
if (drivers.empty()) return {std::format("board: {}; no other driver running{}", path.string(), width)};
std::vector<std::string> each;
for (auto it = drivers.begin(); it != drivers.end(); ++it) {

View file

@ -1009,6 +1009,6 @@ namespace Crafter::Review::Seats {
FormatWhen(NumberOr0(Dyn::Get(d, "since")), now, zone), Dyn::Str(Dyn::Has(d, "inflight") ? d.at("inflight") : zero), held));
total += static_cast<std::int64_t>(NumberOr0(Dyn::Get(d, "inflight")));
}
return std::format("drivers ({} in flight on this machine; the width is the machine's cap): {}", total, Dyn::Join(parts, "; "));
return std::format("drivers ({} in flight on this machine; a width is that driver's cap): {}", total, Dyn::Join(parts, "; "));
}
}

View file

@ -193,7 +193,9 @@ export namespace Crafter::Review::Pool {
struct SchedulerOptions {
Runner* runner = nullptr;
std::int64_t concurrency = 8;
// Passes in flight at once on this machine (with a board, every
// driver's counted); none: only each seat's own cap.
std::optional<std::int64_t> concurrency;
Log log;
std::optional<std::filesystem::path> progressPath;
double pauseAt = 85;

View file

@ -46,7 +46,11 @@ export namespace Crafter::Review::Run {
// --units, split on commas.
std::optional<std::set<std::string>> units;
bool warm = false;
std::int64_t concurrency = 128;
// A cap on the passes in flight on this machine, every driver's
// counted; none by default (2026-10-31, Jorijn: "there is no need to
// do a global cap anymore, the cap should be 128 per available
// account"): each account's own cap is the limit.
std::optional<std::int64_t> concurrency;
std::string effort;
std::int64_t reads = Plan::Reads;
bool noReuse = false;

View file

@ -103,8 +103,9 @@ side (`slots.*`). Each instrument is shared in its own way:
at what they were admitted at, and the budget's free memory is the
reading from before they ran (`Board::MemoryBase`) — "nothing in flight"
means nothing on the machine; their passes count in each seat's
projection and in an API key's slots; and `--concurrency` is the
machine's width, every driver's passes counted. A driver waiting on
projection, in an account's 128 and in an API key's slots; and
`--concurrency`, when given, is the machine's width, every driver's
passes counted. A driver waiting on
what another holds re-checks every 30 s. A driver that died is swept by
its lifetime lock under the board's `drivers/`, a pass by its deadline. A
driver takes its passes with it however it ends: each pass runs under a
@ -241,8 +242,13 @@ therefore the hash: re-freeze with `--replace` into the same dir
| 4 | the owner | approves this exact hash, then the push happens |
Every lens runs on Opus 5.5 at effort `xhigh` (`--effort` overrides).
`--concurrency` is 128 by default (2026-10-15), the width of the machine:
two drivers running at once share it (2026-10-31, the board). One full run
There is no width of the machine by default (2026-10-31, Jorijn: "there is
no need to do a global cap anymore, the cap should be 128 per available
account"): every OAuth account takes at most 128 passes at once, every
driver's counted (`REVIEW_GATES_OAUTH_CONCURRENCY`), an API seat its
`REVIEW_GATES_API_CONCURRENCY`, and free memory the rest; `--concurrency N`
still caps the machine when given (it was 128 by default from 2026-10-15,
the width two drivers running at once shared). One full run
per hash: a second on the same hash is refused while the first is live
(`State::RunLock`; a test run takes a run number of its own), and no two
drivers share a scratch dir. `flatpak run net.catcrafts.CrafterReview seats` shows the
@ -583,4 +589,6 @@ repository.
account takes at most 128 passes at once (every run measured with 149 or
more in flight on one account drew usage-limit rejections); a 'concurrents' rejection backs the seat off for seconds,
not until a window's reset; `seats` shows `no limit` in such an account's
5-hour column.
5-hour column. Then (Jorijn: "there is no need to do a global cap anymore,
the cap should be 128 per available account") the machine-wide width
went: `--concurrency` caps only when given.

View file

@ -341,11 +341,21 @@ echo "{\"type\":\"result\",\"subtype\":\"success\",\"is_error\":false,\"result\"
// Stopped by `crafter-review stop`'s file, and by SIGTERM, while the
// first gate's whole pass is under way: the pass is killed, no later
// gate starts, nothing is left running.
std::vector<std::string> widths;
for (const bool bySignal : {false, true}) {
before = calls();
const std::string name = bySignal ? "term" : "file";
const Process::Child child = start({"--no-reuse", "--no-salvage"}, {{"FAKE_SLEEP", "573"}}, name);
const bool running = WaitFor([&] { return calls() > before; }, std::chrono::seconds(30));
// While it runs, the board shows its width: per account, there
// being no machine-wide cap.
std::string width;
try {
const Json drivers = Dyn::LoadJsonFile(base / "board.json").at("drivers");
for (auto it = drivers.begin(); it != drivers.end(); ++it) width = Text(Dyn::Get(Dyn::Get(it.value(), "settings"), "concurrency"));
} catch (const std::exception&) {
}
widths.push_back(width);
std::string scratch;
for (const std::string& line : Lines(base / std::format("{}.out", name))) {
if (const std::size_t at = line.rfind("; scratch "); at != std::string::npos) scratch = line.substr(at + 10);
@ -368,6 +378,15 @@ echo "{\"type\":\"result\",\"subtype\":\"success\",\"is_error\":false,\"result\"
dump.is_object() ? Text(dump.at("stoppedAt")) : Read(base / std::format("{}.err", name))));
}
// No machine-wide width by default (2026-10-31): the dump's args record
// none, a live run's board entry shows the per-account one, and a
// width below 1 is refused.
const Process::Captured noWidth = run({"--concurrency", "0"});
Check("no global cap by default: the dump records none, a live run shows `128 per account` on the board; --concurrency below 1 is a usage error",
dumpB.is_object() && dumpB.at("args").at("concurrency").is_null() && widths == std::vector<std::string>{"128 per account", "128 per account"} && noWidth.returnCode == 3
&& noWidth.err.contains("argument --concurrency: a number from 1 up"),
std::format("{} | {}", Dyn::Join(widths, ", "), noWidth.err));
// An env file that is not UTF-8: the preflight stops the run.
Write(base / "bad.env", std::string("\xff\n", 2));
const Process::Captured badEnv = run({"--no-reuse", "--no-salvage", "--env-file", (base / "bad.env").string()});

View file

@ -145,6 +145,43 @@ namespace {
Check("a no-room wait ends on the next completion, not at the window reset",
r.DoneList().size() == 3 && Seconds(t0) < 5 && AnyContains(waited, "no seat has room") && AnyContains(waited, "seat room is re-checked") && r.PauseUntil() == 0, Dyn::Join(waited, " | "));
// 2026-10-31: no machine-wide width by default - each account takes at
// most 128 passes at once, every driver's counted.
{
std::vector<Json> two;
for (const std::string_view id : {std::string_view("1"), std::string_view("2")}) {
Json s = SeatPool::NewSeat(std::string(id), "oauth", std::format("t{}", id));
s["state"] = "ok";
s["windows"] = {{"five_hour", {{"utilization", 10}, {"resetsAt", now + 3600}}}, {"seven_day", {{"utilization", 10}, {"resetsAt", now + 86400}}}};
two.push_back(std::move(s));
}
std::mutex lock;
std::map<std::string, std::int64_t> inSeat;
std::map<std::string, std::int64_t> peakSeat;
std::int64_t inFlight = 0;
std::int64_t peak = 0;
Pool::Scheduler wide({.log = [](const std::string&) {}, .seats = &two, .seatConfig = SeatPool::PoolConfig({}), .apiConfig = SeatPool::ApiConfig({})});
std::vector<Json> many;
for (std::int32_t i = 0; i < 300; ++i) many.push_back({{"label", std::format("w{}", i)}});
wide.Run(many, [&](const Json& j, std::int64_t) {
const std::string seat = Dyn::Str(j.at("seat"));
{
const std::lock_guard guard(lock);
peakSeat[seat] = std::max(peakSeat[seat], ++inSeat[seat]);
peak = std::max(peak, ++inFlight);
}
std::this_thread::sleep_for(300ms);
{
const std::lock_guard guard(lock);
--inSeat[seat];
--inFlight;
}
return std::pair<Json, Json>({{"label", j.at("label")}, {"status", "ok"}}, nullptr);
}, NoFollowUp);
Check("no machine-wide width by default: two accounts take 128 passes each, 256 at once, and all 300 finish",
wide.DoneList().size() == 300 && peak == 256 && peakSeat["1"] == 128 && peakSeat["2"] == 128, std::format("peak {}, per seat {} and {}", peak, peakSeat["1"], peakSeat["2"]));
}
std::map<std::string, double> started;
std::map<std::string, double> answered;
logs.Clear();