A hive learned the knowledge repository had changed only by registering
its own forge webhook. This subscribes to the per-hive subject the
controller now publishes on and calls the pull this daemon already runs
at boot.
Shares the hive's ONE queue connection rather than opening a second: a
second connect would double the auth-callout traffic against authelia and
give the two paths independent reconnect state, so one could be serving
while the other was still down. Same argument as the controller side.
No payload is read, because there is none to read — the webhook handler
this replaces took two fields from Forgejo and used neither, then ran
`git pull`, which re-derives everything from the repository.
At-most-once, and that is not a regression: a webhook delivery to a hive
that is down is lost identically today, and the boot pull covers it.
JetStream would require this end to publish to
`$JS.API.CONSUMER.CREATE.<stream>`, which the callout policy does not
grant, so durability would cost grants on both sides to remove a failure
the boot pull already handles.
⚠️ Documented at the call site rather than left implicit: a refused
subscription is indistinguishable from a quiet one, because NATS reports
authorization violations asynchronously on the connection. If hives stop
hearing events, the server log is the thing that knows.
futures-util comes from the workspace (same version swarm-controller
already uses), not a new dependency version.
Review finding from argus. The new endpoint test carried a SAFETY comment
claiming no other test in its module asserts on the variables it perturbs —
the wrong boundary. The module is not the unit that shares the environment,
the process is: meta.rs's render_flake_injects_otel_when_signalled mutates
the same HYPERHIVE_OTEL_ENDPOINT, both land in the one hive-c0re test binary,
and cargo runs it at default parallelism with no serialisation anywhere in
the crate. Each test independently claimed exclusive ownership of shared
global state, which is the instrument-that-looks-solid class the endpoint
change's own gate reasoning warns about.
Adds test_env with a single ENV_LOCK, taken by both. No new dependency: this
is the pattern hive-bash-mcp and hive-agent already use, and hive-bash-mcp's
helper records why it has to be crate-wide rather than per-module — two
per-module mutexes serialise nothing against each other, which produced a
CI-only flake there.
The asymmetry that makes this hard to see locally is worth stating: an agent
container has the hyperhive variables ambient-set, so a losing race still
finds a plausible value and the test passes; the nix sandbox strips them, so
only there can one thread delete a variable out from under another. Verified
in that shape with `env -u HYPERHIVE_OTEL_ENDPOINT -u
OTEL_EXPORTER_OTLP_ENDPOINT`, five consecutive runs green — a sanity check,
not a proof, since a race cannot be shown absent by running. What makes it
correct is structural: both tests take the same lock.
Deliberately scoped to the pair that overlaps. meta.rs has three further
env-mutating tests (HIVE_FORGE_URL twice, the TLS CA pair) that race with
each other, untouched here and tracked separately, because the fix is not
the mechanical one it looks like: std::sync::Mutex is not reentrant, so
adding a lock to a test whose helpers also lock deadlocks. That needs
reading per test rather than a sweep.
The controller reads per-hive status out of a JetStream KV bucket and
nothing was writing one, so every hive rendered `never_reported`. This is
the half that makes the read path mean anything.
A hive offers; the controller never reaches down to collect. The gateway
has gone down in a way where every recovery channel ran through the one
broken thing, so a status path that depended on the controller would go
dark exactly when it is needed to diagnose the controller's own network.
What it publishes is what the hive already says about itself —
`warnings::readiness()`, the same value `/health/ready` serves. Nothing
here stamps a time: freshness is derived by the reader from when the value
landed, so a hive cannot make itself look fresher than it is, and a hive
with a wrong clock skews only its own payload.
The key is this hive's `hiveName`, which `swarm.nix` already asserts is a
key of `swarm.hives` — so a hive that evaluates at all publishes under a
name the roster knows, rather than by convention.
Publish first, then wait: a hive that has just come up is the one whose
status someone is looking at, and sleeping first would make every restart
read stale for a full interval. The interval is one decision with the
controller's staleness threshold, not two — a ratio of 2 means one lost
publish still reads fresh and two consecutive misses read stale.
Failures go to the dashboard banner through SweepHealth, debounced, at
`warn` and deliberately not `crit`: `crit` is what makes this hive report
itself degraded, and a hive that cannot reach the queue is not unhealthy —
the swarm's view of it is. Publishing `degraded` because the publish
failed would be both false and self-erasing on the next tick.
Takes the crate from 26 rustdoc warnings to 1, on top of the ten in the
previous commit.
argus's review findings:
- agent_sockets.rs: [`write`] was still ambiguous (function vs macro).
The previous change narrowed the qualifier and left the ambiguity;
[`write()`] is what resolves it.
- forge/users.rs <hex> and stats/container_stats.rs <name>: unclosed
HTML tags in prose, now backticked.
The rest of the crate, so the count actually reaches zero:
- job_queue/mod.rs: Queue::graph_snapshot -> JobQueue::graph_snapshot
(there is no Queue type), and super::scheduler -> scheduler (mod.rs
*is* job_queue, so super:: pointed outside it)
- job_queue/resource.rs: NodeKind -> super::model::NodeKind
- matrix.rs: password_path(name) -> password_path; and
forge::provision_user_token -> crate::forge::provision_user_token.
Note the path has no `users` segment: forge/mod.rs declares `mod
users` private and re-exports it, so the canonical path comes from the
re-export rather than the directory tree.
- socket_server/lifecycle_handlers.rs: InfraContainer ->
hive_priv_sock::InfraContainer
- stats/otel_metrics.rs: crate::meta::otel_config is a private fn no
path can name from another module, so it becomes prose
- main.rs: redundant explicit link target dropped
coordinator.rs:405 (CrashWatchGuard) is deliberately untouched: #3244
deletes that doc block, so fixing it here would conflict with an open PR
and repair a symbol that is about to stop existing.
The container-resource exporter identified the samples it sent by
container and by hive, but not by machine or by build — so a sample
could not be attributed to the host it came from or the deploy that
produced it.
Both go on the OTEL resource rather than on each data point, for the
same reason `hive` / `swarm` already do: they are constant across one
hive-c0re. Only per-container facts stay per-data-point.
- `host.arch` — mapped to the semconv spelling, not forwarded from
rust's. The two disagree on exactly the architectures this runs on
(`x86_64` / `aarch64` vs `amd64` / `arm64`), and the failure mode is
silent: a dashboard filtering the standard value matches nothing.
A test pins this, since nothing else would catch it.
- `service.version` — the running flake rev, via
`auto_update::current_flake_rev`, the same source the dashboard
snapshot and `get_agent_meta` already use. NOT the crate version:
that's a workspace constant that never moves between deploys, so it
could not answer "which build produced this sample?". Omitted rather
than guessed when the flake ref carries no rev.
`spawn_exporter` takes the flake ref to reach the rev — the string it
needs, not the whole `Coordinator`, so the module's coupling doesn't
widen for one attribute.
The issue's third item, `container`, needs no change: `attrs()` has
emitted `container.name` per data point since this exporter landed.
Checked with clippy (`--all-targets -D warnings`), `cargo test -p
hive-c0re` (321 passed) and `nix fmt`. No option surface is touched, so
no nix-eval gate.
The push side modelled a store per peer hive: a --peer argument, a
swarm.peers.<domain>.snapshotStorePort option, and a swarm_peers module
whose entire job was answering "which peer". A swarm has exactly one
store, so none of that had anything to select between.
The receiver already proved it. It keys destination directories by
agent, not by sending hive, precisely so an agent that migrates keeps
one unbroken incremental chain -- which only makes sense if every hive
pushes to the same place. Per-hive stores would split the chain in two,
the case that keying exists to prevent.
So the destination moves to services.hyperhive.swarm.snapshotStore,
rendered into HYPERHIVE_SNAPSHOT_STORE, and swarm_peers is deleted
rather than adapted. address has no default because it is a
deployment fact this host cannot derive; port defaults because it is a
convention both ends read from the same option docs. An unset or empty
address fails naming the option instead of connecting somewhere
arbitrary, and a test asserts the message suggests no value.
Adds the caller the fd-passing machinery existed for: hivectl agent
<name> subvol snapshot push --peer <hive> resolves the peer, connects
to its snapshot store, writes the agent header, and hands the connected
socket to hive-priv, which runs btrfs send straight into it.
The split keeps the root helper ignorant. Everything that involves
knowing where a peer is, what the wire protocol looks like, and which
hive to trust happens in the unprivileged daemon; hive-priv only ever
receives an already-open descriptor. Once btrfs send starts, neither
process is in the data path, so a multi-gigabyte transfer costs no
per-byte work and survives a hive-c0re restart.
call_with_fd takes the descriptor by value and closes it as soon as the
kernel has it. A socket stays open until every copy closes, so holding
one back would leave the receiver waiting for an EOF that never comes:
btrfs receive blocks and this side reports success for a transfer the
peer never committed. Ownership makes that unrepresentable.
The peer's store port is a new swarm.peers.<domain>.snapshotStorePort
option rather than a constant matching the module default. A pushing
hive cannot read the receiver's configuration, so assuming 51821 would
push at a port nobody promised to listen on; absent, the push fails
naming the option. swarm_peers parses the mesh address the host module
has always rendered into HYPERHIVE_PEERS but nothing read.
Encoding "not configured" as weight 0 worked (the writer omitted the
line) but the type lied: 0 is not a legal cgroup v2 weight, and every
reader had to know the sentinel. Use Option<u32> end to end instead —
wire type, priv_client, HiveEnv, drop-in writer — so "unset" is a state
of the type rather than a magic value.
The nix options become nullOr, keeping their default of 80; null now
expresses "leave the setting out of the drop-in entirely" declaratively,
which is the useful shape on a host whose IO scheduler ignores
io.weight anyway.
Backward compat is unchanged: the fields stay #[serde(default)], so a
request from an older hive-c0re deserialises to None and reproduces the
pre-weights drop-in byte for byte. The test that pins that now passes
None instead of 0.
`CPUQuota=`/`MemoryMax=` are hard caps: they throttle an agent even when
the host is idle, so they are the wrong tool for "be polite under
contention". The cgroup v2 relative shares are, and neither was wired.
Adds `services.hyperhive.{agentCpuWeight,agentIoWeight}` (1..=10000,
default 80) threaded through the existing drop-in path: HiveEnv ->
write_dropins -> WriteResourceLimits -> hyperhive-limits.conf, next to
the caps already there. Hive-wide only, as the operator scoped it on the
issue: no per-agent override, no resource-limits.json field, no
dashboard form.
The default of 80 is below the kernel's 100, so agent containers yield
to everything *not* on this drop-in path -- host services and the infra
containers (hive-ci, hive-forge, hive-gateway, hive-matrix). It does not
rank agents against each other; they all carry the same weight.
`WriteResourceLimits` gains two `#[serde(default)]` fields, and the
writer treats weight 0 as "not configured" and omits the line, so an
older hive-c0re talking to a newer hive-priv still produces the exact
pre-weights drop-in. The body is extracted into `limits_dropin_body` so
that is covered by a test rather than asserted by eye.
The hive applies one `agentCpuQuota` / `agentMemoryMax` to every
container. That's the right default and the wrong ceiling: a build-heavy
agent needs headroom the other twelve don't, and raising the hive-wide
value to suit it hands that headroom to everyone.
Adds a per-agent override, persisted host-side and resolved per-field
against the hive defaults.
Follows the existing `meta/*.json` pattern (`capabilities.json`,
`tool-groups.json`): a host-side map read by `hive-c0re`, staged and
committed in the meta repo so every change lands in the audit trail.
```json
{ "sock": { "cpu_quota": "400%", "memory_max": "8G" } }
```
Fallback is **per field**, not per agent: an entry with only
`memory_max` leaves that agent on the hive-wide CPU quota. Absent file,
absent agent and absent field all resolve to the hive default, so the
feature is inert until someone opts an agent in.
Unlike the other meta files this one is **not** injected into the
container — a limit is something done *to* an agent, not something it
reads about itself.
```
hivectl agents set-limits sock --cpu-quota 400% --memory-max 8G
hivectl agents set-limits sock --reset
```
Values are validated before they're persisted: they go into a systemd
drop-in verbatim, and a typo there makes the unit fail to *start* —
turning a fat-fingered quota into a container that won't come back.
The command is declarative: each call replaces the agent's whole entry.
That makes a forgotten flag a silent revert, so a bare `set-limits
<name>` is rejected at the clap layer and clearing needs an explicit
`--reset`.
`ContainerView` gains `cpu_quota` / `memory_max`, both always populated:
there's no "unset" state to render, only "same as everyone else". They
reflect what the drop-in *says* — what the next start will enforce — not
a live cgroup reading.
The write goes through `meta::commit_resource_limits` rather than the
bare setter, so it's staged and committed under `META_LOCK`. Writing
without committing would leave the meta working tree dirty for the next
`prepare_deploy` to trip over.
Docs: `persistence.md` (the new meta file, and why it isn't injected),
`tools/hivectl.md` (the prose guide), `tools/hivectl-cli.md`
(regenerated clap dump).
Closes: internal/requests issue 25
Fast-follow for #2289 — matrix::ensure_all() (admin user, per-agent
sync, hive Space + chat-room provisioning/invites) ran periodically
every 30 minutes but only ever warn!'d to the journal on failure, so a
persistent problem (missing invites, broken admin token, etc.) was
invisible to the operator.
ensure_all() now returns bool (aggregate ok/fail across every
sub-step) instead of (), and both call sites in main.rs feed that into
a debounced SweepHealth("matrix_ensure_all", warn, threshold=2) —
matches the existing knowledge_pull pattern. A lone bad sweep
self-heals silently; two consecutive failures raise a banner that
clears on the next clean sweep.
forge::ensure_all()'s remaining independent steps are still open —
that sweep only runs once at startup (no periodic loop), so the
debounced pattern doesn't map as directly; left for a follow-up.
hive-c0re already samples each agent container's cgroup load for the
dashboard (stats/container_stats.rs); this rides those gauges out to the
configured OTLP endpoint, reusing the existing services.hyperhive.otel
config (endpoint + auth header) — no new toggle.
- New stats/otel_metrics.rs: exports via the OpenTelemetry Rust SDK (same
crates as hive-metric) with the semconv container.* metric names +
container.name attribute so off-the-shelf OTel/Grafana dashboards work,
plus the hive agent label. container.cpu.time (counter, s, from cumulative
cpu.stat usage_usec), container.memory.usage, container.memory.usage.limit;
memory peak / on-disk storage / instantaneous cpu percent stay hyperhive.*
custom (no semconv equivalent). Observable instruments read a shared
snapshot an async task refreshes (gather() is async; SDK callbacks sync).
- container_stats: expose cpu_time_usec (cumulative) on ContainerResource.
- The OTLP auth header is loaded onto hive-c0re's own unit via systemd
LoadCredential and read from $CREDENTIALS_DIRECTORY/otel-headers.
- docs/observability.md documents the host-emitted semconv metrics.
Host-side export, so it covers containers even when their agent is idle.
Two issues flagged by argus in PR #2388 review:
1. Empty-key fallback: when load_or_generate() failed, webhook_secret was
String::new(). An attacker knowing this could forge deliveries with a
valid HMAC of the empty key. Fix: change to Option<String>; on None,
skip hook registration entirely and return 503 from /webhook/* handlers
(rather than 401 with a misleadingly-verifiable empty-key HMAC).
2. Stale hook cleanup: on upgrade from old code, old loopback hooks
(http://127.0.0.1:.../webhook/knowledge, .../webhook/config-pr) were
left alongside the new domain-URL hook. Fix: during ensure_webhook /
ensure_config_pr_webhook, after listing hooks, delete any that end with
our path suffix but point at a different base URL.
clippy + nix fmt clean.
Both webhook registrations (knowledge push + config-PR pull_request) now
use the public hive domain instead of loopback:
https://<HYPERHIVE_HIVE_DOMAIN>/webhook/{knowledge,config-pr}
This routes deliveries through the gateway, bypassing the Forgejo SSRF
guard that blocked loopback delivery and silently broke the config-PR
merge flow since launch.
Changes:
- webhook_secret: new module — auto-generate + persist a 32-byte HMAC
secret to STATE_ROOT/webhook-secret on first startup; verify
X-Hub-Signature-256 on every incoming webhook POST (HMAC-SHA256).
- forge/mod.rs: ensure_config_pr_webhook now takes hive_domain +
webhook_secret; sets secret in Forgejo hook config.
- workers/knowledge.rs: ensure_webhook same update.
- dashboard/webhook.rs: both handlers read raw Bytes first, verify HMAC,
then parse JSON. Returns 401 on signature mismatch.
- dashboard/mod.rs: AppState carries webhook_secret; serve() takes it.
- main.rs: load/generate secret at startup; pass to registration tasks
+ dashboard; add 5-minute config-PR polling fallback task.
- forge/config_pr_poll.rs: new — scan agent-configs/* for open PRs with
no pending MergeConfigPr approval; queue them. Idempotent.
- stores/approvals.rs: has_pending_merge_config_pr() for poll dedup.
- nix/modules/hive-gateway.nix: remove dashboardAuth from /webhook/
location (HMAC replaces basic auth for webhook endpoints; Forgejo
cannot send HTTP Basic credentials with webhook deliveries).
mara: the background worker is redundant if c0re knows when its own
sockets go missing. damocles: 10s poll latency and redundancy are two
faces of the same issue — poll adds a reconnect window and does
redundant work when c0re could react directly.
design: c0re owns the MCP listener lifecycle, so the only time a
listener disappears without c0re knowing is when c0re itself restarts.
- replace spawn_poll (recurring 10s loop) with sync_on_start (one-shot
sweep at daemon boot): re-registers all running agents on startup
after /run/hyperhive/agents/ is cleared by the tmpfs reset.
- run_reconcile (reconcile-start path): add coord.register_agent(name)
immediately after start_with_fallback — event-driven, no poll delay.
- run_create already calls register_agent eagerly; kill/destroy paths
already call unregister_agent — no changes needed there.
tracker: #2290
- lifecycle::ensure_agent_runtime_dir(name): pure filesystem op, no
Coordinator dep. Creates /run/hyperhive/agents/<name> without touching
the MCP listener map.
- workers/mcp_sockets::spawn_poll(coord): 10 s reconcile loop (same shape
as agent_sockets::spawn_poll). Converges 'agent running => MCP listener
bound'. First tick is immediate so hive-c0re restarts re-register all
running agents without waiting a full interval. Fixes the dead-listener-
after-daemon-restart gap.
- All ensure_runtime() call sites updated:
- Prebuild/Swap/WriteDropin: Coordinator::agent_dir() (pure, no IO)
- Reconcile-Start: ensure_agent_runtime_dir + agent_dir (dir may be
missing after reboot; listener deferred to supervisor)
- run_create / handle_spawn: ensure_agent_runtime_dir + register_agent
(eager on first spawn so socket ready before harness first turn)
- apply_commit / merge_config_pr: ensure_agent_runtime_dir + agent_dir
- Manager (auto_update): ensure_agent_runtime_dir + agent_dir
(manager has no MCP listener; socket_server::start_manager owns it)
- ensure_runtime() retained in Coordinator with updated doc pointing at
the preferred split form. No callers remain outside tests.
Root cause of the boot outage: container@h-* units try to start before
hive-c0re reaches ensure_runtime, so bind-mount source dirs are missing.
Fix: hive-c0re (via hive-priv, which runs as root) writes
/etc/tmpfiles.d/hyperhive-agents.conf whenever the agent set changes.
systemd-tmpfiles-setup.service (sysinit.target) reads it at every boot
BEFORE any container units start, pre-creating:
/run/hyperhive/agents/<name> — MCP socket dir (bind -> /run/hive)
/run/hive-agent/<name> — web socket dir (bind -> /run/hive-agent)
This alone removes the outage class: even if hive-c0re is slow to start,
the bind-mount sources exist and container units can activate.
Added:
- PrivRequest::SyncAgentTmpfiles { agents } in hive-sh4re
- sync_agent_tmpfiles() in hive-priv: generates content, writes atomically,
calls systemd-tmpfiles --create to apply immediately
- priv_client::sync_agent_tmpfiles() wrapper
- lifecycle::sync_tmpfiles() best-effort helper (list + priv call)
- Call sites: hive-c0re startup, handle_spawn success, destroy success
jobs are now DAGs of primitive nodes (prebuild, stop-for-update, swap,
reconcile, signal, drain, ...) driven by one scheduler with N build
slots + per-agent lifecycle leases. per-agent power intent (wanted
up/offline) is durable in agent_power.sqlite; Reconcile nodes converge
observed state to it. kills the graceful-stop watcher thread, the
deferred-start follow-up, and the cascade pre-enqueue (fan-out on
MetaLock completion instead). tracker: #2166
The per-agent and manager sockets ran two parallel dispatchers with
duplicated lifecycle handlers (agent-side topology-gated, manager-side
ungated) plus a manager-only handler set. Collapse to one parameterized
server in socket_server.rs:
- one serve() + dispatch(req, agent, privileged, coord); start() binds
the per-agent sockets (privileged=false), start_manager() binds the
manager socket (privileged=true).
- each lifecycle/config handler (start/restart/kill/update/init_config/
apply_commit) merges its dual: the topology guard (require_child /
require_new_child) runs only on the !privileged path; init_config
records the requester as parent only when !privileged. restart keeps
the orthogonal, capability-gated + audited infra-container branch.
- the agent-state queries (loose-ends / reminder count + rollup) branch
on privileged: privileged keeps any-target + the "*" hive-wide sweep
(query_agent_state-gated), non-privileged keeps the topology/cap gate.
- the privileged-only verbs (schedules / meta-inputs / get_logs) plus
the submit/schedule/watchdog helpers move into socket_server; they are
reached via dispatch_privileged_only(), which rejects the whole group
on a non-privileged socket.
- delete manager_server.rs; repoint refs; merge the test modules.
No behavior change: the topology guard still applies on every
non-privileged lifecycle call, the privileged socket still acts on any
agent, and privileged-only verbs are still rejected on agent sockets.
Adds a durable, operator-visible audit trail of privileged operations
hive-c0re performs on behalf of an agent — the ones that cross the
agent/operator trust boundary. First entry: infra-container restarts via
the infra_admin-gated `restart` tool, which until now were recorded only
as a hive-priv journal trace.
Backend:
- new `audit_log` module: sqlite-backed store (audit_log.sqlite, same dir
as build_logs) with schema (ts/agent/action/target/outcome/detail),
best-effort `record`, `list_recent` (clamped 500), 90-day `vacuum`, and
a process-singleton handle mirroring build_logs.
- Coordinator opens + installs the handle; main spawns the hourly vacuum.
- agent_server::handle_restart_infra records every attempt (ok, error, and
capability-denied) via the global handle — best-effort, never fails the
underlying action.
- dashboard: `GET /api/audit-log` returns recent entries as JSON.
Scope is deliberately agent-initiated privileged actions only (not every
PrivRequest — token writes + nspawn edits are constant lifecycle noise).
Extensible: future agent-initiated priv ops record via the same handle.
Unit tests cover record/list ordering, the 500 clamp, and retention vacuum.
The dashboard *surface* (an AUDIT view consuming /api/audit-log) is a
frontend follow-up coordinated with iris.
The systemd ExecStart carried every host-level setting as its own flag —
nine of them, including two escaped JSON blobs (the context-window map and
the model-price table). Collapse them into a single `--config <file>` JSON.
- Reuse the existing HiveEnv as the container-injected config shape (add
Deserialize + Default), and add a ServeConfig wrapper = flattened HiveEnv
plus the hive-c0re-local model_prices table (kept out of HiveEnv since it
is never injected into containers). serde(default) lets any field be
omitted and fall back to its canonical default.
- clap: add --config; the per-setting flags become optional overrides
(config file is the base, explicit flags win — preserves hivectl/debug
ergonomics and bare `hive-c0re serve`).
- Coordinator::open and cmd_serve now take the bundled HiveEnv, which drops
their too_many_arguments clippy allows. cmd_serve keeps a single
too_many_lines allow (inherent daemon-boot orchestration, not arg-driven).
- nix: write the config as JSON to the store + pass --config, so ExecStart
is one short line.
- Add a round-trip test proving the flatten + per-field defaults work.
Closes the ExecStart-length issue.
Per operator request on the PR: the built-in/default prices were the old
Claude 3 numbers (opus 15/75, etc.). Update opus + haiku to the current
Claude 4.x family list pricing (cache_write = the default 5-minute cache
TTL); sonnet was already correct:
- opus: input 5, output 25, cache_read 0.5, cache_write 6.25
- sonnet: input 3, output 15, cache_read 0.3, cache_write 3.75 (unchanged)
- haiku: input 1, output 5, cache_read 0.1, cache_write 1.25
Updated in both builtin_prices (hive_stats.rs) and the nix modelPrices
default (hive-c0re.nix), with cross-reference "keep in sync" notes on
both sides. Also addresses the earlier reviewer note: dropped the
over-strong "single source of truth" wording in the --model-prices arg
doc (the nix default does mirror the numbers in production).
- modelPrices submodule fields use lib.types.numbers.nonnegative
instead of lib.types.float: accepts bare ints (15) as well as floats
(15.0) and rejects negative prices for free.
- Collapse the triple-sourced default: hive-c0re serve --model-prices
now defaults to "{}" so builtin_prices() is the single in-code
fallback. The nix option default still carries the full
opus/sonnet/haiku table to self-document prices for operators.
The hive-wide cost estimate on the dashboard's ST4TS tab used a
hard-coded model->price table in hive_stats.rs. Anthropic list pricing
drifts, so move the table to a nix option operators can keep current
without a code change.
- New `services.hyperhive.modelPrices` option: attrset of model-family
short name -> { input, output, cache_read, cache_write } USD per
million tokens. Passed to `hive-c0re serve --model-prices <json>`.
- hive_stats: `Prices` is now public + Deserialize; add `PriceTable`
type and `resolve_prices` (longest case-insensitive substring key
wins) with the old hard-coded table preserved as `builtin_prices`
fallback for any model not covered.
- Coordinator holds the parsed table (hive-c0re-local, not injected
into containers, so not part of HiveEnv); `/api/stats-hive` reads it.
- Docs: dashboard.md ST4TS cost note updated; option self-documents
via nixosOptionsDoc.
Closes#1434
