Approvals + helper events

The approval queue is hyperhive's pivot: nothing that changes the shape of an agent (its config, whether it exists) happens without an operator click. The submitting agent — any agent with the approvals tool group, which manages the config of its direct children (the root agent for top-level agents; a sub-manager for its own subtree) — is the policy gate in front of that queue; helper events are how it stays informed about what happens after a decision lands.

End-to-end approval flow

Config changes flow through a forge pull request on the agent's agent-configs/<name> repo — the same surface agents use for code PRs. There is no bespoke MCP tool for config changes: opening the PR IS the request.

  1. The submitting agent (the child's parent, holding the approvals tool group) edits files in the child's proposed config repo (any tracked path, but agent.nix is the contract entry point), commits with its own git identity, and pushes a branch + opens a PR on agent-configs/<name> with hive-forge. The parent's container has the child's proposed config repo bind-mounted read-write at /agents/<name>/config/ (topology-driven via set_nspawn_flags; the agent's own config at /agents/<self>/config/ is read-only). Branch protection (push/merge whitelist = core, approvals whitelist = operator team; see "Forge mirror" and #1787) makes the agent a write collaborator that cannot merge its own config PR.

  2. hive-c0re's /webhook/config-pr endpoint receives the Forgejo pull_request event (opened / synchronized / reopened) and queues a MergeConfigPr approval; a poll fallback catches any missed webhook. The approval row stores the PR number (commit_ref) and the PR head sha at queue time (fetched_sha — the "reviewed" sha). If the PR head later moves, the stale approval is superseded by a fresh one pinned to the new head, so the operator always reviews what will actually deploy.

  3. The operator reviews the PR on the forge (native diff, threaded comments, CI status) and sees a matching card on the dashboard with a "review PR on forge" deep link. They click ◆ APPR0VE (or hivectl approvals approve <id> on the CLI) once satisfied.

  4. On approve, a deploy DAG runs three phases under a resource-holding DeployWindow root (see Queue templates below):

    • MergeVerify re-reads the live PR head and aborts if it drifted from the reviewed fetched_sha (the submitter must push again, which queues a fresh approval); then fetches that head into the applied repo and eval-verifies it — a flake eval on a throwaway checkout. This is the trust gate: it relies on c0re's own eval, not on any in-repo (agent-forgeable) signal like a CI status. Nothing is mutated in this phase, so a rejection here leaves the forge and the applied repo exactly as they were.
    • DeployApply parks the pre-merge applied/main in refs/hyperhive/rollback/<approval-id>, then fast-forward-merges the reviewed head to the forge config repo's main (this IS the merge — a core-authenticated ff-merge pinned to the reviewed sha, so a moved PR head can't substitute bytes), and runs the deploy proper (deploy_applied_target): ff applied/main, two-phase meta deploy, container rebuild. On success it drops the rollback ref and plants deployed/<id>.
    • DeployTail runs on every outcome, including a cancel-cascade. If the rollback ref survived, the deploy never confirmed good: it rolls applied/main back, resyncs the working tree, and aborts the staged meta lock, so the agent stays on its last-good tree. Then it mirrors the config repo (and its new deploy tag) to the forge.

    The rollback state lives in a git ref, not a local variable, on purpose: hive-c0re can restart between the apply and the tail, and the tail still has to know what to undo when it does.

  5. HelperEvent::ApprovalResolved (and Rebuilt) land in the submitting agent's inbox via notify_submitter, carrying both the canonical sha and the terminal tag (the approval row carries a submitter column recording the agent the change is for).

Withdrawing a pending approval

The submitting agent can call cancel_loose_end(kind: "approval", id) to withdraw an approval that hasn't been acted on yet. The row transitions to ApprovalStatus::Cancelled (distinct from Denied/Failed), the dashboard pulls the card out of the pending pane, and ApprovalResolved { status: "cancelled" } fires on the root agent + dashboard channels. Approvals that have already been approved/denied/failed return an error — the resolution is final once the operator (or a lifecycle failure) acted on the row.

The socket refuses the approval kind with a clear error for any agent that lacks the approvals tool group: only an agent with that group submits approvals (for its direct children), so an agent without it has nothing of its own to withdraw.

InitConfig approvals seed a brand-new agent's proposed config repo. On approve, hive-c0re seeds it with a default agent.nix template and sends HelperEvent::ConfigReady { agent } to the submitting agent's inbox via notify_submitter. The submitting agent reviews, edits, and commits the template; the operator then spawns the agent (the Spawn approval / ◆ R3QU3ST SP4WN button), which creates the container from that config. From then on, config changes go through a forge PR (MergeConfigPr) — never a fresh spawn. This gives the submitting agent (and operator) an explicit review gate on the initial configuration before any container is created.

Approval kinds (wire shapes)

ApprovalKind carries five variants; each maps to a different commit_ref encoding because that field is overloaded as the kind-specific payload carrier.

Scheduled prompts (submit paths)

Two ways a row lands in scheduled_prompts:

No self-target shortcut: even agent-self schedules need approval. The existing remind MCP tool stays the quick self-wake path (no approval, lands directly in the agent's own inbox); this module is the bigger, multi-recipient, operator-visible thing.

Scheduled prompt worker (catch-up clamp)

When hive-c0re comes back from being down, the worker sees rows whose next_fire_at_unix is well in the past. For recurring rows that would mean firing N delayed pulses in a row — spammy and useless. Instead the worker fires once per row and bumps next_fire_at_unix to the next interval slot ≥ now, recording how many cycles were skipped in last_result (per-target). Operators see "fired late, caught up from 17 skipped" instead of 17 wake-up storms.

One-shot rows fire once (if past due, on the next worker pass) and are deleted by the worker; recurring rows survive until cancelled.

targets is its own table (scheduled_prompt_targets) so partial cancellation flips a single row and the dashboard can show last-fired / last-result per recipient. Cancelling every target reaps the parent row on the next worker pass.

Missing-target failure

When a target name doesn't resolve to a known agent (container destroyed, operator typo, etc.) the worker:

  1. Records last_result = "no such agent: <name>" on the per-target row.
  2. Sends a single advisory Message from system to operator naming the schedule, target, and reason.
  3. Continues fanning out to the other live targets.

Transient broker errors (sqlite lock contention, etc.) get the same last_result annotation plus a tracing::warn, and then:

Reminder delivery: file-path semantics

A reminder may carry a file_path (the agent-visible path inside its container, e.g. /agents/<name>/state/foo.md). On delivery hive-c0re:

  1. Translates the container path to the host path (/var/lib/hyperhive/agents/<name>/state/foo.md) so c0re can write from outside the container.
  2. Validates the path: rejects anything outside the agent's own state subtree, containing .. (path traversal), or with an empty relative tail. On rejection the write is skipped and the original message is delivered inline with a warning — the reminder still fires.
  3. Defends against symlink escape: after create_dir_all, the parent dir is canonicalized and re-verified to live under the agent's host state root. The final file is opened with O_NOFOLLOW | O_CREAT | O_TRUNC so an existing symlink at the basename cannot redirect the write to an arbitrary host path.
  4. Writes the body to disk and delivers a short pointer message in its place, keeping the agent's inbox / wake-prompt small while the bulky payload is read out of band.

Atomicity of the inbox INSERT + reminders.sent_at UPDATE is handled inside Broker::deliver_reminders_batch; the scheduler only computes the body strings before calling it.

Destroy semantics

HostRequest::Destroy { name, purge } is the lifecycle tear-down, not an approval. Stops + removes the nspawn container, drops the systemd drop-in, fails any pending approvals. Persistent state (proposed/applied repos, claude credentials, /state/ notes) is kept by default — recreating the agent with the same name reuses prior config + login. With purge = true the agent's /var/lib/hyperhive/{agents,applied}/<name>/ trees are also wiped (config history + creds + notes gone forever). The root/bootstrap container is destroyable like any other — hive-c0re recreates it on the next startup if it's absent, so destroying it is transient.

Meta flake

The hive-c0re-owned repo at /var/lib/hyperhive/meta/ declares one flake input per agent (agent-<n>.url = "git+http://<forge>/agent-configs/<n>.git") and one nixosConfigurations.<n> output per agent. Each output wraps inputs.agent-<n>.nixosModules.default with the identity + HIVE_PORT / HIVE_LABEL / HIVE_DASHBOARD_PORT injection module that setup_applied used to generate inline. Containers run against --flake /var/lib/hyperhive/meta#<n>.

The declared input url is the agent's forge config repo (the same agent-configs/<n> the config-PR flow lands approved changes on), so the meta flake references a reviewable, reproducible source rather than a local checkout. hive-c0re authenticates that git+http fetch via a git credential helper that reads the live forge-core token — no token in the url or the lock. The deploy and manual-rebuild paths, however, do not re-lock from the forge: they --override-input agent-<n> git+file:///var/lib/hyperhive/applied/<n>, locking the exact config that verify_commit gated and applied/<n>/main was fast-forwarded to. That keeps a deploy/rebuild reproducible and independent of forge reachability — rebuilds fire on crash-restart and meta bumps, not just config PRs — while the declared url stays the forge. sync_agents re-renders + re-locks the persistent input; a plain nix flake lock leaves an existing applied override in place (it only re-locks when the declared url itself changes), so the forge-declared / applied-deployed split is stable.

Per-deploy lock flow (two-phase), spread across the deploy subtree's nodes — each phase is its own node, so the queue can show which one is running and a restart resumes at node granularity:

  1. DeployApply → meta::prepare_deploy(name) runs nix flake lock --update-input agent-<n> without committing. Working tree of meta now points the input at applied/<n>/main (which the deploy already fast-forwarded to the reviewed PR head).
  2. The rebuild subgraph DeployApply grows into the DAG builds and swaps the container (Prebuild → StopForUpdate → Swap → PostSwap, plus Reconcile). Nix evaluates against the staged lock.
  3. On success — FinalizeDeploy drops the rollback ref, plants deployed/<id>, then meta::finalize_deploy(name, sha, "deployed/ <id>") stages flake.lock and commits with deploy <n> deployed/<id> <sha12>. Meta's git log gains one entry per successful deploy.
  4. On failure — the DeployTail node runs meta::abort_deploy() (git restore flake.lock) so the meta history shows only successes; the failure stays as an annotated failed/<id> tag in applied/<n>. The tail runs on every outcome, so this also covers a hive-c0re restart mid-build: the staged lock is dropped and applied/main rolled back from the parked refs/hyperhive/rollback/<id>.

Single-phase variants exist for paths without rollback semantics: meta::lock_update_for_rebuild(name) for the manual ↻ R3BU1LD button (commits if the lock changed) and meta::lock_update_hyperhive() for the auto-update flake-rev bump (one shot before per-agent rebuilds, commits if the lock changed).

meta::sync_agents(hyperhive_flake, dashboard_port, &agents) is the idempotent reconciler called by spawn, destroy, rebuild, and the startup migration. Renders flake.nix from the agent list; if it differs from disk, runs nix flake lock + commits as regenerate meta flake (or seed meta from N agent(s) on the very first call).

The root agent has /meta RO-bound inside its container: git -C /meta log --oneline is the swarm-wide deploy log, cat /meta/flake.lock | jq '.nodes["agent-<n>"].locked' resolves which sha each agent is pinned at right now. Dashboard surfaces the same info as a deployed:<sha12> chip per container row.

Two repos per agent

/var/lib/hyperhive/agents/<name>/config/    proposed — submitting agent RW
└── <anything>                              # any files the submitting
                                            # agent wants in the commit.
                                            # agent.nix is the
                                            # convention entry
                                            # point; flake.nix is
                                            # tracked boilerplate
                                            # (submitting agent doesn't
                                            # edit it).

/var/lib/hyperhive/applied/<name>/          applied — core-only
├── .git/                                   # tag-rich history
├── flake.nix                               # tracked, fixed
│                                           # boilerplate exporting
│                                           # nixosModules.default
├── agent.nix                               # working tree of main
└── <other committed files>                 # also tracked

/var/lib/hyperhive/meta/                    swarm-wide flake — core
├── .git/                                   # one commit per successful
│                                           # deploy
├── flake.nix                               # generated from agent set
└── flake.lock                              # pins each agent's sha

Why two physical repos: the submitting agent's /agents/<n>/config/ is RW — a buggy or hostile agent can git clean -fdx its own proposed tree. The applied repo is never bind-mounted (except the read-only .git exposure described below) so a destructive move inside the container cannot reach it.

The container's --flake ref is /var/lib/hyperhive/meta#<name> (see "Meta flake" above). The agent's own applied/<n>/flake.nix is a fixed boilerplate that exports nixosModules.default = import ./agent.nix; the meta flake imports that module and wraps it with identity + HIVE_PORT / HIVE_LABEL / HIVE_DASHBOARD_PORT.

Tag state machine

Each deploy leaves a tag on the underlying commit inside the applied repo:

Tag When Annotated?
deployed/<id> rebuild succeeded — main ff's here no
failed/<id> rebuild failed yes (body = error)

deployed/0 is planted at first spawn. applied/main is always the latest deployed/*. A failed/ tree stays browsable forever — git log --tags in the applied repo is the audit trail. A denied or failed config PR carries no extra state on the forge side: the PR stays open, and the submitter pushes again (or closes it) to retry.

Dispatch via the job queue

Long-running approval work — MergeConfigPr, UpdateMetaInputs, Spawn — no longer runs inline inside actions::approve. Instead the approval handler submits a DAG to the global job queue (docs/coordinator.md::Job queue):

ApprovalKind DAG submitted source
MergeConfigPr rebuild (DeployWindow root + MergeVerify → DeployApply + DeployTail) approval
UpdateMetaInputs meta_update (MetaLock + rebuild fan-out) approval
Spawn spawn (Create → WriteDropin → Reconcile) approval
InitConfig — runs inline (sub-second git seed) —
SchedulePrompt — runs inline (single sqlite insert) —

The DAG carries the originating approval_id, surfaced on the node that owns it — for a deploy that's the DeployWindow root, so the dashboard renders one approval card, not four. Every queued kind resolves through actions::resolve_approval_dag when its DAG settles terminal: the deploy's phases are ordinary queue nodes, so the DAG's own terminal state is the authoritative outcome. That hook fires the matching HelperEvent::* via finish_approval, derives the Rebuilt event's terminal tag (verifying the tag actually resolves in the applied repo — a pre-merge rejection plants none), posts the failing build log back to the config PR, and for a spawn runs the post-spawn forge bookkeeping.

Two visible consequences:

The approval source + approval_id mean a tail-end build failure surfaces back as a failed approval row, not just a silent queue entry. manual (dashboard ↻ R3BU1LD) and auto_update (boot reconcile) DAGs use the same queue but skip the approval plumbing.

Forge mirror

The bundled hive-forge container is mandatory (it deploys with hyperhive), and hive-c0re mirrors every agent's applied repo into a private agent-configs Forgejo org. forge::push_config(<name>) pushes applied/main plus every tag to agent-configs/<name> after each ref mutation: the spawn that seeds deployed/0, every successful deploy (which plants deployed/<id>) or failed build (failed/<id>), and a sweep at startup. Pushes are best-effort — a missing or stopped forge never blocks a deploy.

Each agent is a write collaborator on its own agent-configs/<name> repo — so it can push a branch and open a config PR — but not a member of any other agent's, so it can't reach another agent's config through the forge. Branch protection keeps main push/merge core-only with operator-team approval, so an agent can't fast-forward its own config or self-merge its PR (see the End-to-end flow + #1787). The tokenised push URL is passed inline to git push, never written into applied/<n>/.git/config; that repo is RO-bind-mounted into the root agent, and a stored token would leak core's admin credential to an agent.

The dashboard deep-links into this org — a config repo link per container row and a review PR on forge link per config-PR approval card. See docs/web-ui.md.

Submitting agent's view of config repos

Every parent agent's container has its direct children's proposed config repos bind-mounted read-write (topology-driven: lifecycle.rs calls bind_child_agent_dirs for each entry in topology::children_of(agent_name)). An agent with the approvals tool group can therefore edit, commit, and submit changes for any of its direct children directly inside its container at /agents/<child>/config/.

Agents holding the can_manage_top_level_agents topology role (defined as ROLE_CAN_MANAGE_TOP_LEVEL_AGENTS in hive-c0re/src/agent_config/topology.rs) get additional host-side bind mounts via set_nspawn_flags:

The root agent holds this role; a sub-manager that only manages a subtree does not, and only has its direct children's config dirs.

Each proposed repo (/agents/<n>/config/) is pre-configured with applied as a git remote pointing at /applied/<n>/.git. Useful incantations from inside an agent with the full /applied mount:

git -C /agents/<n>/config fetch applied
git -C /agents/<n>/config log applied/main --oneline
git -C /agents/<n>/config show applied/refs/tags/deployed/<id>
git -C /agents/<n>/config show applied/refs/tags/failed/<id>   # body = build error
git -C /agents/<n>/config show applied/refs/tags/denied/<id>   # body = operator note
git -C /agents/<n>/config rebase applied/main                 # base in-flight work on what's deployed

git -C /meta log --oneline                                    # swarm-wide deploy history
cat /meta/flake.lock | jq '.nodes | with_entries(select(.key | startswith("agent-")))'

The RO binds block push at the kernel level — git plumbing inside the container cannot corrupt either authoritative repo.

Migration from the pre-tag / pre-meta schemes

Both overhauls (tag-driven flow + meta flake) ship in-place migrations that run on every hive-c0re startup. Idempotent; each phase is a no-op once already applied. Behaviour:

No state loss in either migration. claude creds, /state/ notes, the events DB, proposed history, and applied history all survive. The root agent keeps its session; sub-agents stay logged in.

The root/bootstrap container is hive-c0re-managed

The root agent container runs through the same lifecycle as sub-agents. On hive-c0re serve startup, if ruth is missing, hive-c0re creates it. The root agent's flake lives at /var/lib/hyperhive/applied/ruth/; its proposed config at /var/lib/hyperhive/agents/ruth/config/. The root agent can edit its own agent.nix (visible inside the container at /agents/ruth/config/) and open a config PR on agent-configs/ruth for operator approval, same as any other agent.

Differences from sub-agents:

Migration note (for older hosts): drop any containers.root = { ... } block from your host NixOS config. hyperhive creates and updates the root agent itself.

Root-agent policy

The system prompt (hive-ag3nt/prompts/system.md, rendered via hive_ag3nt::prompt::render) is the same for every agent; what varies is which MCP tools are surfaced (gated by tool groups and capabilities in agent.nix). There is no role:manager block that renders only for the root agent. The root agent's approval-gating behaviour comes from its CLAUDE.md / agent-specific instructions, not the system prompt template.

ask(question, options?, multi?, ttl_seconds?, to?) is available to any agent — it queues a question and returns the id immediately. When to is omitted (or "operator") the question shows up on the dashboard; when to is another agent's name, the recipient receives a HelperEvent::QuestionAsked and answers via their own answer tool. Either way the answer arrives back as HelperEvent::QuestionAnswered { id, question, answer, answerer } in the asker's inbox. Storage is hive-c0re::operator_questions (sqlite) — same table, with a nullable target column (NULL = operator). Dispatch goes through hive-c0re/src/questions.rs::{handle_ask, handle_answer}. The answer flow is:

POST /answer-question/{id}                       agent: Answer { id, answer }
  → OperatorQuestions::answer(_, _, "operator")    → questions::handle_answer
  → notify_agent(asker, QuestionAnswered {         → OperatorQuestions::answer(_, _, agent)
       answerer: "operator", ... })                → notify_agent(asker, QuestionAnswered {
                                                       answerer: agent, ... })

Two more paths resolve a pending question with a sentinel answer:

Helper events to the submitting agent

Coordinator::notify_submitter(approval_id, &HelperEvent) routes the event to the agent that originally submitted the approval (looked up from the submitter column on the approvals table). The harness delivers it as a regular system inbox message so it drives a normal claude turn. Legacy approval rows that predate the submitter column fall back to the root agent. Variants (hive_sh4re::HelperEvent):

Optional sha field on ApprovalResolved and Rebuilt carries the canonical hive-c0re-vouched commit sha. Optional tag on the same two carries the deploy bookkeeping tag — deployed/<id> on a successful build or failed/<id> on a failed one, planted by the MergeConfigPr deploy. Both fields are Option: None on the paths that don't deploy a new commit (spawn / init_config / meta-update / deny, and auto_update::rebuild_agent reapplying the existing main, or the dashboard ↻ R3BU1LD button when the lock didn't move). When set, git show <sha> against /agents/<n>/applied.git inside the bootstrap container yields the exact tree that was referenced.

To add a new event: new HelperEvent variant + call sites + update prompts/system.md (<!-- role:manager --> block, the lifecycle- event list) so the root agent knows the new shape.

Auto-update on startup

hive-c0re serve runs auto_update::run in a background task right after opening the coordinator. It enumerates managed containers and rebuilds any whose recorded hyperhive rev differs from the current one — sub-agents and the root agent go through the same lifecycle::rebuild path.

"Rev" = canonical filesystem path of cfg.hyperhiveFlake. Marker file: /var/lib/hyperhive/applied/.<name>.hyperhive-rev. If the flake input has no canonical path (e.g. a github: URL), auto-update is a no-op — rebuild manually.

The dashboard surfaces pending updates per agent: a clickable "needs update ↻" badge appears whenever the marker differs from current rev. The badge POSTs /api/rebuild/<name>, calling the same auto_update::rebuild_agent path so manual triggers and the startup scan can't drift. When at least one container is stale, a top-level ↻ UPD4TE 4LL button appears that loops over every stale container.