grid_runtime 0.1.2

Runtime providers for the_grid — spawns and supervises a coding agent per ready bead, isolating each bead's work in a git worktree.

grid_runtime

M3 runtime providers — the layer that gives the_grid hands.

grid_runtime spawns and supervises a coding agent (a claude subprocess) per ready bead, isolates each bead's work in a git worktree, tracks each session's lifecycle as a bead (bd-only writes through the single write chokepoint, --actor grid-controller, never SQL), and lands finished work as a pushed branch / PR — never an auto-merge. It consumes beads_dart's ready-set seam. grid_engine consumes it from above as the engine's subprocess/git/chokepoint transport.

Status — Tracks 2–5 built

The runtime seam + the subprocess provider are implemented (offline, fakes-not-mocks; dart analyze clean, the offline suite green). The rest of the surface landed across M3; what has since moved is noted per track:

• Track 2 (built) — RuntimeProvider interface + RuntimeConfig / RuntimeEvent / RuntimeCapabilities value types + SubprocessProvider (the Friday dogfood default): Futures for acts (start/stop/interrupt), Streams for observations (sealed RuntimeEvent + live output), plus the point-in-time queries (isRunning/processAlive/peek/listRunning/ lastActivity). SubprocessProvider spawns the agent no-shell, with includeParentEnvironment:false and an explicit env allowlist (forwards the agent's CLAUDE_CODE_OAUTH_TOKEN, drops GC_DOLT_PASSWORD and every other host secret), in a new process group (ProcessStartMode.detachedWithStdio) so stop() can SIGTERM→2s grace→SIGKILL the whole tree with the pgid<=1/self-group guard. Per-incarnation GRID_SESSION_ID/GRID_BEAD_ID/ GRID_INSTANCE_TOKEN/GRID_RUNTIME_EPOCH are injected. CUT (Track 2): the inference-provider abstraction, a per-session Unix control socket, attach/nudge. A TmuxProvider was never built — SubprocessProvider is the only RuntimeProvider (detachedWithStdio spawn, liveness-poll death detection, the watchdog deadline).

  Spawn-mode choice (detachedWithStdio), justified. Dart exposes no Setpgid; setsid is absent on macOS and a sh -c wrapper does not start a new group. detachedWithStdio is the one mechanism that both setsid()s the child into a fresh session+group AND keeps stdio connected for the transcript. Its cost — Process.exitCode is unavailable for detached processes — is paid by polling liveness via the ProcessGroupController seam (the honest signal: a backgrounded grandchild can hold stdout open past the agent's exit). A real detached exit therefore surfaces as RuntimeEvent.died; a precise RuntimeEvent.exited with a code is emitted whenever the spawner can read one (the fake spawner / any non-detached path).

• Track 3 (built) — StationGitService: git-worktree-per-bead isolation + the three-gate fail-closed reaper + land-to-PR. All git I/O runs through the injectable GitRunner seam (real impl SystemGitRunner; tests use the real git binary against temp repos, or a scripted fake for the probe-error paths). Two layers, copied from gc's rig model:

  • Layer 1 — root-checkout registration (registerRootCheckout): records a the_grid-OWNED clone path + the default branch probed from origin/HEAD (a verbatim port of gc's ProbeDefaultBranch, internal/git/git.go:92-106; never hardcodes main). It does not create the real lenny-tgdog clone — that is Nico's one-time out-of-band step (auto-provision is the gascity#1556 stale-ancestor hazard).
  • Layer 2 — per-bead worktree (provisionWorktree): git worktree add -b grid/<beadId> <root>/.grid/worktrees/<substation>/<beadId> <base> (mirrors gc's .gc/worktrees/<rig>/<name>, internal/workdir/workdir.go:76-86). The dir name encodes the bead id so listBeadWorktrees re-binds an orphaned worktree to its lifecycle bead on restart with no external state.

  Ported verbatim — the load-bearing safety: the three-gate pre-removal check (reap refuses if HasUncommittedWork OR HasUnpushedCommits OR HasStashes, all fail-closed on probe error — a GateOutcome.probeError blocks exactly like present; git.go:134-213), with git worktree remove run from the root repo, never inside the worktree; the GIT_* env blacklist stripped on every exec by the runner seam (git.go:285-301); the stale-ancestor guard (validateAncestorWorktreesNotStale, workdir.go:303-359) before every git worktree add; and the scope gate (isStrictlyUnderDir) so the reaper can only ever act inside the worktrees root. Land DIVERGES from gc (no prior art): land commits on grid/<beadId> → git push -u origin grid/<beadId> → opens a PR via the injectable PrOpener seam → returns the PullRequestRef for the caller to record on the lifecycle bead. Never auto-merges. CUT (Track 3): per-bead retry-on-new-branch; an offline periodic reaper sweep; registry- removal-deletes-disk (registry removal ≠ disk deletion, mirror gc).

• Track 4 (built) — lifecycle-as-beads + the single StationBeadWriter bd write chokepoint. Three pieces:

  • session_state.dart — a Dart port of gc's session state transition table (internal/session/state_machine.go:106-144): start_pending → spawning → active → {asleep, draining, quarantined, closed}, with close legal from any non-none state (gc's anyState sentinel) and restart from asleep/quarantined/draining. A pure, total reducer (transition / transitionOrNull / allowedCommands), tested before any IO. LifecycleState is an extension type over the wire string (like IssueType) so a gc-written state the_grid does not model is preserved verbatim, never dropped.
  • BeadOwnershipPredicate — the bead-shaped ownership gate the chokepoint and (Track 5) the dispatcher share. M2's OwnsSubstations.owns(Convergence) reads convergence.metadata.rig, a key gc stamps only into convergence beads, so it is structurally uncallable on a plain Bead; this predicate derives a bead's rig from the issue-id prefix (primary) and/or metadata.rig. The shared artifact with the M2 actuator is the rig allow-set Set<String>, not the predicate object (seeded from the station's owned set) so the two gates cannot drift. A no-rig/no-prefix bead is not owned, fail-closed.
  • StationBeadWriter — the single bd write chokepoint wrapping the M2 BdCliService. Before every create / update --metadata / close / delete it re-checks ownership fail-closed on the target rig and refuses + logs loudly ([OwnershipRefused]) any write whose rig is absent / not owned — the second line of defense behind the dispatch predicate (session/recovery writes never flow through a Convergence, so a convergence-side ownership gate cannot cover them). createSession mints + stamps metadata.rig from birth (the M2 BdCliService.create carries no --metadata, so the mint is create + a stamping merge update). --actor grid-controller, merge semantics (works on closed beads), batch only for grouped close+dep, never SQL, never bd show on this path.
  • RuntimeActuator — consumes Track-2 RuntimeEvents and writes session beads through the chokepoint: a spawn mints the bead at start_pending; SessionStarted/ActivityChanged drive it to active; a clean Exited(0) parks it asleep; a crash (Died or non-zero Exited) trips the crash-loop machinery — under the threshold it sets restart_requested (a fresh restart, bead left open, gc manager.go:867-879), at/over the threshold it quarantines (state=quarantined, quarantine_cycle, quarantined_until, gc lifecycle_transition.go:468-497). Crash bookkeeping is in-memory (the live supervisor); the durable counters it writes mirror it. CUT (Track 4): the convergence RecoveryAction actuator (now Track 4b, off the Friday path).

• Track 5 (built; the interactor since deleted) — DispatchInteractor (ready bead → spawn) + the ReadyWorkSource read seam. The interactor was deleted with grid_reconciler (RS-8) — dispatch is now grid_engine's tree reconcile (mount = spawn); ReadyWorkSource + GridReadyWorkSource remain this package's read seam. As built, it attached as a SECOND consumer of the same observable surface M2 uses (a ReadyWorkSource over beads_dart's GraphEvent stream + readyBeads, mirroring the M2 ConvergenceSource) and does not go through reduce→gate→actuate. Two halves:

  • Dispatch. On GraphEvent.readySetChanged.entered (and a start-time reconcile of the current readyBeads), each entered id is resolved to its Bead and gated on the Track-4 BeadOwnershipPredicate (the shared allow-set). A non-owned bead is observed read-only, NEVER dispatched and NEVER mutated (OwnsSubstations is structurally uncallable on a plain Bead). On accept the pipeline runs StationGitService.provisionWorktree (Track 3) → RuntimeProvider.start (Track 2) → RuntimeActuator.spawnSession (Track 4, the session bead minted through the StationBeadWriter chokepoint). Idempotent + single-flight per bead: a PerBeadQueue (reused from the M2 runtime) serializes per work-bead id and a synchronous slot reservation is taken before the first await, so a re-fired ready event — even a burst racing the first spawn — never double-spawns and concurrent different-bead dispatches cannot over-commit the max-in-flight cap.
  • Supervision. It listens to the actuator's CrashDecision stream: a RestartSession re-spawns the same bead in its existing worktree (the session bead is left open); a QuarantineSession parks it (no respawn, no reap); a SessionParked (clean exit) drives the removal trigger — close the session bead through the chokepoint, then StationGitService.reap, which only removes the worktree once the three fail-closed gates pass (an unpushed/uncommitted/stashed worktree is refused and kept for the land step).
  • --dry-run is observe-only: no worktree, no spawn, no bd write — the safe default for the first live run (the live arm is Track 7 + its live gate). CUT (Track 5): the demand-spawned pool / backpressure beyond the max-in-flight cap (the full pool is M4).

• Track 6 (shipped) — the exploration-attach (plugins→extensions) wire-key fix (grid_exploration's kExtensionsKey is extensions).

• Track 7 (built; since deleted) — the grid run CLI + dogfood wiring (composed the Track-5 dispatcher with the M2 reconciler runtime + the chosen provider). Deleted in RS-8 with the interactor — composition now lives in grid_sdk's runGrid.

All Track-4 bd calls go through the StationBeadWriter chokepoint over the M2 BdCliService; the offline suite drives them with a recording fake bd runner (records argv + stdin), asserting the exact --actor grid-controller + metadata-merge commands, no SQL, no bd show, and the fail-closed refusal of a wrong/absent rig. No live state is touched.

Safety

• Single writer per bead. the_grid spawns/supervises/actuates only the disjoint, prefixed rig set it owns; dispatch and the bd write chokepoint share one rig allow-set so they cannot drift. Non-owned ready beads are observed read-only, never spawned, never mutated.
• bd-only writes. All session/lifecycle/recovery mutations flow through the one StationBeadWriter over the M2 BdCliService (--actor grid-controller). Never raw SQL; never bd show on a controller/re-query path; never touch .beads/hooks/ (gc owns them).
• Auth: flag-not-extract, explicit allowlist. The agent OAuth token is an inherited env var on an allowlist — never on argv, never read/printed.
• Tests use fakes, not mocks; pure logic is tested before IO is wired. Process/worktree tests use temp git repos and stub commands, never live repos or the real agent.