diff --git a/docs/adr/0013-task-state-machine-ownership.md b/docs/adr/0013-task-state-machine-ownership.md new file mode 100644 index 0000000..eefbd99 --- /dev/null +++ b/docs/adr/0013-task-state-machine-ownership.md @@ -0,0 +1,254 @@ +# ADR 0013 — Task state-machine ownership: CORE, DAEMON, and the shared `paused` + +**Status:** accepted by CORE and PROTO; all four open items resolved · **Date:** 2026-09-09 +· **Lane:** DAEMON, drafted at PROTO's request (D1 in `core/docs/proto-requests-m1.md`). +**CORE's response:** `core/docs/adr-0013-core-response.md` (`lane/core`, commit `c65e664`) +— accepted as written, no amendments, plus the pinned `tasks_starved` definition and +`pause()`/`resume()` idempotency contract folded in below. +**PROTO's response:** the `error`-on-`paused` widening landed as `contracts/` **1.3.0**, +`lane/proto` commit `6db304a` — no retype, no new field; only the presence condition on +the existing `error: TaskError | null` field widens to include a daemon-initiated pause. +Not yet merged to `main` (main is at 1.1.0 as of this writing) — `daemon/src/sched/`'s +pause/resume logic should be written against `1.3.0` once that merge lands, not before. + +## Context + +`docs/04` §1 hands CORE the download lifecycle (`probing → connecting → downloading ⇄ +paused → retry_wait → assembling → verifying → complete | failed`), but `queued` and every +scheduler transition are DAEMON's, and neither brief says where the seam is. PROTO raised +this as D1 and proposed a split; CORE agreed (`contracts/proto-answers-m1.md` D1); both +declined to write it down alone — a two-lane decision recorded by the lane that owns +neither half is exactly the failure mode `contracts/README.md` rule 4 exists to prevent. +DAEMON drafts it because DAEMON is the one lane touching both halves (it owns the SQL row +and the RPC surface; it calls into CORE). + +The wire side is already frozen and does not need to change: `TaskState` +(`contracts/schema/types/TaskState.schema.json`) has the twelve values with no field +naming who drove a transition, and `event.task.state` carries `previousState` — a client +renders what it's told and keeps no state machine of its own. This ADR is about who +*decides* a transition, not the wire shape of one. + +This is also where ADR 0011 (admission control) and the state machine meet: `queued → +connecting` is the admission event, and `retry_wait` looks identical to starvation from +inside the segment-budget accounting. Both are addressed below because getting them wrong +independently produces the same false alarm from two different directions. + +## Decision + +### 1. Ownership, by state + +| State | Driven by | Entered from | Notes | +|---|---|---|---| +| `new` | DAEMON | (creation) | `download.add` persists the row before CORE knows the task exists. | +| `queued` | DAEMON | `new`, `paused` (resume, subject to §3) | Scheduler admission pool. CORE has no concept of "queued" — `retry_wait` re-enters `connecting` directly inside CORE and never passes back through `queued`. | +| `probing` | CORE | `queued` | DAEMON calls CORE's `start()` once admitted (ADR 0011 §1); CORE owns everything from here until it hands back a terminal or `paused` state. | +| `connecting` | CORE | `probing`, `retry_wait` | | +| `downloading` | CORE | `connecting` | | +| `paused` | **shared** | any CORE state, `queued` | See §2 — this is the one state either side may enter unilaterally. | +| `retry_wait` | CORE | `connecting`, `downloading` | CORE's own backoff timer (`docs/04` §7); DAEMON does not schedule retries. | +| `assembling` | CORE | `downloading` | Normally a no-op rename; a real state for HLS/DASH muxing. | +| `verifying` | CORE | `assembling` | Checksum verification when requested. | +| `complete` | CORE | `verifying` | Terminal. | +| `failed` | CORE | any CORE state | Terminal. `max_retries_exhausted` after `retry_wait`, or any non-retryable error. | +| `cancelled` | DAEMON | any state | Terminal. Always user- or policy-initiated (`download.cancel`, category/queue removal) — CORE never cancels on its own; it only executes the teardown DAEMON asked for and confirms. | + +Reading the table as a picture: + +``` +DAEMON: new ──▶ queued ──▶[admit]──▶ (hand to CORE) + ▲ │ + │ resume CORE: probing ─▶ connecting ─▶ downloading + pause────┤ │ │ ▲ │ │ + (either side)│ retry_wait◀────────┘ │ │ │ + ▼ │ (backoff done)│ ▼ ▼ + paused ◀──────────(auto-pause)─────┘ assembling paused + │ (auto) + verifying + │ + complete + (from anywhere, DAEMON-driven) ──────────────────────▶ cancelled + (from any CORE state, CORE-driven, non-retryable) ───▶ failed +``` + +### 2. `paused` is shared, and idempotency is the whole contract + +Both sides can put a task in `paused`, for disjoint reasons: + +- **DAEMON-initiated**: user clicks pause, a schedule window closes, a queue is stopped, + `Queue.onComplete`, or the admission governor reconciling a lowered `maxActiveSegments` + (ADR 0011 §2 — pausing the lowest-priority excess). DAEMON calls CORE's `pause(TaskId)`. +- **CORE-initiated ("auto-pause")**: `auth_required` (401/407, `docs/04` §7), + `server_file_changed` (F1's carrier, `contracts/proto-answers-m1.md`), disk full. CORE + transitions to `paused` on its own and reports it up through the existing + state-change callback — the same path every other CORE-driven transition uses. This is + not a new mechanism. + +The contract that makes this safe, per CORE's sign-off (`core/docs/adr-0013-core-response.md`): + +- **`pause(TaskId)` is idempotent: a no-op success** on a task already `paused` (however + it got there), and **also a no-op success on a terminal task** (`complete`/`failed`/ + `cancelled`) — a pause racing a completion is not an error. **`resume(TaskId)` is a + no-op success on a task that is not paused.** The only error either returns is + `task_not_found`. CORE emits no state-change event for a no-op call — DAEMON reads the + resulting state from the normal state-change callback / `TaskDetail`, never from + `pause()`'s or `resume()`'s return value. If CORE is mid-transition into `paused` (its + own auto-pause) when DAEMON's `pause()` arrives, the task ends up `paused` once, with + exactly one state-change event. +- **DAEMON persists *why* a task is paused**, in the `tasks` table, not in `TaskState` + itself (the wire type stays a flat enum — this is DAEMON-local bookkeeping, not a + contract change). A `pauseReason` distinguishing at least `user`, `schedule`, + `queue_stopped`, `admission_reconcile`, and CORE's `error.code` when auto-paused — + now readable off `event.task.state.error` since PROTO's 1.3.0 widening (see "A + contract gap this ADR surfaced, now closed" below). This is what makes §3's resume + rule possible. +- **CORE does not need to track why it's paused past the current occurrence.** Once + paused, CORE's job is done; DAEMON is the only side that later decides whether to + resume, and DAEMON is also the only side with persistent storage to remember the + reason across a restart. + +### 3. Resume must not cross reasons + +The bug this section exists to prevent: a schedule window opens, DAEMON blindly resumes +every `paused` task in that queue, and it resumes a task CORE paused because it's waiting +on credentials that were never provided. The task immediately re-fails or re-pauses, looks +like a flapping bug, and burns a retry. + +Rule: **DAEMON resumes a task only when the reason it recorded matches the event that +justifies resuming.** A schedule window opening resumes tasks paused for `schedule`. A +user clicking "resume" resumes anything (explicit user intent overrides any reason). +`auth_required` and `server_file_changed` are resumed only by the paths that actually +address them — `download.provideAuth` (F2) and the user's restart-decision response (F1) — +never by the scheduler. This means `queued` is not the only state a schedule can put a +task back into a run cycle from; the scheduler's "should this task be running right now" +check must skip tasks paused for a reason it doesn't own. + +### 4. `retry_wait` is not starvation + +ADR 0011 §3.6 defines a starved task as a running task with `segments_active(id) == 0`, +and asserts `tasks_starved == 0` in steady state past a bounded delay. A task in +`retry_wait` also holds zero segments, deliberately, for up to 60 s (`docs/04` §7's backoff +cap) — and it is **not** admission-starved, it is CORE's own policy holding it idle. + +**Pinned by CORE's sign-off**, tightening ADR 0011's definition rather than special-casing +it: `tasks_starved` counts only tasks whose `TaskState` is `connecting` or `downloading` +**and** `segments_active == 0` — precisely "the allocator has not granted a slot to a task +that is asking for one." + +| Task state | In `tasks_starved`? | Why | +|---|---|---| +| `probing` | no | uses the probe pool (ADR 0011 §5), not the segment budget | +| `connecting`, 0 segments | **yes** | admitted + probed, waiting on the allocator's first grant — the real starvation case | +| `connecting`/`downloading`, ≥1 segment | no | a segment in its own `connecting` sub-state counts as held (ADR 0011 amendment A3) | +| `downloading`, 0 segments | **yes** | held slots and lost them all (e.g. every segment failed and is being re-requested) — transient, still real | +| `retry_wait` | no | CORE's backoff timer holds it at zero *deliberately*; not asking the allocator for anything until it re-enters `connecting` | +| `paused` (either-initiated) | no | not asking for a slot | +| `new`, `queued`, `assembling`, `verifying`, terminal | no | outside the counted state set | + +So `retry_wait` and auto-paused tasks are excluded **structurally, by not being in the +`{connecting, downloading}` state set** — not via a special case that could rot as the +engine evolves. `starved_tasks()` returns exactly the TaskIds in this count; +`starved_since(id)` is defined only for them. A task in `retry_wait` still counts against +`connection.maxConcurrentDownloads` from DAEMON's side (it is running, not requeued) but +contributes nothing to the segment allocator's guarantee pass until it re-enters +`connecting`. + +### 5. Restart — CORE holds no persistent state, DAEMON reloads to `queued` + +Per the M1 DoD ("kill and restart the daemon mid-download: all tasks reload with correct +state and resume") and the layering rule (CORE has no SQL, no state survives a CORE +restart except what's on disk in `.veloxpart.meta`, which CORE reads back itself): on +daemon startup, DAEMON loads every non-terminal task from SQLite. Any task whose persisted +`TaskState` was a CORE-owned state (`probing` through `verifying`) is **rewritten to +`queued`** in memory before the scheduler sees it — the on-disk `TaskState` is a +last-known-value, not a resumable position, because CORE's in-process state died with the +process. The scheduler re-admits it exactly like any other queued task; CORE re-derives +where to actually resume from `.veloxpart.meta` and re-validates with `If-Range` +(`docs/04` §5), independent of what SQL said the state was. + +`paused` tasks reload as `paused`, with their `pauseReason` intact, and are not +auto-admitted — §3 applies identically after a restart as it does live. + +**Confirmed by CORE, fully.** `start(TaskId)` transparently checks for a valid +`.veloxpart.meta` sidecar (CORE's stage 5), re-validates with `If-Range` (`docs/04` §5), +and either resumes from the recorded offsets or restarts if the sidecar is missing or +fails validation. DAEMON does nothing special beyond rewriting CORE-owned states to +`queued` and re-admitting, as this section already said. + +**Two notes from CORE, not objections to the ADR:** the *transition* into `paused` is +honoured from any CORE state per §1's table, but the *work interruption* is best-effort — +a pause during `verifying` discards the in-progress hash and re-hashes from the start on +resume (cheap, bounded); a pause during `assembling` (HLS/DASH mux) is an M4 concern and +may not be cleanly interruptible mid-mux. Neither changes this ADR's API shape. + +## Consequences + +- `daemon/src/sched/` calls `start(TaskId)` exactly once per admission (`queued → + probing`), never re-enters a CORE-owned state directly, and treats every CORE-owned + state as opaque past that call except for reading it back for projection. +- The `tasks` table needs a `pauseReason` column (DAEMON-local; not a wire type) before + `sched/`'s pause/resume logic can be written correctly — flagging as a concrete + follow-up, not blocking this ADR's acceptance. +- CORE's `starved_tasks()` (ADR 0011) excludes `retry_wait` and CORE-auto-paused tasks by + construction — pinned in CORE's sign-off as the `{connecting, downloading} ∧ + segments_active == 0` definition in §4's table — so DAEMON's governor-invariant warning + will not false-positive on a normal backoff cycle. +- No `contracts/` change. `TaskState`, `event.task.state`, and `previousState` are already + sufficient; this ADR is entirely about which process calls which function when. + +## Alternatives considered + +**A single owner drives every transition (CORE, told about queues).** Rejected — this is +the D1 problem restated with CORE holding the SQL-shaped concepts (`queued`, schedules, +priority) that the layering rule (`CLAUDE.md` §3) forbids it from touching. + +**DAEMON drives every transition, treating CORE as a dumb byte-mover.** Rejected — CORE's +internal states (`retry_wait`, `assembling`, `verifying`) depend on engine internals +(backoff timers, mux completion, streaming hash state) DAEMON has no visibility into +without CORE reporting them; forcing DAEMON to poll or reimplement that timing duplicates +`docs/04` §7 in two places and they will drift. + +**A `pauseReason` on the wire (`TaskState` split into `paused_user` / `paused_auto` / +etc.).** Rejected — it roughly doubles the enum for a fact only DAEMON's resume logic +needs, and adding wire cardinality for internal bookkeeping is the kind of thing that +becomes a compatibility problem the moment a client starts branching on it. The value +DAEMON needs (CORE's reason) travels on the existing `error` field instead — see the +closed contract gap below — without touching `TaskState` itself. + +## A contract gap this ADR surfaced, now closed + +`event.task.state.schema.json`'s description scoped `error` to "whenever the new state is +failed or retry_wait" — **not** `paused`, at the time this ADR was drafted, and the one +fixture (`event.task.state.json`) only exercised the `failed` case. So CORE auto-pausing +for `auth_required` or `server_file_changed` had no wire signal telling DAEMON *why* — §2/ +§3 of this ADR were unbuildable without one. + +**Closed by PROTO**, `contracts/` 1.3.0 (`lane/proto` commit `6db304a`, see the status +line at top for the merge caveat): minor widening of the existing field's presence +condition, per `contracts/README.md` rule 4 — no new field, no retype, `error` stays +`TaskError | null`. It is now populated on a `paused` transition whenever CORE entered it +unilaterally; `error: null` on a DAEMON-initiated pause is unchanged. New fixture +`event.task.state.auto-paused.json` exercises the auto-pause case; the existing +`event.task.state.json` fixture's stale "exactly when failed or retry_wait" assertion was +corrected in the same change. + +## Resolution of the four open items + +1. **`starved_tasks()` excludes `retry_wait`/auto-paused by construction — confirmed.** + Pinned as a design commitment in CORE's sign-off (§4's table above): the counted set is + `{connecting, downloading} ∧ segments_active == 0`, so exclusion is structural, not a + special case. +2. **`pause()`/`resume()` idempotency — confirmed as specified in §2 above**, plus two + details this ADR hadn't anticipated: a no-op call also succeeds against a *terminal* + task (pause racing completion isn't an error), and no state-change event fires for a + no-op — DAEMON reads resulting state from the callback/`TaskDetail`, never from the + call's return value. +3. **Landed — PROTO, `contracts/` 1.3.0** (`lane/proto` commit `6db304a`, not yet merged + to `main`). `event.task.state.error` and `TaskSummary.error` are now populated "on + every `failed` or `retry_wait` transition, and on a `paused` transition the daemon + entered unilaterally" — a deliberate pause still carries `error: null`. New fixture + `event.task.state.auto-paused.json` exercises an `auth_required` auto-pause directly. + DAEMON's §3 resume rule can be implemented once `sched/` is built against `main` at + 1.3.0 or later. +4. **CORE adopts "auto-pause".** No new wire or API term — the discriminator stays + `state == paused` plus the presence of an `Error` (present ⇒ CORE-initiated, absent ⇒ + DAEMON-initiated), exactly as §2 already specified.