@plurnk/plurnk-service 1.3.12 → 1.4.0

package/dist/core/Engine.sql

@@ -1,15 +1,15 @@
--- Engine SQL. SPEC §arch (architecture), §scheme (op dispatch + log).
+-- Engine SQL. SPEC {§arch} (architecture), {§scheme} (op dispatch + log).
 
 -- PREP: engine_loop_status
 SELECT status FROM loops WHERE id = $loop_id;
 
 -- PREP: engine_worker_has_live_child
--- A non-terminal CHILD run (worker:// spawn/fork set parent_worker_id) — a "live thing the worker holds",
--- like an open stream. A SEND[200] while one exists is a premature-terminate (§send-premature-terminate).
+-- A non-terminal child worker (worker:// spawn/fork set parent_worker_id) — a "live thing the worker holds",
+-- like an open stream. A SEND[200] while one exists is a premature-terminate ({§send-premature-terminate}).
 -- Live = a child whose LATEST loop is still pending/running/parked (100/102/202) — the SAME definition
--- engine_child_workers_live uses for the Child Runs orientation, so the 409 gate and the section the model
+-- engine_child_workers_live uses for the Active Child Workers orientation, so the 409 gate and the section the model
 -- reads NEVER disagree: a refused termination is always backed by a child the model can SEE and KILL
--- (§child-orientation). An inherited/historical loop (a fork copies the parent's loops) is not the
+-- ({§child-orientation}). An inherited/historical loop (a fork copies the parent's loops) is not the
 -- latest, so it never makes a concluded child look forever-live.
 SELECT 1 AS live FROM workers r
 JOIN loops l ON l.id = (SELECT id FROM loops WHERE worker_id = r.id ORDER BY sequence DESC, id DESC LIMIT 1)
@@ -22,46 +22,34 @@ WHERE r.parent_worker_id = $worker_id AND l.status IN (100, 102, 202) LIMIT 1;
 SELECT COUNT(*) AS n FROM loops WHERE worker_id = $worker_id AND status = 102;
 
 -- PREP: engine_get_loop_flags
--- Loads the loop's persisted flags (json). Default '{}'; auto listener and
--- SchemeRegistry.resolveForLoop merge over DEFAULT_LOOP_FLAGS for missing
--- fields. Migration 014.
+-- Raw persisted partial policy. LoopFlagsReader is the sole runtime expansion
+-- and validation path ({§loop-flags-effective-read}).
 SELECT flags FROM loops WHERE id = $loop_id;
 
 -- PREP: engine_set_loop_flags
--- Updates the loop's persisted flags (json). Called by loop.run RPC handler.
+-- Updates the loop's persisted flags (json). Called by the runLoop seam.
 UPDATE loops SET flags = $flags WHERE id = $loop_id;
 
--- PREP: engine_tag_loop_attributions
--- #249 — tag the loop with its active plugins' attribution tags (string[] JSON), write-once: the
--- active set is loop-stable, so the `= '[]'` guard keeps later turns from rewriting it.
-UPDATE loops SET attributions = $attributions WHERE id = $loop_id AND attributions = '[]';
-
 -- PREP: engine_set_loop_open_paths
--- #260 — persist the loop.run-passed @file paths (string[] JSON) on the loop before the drain
--- starts, so runTurn foists a turn-0 READ of each (same seam as the #250 AGENTS.md auto-read).
+-- {§methods-loop-run-open-paths}: persist the initial prompt frame's selected
+-- paths (string[] JSON) until turn 1 materializes that frame.
 UPDATE loops SET open_paths = $open_paths WHERE id = $loop_id;
 
--- PREP: engine_get_loop_open_paths
--- #260 — runTurn reads the loop's @file foist-paths at turn 0.
-SELECT open_paths FROM loops WHERE id = $loop_id;
-
 -- PREP: engine_get_loop_prompt
--- Loop's prompt + sequence — runTurn reads it on turn 1 to foist a
--- system-origin EDIT against plurnk://prompt/<run>/<loop>/1 (§packet), at the
--- turn's first action sequence. Prompts are first-class log entries
--- (no synthetic / shim layer).
-SELECT prompt, sequence FROM loops WHERE id = $loop_id;
+-- Initial prompt frame. On turn 1, runTurn stores the owner-keyed
+-- prompt:///<loop>/1 entry with its selected paths and publishes it.
+SELECT prompt, sequence, open_paths FROM loops WHERE id = $loop_id;
 
 -- PREP: engine_reclaim_queued_loop
--- §worker-lifecycle-wake-requeue-not-terminal — atomic 100→102 re-claim by loop id. A wake
+-- {§worker-lifecycle-wake-requeue-not-terminal} — atomic 100→102 re-claim by loop id. A wake
 -- re-queued this loop while ITS OWN live drain was between turns; the drain re-claims and
 -- keeps running (the injected prompt is already the next turn). Conditional so a racing
 -- claimant can never double-claim.
 UPDATE loops SET status = 102 WHERE id = $loop_id AND status = 100;
 
 -- PREP: workspace_get_settings
--- #231 — the workspace's client-chosen open-context bag ({ manifestItems?, mdDocs? }),
--- read at turn-0 with precedence over env.
+-- {§operator-config} — the workspace's validated client settings bag, read at
+-- each owning use site with its declared composition semantics.
 SELECT settings FROM workspaces WHERE id = $workspace_id;
 
 -- PREP: engine_target_diverged_this_turn
@@ -75,26 +63,29 @@ WHERE worker_id = $worker_id AND turn_id = $turn_id
   AND scheme IS $scheme AND pathname = $pathname
 LIMIT 1;
 
--- PREP: engine_list_workspace_entry_tags
--- #note13 — every (entry, tag) in the workspace, for the manifest catalog's tags field.
-SELECT et.entry_id, et.tag
-FROM entry_tags et
-JOIN entries e ON e.id = et.entry_id
-WHERE e.workspace_id = $workspace_id
-ORDER BY et.entry_id, et.tag;
-
--- PREP: engine_list_worker_entries
+-- PREP: engine_list_owner_entries
 -- {§entry-owner} — one principal's entries (catalogRowsFor source for an owner-scoped FIND/foist):
 -- the commons, a worker's own space, or a named space — exactly one owner's rows, its perspective.
 SELECT e.id AS entry_id, e.scheme, e.pathname, ec.name AS channel, ec.content, ec.mimetype, ec.tokens AS tokens, e.deep_hash,
-    CAST(unixepoch('now') - unixepoch(s.opened_at) AS INTEGER) AS seconds
+    s.id AS subscription_id,
+    CASE WHEN s.closed_at IS NULL
+        THEN CAST(unixepoch('now') - unixepoch(s.opened_at) AS INTEGER)
+        ELSE NULL
+    END AS seconds,
+    s.close_status
 FROM entries e
 JOIN entry_channels ec ON ec.entry_id = e.id
-LEFT JOIN subscriptions s ON s.entry_id = e.id AND s.closed_at IS NULL
+LEFT JOIN subscriptions s ON s.id = (
+    SELECT latest.id
+    FROM subscriptions latest
+    WHERE latest.entry_id = e.id
+    ORDER BY latest.id DESC
+    LIMIT 1
+)
 WHERE e.workspace_id = $workspace_id AND e.owner_id = $owner_id
 ORDER BY e.updated_at ASC, e.id ASC, ec.name;
 
--- PREP: engine_list_worker_entry_tags
+-- PREP: engine_list_owner_entry_tags
 -- {§entry-owner} — (entry, tag) for one principal's entries (the owner-scoped catalog's tags field).
 SELECT et.entry_id, et.tag
 FROM entry_tags et
@@ -102,33 +93,63 @@ JOIN entries e ON e.id = et.entry_id
 WHERE e.workspace_id = $workspace_id AND e.owner_id = $owner_id
 ORDER BY et.entry_id, et.tag;
 
--- PREP: engine_worker_scratch_count
--- §worker-scheme — distinct entry count owned by the building worker, to decide whether the
--- turn-0 catalog foists a FIND(worker://~/**) (a worker with no private space foists nothing).
-SELECT COUNT(DISTINCT e.id) AS entries
-FROM entries e
-WHERE e.workspace_id = $workspace_id AND e.owner_id = $owner_id;
-
 -- PREP: engine_next_turn_sequence
 SELECT COALESCE(MAX(sequence), 0) + 1 AS next FROM turns WHERE loop_id = $loop_id;
 
 -- PREP: engine_loop_usage
--- Per-loop usage totals — SUM the loop's turns (§tokenomics stores usage per turn).
--- Surfaced on loop.run + loop/terminated (#197). `context` is the LAST turn's prompt tokens — the
--- honest window-occupancy numerator (#263), distinct from the summed `prompt` (which overcounts a
--- context that grows across turns).
+-- Per-loop usage totals and latest-turn gauge ({§tokenomics-client-gauge}),
+-- surfaced on {§notifications-loop-terminated}. `context` is the latest
+-- provider attempt on the latest turn, distinct from summed billed `prompt`.
 SELECT COALESCE(SUM(usage_prompt), 0)     AS prompt,
        COALESCE(SUM(usage_completion), 0) AS completion,
-       COALESCE(SUM(usage_cost_usd), 0)  AS cost_usd,
-       (SELECT usage_prompt FROM turns WHERE loop_id = $loop_id ORDER BY sequence DESC LIMIT 1) AS context,
-       -- #274 — the LAST turn's model window (denominator), so numerator + denominator come from
-       -- the same loop/model; NULL when the provider reports no window.
+       CASE
+           WHEN COUNT(*) = 0 THEN 0
+           WHEN COUNT(usage_cost_usd) = COUNT(*) THEN SUM(usage_cost_usd)
+           ELSE NULL
+       END AS cost_usd,
+       COALESCE((
+           SELECT json_group_array(json(cost.value))
+           FROM turns charged, json_each(charged.usage_cost) AS cost
+           WHERE charged.loop_id = $loop_id
+       ), '[]') AS costs,
+       (
+           SELECT a.usage_prompt
+           FROM turn_attempts a
+           WHERE a.turn_id = (
+               SELECT latest.id
+               FROM turns latest
+               WHERE latest.loop_id = $loop_id
+               ORDER BY latest.sequence DESC
+               LIMIT 1
+           )
+           ORDER BY a.sequence DESC
+           LIMIT 1
+       ) AS context,
+       -- Latest turn's effective packet allowance; NULL when uncapped or unknown.
+       -- {§tokenomics-client-gauge}
        (SELECT usage_prompt_budget FROM turns WHERE loop_id = $loop_id ORDER BY sequence DESC LIMIT 1) AS context_size,
-       -- #252 — the opaque provider meta blob from the LATEST turn (for example, a
-       -- point-in-time snapshot; latest wins). Service-unenforced passthrough to the client.
+       -- Latest turn's opaque provider metadata. {§meta-passthrough}
        (SELECT meta FROM turns WHERE loop_id = $loop_id ORDER BY sequence DESC LIMIT 1) AS meta
 FROM turns WHERE loop_id = $loop_id;
 
+-- PREP: engine_loop_attributions
+-- {§attribution} — derive the loop projection from exact response-attempt
+-- evidence plus each turn's latest request (which also covers a call that
+-- failed without response evidence).
+SELECT attribution
+FROM (
+    SELECT value AS attribution
+    FROM turns t, json_each(t.packet, '$.attributions')
+    WHERE t.loop_id = $loop_id
+    UNION
+    SELECT value AS attribution
+    FROM turn_attempts a
+    JOIN turns t ON t.id = a.turn_id,
+         json_each(a.attributions)
+    WHERE t.loop_id = $loop_id
+)
+ORDER BY attribution;
+
 -- PREP: engine_loop_turn_seqs
 -- Look up (loop_seq, turn_seq) for a given (loop_id, turn_id). Used by
 -- #writeLog when an op needs to address itself or its output by log
@@ -138,14 +159,14 @@ FROM loops l, turns t
 WHERE l.id = $loop_id AND t.id = $turn_id;
 
 -- PREP: engine_open_turn
--- Turn-as-container model: insert a turn row at runTurn open with a
--- placeholder packet and status=102 (in-progress). Pre-model writes
--- (the user prompt; later, system signals/telemetry events) land into
+-- Turn-as-container model: insert a turn row at runTurn open with no assembled
+-- model request and status=102 (in-progress). Pre-model writes
+-- (the user prompt; later, system signals/notices) land into
 -- this row before the provider is called. The turn is then "closed"
 -- via engine_close_turn with the final packet + status + usage stats
 -- after dispatch completes.
-INSERT INTO turns (loop_id, sequence, status, packet)
-VALUES ($loop_id, $sequence, 102, '{}')
+INSERT INTO turns (loop_id, sequence, status)
+VALUES ($loop_id, $sequence, 102)
 RETURNING id;
 
 -- PREP: engine_close_turn
@@ -159,6 +180,7 @@ UPDATE turns SET
     usage_completion = $usage_completion,
     usage_reasoning = $usage_reasoning,
     usage_cached = $usage_cached,
+    usage_cost = $usage_cost,
     usage_cost_usd = $usage_cost_usd,
     usage_prompt_budget = $usage_prompt_budget,
     finish_reason = $finish_reason,
@@ -166,81 +188,161 @@ UPDATE turns SET
     meta = $meta
 WHERE id = $id;
 
+-- PREP: engine_record_turn_attempt
+-- One provider attempt carrying response evidence beneath a turn. Rejected or
+-- interrupted attempts remain forensic evidence but never become packet history.
+INSERT INTO turn_attempts (
+    turn_id,
+    sequence,
+    accepted,
+    response,
+    parse_errors,
+    attributions,
+    usage_prompt,
+    usage_completion,
+    usage_reasoning,
+    usage_cached,
+    usage_cost,
+    usage_cost_usd,
+    finish_reason,
+    model
+)
+VALUES (
+    $turn_id,
+    $sequence,
+    $accepted,
+    $response,
+    $parse_errors,
+    $attributions,
+    $usage_prompt,
+    $usage_completion,
+    $usage_reasoning,
+    $usage_cached,
+    $usage_cost,
+    $usage_cost_usd,
+    $finish_reason,
+    $model
+);
+
 -- PREP: engine_list_workspace_entries
--- Every entry of a workspace — all schemes, all channels — the source behind the entry
--- catalog (catalogRowsFor / FIND) and the per-turn derivation pump (maintainDerivations).
--- Workspace-scoped (persists across runs); FOLD doesn't drop from the catalog.
--- `seconds` is the live age of an active stream: now − the open subscription's
--- opened_at (closed_at IS NULL). NULL for static entries. unixepoch parses the
--- stored '...%fZ' timestamp directly; re-evaluated every render like tokens.
+-- Every owner-held entry of a workspace — all schemes, all channels — for
+-- internal indexing and aggregate inspection. Addressable catalogs use the
+-- owner-filtered statements above.
+-- The latest subscription carries stream lifecycle into the catalog. `seconds`
+-- is the live age of an open stream; close_status is the exact terminal status
+-- of a closed one. Entries with no subscription remain ordinary static entries.
 SELECT e.id AS entry_id, e.scheme, e.pathname, ec.name AS channel, ec.content, ec.mimetype, ec.tokens AS tokens, e.deep_hash,
-    CAST(unixepoch('now') - unixepoch(s.opened_at) AS INTEGER) AS seconds
+    s.id AS subscription_id,
+    CASE WHEN s.closed_at IS NULL
+        THEN CAST(unixepoch('now') - unixepoch(s.opened_at) AS INTEGER)
+        ELSE NULL
+    END AS seconds,
+    s.close_status
 FROM entries e
 JOIN entry_channels ec ON ec.entry_id = e.id
-LEFT JOIN subscriptions s ON s.entry_id = e.id AND s.closed_at IS NULL
--- entries are workspace-scoped, shared across runs — §machine-processes-one-filesystem
+LEFT JOIN subscriptions s ON s.id = (
+    SELECT latest.id
+    FROM subscriptions latest
+    WHERE latest.entry_id = e.id
+    ORDER BY latest.id DESC
+    LIMIT 1
+)
 WHERE e.workspace_id = $workspace_id
 -- User Note 5 — mtime-ascending: dormant entries hold the stable prompt-cache prefix; churn clusters at the tail.
 ORDER BY e.updated_at ASC, e.id ASC, ec.name;
 
 -- PREP: engine_scheme_catalog_summary
--- Per-scheme entry tally (distinct count; scheme=null → file). Sources the turn-0 per-scheme
--- foist so the model sees which schemes hold content without running FIND(known://**) itself;
--- the foist FIND carries each scheme's live token weight.
+-- Per-scheme entry tally plus one-level catalog-row count (direct entries and
+-- distinct first-segment `dir/**` summaries). The engine uses the latter to
+-- encode a valid first-N file preview range without guessing how many rows the
+-- shallow projection will produce.
 SELECT e.scheme AS scheme,
-    COUNT(DISTINCT e.id) AS entries
+    COUNT(DISTINCT e.id) AS entries,
+    COUNT(DISTINCT CASE
+        WHEN instr(ltrim(e.pathname, '/'), '/') = 0
+            THEN 'entry:' || ltrim(e.pathname, '/')
+        ELSE 'scope:' || substr(ltrim(e.pathname, '/'), 1, instr(ltrim(e.pathname, '/'), '/'))
+    END) AS shallow_items
 FROM entries e
 JOIN entry_channels ec ON ec.entry_id = e.id
 WHERE e.workspace_id = $workspace_id
 GROUP BY e.scheme
 ORDER BY e.scheme;
 
--- PREP: engine_worker_prior_turn_time
--- §env-delta — timestamp of this worker's most recent turn BEFORE the current one
--- (the "since I last looked" boundary). NULL on the worker's first turn → no deltas.
-SELECT MAX(t.timestamp) AS since
-FROM turns t JOIN loops l ON l.id = t.loop_id
-WHERE l.worker_id = $worker_id AND t.id != $turn_id;
-
--- PREP: engine_pull_env_deltas
--- §env-delta — other actors' resolved EDITs on shared entries since this worker last
--- looked. Real edits (origin model/client) AND the plurnk worker's fs-sync fictions
--- (origin=plurnk on the reserved 'plurnk' run); excludes this worker's own rows and
--- other workers' already-materialized deltas (origin=plurnk on a real run). plurnk:///
--- entries (manifest/prompt/doc) never surface. This is the environment door (§actor-boundary-two-doors); the voice door is inject.
-SELECT le.worker_id, le.scheme, le.pathname, le.rx, le.source, le.attrs
-FROM log_entries le
-JOIN workers r ON r.id = le.worker_id
-WHERE r.workspace_id = $workspace_id
-  AND le.op = 'EDIT'
-  AND le.state = 'resolved'
-  AND le.status_rx IN (200, 201)
-  AND (le.scheme IS NULL OR le.scheme != 'plurnk')
-  AND le.worker_id != $worker_id
-  AND le.at > $since
-  AND (le.origin != 'plurnk'
-       OR le.worker_id = (SELECT id FROM workers WHERE workspace_id = $workspace_id AND name = 'plurnk'))
-ORDER BY le.at;
-
--- PREP: engine_insert_env_delta
--- §env-delta — materialize a pulled cross-actor edit as a FOLDED delta (expanded=0)
--- in this worker's log. origin=plurnk; source carries the cause (sibling worker id or
--- 'file'); rx reuses the originating row's result span (§edit-result-render).
+-- PREP: engine_initialize_ambient_cursor
+-- A worker's first packet reads current shared state directly, so pre-existing
+-- occurrences are its baseline rather than historical deltas. The NULL guard
+-- makes this safe on every turn and preserves a fork's inherited progress.
+UPDATE workers
+SET ambient_event_cursor = COALESCE((
+    SELECT MAX(ae.id) FROM ambient_events ae WHERE ae.workspace_id = $workspace_id
+), 0)
+WHERE id = $worker_id
+  AND workspace_id = $workspace_id
+  AND ambient_event_cursor IS NULL
+RETURNING ambient_event_cursor;
+
+-- PREP: engine_pull_ambient_events
+-- One SQLite snapshot captures both ends of the closed observation window.
+-- The LEFT JOIN returns the boundary even when the window contains no event.
+WITH observation AS (
+    SELECT w.ambient_event_cursor AS cursor,
+           COALESCE(
+               (SELECT MAX(ae.id) FROM ambient_events ae WHERE ae.workspace_id = $workspace_id),
+               w.ambient_event_cursor,
+               0
+           ) AS boundary
+    FROM workers w
+    WHERE w.id = $worker_id AND w.workspace_id = $workspace_id
+)
+SELECT o.cursor, o.boundary,
+       ae.id AS event_id, ae.producer_worker_id, producer.name AS producer_worker_name,
+       ae.kind, ae.source,
+       ae.op, ae.scheme, ae.hostname, ae.pathname, ae.rx, ae.attrs,
+       ae.status_rx, ae.prompt, ae.terminated_by
+FROM observation o
+LEFT JOIN ambient_events ae
+  ON ae.workspace_id = $workspace_id
+ AND ae.id > o.cursor
+ AND ae.id <= o.boundary
+ AND ae.producer_worker_id != $worker_id
+LEFT JOIN workers producer ON producer.id = ae.producer_worker_id
+ORDER BY ae.id;
+
+-- PREP: engine_insert_ambient_delta
+-- Materialize one occurrence into the observer's self-contained log. The
+-- targeted conflict rule makes crash replay idempotent without swallowing any
+-- unrelated sequence, FK, or shape violation.
 INSERT INTO log_entries (
-    worker_id, loop_id, turn_id, sequence, origin, source,
-    op, scheme, pathname, tx, mimetype_tx, rx, mimetype_rx, status_rx, expanded, attrs
+    worker_id, loop_id, turn_id, sequence, origin, source, ambient_event_id,
+    op, scheme, hostname, pathname, tx, mimetype_tx,
+    rx, mimetype_rx, status_rx, expanded, attrs
 ) VALUES (
-    $worker_id, $loop_id, $turn_id, $sequence, 'plurnk', $source,
-    'EDIT', $scheme, $pathname, '', 'text/plain', $rx, 'application/json', 200, 0, $attrs
-);
+    $worker_id, $loop_id, $turn_id, $sequence, 'plurnk', $source, $event_id,
+    $op, $scheme, $hostname, $pathname, '', 'text/plain',
+    $rx, $mimetype_rx, $status, $expanded, $attrs
+)
+ON CONFLICT(worker_id, ambient_event_id) WHERE ambient_event_id IS NOT NULL DO NOTHING
+RETURNING id;
+
+-- PREP: engine_advance_ambient_cursor
+-- Advance only from the snapshot that was actually materialized. A concurrent
+-- pull winning the CAS is safe; a loser replays idempotently on its next turn.
+UPDATE workers
+SET ambient_event_cursor = $boundary
+WHERE id = $worker_id
+  AND workspace_id = $workspace_id
+  AND ambient_event_cursor IS $cursor
+  AND $boundary >= ambient_event_cursor
+RETURNING ambient_event_cursor;
 
 -- PREP: engine_worker_stream_channels
--- §exec-stream — every stream channel the worker owns (an EXEC's stdout/stderr live on the
+-- {§exec-stream} — every stream channel the worker owns (an EXEC's stdout/stderr live on the
 -- runtime-tag entry), with content + state + coordinate, so the per-turn injector can emit the
 -- channel's unshown byte-delta. Stays listed until its last delta is shown (cursor == content len).
 SELECT s.id AS subscription_id, e.scheme AS runtime, e.pathname AS coord,
     ec.name AS channel, ec.content AS content, ec.state AS state, s.close_status AS close_status,
-    s.published_channel AS published_channel
+    s.close_result AS close_result, s.published_channel AS published_channel
 FROM subscriptions s
 JOIN entries e ON e.id = s.entry_id
 JOIN entry_channels ec ON ec.entry_id = s.entry_id
@@ -249,7 +351,7 @@ WHERE s.worker_id = $worker_id
 ORDER BY s.id, ec.name;
 
 -- PREP: engine_stream_cursor
--- §exec-stream — bytes of this channel already shown to the worker: the streamEnd recorded on its
+-- {§exec-stream} — bytes of this channel already shown to the worker: the streamEnd recorded on its
 -- latest foisted delta (the caller defaults to 0 when none exists yet).
 SELECT attrs
 FROM log_entries
@@ -258,44 +360,16 @@ WHERE worker_id = $worker_id AND origin = 'plurnk' AND op = 'READ'
 ORDER BY id DESC LIMIT 1;
 
 -- PREP: engine_insert_stream_delta
--- §exec-stream / §env-delta — materialize a channel's unshown byte-delta as a
--- foisted READ@200 row (the model READs the stream it never typed). origin=plurnk; fragment is
+-- {§exec-stream} / {§env-delta} — materialize a channel's unshown byte-delta as a
+-- foisted READ row (the model READs the stream it never typed). origin=plurnk; fragment is
 -- the channel; attrs.streamEnd is the next turn's cursor; expanded=1 when the channel has CLOSED
--- (the terminal delta auto-OPENs), 0 while it streams (ongoing deltas fold). §exec-stream
+-- (the terminal delta auto-OPENs), 0 while it streams (ongoing deltas fold). {§exec-stream}
 INSERT INTO log_entries (
     worker_id, loop_id, turn_id, sequence, origin, source,
     op, scheme, pathname, fragment, tx, mimetype_tx, rx, mimetype_rx, status_rx, attrs, expanded
 ) VALUES (
     $worker_id, $loop_id, $turn_id, $sequence, 'plurnk', NULL,
-    'READ', $scheme, $pathname, $fragment, '', 'text/plain', $rx, 'application/json', 200, $attrs, $expanded
-);
-
--- PREP: engine_pull_loop_terminations
--- §worker-scheme — sibling runs' loops that reached a terminal status since this worker last
--- looked (the loop-termination ambient delta). Carries terminal_message — the SEND[200]
--- deliverable or the abandonment reason — plus terminated_by so a cancellation renders
--- its marker (#379). Excludes this worker's own loops.
-SELECT l.worker_id, r.name AS worker_name, l.status, l.prompt, l.terminal_message, l.terminated_by
-FROM loops l
-JOIN workers r ON r.id = l.worker_id
-WHERE r.workspace_id = $workspace_id
-  AND l.terminated_at IS NOT NULL
-  AND l.terminated_at > $since
-  AND l.worker_id != $worker_id
-ORDER BY l.terminated_at;
-
--- PREP: engine_insert_loop_termination_delta
--- §worker-scheme — materialize a sibling's loop-termination as a delta: a SEND from
--- worker:///<name> carrying the terminal status + message (the deliverable). origin=plurnk,
--- source=the terminated run — uniform with the env-delta. Born OPEN for a 2xx deliverable
--- (a child's success must reach the parent open + awakening), folded otherwise.
-INSERT INTO log_entries (
-    worker_id, loop_id, turn_id, sequence, origin, source,
-    op, scheme, pathname, tx, mimetype_tx, rx, mimetype_rx, status_rx, expanded
-) VALUES (
-    $worker_id, $loop_id, $turn_id, $sequence, 'plurnk', $source,
-    'SEND', 'worker', $pathname, '', 'text/plain', $rx, 'text/markdown', $status,
-    CASE WHEN $status >= 200 AND $status < 300 THEN 1 ELSE 0 END
+    'READ', $scheme, $pathname, $fragment, '', 'text/plain', $rx, 'application/json', $status, $attrs, $expanded
 );
 
 -- PREP: engine_entry_tags
@@ -303,7 +377,7 @@ SELECT tag FROM entry_tags WHERE entry_id = $entry_id ORDER BY tag;
 
 -- PREP: engine_child_workers_live
 -- The worker's LIVE child workers — latest loop non-terminal (100 pending / 102 processing / 202 parked).
--- Powers the Child Runs orienting section (§child-orientation): terse `* <status> worker://<name>`
+-- Powers the Active Child Workers orienting section ({§child-orientation}): terse `* <status> worker://<name>`
 -- pointers so the model SEES what it holds live and reasons for itself (READ/KILL), never told to.
 -- Empty → section omitted.
 SELECT r.name, l.status
@@ -314,7 +388,7 @@ ORDER BY r.name;
 
 -- PREP: engine_child_streams_open
 -- The worker's OPEN streams (subscriptions not yet closed) — their addressable coord. Powers the Child
--- Streams orienting section (§child-orientation): terse `* active <runtime>:///<coord>` pointers the
+-- Streams orienting section ({§child-orientation}): terse `* active <runtime>:///<coord>` pointers the
 -- model OPENs/READs/KILLs. Empty → section omitted.
 SELECT s.scheme, e.pathname
 FROM subscriptions s
@@ -322,14 +396,14 @@ JOIN entries e ON e.id = s.entry_id
 WHERE s.worker_id = $worker_id AND s.closed_at IS NULL
 ORDER BY e.pathname;
 
--- PREP: engine_render_telemetry_errors
--- SPEC §telemetry: 4xx/5xx log rows are mirrored into the packet's telemetry as
+-- PREP: engine_render_errors
+-- SPEC {§operation-results}: 4xx/5xx log rows are indexed in the packet's errors as
 -- LogCoordinate pointers, forcing the model to confront failures instead of letting
 -- them rot in log:///. Window = the current turn AND the immediately-prior one
 -- (>= current_turn_seq - 1): prior-turn for action failures the model just caused,
 -- current-turn so a same-turn engine error (the grinder's budget-overflow row, minted
 -- pre-generate then re-derived) surfaces THIS turn rather than a turn late.
--- §telemetry-uniform-error-channel
+-- {§operation-result-uniform-error-channel}
 SELECT
     le.op, le.sequence, le.status_rx, le.rx, le.mimetype_rx,
     le.scheme, le.pathname,
@@ -339,71 +413,72 @@ JOIN turns t ON t.id = le.turn_id
 JOIN loops l ON l.id = le.loop_id
 WHERE le.loop_id = $loop_id
   AND le.status_rx >= 400
-  -- §log-row-self-explains: every >=400 row points at ITSELF — the op row carries its failure
+  -- {§log-row-self-explains}: every >=400 row points at ITSELF — the op row carries its failure
   -- message on its meta line (packet-wire), so the pointer leads to a record that states its why.
   AND t.sequence >= $current_turn_seq - 1
 ORDER BY t.sequence, le.sequence;
 
--- PREP: engine_grinder_fold_newest_turn
--- §grinder-layer1-rollback — THE DOCTRINE: the log is the model's memory and the model ALONE
+-- TX: engine_grinder_fold_newest_turn
+-- {§grinder-layer1-rollback} — THE DOCTRINE: the log is the model's memory and the model ALONE
 -- curates it (FOLD/KILL). The grinder never touches history — it only blocks NEW memories from
 -- landing when there is no room: one set-op folds the still-open rows of the newest turn
 -- boundary — the immediately-prior turn's emissions and the current turn's pre-model rows
 -- (foists, wake surfaces). Turn 1 is the same rule (no prior turn; its foists are the newest).
--- Folded, never deleted; THREE exemptions (§grinder-errors-exempt): op='error' rows, the
--- user PROMPT (#382 — the task frame the engine foisted is not the model's curatable memory;
--- the engine never reclaims the definition of the task it set), and PLAN rows (#465, owner
--- ruling — the checklist is the model's orientation surface at exactly the moment the grinder
--- fires; plans are concise by rule, so exempting them reclaims almost nothing and preserves
--- the reasoning thread a recovery turn steers by).
-UPDATE log_entries SET expanded = 0
+-- Folded, never deleted; THREE exemptions ({§grinder-errors-exempt}): op='error',
+-- the user prompt, and PLAN. The temporary set captures the
+-- selection once so folding and additive `overflow` tagging cannot diverge.
+CREATE TEMP TABLE IF NOT EXISTS engine_grinder_fold_set (id INTEGER PRIMARY KEY);
+DELETE FROM engine_grinder_fold_set;
+INSERT INTO engine_grinder_fold_set (id)
+SELECT id FROM log_entries
 WHERE loop_id = $loop_id AND expanded = 1 AND op NOT IN ('error', 'PLAN')
   AND COALESCE(scheme, '') != 'prompt'  -- the task frame ({§prompt-self-only}); NULL-safe: a model row's scheme is NULL
   AND (turn_id = $turn_id
        OR turn_id = (SELECT MAX(id) FROM turns WHERE loop_id = $loop_id AND id < $turn_id));
 
+UPDATE log_entries SET expanded = 0
+WHERE id IN (SELECT id FROM engine_grinder_fold_set);
+
+INSERT OR IGNORE INTO log_tags (log_entry_id, tag)
+SELECT id, 'overflow' FROM engine_grinder_fold_set;
+
+DELETE FROM engine_grinder_fold_set;
+
 -- PREP: engine_fold_log_entry
--- §prompt-fold (User Note 6): fold a single log row by id — collapse to its
--- coordinate, body elided in the render, re-OPENable. Used for the foisted prompt
--- EDIT, which duplicates packet.user.prompt: logged for forensics, born folded.
+-- Fold one known log row by id. Model mirror rows use this after insertion so
+-- the complete admitted emission remains available without opening by default.
 UPDATE log_entries SET expanded = 0 WHERE id = $id;
 
 -- PREP: engine_render_log
--- Render-time log assembly (SPEC §packet packet.system.log).
--- Yields log_entries for the whole RUN — the conversation's working
--- memory carries across loops within a workspace's run, not just the
+-- Render-time log-section assembly ({§body-projection}).
+-- Yields log_entries for the whole worker — the conversation's working
+-- memory carries across loops within a worker, not just the
 -- current loop. Coordinate is log:///<loop_seq>/<turn_seq>/<sequence>/<op>.
 -- Status 202 entries in state='proposed' are model-invisible until resolved.
 -- `expanded = 0` rows are FOLDED — listed but collapsed to their coordinate
--- (FOLD); the renderer elides the body. §open-fold: folded rows stay listed, re-OPENable.
+-- (FOLD); the renderer elides the body. {§open-fold}: folded rows stay listed, re-OPENable.
 SELECT
     l.sequence  AS loop_seq,
     t.sequence  AS turn_seq,
     le.sequence,
-    -- le.origin is attribution, never a render filter; the worker's actor — §actor-boundary-origin-not-filter §machine-processes-worker-origin
+    -- le.origin is attribution, never a render filter; the worker's actor — {§actor-boundary-origin-not-filter} {§machine-processes-worker-origin}
     le.origin,
     le.op, le.suffix, le.signal,
     le.scheme, le.username, le.password,
     le.hostname, le.port, le.pathname,
-    le.params, le.fragment,
+    le.query, le.fragment,
     le.status_rx, le.rx, le.mimetype_rx,
     le.tx, le.mimetype_tx,
     le.state, le.outcome, le.expanded, le.source, le.attrs
 FROM log_entries le
 JOIN turns t ON t.id = le.turn_id
 JOIN loops l ON l.id = le.loop_id
--- WHERE renders exactly one worker's log — §actor-boundary-isolation §machine-processes-worker-is-its-log
--- the AND NOT clauses keep proposed (202) rows hidden until resolved (§proposal-proposed-hidden),
--- and SUCCESSFUL log-curation receipts out of the packet (#382 — recorded in the DB for forensics,
--- suppressed from materialization so a curation act rents zero packet space; the successful fold
--- that hid the task frame in run43 left NO trace, the database dig). Two receipts, one principle:
---   OPEN/FOLD — render directives, always log-targeted, suppressed on success.
---   KILL of a LOG item — a real deletion whose tombstone is spent once executed; suppressed on
---   success too (run61: a floor model KILLed log rows ~2000×, and the tombstones — bodyless, but
---   ~130-char meta-lines apiece — became 95% of its packet, a self-growing curation trap). SCOPED
---   to le.scheme='log': a KILL of a worker:// note / sh:// stream is a world mutation, NOT log
---   housekeeping, and stays visible. A FAILED op of any kind still renders — errors are signals
---   (§telemetry-uniform-error-channel).
+-- WHERE renders exactly one worker's log — {§actor-boundary-isolation} {§machine-processes-worker-is-its-log}
+-- the AND NOT clauses keep proposed (202) rows hidden until resolved ({§proposal-proposed-hidden}),
+-- and successful log-curation receipts out of the packet. OPEN/FOLD follow
+-- {§fold-open-meta-operations}; only log-target KILL follows
+-- {§kill-log-receipt-suppressed}. Every failed operation remains visible through
+-- {§operation-result-uniform-error-channel}.
 WHERE le.worker_id = $worker_id
   AND NOT (le.status_rx = 202 AND le.state = 'proposed')
   AND NOT (le.op IN ('OPEN', 'FOLD') AND le.status_rx < 400)
@@ -414,19 +489,19 @@ ORDER BY l.sequence, t.sequence, le.sequence;
 -- Default state='resolved' covers the common path (non-proposing schemes
 -- return their final status immediately). Status 202 + state='proposed'
 -- triggers the proposal lifecycle (engine pauses dispatch; client resolves
--- via loop/resolve RPC; entry transitions through engine_resolve_log_entry).
+-- via proposal resolution; entry transitions through engine_resolve_log_entry).
 INSERT INTO log_entries (
     worker_id, loop_id, turn_id, sequence, origin, source,
     op, suffix, signal,
     scheme, username, password, hostname, port,
-    pathname, params, fragment, lineMarker,
+    pathname, query, fragment, lineMarker,
     tx, mimetype_tx, rx, mimetype_rx, status_rx, tokens,
     state, outcome, attrs
 ) VALUES (
     $worker_id, $loop_id, $turn_id, $sequence, $origin, $source,
     $op, $suffix, $signal,
     $scheme, $username, $password, $hostname, $port,
-    $pathname, $params, $fragment, $lineMarker,
+    $pathname, $query, $fragment, $lineMarker,
     $tx, $mimetype_tx, $rx, $mimetype_rx, $status_rx, $tokens,
     $state, $outcome, $attrs
 )
@@ -434,18 +509,32 @@ RETURNING id;
 
 -- PREP: engine_resolve_log_entry
 -- Transitions a proposed log entry to its terminal state. Used by the
--- proposal lifecycle (loop/resolve RPC, auto resolution, timeout, abort).
+-- proposal lifecycle (client resolution, auto resolution, timeout, abort).
 -- Updates status_rx + rx + state + outcome atomically.
 UPDATE log_entries
    SET state = $state,
        outcome = $outcome,
        status_rx = $status_rx,
-       rx = $rx
+       rx = $rx,
+       deep_hash = NULL
  WHERE id = $id
    AND state = 'proposed';
 
+-- PREP: engine_log_entry_coordinate
+-- Resolve a durable operation occurrence after a proposal settles so a
+-- Problem Details instance names the same model-facing log URI as every
+-- immediately failed operation.
+SELECT l.sequence AS loop_seq,
+       t.sequence AS turn_seq,
+       le.sequence AS sequence,
+       le.op AS op
+  FROM log_entries le
+  JOIN turns t ON t.id = le.turn_id
+  JOIN loops l ON l.id = le.loop_id
+ WHERE le.id = $id;
+
 -- PREP: engine_turn_retrievals
--- §send-premature-terminate — the pending set's retrieval leg: THIS turn's READ/FIND/OPEN rows,
+-- {§send-premature-terminate} — the pending set's retrieval leg: THIS turn's READ/FIND/OPEN rows,
 -- whose results the model cannot have seen (they fold back next packet). A [200] over them is
 -- discarding answers it asked for.
 SELECT id FROM log_entries
@@ -475,36 +564,38 @@ WHERE s.worker_id = $worker_id
 LIMIT 1;
 
 -- PREP: engine_turn_failures
--- §send-200-failed-ops — THIS turn's failed op results (the model's own ops, status >= 400), whose
--- errors the model cannot have seen (they land next packet). A [200] over them concludes blind
--- past a failure — refused 409; [499] abandons regardless (declaring failure IS weighing it).
--- op='error' rows are EXCLUDED: the error CHANNEL is already-surfaced signal (the grinder's
--- overflow row mints pre-packet — the recovery turn SAW it; §grinder-hard-413-recovery's
--- concluding-is-legitimate stands), and the current emission's parse errors ride the threaded
--- count instead (they mint as rows only after dispatch).
+-- {§send-premature-terminate} — THIS turn's failed op results (the model's own ops, status >= 400), whose
+-- errors the model cannot have seen (they land next packet). A [200] or already-drained [202] over
+-- them concludes blind past a failure — refused 409; [499] abandons regardless (declaring failure
+-- IS weighing it).
+-- Actionless engine errors are excluded: the grinder's overflow row mints pre-packet, so the
+-- recovery turn already saw it ({§grinder-hard-413-recovery}). A model-authored statement that
+-- failed grammar parsing is different: source='grammar' records a bounded operation failure
+-- from the accepted emission, unseen until the next packet, so it gates completion like every
+-- other failed model operation.
 SELECT id FROM log_entries
-WHERE turn_id = $turn_id AND origin = 'model' AND status_rx >= 400 AND op != 'error';
-
--- PREP: engine_demote_turn_status
--- §send-premature-terminate — a terminal REFUSED at dispatch (the pending-set 409) demotes the
--- turn to a continue AFTER the close already persisted the provisional status (the close writes
--- packet+usage as soon as the model responds; ops dispatch after). The record must match the
--- truth the return value carries: the loop never went terminal, so the turn didn't either.
+WHERE turn_id = $turn_id
+  AND origin = 'model'
+  AND status_rx >= 400
+  AND (op != 'error' OR source = 'grammar');
+
+-- PREP: engine_reconcile_turn_status
+-- A SEND's signal is only provisional until dispatch adjudicates live obligations and failures.
+-- The close writes packet+usage before ops dispatch, so reconcile the persisted turn whenever
+-- dispatch changes that disposition (for example refused 200 -> 102 or drained 202 -> 200).
 UPDATE turns SET status = $status WHERE id = $id;
 
 
 
 -- PREP: engine_loop_sequence
--- The loop's PER-RUN sequence — the model-facing coordinate (prompt/<run>/<loop-seq>/<turn-seq>,
+-- The loop's per-worker sequence — the model-facing coordinate (prompt/<worker>/<loop-seq>/<turn-seq>,
 -- matching the log's loop-relative numbering). The raw db id leaked into prompt paths and the
 -- model's first loop read as prompt/2/1 (the docs loop holds id 1). Owner: minor but annoying.
 SELECT sequence FROM loops WHERE id = $loop_id;
 
 -- PREP: engine_worker_lineage_root
--- #522 — the PRIMARY worker: the no-parent root the lineage descends from (walk parent_worker_id
--- up). A worker with no parent returns ITSELF (the primary's own turns stamp Worker-Primary ==
--- Worker-Id). The endpoint routes primary→strong, spawned→cheap by equality, so this is the
--- root-of-tree grouping key. Always resolvable (the chain is finite + acyclic — parent != id CHECK).
+-- The no-parent root of a worker lineage; a root worker returns itself.
+-- {§worker-primary}
 WITH RECURSIVE lineage(id, parent_worker_id) AS (
     SELECT id, parent_worker_id FROM workers WHERE id = $worker_id
     UNION ALL
@@ -513,14 +604,15 @@ WITH RECURSIVE lineage(id, parent_worker_id) AS (
 SELECT id FROM lineage WHERE parent_worker_id IS NULL;
 
 -- PREP: engine_worker_has_undelivered_child_term
--- A child worker whose loop TERMINATED after the current turn's timestamp — its deliverable
--- (the §worker-scheme collect delta) is queued for the NEXT packet build and has not been seen.
--- The 1ms-wide fan-out race: workers concluding DURING the parent's generation are not "live"
--- (the wait's J leg misses them) but their results are pending — concluding or ∅-collapsing
--- over them silently discards deliverables the model spawned. {§send-undelivered-child-term}
-SELECT 1 AS pending FROM loops l
-JOIN workers r ON r.id = l.worker_id
-WHERE r.parent_worker_id = $worker_id
-  AND l.terminated_at IS NOT NULL
-  AND l.terminated_at > (SELECT timestamp FROM turns WHERE id = $turn_id)
+-- A child conclusion newer than the parent's observation cursor is complete but
+-- not delivered. This is the same durable boundary the next packet consumes,
+-- not a second timestamp race. {§send-undelivered-child-term}
+SELECT 1 AS pending
+FROM ambient_events ae
+JOIN workers child ON child.id = ae.producer_worker_id
+JOIN workers parent ON parent.id = $worker_id
+WHERE ae.workspace_id = parent.workspace_id
+  AND ae.kind = 'loop_termination'
+  AND child.parent_worker_id = parent.id
+  AND ae.id > COALESCE(parent.ambient_event_cursor, 0)
 LIMIT 1;