@sema-agent/sdk 26.1.0 → 27.0.0

package/openapi.yaml

@@ -2929,6 +2929,7 @@ paths:
             server >= 7.135.0: `503 state.file_history_unreadable` — this session's file history cannot be read WHOLE (a boundary torn / unreadable, the tracked set unreadable); nothing is exported this time. 🔴 It is NOT "no history":
             with nothing to export the manifest still answers `200` with `fileHistory: null`. May carry `permanence` (`transient` / `terminal`; absent = the store did not judge — see `ErrorResponse.permanence`). Older servers read an
             unreadable history as "none" (manifest 200 / blob 404).
+            Server >= 7.136.0: the SAME code with `permanence: "transient"` also answers a database / store connection failure on this leg (the entry-id set, policy, anchors, session row, the history read — the code's `file_history` is its origin; the action reads `permanence`); before 7.136.0 that fell to a generic `500 internal.error`. The SDK retries this 503 up to `maxRetries` times BEFORE the first byte (the same retrier on all three pull legs), except `permanence: "terminal"` (the store judged retrying useless until it is repaired).
           content:
             application/json:
               schema: { $ref: '#/components/schemas/ErrorResponse' }
@@ -2969,6 +2970,13 @@ paths:
         '401': { $ref: '#/components/responses/Unauthorized' }
         '404': { $ref: '#/components/responses/NotFound' }
         '501': { $ref: '#/components/responses/NotImplemented' }
+        '503':
+          description: >-
+            server >= 7.136.0: `503 state.file_history_unreadable` + `permanence: "transient"` — a database / store connection failure BEFORE the response head is sent (nothing is streamed; before 7.136.0 a generic `500 internal.error`). The code's `file_history` is its origin; the action reads `permanence` (see `ErrorResponse.permanence`).
+            A disconnect AFTER the 200 head can only truncate the stream — no `{"__sync":"end"}` trailer — and is never a 5xx. The SDK retries the 503 up to `maxRetries` times BEFORE the first byte (the same retrier on all three pull legs), except `permanence: "terminal"` (the store judged retrying useless until it is repaired); a truncated body is not retried (resume with `afterSeq`).
+          content:
+            application/json:
+              schema: { $ref: '#/components/schemas/ErrorResponse' }
 
   /v1/sessions/{sessionId}/sync/history/blobs/{hash}:
     parameters:
@@ -3011,6 +3019,7 @@ paths:
             server >= 7.135.0: `503 state.file_history_unreadable` — the history cannot be read WHOLE, so the server cannot tell whether this hash is one the session references — it does NOT answer "no such blob" (that is `404 not_found.blob`) and gives no bytes. 🔴 It is NOT "no history":
             with nothing to export the manifest still answers `200` with `fileHistory: null`. May carry `permanence` (`transient` / `terminal`; absent = the store did not judge — see `ErrorResponse.permanence`). Older servers read an
             unreadable history as "none" (manifest 200 / blob 404).
+            Server >= 7.136.0: the SAME code with `permanence: "transient"` also answers a database / store connection failure on this leg (the history read or the blob bytes read — the code's `file_history` is its origin; the action reads `permanence`); before 7.136.0 that fell to a generic `500 internal.error`. The SDK retries this 503 up to `maxRetries` times BEFORE the first byte (the same retrier on all three pull legs), except `permanence: "terminal"` (the store judged retrying useless until it is repaired).
           content:
             application/json:
               schema: { $ref: '#/components/schemas/ErrorResponse' }
@@ -3375,7 +3384,10 @@ paths:
             schema: { $ref: '#/components/schemas/ApprovalDecision' }
       responses:
         '200':
-          description: Decision accepted; the run resumes.
+          description: >-
+            Decision accepted. Usually the run (or the parked background agent) resumes; on the two parked-agent forms that can never revive,
+            a DENY (server >= 7.136.0) is answered `{ taskId, status: "failed", decision: "deny" }` instead — the card is voided and nothing resumes
+            (see `ParkedDecideAccepted`, `ApprovalDecisionResult`).
           content:
             application/json:
               schema: { $ref: '#/components/schemas/ApprovalDecisionResult' }
@@ -3550,6 +3562,10 @@ paths:
             the claim and rolls back when it is gone or different) — a 7.104.0 session delete retires such cards, so this form
             mostly meets cards left by a pre-7.104.0 delete; `decide.row_unrevivable` (DecideError) now also answers when the
             parked row carries no root-session anchor (refused before the claim, nothing consumed).
+            server >= 7.136.0 — those two never-revive forms (a parked child whose row carries no agent name; a row with no root-session anchor) keep answering APPROVE with this 409 (the sentence now ends
+            by pointing at deny), but DENY answers `200 { taskId, status: "failed", decision: "deny" }`: the card is voided under its own scope CAS and leaves the pending queue and the
+            `/v1/approvals/stream` snapshot, the child ends `failed`, the gated tool never runs, no `remember` is granted; a lost void CAS is the existing 409 `decide.gate_moved`. Branch on
+            `status` / `errorCode`, not on the sentence.
           content:
             application/json:
               schema: { $ref: '#/components/schemas/ResumeLegConflictError' }
@@ -7375,10 +7391,16 @@ components:
             `egress:true` + `effect:"write"` (core's ruling); `toolAxes` is the caller's explicit override face.
             🔴 A malformed entry drops the WHOLE peer (named in a warn), never "drop the bad key and keep the
             peer": `allowSkills` is a NARROWING field whose absence means "no restriction", so a half-applied
-            entry would turn a narrowing intent into a widening. The request itself is not rejected.
-            The center/config baseline wins on a name clash (the peer name IS the tool namespace segment). Over
-            the 32-entry cap the surplus entries are DROPPED and named too — same disposition as a malformed
-            entry, not a 400.
+            entry would turn a narrowing intent into a widening.
+            🔴 server >= 7.136.0 — on a FRESH submit (the three submit doors) a badly written declaration is REFUSED, not dropped, and no longer answers a 200
+            that reads as "in effect": a list that is not an array / an entry that is not an object / a key value the judge cannot read as usable
+            (`url` not an http(s) string, `headers` values that are not strings, `allowSkills` not a string array …) ⇒ `400 request.field_invalid` with the sentence only
+            (no `field` / `reason`); a key outside the peer vocabulary (a mis-spelled `allowSkill` …) ⇒ `400 request.body_shape` + `unknownKeys` (full coordinates, e.g.
+            `a2aPeers[0].allowSkill`); a `url` / `cardUrl` that carries a credential ⇒ `400 request.field_invalid` WITH `field` / `reason` (`url_userinfo` / `url_credential_param`);
+            more than 32 entries ⇒ `400 request.field_invalid` WITH `field: "a2aPeers"` / `reason: "over_cap"` / `count` / `cap`. Branch on `errorCode` / `unknownKeys` / `field`, never on the sentence.
+            The durable-resume REPLAY leg (a stored body re-read on resume) does NOT refuse: it keeps the reader's degradation — a bad entry is not mounted and is named in an operator warn, the request is not 400.
+            Servers <= 7.135.x: a malformed entry drops the whole peer (warn), the request is not rejected, and over the cap the surplus is dropped (7.135.0 already answers 400 `over_cap`).
+            The center/config baseline wins on a name clash (the peer name IS the tool namespace segment).
           items: { $ref: '#/components/schemas/A2aServerSpec' }
         resumeAt:
           type: string
@@ -7597,7 +7619,7 @@ components:
             not exceed 5 MiB — otherwise the engine refuses BEFORE the first model call (after the UserPromptSubmit hook; not one model token spent) with 400 `config.image_invalid`, whose
             `error` starts with `TaskSpec.images: `. Before 7.117.0 such an image was sent as-is to the model provider and the whole first request failed there (no code to branch on).
             A base64 string longer than 5 242 880 characters (the model API's 5 MB image cap; the server reads the engine's exported constant) is NOT this code: the submit gate refuses it first
-            with 413 `request.payload_too_large` (zero model calls, no submit hook, no run row). The `{ url }` arm is unchanged. The model must accept image input (`ModelInfo.vision`; absent = unknown, send anyway).
+            with 413 `request.payload_too_large` (zero model calls, no submit hook, no run row). The `{ url }` arm: from server >= 7.136.0 a URL that carries a credential (`user:pass@host`, or a credential-named query / fragment parameter) is refused on a FRESH submit with 400 `request.field_invalid` + `field: "images[i].url"` + `reason` `url_userinfo` / `url_credential_param` (the sixth seat of the URL-credential judge; send the bytes inline as `{ data, mimeType }` instead — an inline image never lands in a URL); the durable-resume replay of a stored body is not judged. Otherwise unchanged (the server has long recognised the type by signature). The model must accept image input (`ModelInfo.vision`; absent = unknown, send anyway).
           items:
             oneOf:
               - type: object
@@ -13299,7 +13321,9 @@ components:
            errorCode?, errorMessage?, retriable?}`, plus `rememberApplied` when the request carried
            `remember:"session"` AND the worker has an exemption store;
          · the PARKED-BACKGROUND-AGENT shape (`parked-decide.ts`, server >=1.267) — `{taskId, status:"resuming",
-           decision}`, an ACCEPTED (not completed) receipt; the revive drives asynchronously.
+           decision}`, an ACCEPTED (not completed) receipt; the revive drives asynchronously. Since server >= 7.136.0 this
+           shape has a SECOND ARM, the VOIDED receipt `{taskId, status:"failed", decision:"deny"}` — a DENY on a parked row that can
+           never revive (no agent name / no root-session anchor) voids the card; nothing resumes (see `ParkedDecideAccepted`).
         A FOURTH shape, the #368 IDEMPOTENT REPLAY (server >= 7.4x; declared 2026-09-08 — it shipped without
         ever reaching this schema): `{sessionId, idempotent:true, decision, bindingEnforced:true,
         executionOutcome?}` — it carries NO `status` and NO `taskId` (nothing was accepted this time). So no key
@@ -13360,7 +13384,7 @@ components:
           description: >
             OPEN string — do NOT narrow. Task-level acceptance (server >= 7.37): the literal `resuming` (= accepted,
             resume in flight — follow the run stream). Terminal leg: the run's resulting status (`completed`/`failed`/
-            `suspended`/`needs_review`). Parked leg: the literal `resuming` (= accepted, revive in flight).
+            `suspended`/`needs_review`). Parked leg: the literal `resuming` (= accepted, revive in flight) — or, server >= 7.136.0, the literal `failed` on the VOIDED receipt (a deny that closed a card which could never revive; nothing resumes).
         decision:
           type: string
           enum: [approve, deny]
@@ -13581,7 +13605,14 @@ components:
           description: 'server >= 7.100.0 (S-700): WHY the park row offers no rule — the same narrow read and closed set (four words from server >= 7.129.0) as the live card frame and `card_json`.'
         boundGrantAbsence:
           allOf: [{ $ref: '#/components/schemas/BoundGrantAbsence' }]
-          description: 'server >= 7.129.0 (core 7.51.0 #1688; park-row fact eight): why this parked card has no grant-binding digest — the park-row twin seat `PendingAction.tool_approval.boundGrantAbsence`, the same narrow read and the same name / value as the live frame and the inbox row. ABSENT on this face = a skill card (it carries the digest; the digest itself is not on this face) / a row parked before engine 7.51.0 (says neither — the inbox face projects `unreadable` for the same row) / a word outside the set / an old server — NOT `not_bindable`.'
+          description: 'server >= 7.129.0 (core 7.51.0 #1688; park-row fact eight): why this parked card has no grant-binding digest — the park-row twin seat `PendingAction.tool_approval.boundGrantAbsence`, the same narrow read and the same name / value as the live frame and the inbox row. ABSENT on this face = a skill card (it carries the digest) / a row parked before engine 7.51.0 (says neither — the inbox face projects `unreadable` for the same row) / a word outside the set / an old server — NOT `not_bindable`.'
+        boundGrantDigest: { type: string, description: 'server >= 7.136.0 — a parked `Skill` card''s grant-binding digest; the same value as the inbox row''s and the live frame''s `boundGrantDigest` (one projection feeds this operator queue row, the `/v1/approvals/stream` `pending` frame and the inbox row — byte-equal for the same parked row). Present only on a skill card; absent on an ordinary card. `card_json` does not carry it. Server <= 7.135.x: not on this face.' }
+        boundGrants:
+          allOf: [{ $ref: '#/components/schemas/BoundSkillGrants' }]
+          description: 'server >= 7.136.0 — the value `boundGrantDigest` hashes, published ONLY when it hashes to it (digest only when they disagree; no digest, no list). DEPLOYMENT TEXT (skill names and list entries are written by the deployer; the server neither redacts nor escapes them): render with contextual escaping.'
+        classifierNoRuling:
+          allOf: [{ $ref: '#/components/schemas/AutoModeFallbackCause' }]
+          description: 'server >= 7.136.0 — only on a parked `classifier_no_ruling` card: why the auto-mode classifier did not rule (closed set, echoed only when a member). Same value as the inbox row''s and the live frame''s.'
         askClass:
           allOf: [{ $ref: '#/components/schemas/AskClass' }]
           description: 'server >= 7.109.0 (S-934 ② / core 7.37.0 #1214): the park-row twin seat `PendingAction.tool_approval.askClass`, read through `readAskClassPair` — the SAME narrow read as the live frame and the card. Present together with `classes` (`classes[0] === askClass`); the basis a durable approver answers by. Absent = an old row (parked before core 7.37.0; resume does not backfill) or an unreadable pair — NOT "no class". The inbox row (InboxRow) does not carry it.'
@@ -13735,7 +13766,7 @@ components:
         classifierNoRuling:
           allOf: [{ $ref: '#/components/schemas/AutoModeFallbackCause' }]
           description: 'server >= 7.118.0 — only on a parked `classifier_no_ruling` card: why the auto-mode classifier did not rule.'
-        objective: { type: ['string', 'null'], description: 'task objective, null if unavailable.' }
+        objective: { type: ['string', 'null'], description: 'task objective, null if unavailable. server >= 7.136.0: the run row''s `objectivePreview` (redacted when written, about 120 characters — the same column as `GET /v1/assistant/tasks`; looked up through the gate''s session: its active run, else its latest run; a delegated child''s gate — whose session has no run of its own — falls back to the run that spawned it, via the parked agent row''s `parentTaskId`, hop by hop, at most 8 hops) — null when no preview could be read (a deployment without a run ledger is one cause; a failed read of the row, a delegation chain that breaks or exceeds 8 hops, or a run row without a preview also answer null — the key does not say which). Before 7.136.0 this was the objective ORIGINAL stored in the parked context (not truncated, not redacted); a consumer that used it as the full objective should know that server 7.136.0 has no full-objective read face (`GET /v1/runs/:id` carries neither `objective` nor `objectivePreview`) — this key IS the preview; a full read face comes with a later server version and this package locksteps then (no substitute path is declared here).' }
         input:
           description: >
             REDACTED tool-args preview for display (null on gates without one, e.g. plan_review). Untyped by
@@ -14153,7 +14184,9 @@ components:
         non-object is `400 request.body_shape`. `title`: a string renames (1..120 CHARACTERS counted as Unicode characters = the storage column width, no leading /
         trailing whitespace, stored verbatim — not rewritten, not redacted); `null` CLEARS it back to the automatic title (the next submit lets the auto-titler fill
         it again). `""` is NOT a clear — it is `400 request.body_shape`. A string of the wrong form / a non-string-non-null `title` / a non-boolean `pinned` is
-        `400 request.field_invalid`.
+        `400 request.field_invalid`; from server >= 7.136.0 a title made only of invisible characters (zero-width, bidi controls, soft hyphen, C1 controls — empty after the cross-session peer directory's own
+        usable-name normalization) is `400 request.field_invalid` too (`title must contain at least one visible character`; the session-sync 2c import's `title` reads the same rule). Such titles written before
+        7.136.0 are not migrated.
       additionalProperties: false
       minProperties: 1
       properties:
@@ -15067,8 +15100,8 @@ components:
           type: string
           enum: [transient, terminal]
           description: >-
-            server >= 7.135.0, on `503 state.file_history_unreadable` and `422 adopt_failed` only: the file-history store's own judgment of the fault (the engine's `GovernanceRetryClass`) —
-            `transient` = nothing was changed, retrying the same shape later may succeed; `terminal` = repair the store first. ABSENT = the store did not judge (retry later once; still red ⇒ look at the store). SDK: `err.extra.permanence`.
+            server >= 7.135.0, on `503 state.file_history_unreadable` and the file-history store's fault answers (`422 adopt_failed`; server >= 7.136.0 also `restore_failed` and the like, and the session-sync pull legs' database-disconnect `503`, where `transient` is judged for connection-class faults): the file-history store's own judgment of the fault (the engine's `GovernanceRetryClass`) —
+            `transient` = nothing was changed, retrying the same shape later may succeed; `terminal` = repair the store first. ABSENT = the store did not judge; not terminal — retried by status like any retryable 5xx (the SDK does so automatically on the pull legs, up to `maxRetries`); still red ⇒ look at the store. SDK: `err.extra.permanence`.
         ordinal:
           type: integer
           description: >-
@@ -15079,7 +15112,7 @@ components:
             A per-code closed word — read it by `errorCode`, never across codes. On 400 `request.field_invalid` for a URL that carries a credential (server >= 7.135.0): `url_userinfo` (`user:pass@host`) / `url_credential_param` (a credential-named
             query / fragment parameter — the names the service's log redactor treats as credentials), beside `field` = the body coordinate (`mcpServers[i].transport.url`, `a2aPeers[i].url`, `a2aPeers[i].cardUrl`,
             `settings.hooks.<event>[g].hooks[h].url` — the FULL coordinate, not `field: "settings.hooks"` + `detail` — and `settings.webSearch.endpoint`; a `settings.webSearch.endpoint` with userinfo keeps its sentence and gains `field` / `reason`);
-            `a2aPeers` over 32 entries answers `over_cap` + `count` + `cap` on `field: "a2aPeers"`. On 503 `fleet.lease_unavailable` (server >=
+            `a2aPeers` over 32 entries answers `over_cap` + `count` + `cap` on `field: "a2aPeers"`. Server >= 7.136.0: the same URL-credential words on a SIXTH seat, `images[i].url` (`field` = `images[i].url`, an `images` `{ url }` arm — fresh submits only); `mcpServers` over 32 entries answers `over_cap` + `count` + `cap` on `field: "mcpServers"` (the same body as `a2aPeers`; the probe door `POST /v1/capabilities/mcp/probe` answers the same sentence and keys — before 7.136.0 a sentence without machine keys); a malformed `a2aPeers` list / entry / key value carries NO `reason` (sentence only), a key outside the peer vocabulary is `request.body_shape` + `unknownKeys`. On 503 `fleet.lease_unavailable` (server >=
             7.106.0 / S-866 ③): `auth-rejected` (the center refused this engine's credential, or the credential file is
             unreadable) / `unreachable` (unreachable / 5xx / timeout / bad body / any status other than 200 / 409 / 400 / 401 / 403 —
             center 404 / 429 / 3xx included). On 400 `request.field_invalid` for a `settings.permissions` bucket (server >= 7.106.0 /
@@ -15117,6 +15150,10 @@ components:
             `[i]`, record members `.key`: `mcpServers[0].transport.headers.Authorization`, `settings.hooks.PreToolUse[0].hooks[0].headers.Authorization`; the key
             name is not a secret, the value never enters the refusal); on 400 `request.field_invalid` for the request key `settings.approvalFrameFamilies` it is
             that literal (with `reason: "bad_entry"` and the bad `index`; no `index` for a non-array).
+            server >= 7.135.0: on 409 `conflict.checkpoint_ctx_unsealable` with `reason: "unrecognized_seal"` the arm of the stored credential-slot GENERATION newer than this server's (or not a positive integer) reads `field: "credentialSlotGeneration"` — not an in-body coordinate.
+            server >= 7.136.0: the credential-slot set gains four seats (generation 3) — `mcpServers[i].transport.command` / `mcpServers[i].transport.args` (stdio transport) and `settings.hooks.<event>[g].hooks[h].command` / `.args` (command hook); `args` is sealed as ONE value (the whole array); on 409 `conflict.checkpoint_ctx_unsealable` `field` names that coordinate (which doors can answer which `reason`, and how the refusal arrives, is written once — in the SDK documentation of `CHECKPOINT_CTX_UNSEALABLE_REASONS` — not repeated here).
+            server >= 7.136.0: on 400 `request.field_invalid` for a URL that carries a credential, `field` also names `images[i].url` (the sixth seat); on `mcpServers` over 32 entries `field` is `mcpServers` (beside `reason: "over_cap"` / `count` / `cap`); a key outside the `a2aPeers` entry vocabulary is `400 request.body_shape` whose `unknownKeys` carry the full coordinates (`a2aPeers[0].allowSkill`).
+            server >= 7.135.0: on `a2aPeers` over 32 entries `field` is `a2aPeers` (beside `reason: "over_cap"` / `count` / `cap`).
         detail:
           description: >
             🔴 SHAPES differ by code — branch on `field` (or the `errorCode`), never guess from the type of `detail`. On 503
@@ -18100,7 +18137,7 @@ components:
             second card word `contentKind` that server <= 7.125.x derived from the tool name (deleted at 7.126.0, no alias: a consumer reading `contentKind === "content_ask"` reads
             `approvalCard === "question"`). ABSENT = the engine minted none (a host-built ask / an older engine / a non-approval gate) — never `generic` by default. Display / triage only; it NEVER
             takes part in resume / gate / CAS. The same word rides the `livePending` row top level (= this frame's own value), the operator queue row and the `/v1/approvals/stream` `pending` /
-            `live_pending` frames; the `card_json` card face and the `approval_request` frame's `card` do NOT carry it (`ApprovalCard` is `.strict()`).
+            `live_pending` frames; from server >= 7.136.0 the `card_json` card face and the `approval_request` frame's `card` (live mint and reconnect replay alike) carry the same word too (`ApprovalCard.approvalCard`: one read, three faces — before 7.136.0 the card did not carry it, so a replayed frame had none and a reconnecting shell fell back to the generic card).
         inputHasBidi:
           type: boolean
           enum: [true]
@@ -18379,6 +18416,9 @@ components:
         argsOmitted: { type: boolean, enum: [true], description: 'Present (true) only when args exceeded the byte cap. Never false.' }
         toolCallId: { type: string, description: 'The engine''s tool-call id (the `call_…` on the assistant message''s tool_use block) — same domain as ToolApprovalFrame.toolCallId.' }
         risk: { $ref: '#/components/schemas/ApprovalRiskAxes' }
+        approvalCard:
+          allOf: [{ $ref: '#/components/schemas/ToolApprovalCardWord' }]
+          description: 'server >= 7.136.0, ADDITIVE: the engine-minted card word — the SAME word, read once, as the live frame''s top-level `ToolApprovalFrame.approvalCard`; it now rides the live-minted `approval_request` frame''s `card`, the stored `card_json` envelope and the reconnect-replay frame''s `card` (before 7.136.0 the card did not carry it, so a replayed frame had none). ABSENT = the engine minted none / an out-of-set word (the server narrows it again when it mints the card) / a card stored before 7.136.0 — never `generic` by default. Display / triage only; it NEVER takes part in resume / gate / CAS. The stored envelope is strict: a 7.135.x replica cannot read a row carrying this key (replay skips it, the idempotent re-entry goes to the durable park) — roll the fleet in one pass.'
         rememberRefused:
           type: boolean
           enum: [true]
@@ -20212,6 +20252,13 @@ components:
         boundGrantAbsence:
           allOf: [{ $ref: '#/components/schemas/BoundGrantAbsence' }]
           description: '`pending` frames only (server >= 7.129.0 / core 7.51.0 #1688) — see PendingCheckpoint.boundGrantAbsence. A `live_pending` frame carries it only inside its `frame`.'
+        boundGrantDigest: { type: string, description: '`pending` frames only (server >= 7.136.0) — see PendingCheckpoint.boundGrantDigest. A `live_pending` frame carries it only inside its `frame`.' }
+        boundGrants:
+          allOf: [{ $ref: '#/components/schemas/BoundSkillGrants' }]
+          description: '`pending` frames only (server >= 7.136.0) — see PendingCheckpoint.boundGrants. DEPLOYMENT TEXT: render with contextual escaping.'
+        classifierNoRuling:
+          allOf: [{ $ref: '#/components/schemas/AutoModeFallbackCause' }]
+          description: '`pending` frames only (server >= 7.136.0) — see PendingCheckpoint.classifierNoRuling.'
         askClass:
           allOf: [{ $ref: '#/components/schemas/AskClass' }]
           description: '`pending` frames only (server >= 7.109.0 / S-934 ②) — see PendingCheckpoint.askClass. A `live_pending` frame carries the pair only inside its `frame`.'
@@ -20222,20 +20269,35 @@ components:
           description: '`pending` frames only (server >= 7.109.0) — see PendingCheckpoint.classes.'
 
     ParkedDecideAccepted:
-      type: object
       description: >
         The 200 body of POST /v1/approvals/{sessionId}/decide when the pending checkpoint belongs to a
         PARKED background agent (server 1.267, ASSISTANT-WIRE-CONTRACT §4a parked variant; server
-        parked-decide.ts — exact literal shape). ACCEPTANCE semantics: the revive drives
-        ASYNCHRONOUSLY — 200 means accepted, not completed; a failed drive honestly re-parks the row and
-        the pending re-appears on the list. Task-level suspends keep the legacy resumed shape —
-        discriminate by `status === "resuming"`.
-      required: [taskId, status, decision]
-      additionalProperties: true
-      properties:
-        taskId: { type: string, description: 'The background agent handle (a* domain).' }
-        status: { type: string, enum: [resuming] }
-        decision: { type: string, enum: [approve, deny] }
+        parked-decide.ts — exact literal shapes). TWO ARMS, read by `status` (the SDK type `ParkedDecideAccepted` is their union):
+        (1) `status: "resuming"` — the ACCEPTANCE receipt. The revive drives ASYNCHRONOUSLY — 200 means accepted,
+        not completed; a failed drive honestly re-parks the row and the pending re-appears on the list.
+        (2) `status: "failed"` (server >= 7.136.0; `decision` is always `deny`) — the VOIDED receipt: on the two forms that can
+        never revive (a parked child whose row carries no agent name — an unnamed `Agent` call — or no root-session anchor)
+        a DENY voids the card under its own scope CAS; the card leaves the pending queue and the `/v1/approvals/stream` snapshot,
+        the child ends `failed` (the engine's reconcile), the gated tool never runs, no `remember` is granted — and NOTHING is
+        resuming. Do not wait for a revive that never comes. APPROVE on those forms stays `409 decide.row_unrevivable`; a lost
+        void CAS is the existing `409 decide.gate_moved`.
+        Task-level suspends answer their own shapes (see `ApprovalDecisionResult`): discriminate a parked receipt by `decision`
+        present and `sessionId` absent, then read `status`.
+      oneOf:
+        - type: object
+          required: [taskId, status, decision]
+          additionalProperties: true
+          properties:
+            taskId: { type: string, description: 'The background agent handle (a* domain).' }
+            status: { type: string, enum: [resuming] }
+            decision: { type: string, enum: [approve, deny] }
+        - type: object
+          required: [taskId, status, decision]
+          additionalProperties: true
+          properties:
+            taskId: { type: string, description: 'The background agent handle (a* domain).' }
+            status: { type: string, enum: [failed] }
+            decision: { type: string, enum: [deny] }
 
     FleetMeta:
       type: object

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sema-agent/sdk",
-  "version": "26.1.0",
+  "version": "27.0.0",
   "description": "Typed, zero-runtime-dependency SDK for the Sema agent fleet usage plane. The shared substrate for all doors (CC/Codex MCP façade + web). Server-side only — tokens never enter the browser.",
   "type": "module",
   "license": "BUSL-1.1",
@@ -55,6 +55,6 @@
     "prepublishOnly": "rm -f tsconfig.tsbuildinfo && npm run build && npm run typecheck --workspaces=false --prefix ../.. && npm test --workspaces=false --prefix ../.. && npm run prepack"
   },
   "devDependencies": {
-    "@sema-agent/settings-schema": "^14.4.0"
+    "@sema-agent/settings-schema": "^14.5.0"
   }
 }