DeepSeek Harness compared to our session event catalog
Part of Session Store Research. Produced by running RESEARCH_PROMPT_COMPARISON. Stage-one dossier: DeepSeek Harness. Compared against proto/trogonai/session/sessions/v1alpha1/ and ADR#0035 on 2026-08-20. Upstream is pinned to commit 141eb6fef83422698aef7a981029e843e8161534, tagged dsh-v0.1.0-rc.8.
Store maturity: 5/12 - evolution scars 1/3 (the logical format remains version 0, limited old records are normalized in memory, SQLite is already at physical schema 17, and incompatible schemas are rejected rather than migrated: types.ts:33-56, README.md:38-40, schema.ts:17-20, schema.ts:107-149); operational age 0/3 (the persistence seam first landed on 2026-06-15 in df4b7d3, only 66 days before this retrieval, entirely within the prerelease window); exposure 1/3 (it is officially distributed through npx, but remains a compatibility-breaking developer preview and its first-party stores are local, single-owner designs: README.md:9-23, session-persistence-jsonl/README.md:70-77); design independence 3/3 (inference from repository history: the persistence code was introduced as first-party DeepSeek Harness code in df4b7d3, with JSONL and later SQLite implementations, not inherited from a store fork).
Because this score is under 6, a recommendation supported only by DeepSeek Harness is labelled thin evidence. It is not treated as an industry norm.
The one structural difference everything else follows from
DeepSeek Harness has two commit points. Session.append() synchronously commits a fact to the process-local live session, then observers enqueue it for write-behind persistence. A persistence append() acknowledges only after the batch is durable, but failure there does not undo the already-acknowledged live fact. packages/core/session/src/index.ts:569-655packages/session/session-persistence/src/write-behind.ts:18-56
Our Session has one authoritative commit point: the durable JetStream append to one session subject. ADR#0035 decision 2 classifies each command as NoStream, At(current_position), or Any; success means the selected durable write completed. This difference explains the rest:
- DeepSeek sequence allocation and live exclusion are process-local first, with backend tail checks as a second defense. Our optimistic concurrency is at the authoritative append boundary.
- DeepSeek repairs a cold durable log to catch it up to semantic state after a process crash. Our log already contains every acknowledged fact, so recovery is a reconciler over durable history rather than a second commit domain.
- DeepSeek can use a mutable in-memory surface and copy fork prefixes because a
Sessionobject owns the immediate view. Our projections must be deterministic over immutable cross-stream facts.
DeepSeek narrows the risk of its two commit points with a separate checkpoint policy. First-party persisted runtimes flush before model dispatch, before a top-level tool body may cause an external side effect, and before the next step. The plugin fails those boundaries closed, but a deployment may omit it. packages/session/session-checkpoint-policy/README.md:5-23 Our durable operation ordering is structural instead: OperationReserved records operation_id, request_digest, and operation_kind before a guarded side effect, at the same authoritative append boundary used by every command.
Mapping
The stage-one field table is the field-level source of record: it transcribes each event's complete data payload and expands aliased nested values immediately below the table. The rows here map every one of those event types and call out the payload fields that do not map together. Physical SQLite packing adds no logical facts. docs/persistence-catalog.md:1-20
Header, envelope, and store mechanics
| DeepSeek Harness persisted fact | Our equivalent | Verdict |
|---|---|---|
Header version | Concrete protobuf type name plus v1alpha1 package policy in proto/trogonai/session/sessions/v1alpha1/events.proto | Semantic mismatch. We have no per-session format integer; promotion requires a written compatibility policy. |
Header id | session_id and the session subject | Equivalent opaque identity. |
Header createdAt | No payload equivalent in proto/trogonai/session/sessions/v1alpha1/session_started.proto | Gap. Stream append time is not the same occurrence fact. |
Header cwd | SessionStarted.checkout_ref as CheckoutRef | Ours is typed and portable rather than a raw local path. |
Header parentSession | ParentLinked.parent_session_id for delegation, or SessionForked.source_session_id for branching | Semantic mismatch. DeepSeek overloads one header field; we distinguish delegation from fork ancestry. |
Header seedLength | SessionForked.context_prefix_boundary | Semantic mismatch. DeepSeek counts copied child events; ours addresses an immutable source prefix by reference. |
Header origin | Presence of ParentLinked | Ours derives the classification from a durable relationship. |
Header delegationDepth | Projection over ParentLinked ancestry | DeepSeek stores a denormalized depth; ours can derive it and avoids drift. |
Header agentPreset | SessionStarted.execution_plan.plan_bytes and plan_digest | Ours stores the complete immutable execution plan, not a preset label. |
Event type, data | One concrete protobuf event type and its typed fields | Ours is closed and schema-validated; DeepSeek is declaration-merge extensible. |
Event seq | SessionOrdinal.value in proto/trogonai/session/sessions/v1alpha1/session_ordinal.proto, fold-derived and 1-indexed | Equivalent ordering, different base and ownership. DeepSeek has no independent event id. |
Event time | Append metadata, plus typed occurrence timestamps only where event time matters | Ours deliberately does not put a generic wall clock in every payload. |
Event ignorable | No direct equivalent | DeepSeek lets unknown informational events be skipped. Our concrete event type and compatibility policy fail closed at the storage boundary. |
Event sourceEventSeqs | Domain-specific joins and digests, such as Compacted.covered_input_digest | Ours rejects a generic provenance list in favor of typed evidence. |
Event surfaceOp | No generic equivalent. Compacted, SessionRewound, and RedactionApplied alter effective model-visible history through separate typed semantics. | Semantic mismatch. DeepSeek has a general append-or-replace surface algebra; ours keeps compaction, rewind, and privacy masking distinct. |
SQLite incarnation, revision, and JSONL source-qualified revision | JetStream stream identity and current position | Both are backend freshness values, not session facts. DeepSeek revisions protect prepare/inspect freshness, not command-level CAS. |
| Checkpoint-policy flush before model dispatch, top-level tool side effects, and the next step | Durable fact append, especially OperationReserved before a side effect | Same ordering goal, different enforcement. DeepSeek policy is an optional plugin over write-behind; ours makes the reservation an aggregate fact. checkpoint-policy/README.md:5-23 |
Projection cache domain version, unit stateVersion, watermark, and payload | Rebuildable projection state outside the Session stream | Equivalent. Both treat projection state as disposable. session-projection-cache/src/spec.ts:16-69 |
| FTS query rows | External read model | Equivalent. Neither makes search index state authoritative. |
Event payload catalog
| DeepSeek Harness event type | Our equivalent | Verdict |
|---|---|---|
agent/inbox/spliced {target, start, removedCount?, inserted, outcome?} | No equivalent | Every payload field belongs to a mutable pending-inbox splice; our stream records admitted messages instead. catalog:101-122 |
agent-preset/selected {agentPreset} | Immutable StoredSessionExecutionPlan in SessionStarted | Semantic mismatch: changing execution identity means a new session or fork for us. catalog:124-140 |
approval/asked {id, toolName, callId?, reason?} and approval/decided {id, outcome} | No direct equivalent. ToolCallApproved and ToolCallDenied overlap only approved and denied tool outcomes. | DeepSeek owns a separate ApprovalRequestId audit lifecycle whose outcomes also include cancelled and unavailable. ToolCallRequested is model intent, not an approval request. catalog:142-183 |
approval/policy {policy, source?} | No typed Session event; at most an opaque policy/configuration binding inside the immutable plan | Semantic mismatch. DeepSeek can change an explicit approval-policy snapshot in one session and records delegation as the source. catalog:185-207 |
assistant/chunk {turn, step, chunk} and assistant/message {turn, step, message, usage?, interrupted?} | turn_id, AssistantMessageStarted, AssistantMessageCompleted, AssistantMessageFailed, and TokenUsage; no step number or token delta | Completed message and usage map. DeepSeek can retain a provider-visible partial message and mark it interrupted in one event; our mutually exclusive completed or failed terminal facts cannot express both, so partial interrupted content is a gap. Ours deliberately omits raw chunks. catalog:209-244 |
command/run {commandId, name, args?, source} and command/done {commandId, kind, text?, sourceEventSeq?} | No direct equivalent | These are slash-command handler audit facts, not tool execution. Mapping them to ToolCallRequested would invent model intent. catalog:246-287 |
compaction/start {compactionId, sourceCommandId?, turn}, summary {summary, shadowedRange, shadowedSeqs, shadowedTokenCount, provider, model, maxTokens?, usage?, rawOutput?, llmStreamCall?}, prune {shadowedRange, shadowedSeqs, shadowedTokenCount}, and end {compactionId, sourceCommandId?, turn, error?} | Compacted maps summary, inclusive range, model, and usage; summary_id, trigger, context_root, producer, and covered_input_digest are ours only | DeepSeek records the procedure and mutable replacement provenance. We record one validated result and do not retain command id, raw output, or a start/end bracket. catalog:289-398 |
feedback/record {text} | No typed equivalent | Gap. SystemNoticeRecorded would preserve text but lose feedback semantics. catalog:400-414 |
goal/change with version, operation, goal snapshot and timestamps, or clear tombstone | No Session event equivalent | Deliberate boundary until goal ownership is assigned; none of its durable goal identity, revision, round, or tombstone fields map. goal/domain.ts:13-44 |
hook/invoked {turn, point, dialect, matcher?, handlerId} and hook/result {turn, point, handlerId, decision, exitCode?, stderrSummary?, durationMs} | Only turn_id correlates; no typed hook lifecycle | Gap if hook identity, policy decision, exit, stderr, and duration must be session audit facts. catalog:431-479 |
llm/retry {retryId, turn, step, provider, mode, policyKey, retry, maxRetries?, delayMs, failure} and llm/retry-started {retryId, turn, step, retry} | turn_id and assistant failure overlap; no model-request retry event | Execution attempts are harness attempts, not provider retries. Retry identity, policy, attempt count, delay, and transition are gaps, but this store is not evidence for copying their exact schema. llm-retry/types.ts:15-48 |
permission/preset {preset}, plan/mode {active}, sandbox/mode {mode, source?} | No typed Session events; the immutable execution plan and configuration may bind related policy | Deliberate semantic mismatch: DeepSeek mutates explicit runtime modes inside one session; our catalog fixes plan identity without naming equivalent mutable fields. catalog:505-538 catalog:570-591 |
request/context {provider, model, contextWindow?} and request/header {header, reason}, where header has config, adapterDefaults?, system?, tools? | The immutable SessionExecutionPlan maps the resolved provider route, protocol, driver, connection, and non-secret binding; AssistantMessageStarted maps model and settings | Gaps are advertised context window, exact rendered prompt, effective adapter defaults, tool schemas, and initial/resume/change reason. types.ts:196-228 |
schedule/change version-1 create, delete, or dispatch over {id, kind, prompt, scheduledAt} plus after/every interval and dispatch time | No Session event equivalent | Deliberate boundary: none of the schedule rule, identity, or dispatch fields are part of this aggregate today. schedule/types.ts:9-105 |
session/end-seed {} | No general equivalent. SessionForked records the fork case only. | DeepSeek marks the end of a constructor seed for resume, fork, or replay. Our resume and replay are reads and append no lifecycle fact; fork is explicit atomic creation. types.ts:315-336 |
session/title {title, messageSeqs, source} and session/title-llm-request {titleProvider, messageSeqs, route, system, messages, maxTokens} | SessionRenamed maps the final title only | DeepSeek also retains source-message provenance and the complete derived title request; ours has no title-generation request audit. catalog:643-672 |
step/start {turn, step}, step/end {turn, step}, turn/start {turn}, turn/end {turn, reason} | turn_id on conversation and tool events; terminal message and tool facts | Turn identity maps, but our correlation does not store numeric step boundaries or one aggregate turn reason. catalog:674-696 catalog:943-979 |
subagent/descriptor {version, mode, provider, label?, agentProvider?, agentModel?, persona?, toolFilter?} | Child SessionStarted.execution_plan maps composition and ParentLinked maps lineage; mode and label have no exact field | Ours separates executable identity from lineage. DeepSeek preserves one-shot or continuable mode and an enumeration label but has no parent-side dispatch fact. descriptor.ts:41-88 |
team/member {version, teamId, member}, team/task {version, teamId, task}, team/message/queued {version, teamId, message}, team/message/delivered {version, teamId, messageId, targetId} | No Session event equivalent | The nested snapshots retain member identity, provider, context and phase; task revision, ownership, dependencies and write scopes; and mailbox sender, target, delivery and content. These are a deliberate team aggregate boundary for us. agent-team/types.ts:43-116 agent-team/types.ts:203-218 |
todo/write {todos[{content, status}]} | TodoUpdated.items[{id, content, status}] plus revision | Equivalent whole-list snapshot; ours adds stable item ids and a monotonic revision. catalog:776-789 |
tool/call {turn, step, callId, name, arguments} and tool/result {turn, step, message, error?, meta?} | ToolCallRequested, ToolCallStarted, ToolCallCompleted, ToolCallFailed | Call identity, name, raw arguments, content, and error map. Ours adds execution id, approval, application-error status, duration, termination, artifacts, observed resources, and operation id; DeepSeek adds opaque tool presentation metadata and step number. catalog:791-806 catalog:856-883 |
tool/code-dispatch-start {rootCallId, parentCallId, subCallId, name, arguments} and tool/code-dispatch with the same ids plus isError, content | ToolCallRequested.parent_tool_use_id plus nested tool lifecycle | Parent-child call correlation maps; DeepSeek also stores root identity and a specialized combined result. catalog:808-854 |
tool-workflow/run-start {runId, name}, agent-start {runId, seq, label, phase?, childId}, agent-end {runId, seq, outcome}, run-end {runId, stopReason} | Operation ledger and delegation facts map child dispatch and outcomes, but no workflow-run aggregate | Run identity, step sequence, label, phase, and aggregate stop reason have no direct equivalent. catalog:885-941 |
user/message with id, role, content[], source | UserMessageRecorded.message maps identity, role, and typed content; turn_id is ours only | DeepSeek's source attribution has no field in CanonicalMessage; ours adds message created_at and explicit turn correlation. catalog:981-998 |
web/deepseek-search-llm-request {endpoint, apiVersion, body{model, max_tokens, messages, tools}} | No provider-specific request event | Every provider request field is absent today; a generic request-envelope artifact would cover it without adding provider-specific schema. catalog:1000-1007 |
Facts only our catalog records
| Our facts | DeepSeek Harness position | Assessment |
|---|---|---|
SessionClosed, SessionCancelled, SessionFailed, SessionHidden, archive and unarchive | No durable session lifecycle or hide/archive API | Our lifecycle is materially stronger. |
SessionForked and SessionRewound | Fork is copied header state; no rewind event or active-history pointer | Our history control is explicit and append-only. |
ExecutionAttemptStarted, ExecutionAttemptReady, ExecutionAttemptEnded, CheckpointProduced | Process resume and projection checkpoints exist, but no durable execution-attempt aggregate or attested checkpoint | Our execution evidence is materially stronger. |
OperationReserved, OperationCancellationRequested, OperationOutcomeRecorded | No general side-effect ledger | Our crash reconciliation does not infer operation identity from tool prose. |
DelegationDispatched, ParentLinked, termination, history invalidation, and two-sided detach facts | One child header points at one parent; orderly process-local teardown drains descendants outside the durable store | Our durable lineage and cascade semantics are materially stronger. |
RedactionApplied, ArtifactErased | No retention, redaction, or deletion API | Our decision 7 defines read-time privacy masks while retaining audit history. |
ArtifactRecorded, FileChanged, ResourceObservation | Tool-private payloads can preserve some evidence, but no shared typed artifact and file ledger | Our event sizes, integrity checks, and erasure joins are stronger. |
ExternalDelegationDispatched | A non-session-backed provider may leave no local child record | Our external delegation still records authorization and request evidence. |
Semantic mismatches that must not be mapped by name alone
Append. DeepSeek's live append means accepted in memory; its persistence append means durable. Our append acknowledgement is the authoritative durable commit. A bridge must select one DeepSeek commit point explicitly.
Fork. DeepSeek's seedLength addresses physically copied child events. Our SessionForked.context_prefix_boundary addresses source-stream history that the child inherits by immutable reference. Reusing the integer without the source id changes its meaning.
Compaction. DeepSeek's summary becomes active through a later generic surfaceOp.replace. Our Compacted is itself a self-sufficient fact with exact covers_from, covers_through, context_root, producer, and covered_input_digest. A converter cannot map only the summary text.
Revision and OCC. DeepSeek revisions invalidate stale prepared objects, and SQLite checks the durable tail inside a transaction. Neither is a caller-visible expected-version command precondition. Our At(current_position) is. packages/session/session-persistence/README.md:46-59
Projection checkpoint. DeepSeek projection checkpoints cache fold state. Our Checkpoint is an attested, digest-bound execution checkpoint that may be restored by a later execution attempt. The names are not interchangeable.
What we should consider changing
1. Persist evidence of the exact rendered model request
- Change: Add an optional
ArtifactRef request_envelope_reftoproto/trogonai/session/sessions/v1alpha1/assistant_message_started.proto, with canonical bytes for the rendered system prompt, tool schemas, effective adapter defaults, and other request-scoped envelope values not already bound by the immutable plan or the event. Keepmodelandsettingsqueryable in the event and resolve route identity from the session's stored plan. - Evidence: DeepSeek Harness, 5/12, persists those exact values in
request/header.packages/core/session/src/types.ts:196-210 - Blast radius: Additive.
- Judgment: Consider it. This is thin evidence from DeepSeek alone, but it identifies a real reproducibility gap: plan bytes plus model settings do not prove which prompt and tool schemas were actually rendered for a call.
- Cost: One artifact write and digest verification per distinct request envelope, privacy filtering for prompt content, artifact retention policy, and a new failure mode when envelope storage succeeds or fails separately from model dispatch.
2. Add an assistant process-loss reason and safe crash-closure contract
- Change: Add
ASSISTANT_MESSAGE_FAILURE_REASON_PROCESS_LOSTtoAssistantMessageFailureReasoninassistant_message_failed.proto, then add a recovery row to the ADR command matrix and backend conformance tests. A reconciler should preserve every valid fact and may close an orphaned assistant generation with that reason. It must not map an unmatched recorded tool call toToolCallFailed: when a reserved tool operation has no known outcome, appendOperationOutcomeRecorded.unknownand reconcile it before claiming success or failure. Never truncate a complete fact. The current assistantINTERRUPTEDvalue means user steering, soERRORis the only honest existing assistant value until the new reason exists. - Evidence: DeepSeek Harness, 5/12, drops only a torn final fragment and appends semantic closure for a cold interrupted log. Its repair distinguishes an assistant-declared call with no durable
tool/callasTOOL_NOT_STARTEDfrom a durabletool/callwith no result whose side effect may have occurred, rendering the latter asTOOL_OUTCOME_UNKNOWNrather than declaring failure.packages/session/session-persistence/README.md:25-30packages/core/session/src/repair.ts:89-123 - Blast radius: Additive, with no migration: one protobuf enum value plus new recovery-command, fold, projection, and operation-reconciliation behavior.
- Judgment: Do it. The product evidence is thin, but the rule follows our existing first-terminal-outcome fold and removes ambiguity from recovery.
- Cost: One enum value, a recovery detector, idempotency tests across repeated repair, and an operation reconciliation path that may leave the model-facing tool lifecycle open until the side effect is classified.
3. Require one OCC conformance suite for every store implementation
- Change: Make ADR#0035 decision 2 acceptance criteria require each backend to pass the same
NoStream,At, andAnyrace tests, including first creation, stale invariant transition, commuting terminal facts, and cross-process writers. - Evidence: DeepSeek Harness, 5/12, shares a coordinator but still documents different concurrency limits: JSONL permits one live writer, while SQLite uses
BEGIN IMMEDIATEand tail checks without an application lease.session-persistence-jsonl/README.md:70-77session-persistence-sqlite/src/store.ts:173-239 - Blast radius: Additive tests and release criteria.
- Judgment: Do it. DeepSeek alone is thin evidence, but it reinforces the stronger cross-product conclusion already recorded in the stage-two synthesis.
- Cost: A reusable adversarial harness, at least two backend instances in each test, deterministic race orchestration, and blocking any backend whose advertised interface is stronger than its actual atomicity.
4. Do not replace reference forks with copied prefixes
- Change: Reject changing ADR#0035 decision 5 or
proto/trogonai/session/sessions/v1alpha1/session_forked.prototo store the source prefix again on the child stream. - Evidence: DeepSeek Harness, 5/12, copies an inclusive source prefix into the child's seed and records
parentSessionplusseedLength.packages/core/session/src/index.ts:1067-1138 - Blast radius: Breaking the decision, not the schema: decision 5, if adopted. No change is needed to retain the current design.
- Judgment: Do not do it. The copied prefix makes each child self-contained, but duplicates O(history) facts and makes privacy and artifact reachability harder to reason about. Low maturity does not justify reversing our decision.
- Cost: Keeping the current design requires a cross-stream context projection and source-prefix availability for the lifetime of the fork.
5. Do not generalize Compacted into a mutable surface operation
- Change: Reject a generic
surfaceOp { append | replace(start, end) }on conversation events. KeepCompactedinproto/trogonai/session/sessions/v1alpha1/compacted.protoas the only compaction operation and retain its digest, producer, context root, and inclusive ordinal range. - Evidence: DeepSeek Harness, 5/12, makes the model-visible summary active by appending a
user/messagewith a range replacement after the summary event.packages/compaction/compaction-basic/src/region.ts:152-254packages/compaction/compaction-basic/src/region.ts:426-477 - Blast radius: Breaking the decision, not the schema: decision 4, if adopted. No change is needed to retain the current design.
- Judgment: Do not do it. DeepSeek's surface algebra is compact, but a generic replacement can become valid without proving which masked input was summarized or which plan and attempt produced it.
- Cost: Our typed marker is larger and makes every compaction recalculate
covered_input_digest; its projection is less generic than DeepSeek's surface fold.
What our design already does better
One durable authority. A successful command cannot exist only in process memory. NoStream, At, and Any express command intent at the same boundary that assigns order.
Reference fork with explicit meaning. SessionForked carries source_session_id, inclusive context_prefix_boundary, and reason; the child folds only its own stream. DeepSeek stores a complete copied log and a coarse header boundary.
Self-sufficient compaction. Compacted stores the summary, exact coverage, trigger, context root, producing attempt and plan digest, and covered-input digest. DeepSeek's replacement records range and source sequence provenance but does not bind the active surface replacement to equivalent privacy-masked input or execution-plan evidence.
Two-sided lineage and cascade. Parent DelegationDispatched and child ParentLinked join by operation_id, then typed parent termination, history invalidation, and detach facts preserve why a child stopped or became independent. DeepSeek has one child-to-parent header pointer.
Privacy lifecycle on a keep-forever stream. SessionHidden, RedactionApplied, and ArtifactErased define read-time masking and content erasure without rewriting event history. DeepSeek has no delete or retention operation at all.
Typed recovery and operation evidence. Execution attempts, checkpoints, and the operation ledger let recovery distinguish a retry from a new side effect. DeepSeek repairs conversation shape well, but does not provide a general side-effect reservation and reconciliation protocol.
Trade-offs, not gaps
Live responsiveness versus one commit boundary. DeepSeek can update its UI and model-visible surface without waiting for storage. It pays with an acknowledged-live versus durable distinction. We pay durable append latency for one unambiguous authority.
Copied fork versus reference fork. A DeepSeek child can be read after its parent artifact disappears. Our child stays small and preserves one copy of the facts, but the context projection must resolve the source prefix.
Raw chunks versus coarse facts. DeepSeek can replay token-level generation. We avoid making high-volume deltas durable and accept that exact streaming animation is not reconstructable from the Session stream.
Mutable runtime modes versus immutable execution identity. DeepSeek can change presets, permissions, plan mode, and sandbox mode inside one session. We make an execution plan immutable, which improves audit and authorization at the cost of requiring a new session or fork for a materially different runtime.
What not to copy
- Do not present process-local acceptance as durable success.
- Do not use copied history as the persisted meaning of fork.
- Do not let a generic surface replacement bypass compaction provenance and digest checks.
- Do not treat a schema-version integer as a migration system. DeepSeek's version
0and SQLite schema17currently reject incompatibility. - Do not mistake best-effort process-local descendant drain for durable cascade, or leave retention to out-of-band file or row deletion.
- Do not expose one pluggable store interface while allowing its backends to disagree silently on concurrency guarantees.
The two gaps the industry has not closed
Subagent cascade
DeepSeek Harness does not challenge ADR#0035 decision 6. A session-backed child is an independent session with parentSession in its header, and child discovery follows that pointer. During orderly process-local shutdown, SubagentRuntime closes new admission, cancels continuable descendants top-down, releases them child-first, and attempts a final flush. That is real runtime cascade, but it appends no durable parent disposition and does not delete child artifacts. A hard crash skips it. The persistence API has no delete cascade, parent-termination, rewind-invalidation, or detach operation; out-of-band parent removal can therefore still leave an orphan. packages/subagent/subagent/src/child-agent.ts:85-120packages/subagent/subagent/src/index.ts:294-325packages/subagent/subagent/src/continuation.ts:746-841packages/subagent/subagent/src/continuation.ts:1332-1394packages/session/session-persistence/src/index.ts:78-240
Result: DeepSeek validates top-down cancel plus child-first release as an orderly runtime policy, but decision 6 remains necessary for durable recovery. Its answer is worse after hard crash, delete, or rewind, and has no analog to history invalidation.
Retention on an unbounded log
DeepSeek Harness separates transcript compaction from storage compaction, which supports one premise of ADR#0035 decision 7: model-visible replacement should not silently delete audit history. It offers no answer to growth, however. JSONL files and SQLite rows accumulate; there is no delete, retention, or background historical compaction API. session-persistence-jsonl/README.md:70-77session-persistence-sqlite/README.md:55-63
Result: the product validates the distinction between context compaction and physical retention, but neither validates nor challenges our read-time redaction and artifact-erasure contract. Its unbounded-growth answer is worse.
Open questions for the ADR
- Must
AssistantMessageStartedcommit to the exact rendered model request, or is immutable plan identity plusmodelandsettingsintentionally enough? - Which durable fact proves that a recovery reconciler, rather than a live writer, closed an interrupted message or tool call?
- What exact compatibility artifact is required when
v1alpha1becomesv1: a per-stream format marker, a type-registry version, explicit migration events, or only stable concrete type names and readers? - Must every future store backend support cross-process writers, or may a local backend advertise a weaker capability while still implementing the same interface?
