implementation-transcript-recovery

# Implementation Transcript Recovery ## Workflow 1. Resolve the frozen application-home and workspace identities. For a conversation, acquire the `TX-011` session directory and persist accepted semantic events incrementally; for inventory, open only an existing workspace partition without materializing it. 2. Preserve the transcript as a graph: distinguish sequential parents, logical compact parents, streamed assistant siblings, tool-result source parents, branches, and sidechains. 3. Flush and materialize files under explicit durability, security, deduplication, and remote-ingress rules. 4. Load defensively, repair legacy and parallel-tool topology, choose the intended live leaf, and restore associated snapshots and metadata. 5. Use the authoritative native-session inventory for resume, continue, fork, creation, and provider-free deletion eligibility. 6. Verify every selected attachment manifest/blob before resume, and copy immutable referenced media into a fork's destination store before publishing it. 7. Apply resume or fork ownership rules atomically before accepting a new turn, or detach and clean one deletion target through its recoverable state machine. Read [the complete transcript and recovery contract](references/transcript-recovery-contract.md) before implementing or auditing this domain. Use the [architecture diagram](assets/architecture.drawio) to inspect the event graph and recovery transformations, and the [native session-management diagram](assets/session-management.drawio) to inspect authoritative inventory, selection guards, atomic deletion detach, and cleanup recovery. For the exact shared JSON-lines parser, malformed-record recovery, and 100 MiB tail-read behavior used beneath transcript loading, read the supporting [portable JSON-lines contracts](../implementation-platform-lifecycle/references/portable-data-primitives.md#json-lines-parsing-and-tail-reads) (`PRIM-030` through `PRIM-035`); this is a supporting contract link, not an actionable skill route. ## Boundaries Own authoritative session history, native-session selection, local-store deletion state, and restore semantics. Do not persist replace-in-place progress as conversation history. Defer active-context reduction to memory/compaction, CLI/JSON presentation to the relevant surface, and deletion of remote copies or presentation caches to their owning products. ## Completion check - Preserve all `TX-*` contracts in the reference. - Round-trip normal turns, parallel tools, legacy progress, rewind branches, compact preserved segments, snips, tombstones, sidechains, non-fork resume, fork resume, missing worktrees, content-bound bounded inventory, atomic receipt publication, partial-stage recovery, and crash-retry deletion. - Confirm a resumed model sees a coherent, paired message sequence and not merely the lines that happened to be last on disk. - Confirm resume, continue, fork, and explicit creation cannot select an incomplete fork, deletion intent, detached cleanup stage, or pending deletion receipt, while a completed receipt does not reserve a genuinely new generation. - Confirm legacy text resumes without migration loss; native transcripts expose manifests only, retain referenced blobs, and fail attributably on missing or tampered durable media.