Context
Assessed against claude-agent-sdk-python v0.2.91 (~/pjv/anthropics/claude-agent-sdk-python) in the 2026-06-05 session, with corrections that reversed the initial verdict:
pond is a legitimate, distinctive SessionStore backend - the lossless, searchable, cross-agent-restoring one. The conformance contract's deep-equal load() is satisfiable because pond already stores every source record verbatim (options.source.raw_record + placement rule 3 catch-all); delete is OPTIONAL in the conformance suite (_OPTIONAL_METHODS), so pond's append-only stance is fully conformant. Cross-agent resume (ingest a Codex/opencode session, restore it as Claude Code JSONL, hand to materialize_resume_session) is the capability no redis/postgres/s3 backend has.
Depends on: live-write ingest (#43).
Work items
- Ordering tiebreaker (LOW). pond orders by
(timestamp, id); the contract wants append order. Real transcripts have distinct timestamps so the orders coincide; strict conformance (timestamp-less synthetic entries) needs options.source.line promoted to a sortable column. Rule: the tiebreaker must be source position, never a write-time counter (keeps re-ingest idempotent per lance-deterministic-pk / additive-sync).
- Key-model mapping (the trickiest bit, not lossy). SDK keys are
(project_key, session_id, subpath); pond models subagents as child sessions linked by parent_session_id. Define the bidirectional mapping.
- Schema fidelity check. Verify the SDK transcript_mirror entry types (
custom-title, tag, agent_metadata, summary sidecars) flow through the claude_code adapter catch-all losslessly. Expected yes (no message field -> System carrier path); prove it against fixtures.
- Python shim. Thin in-process
SessionStore class over pond HTTP (/v1/ingest + load/restore). Reference shape: the tee decorator - delegate store-of-record duties to any real backend, forward entries to pond - which passes conformance by construction and is the right first example regardless.
- Proof. Run the SDK's 14-contract conformance suite against the pond-backed store.
Positioning (decided, do not relitigate)
Both framings ship: the tee example ("mirror SDK sessions into pond for cross-session search") as the low-risk entry, and the direct backend as the differentiated one. The cross-agent resume demo is the headline artifact.
Context
Assessed against claude-agent-sdk-python v0.2.91 (
~/pjv/anthropics/claude-agent-sdk-python) in the 2026-06-05 session, with corrections that reversed the initial verdict:pond is a legitimate, distinctive SessionStore backend - the lossless, searchable, cross-agent-restoring one. The conformance contract's deep-equal
load()is satisfiable because pond already stores every source record verbatim (options.source.raw_record+ placement rule 3 catch-all);deleteis OPTIONAL in the conformance suite (_OPTIONAL_METHODS), so pond's append-only stance is fully conformant. Cross-agent resume (ingest a Codex/opencode session, restore it as Claude Code JSONL, hand tomaterialize_resume_session) is the capability no redis/postgres/s3 backend has.Depends on: live-write ingest (#43).
Work items
(timestamp, id); the contract wants append order. Real transcripts have distinct timestamps so the orders coincide; strict conformance (timestamp-less synthetic entries) needsoptions.source.linepromoted to a sortable column. Rule: the tiebreaker must be source position, never a write-time counter (keeps re-ingest idempotent per lance-deterministic-pk / additive-sync).(project_key, session_id, subpath); pond models subagents as child sessions linked byparent_session_id. Define the bidirectional mapping.custom-title,tag,agent_metadata, summary sidecars) flow through the claude_code adapter catch-all losslessly. Expected yes (nomessagefield -> System carrier path); prove it against fixtures.SessionStoreclass over pond HTTP (/v1/ingest+ load/restore). Reference shape: the tee decorator - delegate store-of-record duties to any real backend, forward entries to pond - which passes conformance by construction and is the right first example regardless.Positioning (decided, do not relitigate)
Both framings ship: the tee example ("mirror SDK sessions into pond for cross-session search") as the low-risk entry, and the direct backend as the differentiated one. The cross-agent resume demo is the headline artifact.