Synapse v2.1.5 — "no more argv"

Patch release fixing a live-observed ingest failure on large vaults. No schema migrations.

Fixed

• CLI provider ingest could fail with "Argument list too long" on large vaults: live evidence — a ~150-file ingest run processed roughly 30 files cleanly, then every subsequent file failed the same way. Root cause: claude-agent-sdk passes a plain-string ClaudeAgentOptions.system_prompt directly as a claude CLI command-line argument, not via stdin. Synapse's block-loop system prompt (the default ingest path) folds in schema.md/purpose.md/wiki/index.md, and index.md is unbounded by design — it grows by one entry per page ever ingested — so as the vault (and thus the prompt) grows, the argument list eventually exceeds the kernel's ARG_MAX, and the spawned claude process fails to start (E2BIG) before a single message is exchanged. This explains why early files in a large ingest run succeed and later ones deterministically fail once the vault crosses the size threshold — it is not intermittent, it is monotonic with vault growth. Fixed by routing system_prompt through the SDK's SystemPromptFile option (a scoped temp file, cleaned up after the session) at all three CliAgentProvider call sites (chat, delegated ingest, block-loop completion) instead of passing it as a raw command-line argument.
• wiki/index.md's contribution to the block-loop prompt was uncapped: as a complementary hardening (independent of the CLI-transport fix above, so it also bounds prompt size/cost on the API and Local provider backends), index.md is now capped in the analysis/generation system prompt the same way the review-stage prompt already caps its equivalent section.

Upgrade notes

No schema migrations. If a prior ingest run failed with "Argument list too long" (or the underlying claude-agent-sdk subprocess failed to start) on a large vault, retry it — the system prompt no longer travels through the command line, so vault size no longer risks tripping the OS argument-list limit.