v0.5.2 — repo memory

The agent now remembers your project across sessions — at zero token cost. Upgrade: pipx upgrade meshapi-code (or just say y when the CLI offers it).

🧠 Repo memory

• Every file the agent writes or reads is structurally mapped (symbols, sizes) into ~/.meshapi/context/ — captured from content already in hand, so it costs zero extra tokens, and it never touches your repo.
• The next session in the same directory starts warm: a compact repo map + your remembered notes ride the system prompt. Ask "what do you know about this repo?" and it answers before reading a single file.
• remember tool — the model persists durable decisions ("uses pnpm", "tests run with pytest -q") for future sessions.
• Read-dedupe — re-reading an unchanged file returns a one-line "already in your context" pointer instead of the full body. Provably safe: sha256-verified against disk, correct against the /optimize pruning lever at any dial, and an immediate re-ask always returns the real body.
• /memory inspect · /memory notes · /memory clear · /memory off.

Also in this release

• web_search fix: prod returns result text in content — it was being silently dropped; the model now sees the actual results (verified live).
• Verified live against the production gateway: /route preview (/v1/router/select) and web_search (/v1/web/search) both work.
• Slash commands with leading whitespace ( /memory clear) now work.
• Server records whose process died out-of-band no longer linger in the toolbar.
• README now has a Features in action section — every feature with real terminal captures.

Full changelog: v0.5.1...v0.5.2

🤖 Generated with Claude Code