Feature: switch CC invocation from `claude -p` to `claude --resume`

[Rustam-Z]

Summary

The harness currently invokes CC in single-shot print mode (claude -p "<prompt>") — a fresh process per turn that receives the entire context as input. Switch to claude --resume so each turn continues an existing CC session instead of starting a new one.

Why

• Token cost. -p resends the full conversation history on every turn (system prompt, project prompt, prior <msg> envelopes, prior tool I/O). With --resume, CC maintains its own session, so each turn only sends the new inbound message + tool results. Especially significant for long-running chats and the daily self-reflection skill.
• Latency. Less to send / parse on each turn → faster first-token.
• Cache hit rate. --resume keeps the prompt cache warm across turns; -p rebuilds prompt-prefix every time. The Anthropic cache TTL is 5 minutes — --resume actually benefits from it, -p rarely does.
• Natural memory model. Working context lives inside CC's session, not reassembled by the harness each turn.

Trade-offs / risks

• Loss of per-turn isolation. With -p, every turn starts from a known, sanitized prompt. With --resume, prior turn's state (including any successful prompt injection) persists. Mitigations: the auto-restart-on-wedge work in Encoded prompt injections (base64, Morse) crash Claude Code sessions #39, plus an explicit "reset session" path the harness can trigger on detected compromise.
• Session lifetime management. Need a strategy for when to actually start a fresh session — daily? On /pause → /resume? On any anomaly? Worth designing alongside Encoded prompt injections (base64, Morse) crash Claude Code sessions #39 and Feature: /pause and /resume slash commands #41.
• --resume requires a session ID. Harness needs to persist the active session ID across container restarts and recover gracefully if the ID is gone.
• Tool output handling. --resume streaming format may differ from -p JSON-mode output — adapter layer needs review.

Suggested implementation steps

1. Start a CC session at harness boot with the system + project prompt. Persist the session ID to disk.
2. Per turn: invoke claude --resume <session_id> with just the new inbound <msg> envelope(s). Stream tool calls + replies as usual.
3. On wedge detection (Encoded prompt injections (base64, Morse) crash Claude Code sessions #39) or /pause+/resume (Feature: /pause and /resume slash commands #41), tear down the session and spin a new one.
4. Add a CLI knob to opt back into -p mode for debugging.

Severity

Feature, but high leverage — touches cost, latency, and the foundation that #39 / #40 / #41 build on. Worth treating as a small redesign, not a one-line patch.