Overhaul agent quality — prompts, tools, config, compression

- Rewrite system prompt: structured sections, explicit tool descriptions
  with full SKILL.md descriptions, multi-agent awareness
- Add write_file skill for creating/modifying workspace files
- Per-template config passthrough: temperature, num_predict, context_size,
  compress settings, max_tool_rounds, max_response_lines
- Bump defaults: 1024 output tokens (was 512), 500-char deque (was 200),
  250-token summaries (was 150), compress threshold 16 (was 12), keep 8 (was 4)
- Cache compression by content hash — no redundant summarization
- Update all 5 templates with tuned settings per role

TODO.md

@@ -13,18 +13,20 @@
 - [x] Network policies, thread safety, trigger fix, race condition fix
 - [x] Install/uninstall scripts, deployed on 2 machines
 - [x] Refactor: firecracker-vm.ts shared helpers, skill extraction
+- [x] Large output handling — save >2K results to file, preview + read_file skill
+- [x] Session persistence — SQLite + FTS5, conversation history survives restarts
+- [x] !logs — tail agent history from workspace
+- [x] Context compression — cached summaries, configurable threshold/keep
+- [x] write_file skill — agents can create and modify workspace files
+- [x] Structured system prompt — explicit tool descriptions, multi-agent awareness
+- [x] Per-template config — temperature, num_predict, context_size, compress settings
+- [x] Response quality — 500-char deque storage, 1024 default output tokens, 250-token summaries
+- [x] update.sh script — one-command rootfs patching and snapshot rebuild
 
 ## Next up (Phase 5 — by priority)
 
-### Quick wins
-- [ ] Large output handling — save >2K results to file, preview + read_file
-- [ ] Iteration budget — configurable max rounds per template, prevent runaway loops
-
 ### Medium effort
 - [ ] Skill registry git repo — shared skills on Gitea, `fireclaw skills pull`
-- [ ] Session persistence — SQLite + FTS5 in workspace
-- [ ] Context compression — summarize old turns when context gets long
-- [ ] !logs — tail agent history from workspace
 
 ### Bigger items
 - [ ] Skill learning — agents create new skills from experience
@@ -34,7 +36,6 @@
 
 ## Polish
 
-- [ ] Agent-to-agent response quality — 7B models parrot, needs better prompting or larger models
 - [ ] Cost tracking per interaction
 - [ ] Execution recording / audit trail
 - [ ] Update regression tests for skill system + channel layout