Update roadmap with Phase 5 priorities from Hermes analysis

ROADMAP.md

@@ -22,49 +22,57 @@
 - [x] Ollama with 5+ models, hot-swappable per agent
 - [x] Agent rootfs — Alpine + Python IRC bot + podman + tools
 - [x] Agent manager — start/stop/list/reload long-running VMs
-- [x] Overseer — !invoke, !destroy, !list, !model, !models, !templates, !status, !help
+- [x] Overseer — !invoke, !destroy, !list, !model, !models, !templates, !persona, !status, !help
 - [x] 5 agent templates — worker, coder, researcher, quick, creative
-- [x] Agent tools — run_command, web_search (searx), save_memory
-- [x] Persistent workspace — 64 MiB ext4 as second virtio drive
-- [x] Agent memory system — MEMORY.md pattern, survives restarts
-- [x] Agent hot-reload — model/persona swap via SSH + SIGHUP
-- [x] Non-root agents — unprivileged `agent` user
-- [x] Agent-to-agent via IRC, DMs, /invite
-- [x] Overseer resilience — crash recovery, health checks, KillMode=process
-- [x] Graceful shutdown — IRC QUIT before VM kill
-- [x] Systemd service, regression tests
+- [x] Discoverable skill system — SKILL.md + run.py per tool, auto-loaded at boot
+- [x] Agent tools — run_command, web_search, fetch_url, save_memory
+- [x] Persistent workspace + memory system (MEMORY.md pattern)
+- [x] Agent hot-reload, non-root agents, agent-to-agent, DMs, /invite
+- [x] Overseer resilience, health checks, graceful shutdown, systemd
 
 ## Phase 4: Hardening & Deployment (done)
 
-- [x] Network policies per agent — full/local/none via iptables
-- [x] Thread safety — lock around IRC socket writes
-- [x] Agent health checks — 30s interval, announces deaths in #control
-- [x] Trigger matching fix — start-of-message only
-- [x] agents.json race condition fix
-- [x] Install script — one-command deployment, battle-tested on Debian + Ubuntu
-- [x] Uninstall script
-- [x] Deployed on GPU server (Xeon + Quadro P5000)
-- [x] Refactor — shared firecracker-vm.ts helpers, -43 lines
+- [x] Network policies, thread safety, trigger fix, race condition fix
+- [x] Install/uninstall scripts, deployed on Debian + Ubuntu + GPU server
+- [x] Refactor — shared firecracker-vm.ts, skill system extraction
 
 ### Remaining
-- [ ] Warm pool — pre-booted VMs from snapshots for instant spawns
+- [ ] Warm pool — pre-booted VMs from snapshots
 - [ ] Concurrent snapshot runs via network namespaces
-- [ ] Thin provisioning — device-mapper snapshots instead of full rootfs copies
+- [ ] Thin provisioning — device-mapper snapshots
 
-## Phase 5: Advanced Features
+## Phase 5: Agent Intelligence
 
-- [ ] Scheduled/cron tasks — agents that run on a timer
-- [ ] !logs command — tail agent interaction history
-- [ ] Persistent agent memory v2 — richer structure, auto-save
-- [ ] Advanced tool use — MCP servers in Firecracker VMs
-- [ ] Cost tracking — duration, model, tokens per interaction
-- [ ] Execution recording — audit trail
+Priority order by gain/complexity ratio.
 
-## Phase 6: Ideas & Experiments
+### High priority (high gain, low-medium complexity)
 
-See IDEAS.md for the full list. Highlights:
-- MCP servers as a single Firecracker VM with podman containers
-- Cron agents, webhook triggers, alert forwarding
-- Agent-written agents, agent debates, dream mode
-- Web dashboard, install script dry-run
-- Persistent agent memory with CLAUDE.md pattern (v2)
+- [ ] **Large output handling** — tool results >2K chars saved to workspace file, agent gets preview + can read the rest. Prevents context explosion. Simple, high impact.
+- [ ] **Iteration budget** — shared token/round budget across tool calls. Prevents runaway loops, especially with GPU server running faster models that chain more aggressively. Add per-template configurable limits.
+- [ ] **Skill registry as git repo** — separate git repo for community/shared skills. Clone into agent rootfs. `fireclaw skills pull` to update. Like agentskills.io but self-hosted on Gitea.
+- [ ] **Session persistence** — SQLite in workspace for conversation history. FTS5 full-text search over past sessions. Agents can search their own history.
+
+### Medium priority (medium gain, medium complexity)
+
+- [ ] **Context compression** — when conversation history exceeds threshold, LLM-summarize middle turns. Protect head (system prompt) and tail (recent messages). Keeps agents coherent in long conversations.
+- [ ] **Skill learning** — after complex multi-tool tasks, agent creates a new SKILL.md + run.py in workspace/skills. Next boot, new skill is available. Self-improving agents.
+- [ ] **Scheduled/cron agents** — template gets a `schedule` field. Overseer spawns agent on schedule, agent does its task, reports to #agents, self-destructs.
+- [ ] **!logs command** — tail agent interaction history from workspace.
+
+### Lower priority (good ideas, higher complexity or less immediate need)
+
+- [ ] **Dangerous command approval** — pattern-based detection (rm -rf, git reset, etc.) with allowlist. Agent asks for confirmation before destructive commands.
+- [ ] **Parallel tool execution** — detect independent tool calls, run concurrently. Needs safety heuristics (read-only, non-overlapping paths).
+- [ ] **Cost tracking** — Ollama returns token counts. Log per-interaction: duration, model, tokens, skill used.
+- [ ] **Execution recording** — full audit trail of all tool calls and results.
+
+## Phase 6: Infrastructure
+
+- [ ] MCP servers in Firecracker VM with podman containers
+- [ ] Webhook triggers — HTTP endpoint that spawns ephemeral agents
+- [ ] Alert forwarding — pipe system alerts into #agents
+- [ ] Web dashboard — status page for running agents
+
+## Phase 7: Ideas & Experiments
+
+See IDEAS.md for the full list.