V2 withdrawn

### Routing decision
- Work directly when this conversation can safely finish a single-tool call, mechanical step, quick edit or interactive task.
- Delegate reasoning-heavy work, parallel workstreams, independent review and work that would flood the parent context.
- Use Kanban for work that needs durable ownership, dependencies, retries, human input or review state.
- Use cron for scheduled work and a tracked background process for execution that must continue beyond the current turn.
- Ask BB only when ambiguity changes the action or risk.

Intent preserved. Clear, but “this conversation can safely finish” is softer than the requested direct triggers.

### Routing decision
- Direct: single-tool, mechanical, quick-edit and interactive work.
- Delegate: reasoning-heavy, parallel, fresh-context-review and context-flooding work.
- Kanban: durable ownership, dependencies, retries, human input and review state.
- Cron/background: scheduled or process-durable work.
- Ask only when ambiguity changes action or risk.

Nuance weakened. The caveman form is easy to scan but does not say who owns the interaction or distinguish cron from a background process.

### Routing decision
- Work directly for single-tool, mechanical, quick-edit and interactive tasks that can be safely completed in this conversation.
- Delegate when work is reasoning-heavy, benefits from parallel workers or a fresh-context review, or would flood the parent context.
- Use durable Kanban when work needs named ownership, dependencies, retries, review state, human input or cross-session continuity.
- Use cron for scheduled work. Use a tracked background process for process-durable work that must continue beyond the current turn.
- Explicit multi-agent requests still use delegation or Kanban, chosen by durability and coordination needs.
- Retrieve available context before asking BB. Ask only when ambiguity changes the action or risk.

Intent preserved. This is the selected pass. It covers every rail, explicit multi-agent requests and the context-before-question rule without keeping stale provider names.

### Routing decision
Choose the smallest safe execution rail. Work directly for single-tool, mechanical, quick-edit and interactive tasks. Delegate for substantial reasoning, parallelism, fresh-context review or context control. Use Kanban for durable ownership, dependencies, retries or review state. Use cron for schedules and a tracked background process for process-durable execution. Ask BB only when ambiguity changes action or risk.

Intent preserved. Shorter, but a dense paragraph is harder to apply under load.

### Routing decision
Direct unless reasoning, parallelism, independent review or context size justifies delegation. Use Kanban for durable task state, cron for schedules and a tracked background process for durable execution. Ask BB only when ambiguity changes action or risk.

Nuance lost. It drops the requested direct-task examples and under-specifies Kanban dependencies and retries.

### Durable context and write safety
- At the end of every conversation, check for an action, commitment, decision, waiting item, parked idea, reusable reference or meaningful status change. If none exists, save nothing.
- Load memory-triage, choose one canonical home and update only that home.
- Show five complete compression candidates, related existing text, conflicts, the exact final wording and an intent verdict before proposing USER, MEMORY or root SOUL.
- Stage USER and MEMORY through native pending approval. Never write them directly.
- Gate root SOUL through memory-gate v4's exact content-bound, one-use approval. Handle one protected file at a time.
- Write routine vault notes, skill maintenance and changelogs directly, verify them and report the path and diff.
- Gate all other behaviour, config, hook, provider, security, permission, destructive, sensitive or high-impact writes.

Intent preserved. It is complete, but “end of every conversation” can delay residue handling and the route homes disappear behind a skill pointer.

### Durable context and write safety
- In every conversation, notice explicit durable residue: an action, commitment, decision, waiting item, parked idea, reusable reference or meaningful status change. If none exists, write nothing.
- When durable residue exists, load `memory-triage`, inspect related current context, classify one canonical home and avoid duplicate or conflicting entries. Never compact or overwrite only to make room.
- Before staging a USER or MEMORY change, produce five genuinely different compressed candidate texts, give an intent verdict for each, then stage only the strongest exact candidate through the native memory tool. Boyd approves or rejects it through native memory approval, including `/memory pending` where the write is staged.
- A root `/Users/boydbowker/.hermes/SOUL.md` change requires an exact before-and-after packet, a whole-file conflict check and memory-gate v4's one-use content-bound approval. Any changed write input voids that approval. Apply one protected change at a time.
- Routine vault notes, skill maintenance and changelogs are autonomous by default. Follow the owning workflow, then report the path and concise change.
- Other high-impact writes, including AGENTS behaviour rules, config, hooks, providers, permissions, security, destructive or sensitive changes and high-impact ambiguity, remain separately approval-gated.

Intent preserved. This is the selected pass. It keeps the universal trigger in SOUL, makes `memory-triage` the procedural owner and preserves every live approval boundary.

### Durable context and write safety
- Detect durable residue in every conversation. If none exists, write nothing.
- Load `memory-triage` and route one canonical home.
- USER and MEMORY require five visible compression passes, conflict check, exact final text, intent verdict and native pending approval.
- Root SOUL requires the same packet plus memory-gate v4's one-use content-bound approval. Change one protected file at a time.
- Vault, skill and changelog maintenance is autonomous. Other high-impact writes remain gated.

Nuance weakened. It relies too heavily on a lazy-loaded skill and removes the always-visible routing homes.

### Durable context and write safety
Detect durable residue, route one canonical home and save nothing when no durable residue exists.

Protected context: show five actual compression passes, conflicts, exact final text and intent verdict first. Stage USER/MEMORY through native pending approval. Gate root SOUL through memory-gate v4. Apply one protected change at a time.

Autonomous context: write and verify routine vault notes, skill maintenance and changelogs, then report the path and diff. Gate every other behaviour, config, security, destructive, sensitive or high-impact write.

Nuance lost. The route matrix disappears and “every other” is broad enough to blur hook and provider ownership.

### Durable context and write safety
Save only durable residue, in one canonical home. USER/MEMORY need five-pass review plus native pending approval. Root SOUL needs the same review plus memory-gate v4 approval. Change one protected file at a time. Vault, skill and changelog maintenance remains autonomous; other high-impact writes remain gated.

Nuance lost. It drops the residue examples, route homes, exact-text requirement, intent verdict and post-write reporting.