| name | bob-protocol |
| description | Multi-persona coordination protocol. Enables AI to switch between specialized personas (Neo, Morpheus, Trin, Oracle, Mouse, Cypher, Bob, Smith) based on task needs. Use for *chat workflow, state management, and cross-agent communication. |
| triggers | ["*chat"] |
| requires | ["chat","bloop","sprint","make"] |
Orchestrates multi-persona AI coordination through a shared chat log using the *chat trigger.
TLDR:
Routes *chat messages to the right specialist persona — by explicit @mention or auto-selection.
Each persona loads state on entry, executes one task, saves state on exit, posts to agents/CHAT.md.
For full workflow chains use loop commands (*fix, *impl, *qa, *review, *plan sprint) — see bloop skill.
Key rule: no third fix attempt without Oracle consult + user sign-off.
Bob Protocol — Multi-Persona Coordination
Overview
One AI dynamically switches between specialized personas. All coordination happens through agents/CHAT.md. On cold start, each persona reads their state files to resume where they left off.
Available Personas
Each persona is defined in agents/<name>.docs/SKILL.md:
| Persona | Role | Prefix | Use When |
|---|
| Neo | Senior SWE | *swe | Implementation, coding, debugging |
| Morpheus | Tech Lead | *lead | Architecture, design decisions |
| Trin | QA Guardian | *qa | Testing, code review |
| Oracle | Knowledge Officer | *ora | Documentation, knowledge queries |
| Mouse | Scrum Master | *sm | Sprint tracking, coordination |
| Cypher | Product Manager | *pm | Requirements, user stories |
| Bob | Prompt Engineer | *prompt | Agent creation, process improvement |
| Smith | HCI Expert | *user | UX review, usability testing, sprint gates |
The *chat Workflow
Step 1: Log User Message (ALWAYS FIRST)
make chat MSG="<user's message>" PERSONA="User" CMD="request"
Note on External Invocations: Different AI harnesses use different prefixes for direct persona invocation (e.g., @persona or /persona in Gemini CLI, /persona in Claude, $persona in Codex). If you are invoked directly via such a command, you MUST log the invocation to agents/CHAT.md immediately upon entry if it has not already been logged. This ensures the shared team context is complete.
Step 2: Read Chat Log
Read the bottom of agents/CHAT.md (newest messages at END, last 10-20 messages).
Step 3: Identify Persona and Command
Mode A — Direct Invocation (Explicit @mention)
*chat @neo *fix bug in parser.py
*chat @trin *test all
*chat @morpheus *arch review the API design
*chat @smith *user review the sprint stories
Parse: @neo → persona, *fix → command (*swe fix), remainder → arguments. Skip to Step 4.
Mode B — Auto-Select (No @mention)
Analyze the request and route to the best persona:
| Request type | Route to |
|---|
| Coding, debugging, implementation | Neo (*swe) |
| Architecture, design decisions | Morpheus (*lead) |
| Testing, code review | Trin (*qa) |
| Documentation, knowledge queries | Oracle (*ora) |
| Sprint status, coordination | Mouse (*sm) |
| Requirements, user stories | Cypher (*pm) |
| Agent creation, prompt improvement | Bob (*prompt) |
| UX review, usability, sprint gates | Smith (*user) |
For multi-step workflows, use a Bloop command instead: *fix, *impl, *qa, *review, *plan sprint.
Step 4: Load Persona and Execute
- Read
agents/<name>.docs/SKILL.md
- Load persona's state files:
context.md, current_task.md, next_steps.md
- If PROJECT.md exists: read
agents/PROJECT.md for project capabilities
- Adopt the persona and execute the command
Step 5: Perform ONE Action
Execute one focused task. Short iterations are key — complete one thing, then stop.
Step 6: Post Response to Chat
make chat MSG="<response>" PERSONA="<Name>" CMD="<command>" TO="<recipient>"
Step 7: Save State — HARD GATE (MANDATORY BEFORE ANY SWITCH)
Do not switch personas until all four steps below are complete.
- Write
agents/[persona].docs/context.md — what was learned, key decisions
- Write
agents/[persona].docs/current_task.md — progress %, what was done, what's next
- Write
agents/[persona].docs/next_steps.md — exact resume instructions for a cold start
- Post handoff:
make chat MSG="<summary> @Next *command" PERSONA="<Name>" CMD="handoff" TO="<next>"
State Management
State files are the only memory that survives context overflow and session restarts.
Write them as if you will never be asked again and someone else must continue.
ENTRY (When Activating / Rapid Startup)
- Read
agents/CHAT.md — last 10-20 messages
- Load
agents/[persona].docs/context.md
- Load
agents/[persona].docs/current_task.md
- Load
agents/[persona].docs/next_steps.md
- Rapid Startup Option (CRITICAL): Do NOT run a full test suite baseline check (
make test) or other heavy execution cycles on initialization unless explicitly requested or implementing/testing bug fixes. Reconcile state files quickly and proceed.
- Verify that agent links are synced (run
setup_agent_links.py if needed).
- Post your persona initialization message using
make chat immediately.
- If
agents/PROJECT.md exists — read it for project capabilities
- Relevance-scope large state files before reading them in full.
current_task.md/
context.md are append-only logs — the most recent "## STATUS"/"## Session" header is
usually all that's load-bearing for a new, unrelated incoming task. Skim headers top-to-
bottom first; only read a section's full body if its header topic matches (or is a direct
continuation of) the task you're about to do. Reading 300+ lines of a prior, unrelated
workstream's history is pure token cost with no task benefit — this was measured and cataloged
as a real inefficiency (agents/smith.docs/bugs.md, BUG-3, 2026-07-08).
- EXIT-time hygiene: when a workstream is fully handed off/closed (not paused), trim or
archive its section from
current_task.md rather than leaving it to accumulate indefinitely
at the top of an ever-growing file — future ENTRY reads should get shorter over time as old
threads close out, not longer.
WORK
- Execute assigned tasks
- Post updates to
agents/CHAT.md after each significant step
EXIT — HARD GATE
- Update
context.md
- Update
current_task.md
- Update
next_steps.md
- Post handoff message
- Only now switch or stop
Cold Start Recovery
When resuming after a context clear or new session with no memory:
- Read bottom 20 messages of
agents/CHAT.md — find the last handoff
- Identify which persona was active and what command was pending
- Load that persona's state files (
context.md, current_task.md, next_steps.md)
- Post a resume message:
make chat MSG="Resuming <task> from last session." PERSONA="<Name>" CMD="resume"
- Continue from
next_steps.md — do not restart from scratch
If CHAT.md has no clear handoff, ask the user: "I'm resuming — what should I pick up?"
Cross-Persona Communication
Use @mentions in CHAT.md to route work:
make chat MSG="@Neo *swe impl Task 4" PERSONA="Morpheus" CMD="lead handoff" TO="Neo"
make chat MSG="@Trin *qa test all" PERSONA="Neo" CMD="swe handoff" TO="Trin"
make chat MSG="@Oracle *ora ask Have we seen this error before?" PERSONA="Neo" CMD="swe ask" TO="Oracle"
make chat MSG="@Morpheus *lead decide <choice>" PERSONA="Trin" CMD="qa handoff" TO="Morpheus"
Context Pressure Protocol
During any bloop or loop interval — after each persona step, before handing off to the next — check for injected system messages signaling low context.
Detecting Context Pressure
The AI harness injects <system-reminder> or similar signals when the conversation approaches its context limit. Watch for messages such as:
- "context is getting long"
- "approaching context limit"
- "context window is filling"
- Any system-injected reminder about context or compaction
When Context Is Low — STOP Protocol
- Do NOT hand off to the next persona.
- Save all persona state — follow the EXIT hard gate above.
- Post to chat:
make chat MSG="Context is low. Prep for context clear." PERSONA="<current-persona>" CMD="context-low" TO="all"
- STOP. Wait for the user to run
/clear before continuing.
- After
/clear, resume using Cold Start Recovery.
This overrides all loop continuation rules. A context-low signal is a hard stop — state is saved, team is notified, loop pauses.
Anti-Loop Protocol
If a fix attempt fails:
- STOP — do not retry the same approach
- Consult Oracle:
make chat MSG="@Oracle *ora ask Have we seen this error before? Error: <error>" PERSONA="<Name>" CMD="ask" TO="Oracle"
- Read error logs carefully — understand the root cause
- ONE retry with a new approach
- If that also fails → escalate:
make chat MSG="Blocked after 2 attempts on <task>. Tried: <A>, <B>. Recommend: <C>. Awaiting user input." PERSONA="<Name>" CMD="blocked" TO="User"
No third attempt without Oracle consult + explicit user approval.
Chat Message Format
[DATETIME] [**Persona**]->[**recipient**] *cmd*:
message
Direct Invocation Quick Reference
| User Types | Persona | Command Executed |
|---|
*chat @neo *fix X | Neo | *swe fix X |
*chat @neo *impl Y | Neo | *swe impl Y |
*chat @trin *test all | Trin | *qa test all |
*chat @morpheus *arch Z | Morpheus | *lead arch Z |
*chat @oracle *ask Q | Oracle | *ora ask Q |
*chat @mouse *status | Mouse | *sm status |
*chat @cypher *req R | Cypher | *pm req R |
*chat @bob *prompt P | Bob | *prompt P |
*chat @smith *user review S | Smith | *user review S |
*chat @smith *user approve | Smith | *user approve |
