SHIFT — Multi-Identity Delegation Skill

One agent. Many specialists. One conversation.

---

⚠️ SECURITY AND PRIVACY — READ BEFORE ENABLING

What SHIFT Accesses

SHIFT reads from your OpenClaw workspace:

• - ~/.openclaw/workspace/MEMORY.md — excerpts attached to delegation context
• INLINECODE1 — your persona and cost settings
• Session files it creates in INLINECODE2

SHIFT writes to:

• - ~/.openclaw/workspace/.shift/sessions/<runId>/ — per-delegation session files
• INLINECODE4 — delegation cost tracking
• INLINECODE5 — your persona overrides

What SHIFT Transmits

When you enable delegation, the following are sent to whatever model providers you configure for each persona:

• - Your message content
• Conversation history (last N turns, controlled by contextBridge.historyTurns)
• Active file contents and MEMORY.md excerpts relevant to the task
• Delegation metadata (runId, timestamps, persona name)

You control which models are used. SHIFT does not ship with or store credentials — it uses the model providers configured in your OpenClaw agent config. Review your provider's data handling policy.

Data Exposure Risk

If your workspace files or MEMORY.md contain secrets (API keys, credentials, private data), and those files are referenced in a delegated task, that content will be transmitted to your configured model provider.

Mitigations:

• - Do not delegate tasks that reference sensitive file paths
• Set contextBridge.historyTurns to a low number (e.g., 3)
• Use costManagement.trackOnly: true during evaluation to observe behavior without enforcement
• Use trusted or self-hosted model providers for sensitive work

Setup Script

INLINECODE9 creates local directories only. It does NOT:

• - Make network calls
• Escalate privileges
• Modify system files
• Access credentials

Cost and Consultations

Consultations spawn additional model calls. Costs are tracked in cost-tracking.json per delegation. Set costManagement.trackOnly: true to monitor before enforcing limits.

Recommendations

1. 1. Start with costManagement.trackOnly: true to observe costs
2. Set contextBridge.historyTurns to a small number (5 or less)
3. Set contextBridge.archiveAfterMinutes: 30 for faster cleanup
4. Do not delegate tasks referencing sensitive file paths
5. Review model providers' data policies before sensitive work

---

Triggers on: any message that might benefit from specialized handling — coding tasks, research, analysis, quick lookups, or anything that warrants delegating to a sub-identity.

When SHIFT triggers, you read and follows the procedures below to: classify the task, delegate to the right sub-identity, handle consultation between sub-identities, and synthesize the final response in his own voice.

---

SHIFT Procedures

Follow these steps for every message that triggers SHIFT.

---

STEP 1 — Fast Path Check

Before anything, check if this message is trivially handled by you directly.

Bypass delegation (return none) if message is ONLY:

• - Greetings: hi, hey, hello, yo, INLINECODE20
• Acknowledgments: thanks, thank you, ty, ok, sure, got it, makes sense, cool, INLINECODE29
• Meta: that was fast, nice one, lol, INLINECODE33

If none of those match, proceed to STEP 2.

---

STEP 2 — Load Config

Read the SHIFT config at:
CODEBLOCK0

If the file doesn't exist, run the setup script first:
CODEBLOCK1

From config, extract:

• - displayMode — hidden or INLINECODE36
• INLINECODE37 — conservative or INLINECODE39
• INLINECODE40, costManagement.costBudgetPerHour, INLINECODE42
• INLINECODE43, INLINECODE44
• INLINECODE45 — enabled personas and their settings

---

STEP 3 — Cost Budget Check

If costManagement.enabled: true:

Read ~/.openclaw/workspace/.shift/cost-tracking.json:

CODEBLOCK2

If current UTC hour > hourStart (new hour):
→ Reset totalSpend: 0, update hourStart to current hour.

If totalSpend >= costBudgetPerHour:
→ You handle the task directly. Say: "Handling this one myself to stay within your cost budget."
Return budget_exceeded.

If totalSpend >= alertThreshold * costBudgetPerHour:
→ Send a quiet warning: "Approaching delegation budget limit for this hour."

---

STEP 4 — Route to Persona

Using the enabled personas from config, classify the user's message.

Keyword Scoring

For each enabled persona, count keyword matches in the user's message (case-insensitive, whole-word). Raw score = matches / total keywords.

If raw score >= minConfidence: Add to candidates.

Runner special case: Runner has requireExplicit: true. It only triggers if the message is predominantly a Runner task, not just containing a keyword. If the message has strong code/research keywords, Runner is NOT a match even if it has Runner keywords.

Tie-Breaking

If multiple personas score above threshold:
→ Higher score wins. If tied: codex > researcher > runner.

Default

If no persona scores above threshold → You handle directly. Return none.

---

STEP 5 — Prepare Delegation Context

Generate a unique runId:
CODEBLOCK3

Create the session folder:
CODEBLOCK4

Write INBOUND.json

CODEBLOCK5

Write CONTEXT.md

Pull from:

• - ~/.openclaw/workspace/MEMORY.md (keyword-matched relevant sections)
• Any active file contents mentioned in the conversation
• Persona-specific context needs (from the persona definition in config)

Format:

# Context for {persona}

## Project/Memory (relevant)
...

## Active Files
...

## Persona Context Needs
...

---

STEP 6 — Build Task Prompt

Construct the prompt for the sub-agent:

CODEBLOCK7

---

STEP 7 — Spawn Sub-Agent

CODEBLOCK8

Track elapsed time. If elapsed > timeout * 0.7:
→ Send status to user: "{persona} is still working on this..."

---

STEP 8 — Wait and Read Result

Wait for sessions_spawn to complete.

Read OUTBOUND.md

From the session folder:
CODEBLOCK9

Parse:

• - Status: complete | error | INLINECODE63
• INLINECODE64: the actual response content
• INLINECODE65: if sub-identity consulted another
• INLINECODE66: token counts if available

Check for ESCALATE.md (Runner only)

If Status: escalation:
→ Runner detected the task was too complex. You take over directly.
→ Read reason, summary, partialAnswer from ESCALATE.md.
→ Incorporate any partial work Runner did.

---

STEP 9 — Update Cost Tracking

After successful delegation, update cost-tracking.json:

CODEBLOCK10

---

STEP 10 — Synthesize Response

Transform the sub-identity's output into your voice.

Detection

If ConsultationLog is present and non-empty:
→ Build consultation mention (ALWAYS shown):
CODEBLOCK11

Transformation Rules

Sub-identity output	your transformation
Code	Explain what it does in plain English, highlight key parts, offer to add tests
Analysis/Research

Pull 2-3 key insights, bullet-point them, note caveats | | Quick answer (Runner) | Relay directly with warmth added |

Display Mode

Hidden (default):
CODEBLOCK12

Transparent:
CODEBLOCK13

Consultation mentions are ALWAYS shown in both modes.

---

STEP 11 — Handle Errors

Scenario	Response
Sub-agent timeout	"Looks like this one is taking longer than expected — let me work through it directly."
Sub-agent error

"Hit a snag with the specialist — let me handle this myself." |
| sessions_spawn fails | Fall back to your handling directly |
| OUTBOUND.md missing | Assume sub-agent failed, You handle directly |
| Model unavailable | Log warning, You handle directly |

---

Consultation Protocol (Master-Orchestrated)

Key insight: Sub-agent sessions cannot cleanly spawn child sessions and wait. The solution: You orchestrate the consultation from the master level.

Flow

CODEBLOCK14

Sub-Agent Side (Codex/Researcher)

When a sub-agent needs to consult another sub-identity:

1. 1. Check: is target in my consults list? Is target NOT in my consultsNever?
2. Write CONSULT-INBOUND.md to the session folder with the question + context
3. Write OUTBOUND.md with:

   ## Status: needs_consultation
   
   ## PartialResult
   (what Codex has done/understood so far)
   
   ## ConsultationLog
   - consulted: researcher
   - question: "..."
   - status: pending
   

1. 4. End response with [NEEDS_CONSULTATION] instead of INLINECODE78
2. Stop — You will take over from here

Master Side

When reading OUTBOUND.md after a sub-agent returns:

STEP 8a — Detect Consultation Need

If Status: needs_consultation is present:

1. 1. Read ConsultationLog to get target and question
2. Read CONSULT-INBOUND.md for full context
3. Send user update (transparent mode): INLINECODE82
4. Proceed to STEP 8b

STEP 8b — Run Consultation

CODEBLOCK16

STEP 8c — Feed Result Back

After consultation completes:

1. 1. Read INLINECODE83
2. Append consultation answer to CONTEXT.md:

   ## Consultation Answer from {target}
   {answer from CONSULT-OUTBOUND.md}
   

1. 3. Re-spawn the original sub-agent with enriched context:

CODEBLOCK18

STEP 8d — Final Result

When re-spawned sub-agent completes:

1. 1. Read final INLINECODE85
2. INLINECODE86 should now be INLINECODE87
3. Proceed to STEP 10 (Synthesis)

Consultation is always mentioned to the user — not gated by display mode.

Depth Limit

Max consultation depth = 1. A sub-agent that was consulted CANNOT consult anyone else. This is enforced by: the consultation re-spawn prompt does NOT include the consultation protocol section. Only the first-level sub-agent has it.

Timeout Handling

• - If consultation times out: You fall back to Codex's partial result + own knowledge
• You warn: "{target} took too long — Codex will continue without that context."
• Partial result (if any) is still used

---

Commands

/shift status

Show current SHIFT status:

• - Which personas are enabled
• Current cost budget status
• Display mode

/shift mode hidden|transparent

Toggle display mode. Updates config.yaml.

/shift fastpath on|off

Toggle fast-path mode.

/delegate \ \

Explicitly delegate a task to a specific persona, ignoring keyword routing.

---

Config Reference

Full schema at: INLINECODE88

Key settings:

CODEBLOCK19

---

Session Files

All delegation session files live at:
CODEBLOCK20

Files:

• - INBOUND.json — delegation input
• INLINECODE90 — project context
• INLINECODE91 — sub-identity result
• INLINECODE92 — Runner escalation (if any)
• INLINECODE93 — consultation input
• INLINECODE94 — consultation result

Auto-archived after contextBridge.archiveAfterMinutes (default: 60min).

---

SHIFT — one brain, many specialists.