| name | signalmode-refresh |
| version | 2.0.0 |
| description | Compiles the active session into a lean, verifiable continuation contract. Use when a session is becoming long, outputs are drifting, work must move to a new session, or a project requires a portable handoff to a human or another agent. Preserves the active objective, confirmed state, decisions, constraints, risks, recovery path, and next executable actions while excluding transcript noise and inaccessible context.
|
| trigger | /signalmode-refresh |
| argument-hint | [lite | full | ultra] [local | portable | auto] [human | agent | auto] |
SignalMode Refresh
Purpose
Produce the minimum sufficient, trustworthy context a recipient needs to continue work without relying on the current conversation, hidden memory, or inaccessible resources.
The output is a continuation contract, not a transcript summary. It must help the recipient orient, act, verify, recover, and finish.
Fast Path
Use this path unless a later rule requires deeper inspection:
- Resolve the objective, level, destination, and recipient.
- Extract the active state and decisions from the current conversation.
- Detect material contradictions.
- Inspect only the resources needed to resolve gaps or verify the first actions.
- Build one ordered action plan with accessible resources and explicit stopping conditions.
- Run the readiness gate and deliver either a ready handoff or a blocked notice.
Do not scan an entire workspace merely because access exists. Expand inspection only when the current evidence is insufficient.
Core Principles
- Continuation over compression. Preserve what enables the next action, not everything that happened.
- Access over location. A resource is useful only when the recipient can resolve and open it.
- Observed state and intended state are different. Never present a requested change as already implemented.
- Verify before asserting. Never invent a file, path, URL, identifier, decision, completion state, date, recovery point, or project detail.
- One fact, one home. State information once, then reference it by a stable action, decision, conflict, or resource ID.
- Context is a budget. Preserve enough to continue while leaving most of the recipient's context available for execution.
- Fail safely. Do not label a handoff executable when a material blocker remains.
- Plan for recovery. Medium-risk work must identify a safe checkpoint. High-risk work requires a verified recovery path before execution.
- Separate certainty levels. Distinguish confirmed facts, observed state, intended state, assumptions, unresolved conflicts, and open questions.
- Minimize exposure. Never include live credentials or unnecessary sensitive information.
- Make resumption cheap. Preserve stable IDs and a checkpoint so the next recipient can continue without rereading the entire handoff.
- Avoid duplicate mutation. Before rerunning a completed or uncertain action, verify whether it is idempotent or whether repetition could create duplicate, conflicting, or destructive effects.
Inputs
Level
lite — rapid reset with only the minimum continuation state. ~500–1,200 tokens.
full — verified operational handoff. Default. ~1,200–3,500 tokens.
ultra — broader continuation audit with risks, rejected approaches, dependencies. ~2,500–6,000 tokens.
Destination
local — recipient works in the same accessible project or workspace.
portable — recipient may use another machine, account, workspace, or organization.
auto — infer from request. Default to local only when continued access to the same resources is reasonably supported.
Recipient
human — optimize for readable context, rationale, and judgment notes.
agent — optimize for deterministic structure, verb-first actions, stable IDs, inputs, outputs, validation, and stop conditions.
auto — infer from request. Use agent only when autonomous execution is explicit.
Risk
low — analysis, drafting, reading, or readily reversible edits.
medium — configuration changes, batch edits, deployments with a known rollback.
high — destructive writes, migrations, production changes, irreversible external actions.
Interaction Rule
Ask at most one clarification question.
- If the goal is clear, do not ask.
- If the goal is missing, ask: "What should the next session or recipient accomplish? Mention whether it is for a person, an autonomous agent, or another workspace if that matters."
- End the turn after asking and wait.
- Do not ask a second question. Resolve remaining ambiguity conservatively.
Lite Mode Workflow
- Analyze active git diff, file changes, and recent session history.
- Separate observed state (actual file changes) from intended state (next steps).
- Look for an existing file at
memory/history/YYYY-MM-DD.md (using today's date). If it exists, append to it; if not, create it.
- Format output using the lite schema below.
- Update
memory/map.md to reflect newly completed changes.
- Conclude by prompting the user with the Token Refresh Protocol message.
Output Schema — Lite
---
schema: signalmode-refresh-handoff
schema_version: 2.0
status: ready
created_at: <ISO 8601 timestamp>
level: lite
destination: <local | portable>
recipient: <human | agent>
risk: <low | medium | high>
---
# Session Refresh: [Short Objective Slug]
Handoff reason: <why this was created>
- **Session Objective**: What were we trying to accomplish?
- **Observed State**: Active file paths modified and verified working state.
- **Key Decisions (D1)**: Architecture boundaries established.
- **Immediate Actions (A1–A3)**: 1–3 verb-first, linear next steps.
- **Open Questions / Blockers**: Remaining gaps or ambiguities.
Output Schema — Full (Human Mode)
---
schema: signalmode-refresh-handoff
schema_version: 2.0
status: <ready | ready_with_warnings>
created_at: <ISO 8601 timestamp>
level: full
destination: <local | portable>
recipient: human
risk: <low | medium | high>
---
# Session refresh: <goal>
Handoff reason: <why this was created>
You are continuing work on <project or workstream>.
Objective: <single concrete objective>
## Current state
<confirmed observed state and completed work>
## Intended state
<target state, only when different from current state>
## Confirmed decisions and constraints
- D1: —
C1:
A1:
A2:
A3:
A4: — Blocked by:
Risk:
Last known-good state:
Rollback:
R1: — ; access: ; freshness: ; verified:
Assumption:
Open question:
Verified:
Warnings:
Unverified:
Output Schema — Full (Agent Mode)
---
schema: signalmode-refresh-handoff
schema_version: 2.0
status: <ready | ready_with_warnings>
created_at: <ISO 8601 timestamp>
level: full
destination: <local | portable>
recipient: agent
risk: <low | medium | high>
checkpoint: <not_started | A# completed | custom checkpoint>
---
# Continuation contract: <goal>
Handoff reason: <why this was created>
Objective: <single measurable objective>
## Execution contract
- Scope: <what may be changed or investigated>
- Allowed: <authorized operations>
- Do not change: <protected areas or >
Approval gates:
Completion report:
D1: | rationale: | evidence:
Inputs:
Depends on:
Expected output:
Validate:
Stop if:
Approval required:
Repeat safety:
A2: | depends on: | output:
S1:
Risk:
Last known-good state:
Rollback validation:
R1 | locator: | purpose: | access: | freshness: | verified
Token Refresh Protocol
After completing the handoff, always conclude with:
⚠️ TOKEN REFRESH PROTOCOL ACTIVATED: Session memory compiled to memory/history/ and maps updated. To drop your context token count to zero, please type /clear now. You can resume immediately by typing: Resume from today's history log.
Readiness Gate
Hard blockers — set status to blocked when:
- The objective is missing or materially ambiguous.
- A material contradiction makes the first action unsafe.
- A critical dependency is inaccessible and cannot be carried.
- A live secret remains exposed.
- A high-risk operation lacks a verified usable recovery path.
- The recipient cannot safely determine where to begin.
Warnings — set status to ready_with_warnings when:
- An optional resource is unverified.
- A noncritical external dependency may be stale.
- A secondary open question remains.
Boundaries
- Never include passwords, API keys, access tokens, or live authentication material in the handoff.
- Replace secret values with environment-variable names or credential labels.
- Portable mode must not expose sender-local absolute paths.
- Never invent a file, path, completion claim, or recovery point.