System Prompt Extraction from Identity Files

Signet's system prompt is being moved from AGENTS.md to the session-start hook to fix invisibility, identity pollution, and duplication issues. The new design injects a tool-focused briefing at every session, ensuring updates take effect immediately without reinstalling connectors.

System Prompt Extraction from Identity Files The Signet system prompt should be infrastructure, not content. Problem Today the Signet system prompt lives inside a < -- SIGNET:START -- / < -- SIGNET:END -- block that gets injected into AGENTS.md and its harness-specific copies like CLAUDE.md . This has three problems: 1. The prompt is invisible to users who already have an AGENTS.md The Signet block is injected when AGENTS.md is generated from scratch during setup. If a user already has an AGENTS.md — either hand-written or from a previous version of Signet — the system prompt is never re-inserted. The user’s agent runs without Signet’s tool instructions, memory context, or identity stewardship guidelines. The connector’s install method calls buildSignetBlock and writes it into the generated file, but it won’t overwrite an existing file. And even if it did, users would rightly object to their hand-crafted instructions being clobbered by auto-generated plumbing. 2. The prompt pollutes the identity layer AGENTS.md is supposed to be the agent’s operating instructions — how it works, what it should do. The Signet system prompt is infrastructure: tool availability, memory commands, file locations, architecture explanations. Mixing the two makes both harder to maintain and harder for models to parse. The user’s intent gets buried under Signet’s self-description. 3. The prompt is duplicated and scattered The system prompt content is defined in @signet/core ’s buildSignetBlock , but each connector has its own copy/paste installation path. Updates to the prompt require rebuilding and reinstalling connectors. The SIGNET-ARCHITECTURE.md file is a separate artifact that duplicates some of the same information. There’s no single source of truth that all connectors read at runtime. Design Move the system prompt to session-start hook output The session-start hook already returns an inject string that gets prepended to the model’s context. This is the natural place for the Signet system prompt. The hook runs on every session, regardless of what’s in AGENTS.md , and it’s controlled by the daemon — so updates take effect immediately without reinstalling connectors. Current flow: AGENTS.md contains Signet block → symlinked to CLAUDE.md → model reads it Session-start hook → returns memories + date/time → prepended to context Proposed flow: AGENTS.md user content only → symlinked to CLAUDE.md → model reads it Session-start hook → returns system prompt + memories + date/time → prepended to context What the system prompt should contain The prompt should be short, tool-focused, and avoid duplicating what models already know. It’s not an architecture document — it’s a briefing. Proposed structure: signet active You have persistent memory managed by Signet. Your primary tool for retrieving memory is: mcp signet memory search — hybrid vector + keyword search across all stored memories. Returns scored results with content, type, tags, and entity graph context. For deeper exploration: mcp signet lcm expand — drill into a session summary node. Returns parent/child lineage, linked memories, and optionally the cleaned transcript. Use this when memory search returns a session reference you want to expand. mcp signet knowledge expand — drill into an entity's aspects, attributes, constraints, and dependencies. mcp signet knowledge expand session — find session summaries linked to a specific entity. Cross-session history is also available through linked summary and transcript artifacts in the Signet workspace. Inspect those artifacts directly when MEMORY.md or recall snippets are not enough. To store a memory explicitly: mcp signet memory store — write a memory with content, type, and optional tags. Your identity files are in your Signet workspace: AGENTS.md — how you operate you maintain this SOUL.md — your personality and values you maintain this IDENTITY.md — who you are you maintain this USER.md — who the user is you maintain this MEMORY.md — auto-generated working memory summary system-managed Secrets are available via mcp signet secret list and mcp signet secret exec. Secrets are injected as environment variables into subprocesses, never exposed as raw values. This is ~1200 chars. It tells the model what tools exist, what they do, and how to use them. No architecture lectures, no philosophy about memory layers, no instructions about “durable substrate.” Models are smart enough to figure out usage patterns from the tool descriptions themselves. What gets removed from AGENTS.md The entire < -- SIGNET:START -- / < -- SIGNET:END -- block is removed from buildSignetBlock in @signet/core/src/markdown.ts . The function either returns an empty string or is removed entirely. SIGNET-ARCHITECTURE.md stays as an on-demand reference file — the system prompt doesn’t need to mention it. If a model or user wants to understand the pipeline internals, they can read it directly. What changes in the session-start hook handleSessionStart in platform/daemon/src/hooks.ts already builds an inject string from memories, date/time, and metadata. The system prompt becomes the first section of that inject string, before memories and other context. The system prompt is built at runtime by the daemon not baked into connector configs , so it’s always current. If tools are added or renamed, the prompt updates on next session start. What changes in connectors Connectors no longer inject buildSignetBlock into generated markdown files. The install methods still create/symlink AGENTS.md → CLAUDE.md etc., but only with the user’s content. For existing users who already have a Signet block in their AGENTS.md , signet install strips it on upgrade remove content between SIGNET:START and SIGNET:END markers as a marker-bounded, idempotent migration. Migration Path New users Setup wizard creates AGENTS.md with only user-authored content or a minimal template . No Signet block. System prompt comes from the session-start hook. Existing users Active migration is required in this phase: signet install detects SIGNET:START / SIGNET:END markers and strips the legacy block from workspace AGENTS.md .- The cleanup is marker-bounded and idempotent, preserving all user-authored content outside the block. - Session-start inject supplies the runtime system prompt on every new session, so users do not need to regenerate identity files. Tool Naming The current tool names mcp signet memory search , mcp signet lcm expand , etc. are functional but not discoverable. The system prompt extraction is an opportunity to establish clearer names or at least clear descriptions. Candidates for renaming non-blocking, can be done separately : | Current | Candidate | |---|---| mcp signet lcm expand | mcp signet session expand | mcp signet knowledge expand | mcp signet entity expand | mcp signet knowledge expand session | mcp signet entity sessions | These are MCP tool names so renaming has compatibility implications. Could be done as aliases first, deprecating the old names. Files to Modify | File | Change | |---|---| platform/core/src/markdown.ts | Remove or empty buildSignetBlock . Keep SIGNET BLOCK START / END constants for migration detection. | platform/daemon/src/hooks.ts | Add system prompt to handleSessionStart inject output | integrations/claude-code/connector/src/index.ts | Stop injecting Signet block into generated files. Optionally strip existing blocks on install. | integrations/opencode/connector/src/index.ts | Same as above | integrations/openclaw/connector/src/index.ts | Same as above | integrations/codex/connector/src/index.ts | Same as above | integrations/oh-my-pi/connector/src/index.ts | Apply the same legacy block cleanup during install | platform/daemon-rs/crates/signet-daemon/src/routes/hooks.rs | Mirror session-start system prompt injection to keep shadow parity | Implementation Decisions This Phase Character budget: accepted for MVP. The injected system prompt is compact and fits inside existing session-start budgets. Per-harness variation: deferred. This phase uses one shared prompt across harnesses. User override flag: deferred. No suppression toggle in this phase. Tool availability detection: deferred. Prompt currently lists the canonical Signet MCP tools unconditionally.