Signet boundary
Type: agentic-system-analyses/types/agentic-system-boundary.md
Boundary and evidence
This run analyses Signet as the shipped memory, identity and context-delivery system, not merely its library API or its Dreaming worker. The intended use is to explain how captured evidence becomes retained state, how model-driven maintenance changes that state, and how later agents receive it. The target is in scope as a memory/knowledge/context-engineering system. Its own background maintenance issues model calls, and its shipped adapters serve retained context to model-using hosts (SRC-1, platform/daemon/src/pipeline/dreaming-worker.ts, integrations/pi/extension/src/index.ts).
The source's declared purpose is:
Signet gives your AI agents a shared memory. You can use it to store, sync, and share memories, system prompts, transcripts, institutional knowledge, and secrets across all the AI tools and models you use. ---
README.md
The boundary is whole-system: it includes evidence capture and imports; agent-scoped persistence and authorization; indexing, recall and context construction; Dreaming prompts, tools and audited mutations; identity and generated harness configuration; shipped CLI, SDK, integrations, browser and desktop clients; backup, recovery, maintenance and evaluation paths. These same-repository callers are part of the target, not external applications. The table below distinguishes boundary inclusion from inspection depth. Inclusion does not imply every implementation file has already been read.
Initial entry-point and loop trace
The npm launcher calls launchSignet, whose implementation resolves and spawns a platform binary (SRC-1, dist/signetai/bin/signet.js, dist/signetai/bin/launch.js). The root workspace scripts identify CLI, daemon, native, connector and extension build surfaces (SRC-1, package.json). The CLI registers memory, sources and Dreaming commands; the daemon registers memory, hook, source, import, transcript-import, knowledge, ontology and pipeline routes (SRC-1, surfaces/cli/src/cli.ts, platform/daemon/src/daemon.ts). Memory routes import the transactional ingest and mutation functions together with recall and database-owner facilities (SRC-1, platform/daemon/src/routes/memory-routes.ts). The ingest implementation inserts memory content, source identity, agent scope and evidence metadata into SQLite (SRC-1, platform/daemon/src/transactions.ts).
The daemon starts a transcript-capture worker and a Dreaming worker through its pipeline startup path. The inspected startup explicitly distinguishes current semantic maintenance from legacy extraction:
logger.info("dreaming", "Dreaming owns all semantic writes; legacy extraction is retired"); ---
platform/daemon/src/daemon.ts
Dreaming's executor calls the inference router with a prompt and tools; this establishes model-dependent operation rather than a passive storage library (SRC-1, platform/daemon/src/pipeline/dreaming-worker.ts). Shipped maintenance instructions are also in scope. Their mutation contract is declared, not treated here as proof that every handler enforces it:
Do not rewrite raw transcript/source artifacts or edit SQLite directly. ---
skills/dreaming/SKILL.md
The consumer side is shipped too. The Pi lifecycle implementation posts session-start and session-end payloads, reads session-file snapshots, and retains returned startup context. Its extension registers prompt, context and compaction handlers (SRC-1, integrations/pi/extension-base/src/lifecycle.ts, integrations/pi/extension/src/index.ts). The context handler returns hidden injected messages alongside the host's messages:
messages: [...event.messages, ...hiddenMessages], ---
integrations/pi/extension/src/index.ts
The SDK, dashboard and browser extension also call daemon memory or hook APIs. The desktop client manages daemon startup rather than being an excluded enclosing application (SRC-1, libs/sdk/src/client-p2.ts, surfaces/dashboard/src/lib/api.ts, surfaces/browser-extension/src/shared/api.ts, surfaces/desktop/src/daemon-manager.ts). MemoryBench's Signet provider calls remember, session-end and recall endpoints; its orchestrator includes ingest, indexing, search, answer and evaluation phases. Evaluation is therefore included as a shipped driver, not inferred from the README's reported score (SRC-1, memorybench/src/providers/signet/index.ts, memorybench/src/orchestrator/index.ts).
Repository coverage
Every top-level tracked entry is classified below. “Included; mapped” means the tracked tree and entry wiring establish a material responsibility, but detailed implementation remains for later analysis. “Excluded” is a deliberate functional exclusion, not missing evidence. There is no identified unavailable material source path at this pin. All rows refer to SRC-1; listed paths are commit-relative anchors.
| Top-level tracked entry | Coverage and material paths | Role and limit |
|---|---|---|
platform/ |
Included; initial code inspection of platform/daemon/src/daemon.ts, platform/daemon/src/routes/memory-routes.ts, platform/daemon/src/routes/hooks-routes.ts, platform/daemon/src/transactions.ts, platform/daemon/src/pipeline/dreaming-worker.ts, platform/daemon/src/pipeline/dreaming.ts, platform/native/src/lib.rs; mapped platform/core/src/migrations/, database, identity, workspace, recall and search modules; daemon source providers, transcript capture/import, DB-owner workers, ontology operations, memory-head publication, maintenance, recovery, inference, auth and MCP paths |
State production, admission, storage, semantic maintenance and retrieval. Individual enforcement, scheduling and failure semantics are not yet established by this boundary. |
integrations/ |
Included; initially inspected integrations/pi/extension/src/index.ts and integrations/pi/extension-base/src/lifecycle.ts; mapped connectors for Claude Code, Codex, Forge, Gemini, Hermes, Kimi, Muse Code, Oh My Pi, OpenClaw, OpenCode and Pi; Codex plugin skills/hooks, Hermes plugin, OpenClaw memory adapter, OpenCode plugin, Oh My Pi extension and integrations/unreal/plugin/ |
Shipped capture, reload and consumption paths remain inside the system. Pi wiring is not evidence of identical lifecycle behavior across all hosts; those paths need their own inspection. |
libs/ |
Included; initially searched libs/sdk/src/client-p2.ts; mapped generated SDK client, transport, AI/OpenAI helpers, libs/connector-base/src/ and libs/lifecycle-proof/src/ |
Shared API consumption, integration installation and lifecycle instrumentation. Instrumentation code alone is not an observed trace. |
surfaces/ |
Included; initially inspected CLI entry registration, dashboard API client, browser API client, desktop daemon manager and tray entry searches; mapped CLI command/features modules, surfaces/cli/templates/, dashboard onboarding/settings/chat/graph views, desktop runtime staging and tray controls |
Operator admission, persistence requests, retrieval and runtime management. surfaces/cli/templates/AGENTS.md.template and surfaces/cli/templates/DREAMING.md.template were read; remaining templates and UI details are included but not yet inspected. |
dist/ |
Included; read launcher and searched dist/signetai/bin/launch.js; mapped wrapper installation/runtime helpers, native-platform manifests, dist/signetai/templates/ and bundled Hermes plugin |
Shipping and bootstrap paths, including prompts and consumers. Actual release binaries and installed package contents are not supplied, preventing binary/source equivalence claims. |
skills/ |
Included; read skills/dreaming/SKILL.md; mapped skills/onboarding/SKILL.md |
Shipped maintenance and setup instructions. Onboarding details not yet inspected; maintenance instructions support doctrine, not observed compliance. |
memorybench/ |
Included; inspected Signet provider and orchestrator excerpts/searches; mapped benchmark datasets, provider prompts, answer/evaluate/search/ingest phases, judges, checkpoints, server and UI | Shipped evaluation driver and evidence-consuming workflow. No benchmark execution or result verification was performed. Other provider adapters are comparison plumbing, not separate analysed systems. |
scripts/ |
Included; mapped scripts/bench-memory.ts, scripts/autoresearch-memorybench.ts, scripts/claim-trace-eval.ts, scripts/observer-belief-eval.ts, scripts/evals/, MCP probes, native/build/staging tools, memory backfill and regression tests |
Build, maintenance and evaluation support. Root scripts establish invocation wiring; detailed scripts are not yet inspected and do not establish performance. |
tests/ |
Included; mapped tests/integration/acceptance/, pipeline LLM, search visibility, web-content indexing, stress and import regression paths |
Test and acceptance contracts; not executed, so no test-pass claim. |
deploy/ |
Included; mapped deploy/docker/ entrypoint, compose, healthcheck/token scripts and deploy/aur/, deploy/channels/ packaging |
Deployment and access setup. Detailed deployment behavior not yet inspected; no live deployment assurance. |
docs/ |
Included; mapped runbooks, architecture diagrams and docs/specs/ approved/complete/draft contracts |
Target-native doctrine and historical design. A spec's status does not establish current implementation. No external analysis is imported. |
web/ |
Included: read web/docs/src/content/docs/architecture/packages-data-flow.md; mapped remaining API, pipeline, hooks, harness, memory, source and benchmark documentation, plus web/marketing/public/install.sh, web/marketing/public/install.ps1, web/marketing/public/skill.md. Excluded: marketing presentation/blog/testimonials, ads and web/workers/reviews/ |
Documentation, installers and operator setup belong to the boundary. Marketing and review-service exclusion prevents conclusions about public-site operation or marketplace reviews, not local memory behavior. Unread installers remain included, not an access gap. |
.agents/ |
Included; read .agents/skills/benchmarking/SKILL.md; mapped benchmarking reference/scripts/results, Dreaming development and other repository-agent instructions |
Target-native evaluation and development procedures. Uninspected results are not independently verified evidence. |
.claude/ |
Included; mapped .claude/skills |
Instruction projection; projection behavior not yet inspected. |
.github/ |
Included; mapped build, release, Docker, Dreaming benchmark gate, migration, event-loop, transcript-import and acceptance workflows. Excluded: issue/PR templates and PR screenshots | Shipped CI/evaluation and release wiring. Exclusion prevents conclusions about issue administration and screenshot fidelity. No CI success inferred. |
.githooks/ |
Included; mapped commit-message and pre-commit hooks | Development gates; enforcement not tested. |
package.json, bun.lock, bunfig.toml, biome.json, tsconfig.json, turbo.json, .bump-level, .dockerignore, .gitignore |
Included; root package.json read; remaining configuration mapped |
Workspace entry points, dependency/build and packaging contracts. Dependency implementations are not frozen by a lockfile alone. |
AGENTS.md, CLAUDE.md, AI_POLICY.md, CONTRIBUTING.md |
Included; mapped | Repository development instructions, not user memory content. Not yet inspected; no claim of development-policy compliance. |
README.md, README.de.md, README.ja.md, README.ko.md, README.zh-CN.md, VISION.md, ROADMAP.md, CHANGELOG.md, repo.map.yaml, SECURITY.md |
Included; English README read, remaining documents mapped | Declared purpose, orientation, plans/history and security guidance. Translations and historical statements are not independently corroborated. |
CODE_OF_CONDUCT.md, LICENSE, NOTICE, THIRD_PARTY_LICENSES.md |
Excluded; tracked root documents | Community and legal materials; no legal or licensing assessment. |
public/, screenshots/ |
Excluded; image/branding assets and UI illustrations | No conclusions about rendered visual fidelity or screenshot authenticity. |
Frozen evidence and external limits
The only frozen source is the acquired repository at full commit 12dba1ce753babda735f2ba7b6f150c6c0b7f95e; checkout HEAD was checked against that pin. The applicability cutoff is 2026-10-14, the supplied run date. This fixes source applicability, not the date of a deployment or a measured execution. Evidence tier is code-grounded, with target-native documentation separately labelled as doctrine/design.
External agent harness implementations, model providers and model weights, remote source services, OS/keyring behavior, downloaded release assets and dependency internals are outside the frozen source. Signet's adapters and provider/service routing remain included. These limits prevent conclusions about an external host's complete agent loop, provider reliability or answer quality, actual remote-service behavior, and cross-platform deployment correctness. The Bun/Node/Rust/native and embedding/search dependencies are visible as manifests and interfaces, not independent implementation sources (SRC-1, package.json, platform/native/src/lib.rs, dist/signetai/bin/launch.js).
No target execution, tests, model calls or live-service inspection was performed. Code supports an account of executable paths; documentation supports attributed intent. Neither supplies observed operation or a causal experiment. The reported benchmark score in the README is not adopted as a verified result. No live documentation page or unpinned repository is additional evidence.
Source register
| source ID | kind | identity/location | revision or capture | evidence layer | inspected scope | citation anchors | access gaps and conclusion prevented |
|---|---|---|---|---|---|---|---|
| SRC-1 | git | https://github.com/signet-ai/signetai |
12dba1ce753babda735f2ba7b6f150c6c0b7f95e | implementation; doctrine/design | Access root: /home/zby/llm/commonplace/.commonplace/worktrees/signetai-64f703d83203/related-systems/signet-ai--signetai. Implementation: tracked tree and root workspace scripts; launcher; daemon route/startup wiring; memory transaction and route excerpts; Dreaming executor and state/prompt searches; Pi extension/context and shared lifecycle excerpts; SDK, dashboard/browser API and desktop/tray entry searches; MemoryBench Signet provider and orchestrator searches; native module entry. Doctrine/design: English README, Dreaming skill, benchmarking skill, CLI AGENTS/DREAMING templates and package/data-flow architecture page. These are initial inspected paths, not an allowlist; all included paths at this commit may supply later evidence. |
Implementation: daemon, Dreaming executor, Pi consumer, evaluation provider. Doctrine/design: README, maintenance instruction, data-flow contract. Additional local anchors appear in the trace and coverage table. | No identified inaccessible material checkout paths. External dependencies and release binaries are not captured. No observed-run or causal-experiment layer established; runtime correctness, benchmark scores, deployment equivalence and provider outcomes cannot be concluded from this initial inspection. Mapped-but-unread files are inspection work remaining, not missing access. |