Synapptic boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

Synapptic is in scope as a memory/knowledge/context-engineering system. Its intended use is to turn AI coding-session transcripts into persistent user preferences, AI failure observations and preventive rules, then publish narrative profiles to coding-assistant instruction or memory files. It issues model calls for observation extraction, narrative synthesis and behavioral benchmarking; it does not own the coding assistant's enclosing reasoning-and-tool loop. This classification follows the shipped CLI and its consumers, not just the package description (SRC-1, pyproject.toml, src/synapptic/cli.py, src/synapptic/extract.py, src/synapptic/synthesize.py, src/synapptic/outputs.py).

synapptic = "synapptic.cli:cli" --- pyproject.toml

def ingest(model, max_tokens, project, min_lines, limit): """Full pipeline: extract → merge → synthesize → integrate.""" --- src/synapptic/cli.py

The boundary is whole-system for the shipped repository, including its optional relay rather than only the profile library API. Initial inspection traced ingest from transcript discovery and processed-session selection through filtering, model extraction, observation persistence, global/project routing, weighted merging, cross-project promotion, narrative persistence and output integration. It also traced the separately installed SessionEnd hook and the standalone synthesize command. Those are included shipped drivers even where their paths differ: the hook calls extract, merge and synthesize, while the inspected standalone synthesize command saves narratives without calling output integration. Inclusion does not imply that every entry point completes the same publication loop (SRC-1, src/synapptic/cli.py, src/synapptic/bundle/hooks/synapptic-session-end.sh).

Persistent content has shipped consumers. Extraction reloads the existing profile as prompt context; synthesis reloads weighted profiles; integration reloads saved archetypes; benchmarks reload profiles and narratives. The Claude Code writer creates user_archetype.md and, when an existing MEMORY.md is present, maintains a reference to it. Other writers publish to assistant-specific project files. These establish production and reuse interfaces inside Synapptic, not proof that each external host actually reads or obeys the output (SRC-1, src/synapptic/cli.py, src/synapptic/state.py, src/synapptic/extract.py, src/synapptic/integrate.py, src/synapptic/outputs.py, src/synapptic/benchmark.py).

output_targets = config.get("outputs", ["claude-code"])
global_archetype = load_archetype(project_slug=None) or ""

--- src/synapptic/integrate.py

The optional relay is also included: the CLI launches external tools with relay endpoint environment variables; FastAPI routes forward Anthropic/OpenAI requests and retain metrics; SQLite stores request/session data and a transcript index; browser/dashboard routes and templates consume those records or local JSONL transcripts. This is a shipped observability and transcript-browsing path, not an excluded enclosing application. Its inclusion does not attribute the launched tools' planning, transcript generation or instruction-loading semantics to Synapptic (SRC-1, src/synapptic/cli.py, src/synapptic/relay/app.py, src/synapptic/relay/routers/anthropic.py, src/synapptic/relay/routers/browser.py, src/synapptic/relay/store.py, src/synapptic/relay/indexer.py).

The frozen evidence is repository https://github.com/appcuarium/synapptic at full commit bcde026770daa5cd9a8ca18ca7182b0aeaa87d1f. The read-only acquired checkout's HEAD matched that pin. The applicability cutoff is 2026-10-15; later commits, live documentation, installed packages and external service behavior are outside this evidence set. The overall tier is code-grounded. Target documentation supplies declared intent, not independent confirmation of its claims.

Repository coverage

Every top-level tracked file or directory is classified below. Included scope covers the paths named whether or not every line has already been inspected. Initial inspection is identified separately from remaining detail; an available but unread file is not an access gap. All entries refer to SRC-1.

External dependencies and evidence limits

External coding assistants supply transcripts, invoke the installed hook, load published files and perform coding work. Their implementations are not captured. This prevents conclusions about actual host admission, instruction precedence, compliance or downstream task improvement. The shipped skill's automatic-loading and cost claims are first-party doctrine, not observed operation (SRC-1, README.md, src/synapptic/bundle/skill/SKILL.md, src/synapptic/outputs.py).

Model inference depends on configured Claude CLI, Anthropic, Ollama or OpenAI-compatible providers. The optional relay additionally depends on FastAPI, httpx and uvicorn. Source inspection establishes adapter responsibilities, not service availability, response quality, compatibility across provider versions or reliable runtime performance (SRC-1, pyproject.toml, src/synapptic/providers.py, src/synapptic/relay/app.py).

No real user transcript corpus, generated profile history, benchmark execution artifact or deployment trace was supplied. Shipped tests and benchmark code are inspectable evaluators, not observed-run or causal-experiment evidence for effectiveness. In particular, the inspected orchestration test mocks all model calls:

class TestRunBenchmarkSmoke: """Smoke test for run_benchmark() orchestration — all LLM calls mocked.""" --- tests/test_benchmark.py

This limits conclusions about learned-profile accuracy, privacy protection in practice, guard efficacy, cost, throughput and long-term adaptation. No target code, tests, examples or providers were executed. There is no known unavailable repository path that forces a partial-loop boundary; remaining source detail is available for later jobs under the same pin.

Source register