Pond boundary

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

Boundary and evidence

This run analyses Pond as a durable archive and recall system for agent sessions, not as an enclosing agent runtime. Its intended use is to preserve prior sessions, retrieve their content for later model context, and restore sessions into supported clients. The target class is memory/knowledge/context-engineering system; this includes archival recall rather than assuming a distilled-fact memory mechanism. The boundary kind is whole-system: the shipped Rust application and library, adapters, prompts, maintenance paths, host plugins, restoration consumers, deployment examples and evaluation material belong to the target. The host integrations are not excluded merely because they call Pond's APIs.

The frozen evidence is SRC-1, repository https://github.com/tenequm/pond at commit 0c440e130faf44052d0d896a585e572090079f1f. The checkout HEAD equals this acquisition pin and the checkout status was clean. The applicability cutoff is 2026-10-14, the supplied run date, not a claim that all reported measurements were made that day. The overall evidence tier is code-grounded. Source inspection supports implementation descriptions; the source's specification and tool instructions supply declared contracts. No target program, test, model provider or deployment was executed in this run.

Qualification and functional loop

The source describes the archival purpose directly:

pond ingests sessions from agentic clients - Claude Code, Codex, and others on the roadmap - into one canonical form, stores them in Lance, and serves search (full-text, with optional semantic search; one arm per query, Section 8.7) over them at message granularity. --- docs/spec.md

This is in scope because its retained state is served to model-driven clients through recall tools and shipped consumption instructions. The bundled skill makes that consumption policy explicit (SRC-1, packages/pond/SKILL.md):

has it: recall first, then answer. Before you say "I don't know" or re-derive something that sounds prior, search pond. --- packages/pond/SKILL.md

The initially traced loop comprises:

• Produce and admit: external harnesses produce native logs; Pond's shipped adapter registry parses them into canonical events. AdapterFactory owns discovery/configuration and serialization; Adapter streams events and offers freshness planning. The CLI's sync pipeline drives these adapters, and ingest_adapter validates/batches their output. HTTP ingest is a second admission surface. Evidence: SRC-1, packages/pond/src/adapter/mod.rs, packages/pond/src/main.rs, packages/pond/src/handlers.rs, packages/pond/src/transport.rs.
• Persist and maintain: the session store filters existing message/part keys, optionally embeds new message rows, appends messages and parts and inserts absent sessions through the storage substrate. Sync has a shared pipeline for CLI and in-server use, including index maintenance and OpenClaw deletion detection. Scheduled sync, explicit optimize, copy and archive restore remain included state-maintenance paths. Evidence: SRC-1, packages/pond/src/sessions.rs, packages/pond/src/substrate.rs, packages/pond/src/main.rs, packages/pond/src/schedule.rs.
• Reload and consume: search handlers select full-text or optional vector retrieval from the store; get/session/message and SQL surfaces expose retained content. MCP tool descriptions route model callers to search and subsequent transcript reads. The Pi, OpenClaw and Hermes packages register recall tools and supervise or connect to a Pond server. Their managed server arguments include periodic sync, so the integrations participate in capture as well as recall. Evidence: SRC-1, packages/pond/src/handlers.rs, packages/pond/src/transport.rs, packages/pi-pond/index.ts, packages/pi-pond/src/service.ts, packages/openclaw-pond/index.ts, packages/openclaw-pond/src/service.ts, packages/hermes-pond/__init__.py, packages/hermes-pond/service.py.
• Restore and continue: the CLI reloads retained lineage and serializes it through a target adapter. Pi's shipped command bridges this write-producing CLI operation to the host's session switch; it is a material later consumer, not an external application excluded from the target. Evidence: SRC-1, packages/pond/src/main.rs, packages/pi-pond/src/resume.ts, packages/pi-pond/index.ts.

The Pi continuation handoff is executable code, not just an advertised use (SRC-1, packages/pi-pond/index.ts):

await ctx.switchSession(outcome.sessionFile, { withSession: async (fresh) => { fresh.ui.notify(note, "info"); }, }); --- packages/pi-pond/index.ts

Shipped prompt material includes both the bundled skill and plugin tool definitions. In particular, the Pi tools attach promptSnippet and promptGuidelines; its entry-point comment's “no prompt hooks” must not be read as absence of instructions supplied to the model (SRC-1, packages/pi-pond/src/tools.ts). These instructions and the tool descriptions are included even though the model's decision to invoke them remains a host/model responsibility.

Repository coverage

The table classifies every top-level tracked file or directory found at the pin. Included means inside the analysis boundary, not that every blob has already been read. “Not yet inspected” is an inspection status, not an access gap. The initial inspection traced material entry points and consumers; later analysts may inspect further files at this same commit under SRC-1.

External dependencies and limits

External agent runtimes and providers supply session records and execute later model/tool loops. Their implementations are outside this repository boundary. Their APIs are visible through shipped plugins and adapters, but this prevents attributing host context selection, provider behavior, upstream compaction or actual restored-session continuation to Pond alone. Pond's own plugin registration, prompts, capture supervision and session-switch handoff remain included. Evidence: SRC-1, packages/pi-pond/index.ts, packages/openclaw-pond/index.ts, packages/hermes-pond/__init__.py, docs/spec.md.

Lance/object-store services supply storage, indexing and concurrency primitives; Candle/model artifacts supply optional embeddings; axum, rmcp, Tokio and host SDKs supply transport/runtime interfaces. The package manifest and source specification establish these dependencies, but their external repositories, cloud services and remotely downloaded model weights are not additional frozen sources. Their absence prevents independently verifying their internals, operational reliability or deployed IAM/encryption. Evidence: SRC-1, packages/pond/Cargo.toml, docs/spec.md.

No required repository evidence was unavailable. Uninspected files listed above are available for later inspection. There is no newly observed execution trace or causal experiment. The benchmark report initially read supplies reported operation on one author's corpus/machine with five selected queries and byte-based token estimates; it cannot establish universal recall quality, current-pin latency, or causal downstream task improvement. Tests and captures in the repository are included evidence candidates, not assumed passing tests or authenticated production outcomes (SRC-1, docs/benchmarks/recall-context-cost.md, ops/e2e/README.md).

Source register