> ## Documentation Index > Fetch the complete documentation index at: https://git.parsica.ai/llms.txt > Use this file to discover all available pages before exploring further. # SECURITY # Security policy ## Reporting a vulnerability Please report suspected vulnerabilities **privately** - do not open a public issue for an unfixed security problem. Use this repository's **GitHub Security Advisories** ("Report a vulnerability" under the Security tab) so the report reaches the maintainers privately. We aim to acknowledge reports within a few business days. ## What Weave does with your data * **Everything stays local.** The store is a single workspace under `~/.claude/parsica-weave/`: shared journal and lessons at the top level, each project's mirror and machine state under `mirror//` and `projects//`. Weave has no network listener, opens no outbound network connections, and sends no telemetry. It is a Claude Code extension, not an air-gapped vault: text injected by hooks becomes Claude Code context, and optional enrichment workers process staged chunks through the configured Claude deployment, so that text may be included in model requests according to your Claude Code and Anthropic settings. * **No live cross-session context, by design.** Weave never shares one session's LIVE context (its active window, its in-progress work) with another session or project, and there is no setting that enables this. What may cross a project boundary is only DURABLE, already-captured, already-scrubbed memory: by default on demand through a query, or (opt-in) a small one-time boot digest of recent captured chunks. A snapshot of the past, never a window into the present. Durable learning is shared; live context is separated. * **Capture is scrubbed before it is written.** High-confidence secrets and PII are redacted at collection, at the single store-write seam every writer crosses, so a raw secret does not reach the durable record. * **Enrichment is local.** Semantic overlays are generated and stored locally as plain text; no embeddings or model artifacts are persisted, and no content is sent to any external service by Weave itself. * **Provenance stays in records, never in output.** Stored records may carry local provenance labels (project keys, session ids) that you own; every output surface - query display, plan headers, session injects, digests - passes the refusal fence before anything is shown, so a fenced term riding a label never reaches a screen or transcript. * **Retention rotates, never deletes.** Files that age past their active window move to `archive/` directories inside the workspace; no code path in Weave unlinks user data. Removal is always the user's own explicit action. ## Security surface Weave is a local Claude Code plugin. Its surface is local, not networked: * **Filesystem containment** - everything lives under one workspace directory; per-project material is keyed by git root and its paths are validated for containment under `mirror/` and `projects/`. * **Session isolation** - `PARSICA_ISOLATED=1` gives a session full read access and no durable writes: nothing is captured, no state advances, no files rotate. * **Staging containment** - the enrichment pass stages content into a temporary inbox that is refused if it resolves inside the store, if any component of its path is a link or junction, or if an existing unmarked directory is non-empty. The inbox must be a directory private to the current OS user; it is cleared only of Weave's own staging files, never a caller's directory. * **Hook commands** - the five continuity hooks run local Python from the plugin directory. ## Threat model Weave defends against malformed data and pre-existing redirected, linked, shared, or non-regular objects beneath its managed paths - a hostile project, a legacy artifact, a pre-planted symlink, junction, or hardlink, or ordinary corrupt input. Every write and read gateway proves the file or directory it is about to touch, not just the path that named it. Weave does not claim isolation from another process running as the same OS user with concurrent write access to Weave's private storage. A process with that access already has enough authority to read, replace, or corrupt the store directly; renaming a directory mid-operation is not a stronger position than doing so. Defending against a concurrently-writing same-user process is tracked as hardening, not a v1.0 boundary - see Known limitations below. ## Known limitations (tracked for v1.0.1) * **Directory authority is proved by path, not held.** Gateways re-verify the full lexical path of what they are about to touch, but between that proof and the operation completing, nothing holds the directory open. A concurrent same-user process that renames a managed directory in that narrow window (and replaces the old name with a link back to it) can redirect the effect before a later check refuses it. Closing this fully means routing every managed-path operation through held directory handles (POSIX `dir_fd`-relative opens; Windows native handles without `FILE_SHARE_DELETE`) instead of repeated path checks. * **The enrichment inbox marker is a sentinel, not authenticated ownership.** `.parsica-weave-inbox` is a known, fixed filename; on platforms or configurations where the inbox's parent directory is not private to the current OS user, another process could pre-create a directory and marker with that name before Weave does. The default inbox path requires a user-private parent for this reason. A fresh, unpredictable run directory recorded in protected Weave state (rather than a static well-known marker) is the planned v1.0.1 replacement. ## Supported versions Security fixes are provided for the latest released version. ## Related topics - [From the first user.](/Agent-Handoff.md) - [TO THE NEXT AGENT](/TO_THE_NEXT_AGENT.md)