Reported issues for rag-rat

Pod holds 23 of 75 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user.

Back to rag-rat.

Most discussed

rag-rat VS Code lens: WS-backed extension (+ optional bundled web-VS-Code appliance)

Idea (north-star)

A read-only VS Code lens over the rag-rat index: a familiar editor with flamegraph/heat-style overlays for duplication, hotspots, and fan-in/out, plus a guided branch-review mode. Not an IDE you edit in — "Google Maps for the repo."

Architecture decision (supersedes the earlier "custom Monaco app" framing): ship a VS Code extension, not a hand-built editor. Monaco is only the editor widget; a web VS Code is the whole shell for free (file tree, search, git, na

Read the thread · 2026-06-18 · closed · 4 comments

SCIP oracle epic — live LSP hardening and additional backends remaining

Tracking — decomposed into sub-issues

This issue is now the epic; the design below is the reference. Slices, in dependency order:

• #67 — Prerequisite: persist callee-identifier byte range on edges (additive, byte-identical; unblocks everything)
• #68 — Phase 1: SCIP protobuf reader + heuristic precision/recall eval (depends #67)
• #69 — Phase 2: oracle run for Rust — Compiler tier, resolved-external, compare_graph_to_scip (depends #68)
• #70 — Phase 3: monik

Read the thread · 2026-06-11 · open · 4 comments

opencode plugin bundle (MCP + hooks)

Context

The repo ships a shared plugin bundle under plugin/ that serves Claude Code (.claude-plugin/plugin.json + hooks/hooks.json) and Codex (.codex-plugin/plugin.json + .mcp.json), with a single shared launcher (plugin/scripts/launch.js), shared skills (plugin/skills/), and the harness-neutral rag-rat agent-hook stdin/stdout handler. Both harnesses use the same hook contract: stdin JSON hook_event_name / tool_name / tool_input → stdout `hookSpecificOutput.addi

Read the thread · 2026-07-19 · closed · 3 comments

chunk_fts (contentless FTS5) docsize corruption is invisible to every integrity check and breaks memory_search via bm25; needs detection + self-heal

Symptom

The repo_memory_fts FTS5 index (memory search) got corrupted on the shared/consolidated DB. memory_search (MCP) returned MCP error -32603: database disk image is malformed on every call, while index_status (MCP) and rag-rat memory list (CLI, fresh connection) worked fine.

The on-disk DB was not structurally corrupt — only the FTS5 shadow index was, which PRAGMA integrity_check doesn't fully validate.

What fixed the disk, and what didn't fix the server

Repairin

Read the thread · 2026-07-10 · closed · 3 comments

fleet hot-upgrade: CLI-won worktree election must carry fleet_bin (auto-trigger shipped in #510)

Observed live: after cargo install --path crates/rag-rat-cli replaced the installed binary (new inode confirmed), two armed long-lived rag-rat mcp servers (RAG_RAT_UPGRADE_BIN present in environ) were never signaled — well past FLEET_DEBOUNCE/FLEET_MAX_LATENCY (0.5s/2s) and across multiple request boundaries. A manual kill -USR1 re-exec'd both immediately, so the arming and the re-exec path work; the trigger side didn't fire.

Likely cause to verify: the per-worktree watcher election can b

Read the thread · 2026-07-08 · open · 3 comments

sync: add the node_anchors/1 content op carrying portable binding facts

The carrier for cross-account anchor replication (#1180): a new /3 op kind snapshotting a memory's portable binding facts.

MemoryOp::NodeAnchors { node_id, anchors }, where each anchor carries exactly the /5 portable column set — binding_kind, binding_id, path, start_line, end_line, commit_hash, tracker, project, item_key, created_at_ms, symbol_kind, signature_hash, moniker_tool, moniker_tool_version. Checkout-local resolution state (logical_symbol_id, `symbo

Read the thread · 2026-08-18 · closed · 2 comments

Point rag-rat init MCP hint at add-mcp (Rust hook config-writer retired; blocked on #545)

Once the plugin + @rag-rat/setup land, retire the Rust config-writing so per-agent config no longer lives in the binary.

• MCP: point rag-rat init at @rag-rat/setup / add-mcp instead of printing the claude/codex mcp add command.
• Hooks (migration hazard): existing installs carry rag-rat claude-hook entries in .claude/settings.json; installing the plugin without removing them double-fires every hook. Staged: keep hooks uninstall --claude + hooks status for a deprec

Read the thread · 2026-07-09 · open · 2 comments

memory_validate 'relocate' leaves stale binding symbol_id/logical_symbol_id (and likely orphan symbol rows after full rebuild)

Symptom

After memory_validate (which reported relocated: 1), a symbol-bound memory still carries stale ids in its binding. In a single find_callers response for install_path:

• live symbol resolves to symbol_id: 37112, logical_symbol_id: 9128717097963335254
• the bound memory (mem_19eae08c32d_6d77d453c833) binding shows symbol_id: 221740, logical_symbol_id: 8973413979112100636, chunk_id: 283741, anchor_status: "relocated"

Both come from the same DB read, so this i

Read the thread · 2026-06-09 · closed · 2 comments

Most recent

sync: demote stale synced anchors in memories_for_* surfaces

The staleness rider for #1180, separable from the rest of the train and shippable independently.

source_text_hash rides the node_anchors op; the drain stamps it onto the synced repo_memories row, which is inserted NULL today. On this carrier neither an anchors/1 spec bump nor its backfill is needed — the two costs the /5 route would have incurred.

Policy, settled: on hash divergence, demote to a marked stale-style presentation, never hide. Exact-text hashing cannot distinguis

Read the thread · 2026-08-18 · open · 1 comment

feat(mcp): declare the per-request worktree parameter in tool schemas

Per-request worktree scoping is implemented but undiscoverable. call_tool_for_config (crates/rag-rat-mcp/src/tools/mod.rs) reads a common worktree field from the request and calls use_worktree_scope; the fallback (worktree_arg) is the MCP server process's own cwd. But worktree is declared in no tool's JSON schema — the schemas are generated from the per-tool arg structs, none of which carry it — so a client scoping reads to a linked worktree has no way to learn the parameter exists,

Read the thread · 2026-08-18 · closed · 0 comments

is_pascal_case treats a SCREAMING_CONST with a method chain as PascalCase, so classify_call returns Wrapper

Description

is_pascal_case cannot distinguish a SCREAMING_CONST followed by a method chain from a genuine PascalCase variant, so classify_call returns Wrapper for a call whose head is a constant.

Concrete case, at crates/rag-rat-mcp/src/server/service.rs:43:

crate::tools::TOOL_NAMES.iter().map(|name| { … })

The head segment here is TOOL_NAMES.iter().map(…). is_pascal_case sees a leading T plus the lowercase characters contributed by the appended method names

Read the thread · 2026-08-02 · closed · outside contributor · 0 comments

sync: /5 table-stream transport with bounded scoped manifest

Goal

Expose the transport-independent table-sync engine over iroh without registering a production table yet.

Part of #892. Follows #1049 / #1095.

Scope

• add a dedicated table-sync ALPN and endpoint dispatch alongside account and content sync;
• exchange a bounded canonical manifest of supported (repo_id, incarnation_ref, scope_id, stream_id) streams after mutual account authorization;
• derive and verify each advertised stream against local current-incarnation authority; the manife

Read the thread · 2026-08-01 · closed · 0 comments

Rotate the discovery tag so a removed device cannot enumerate hosts

Split out of #1080, which now covers sealing announcement payloads and the service-side eviction change. This issue is the half those do not close.

The gap

The discovery tag is sha256(domain || account genesis entry hash). The genesis hash is immutable and already sits in every enrolled device's database, so a device removed from the roster can compute the tag forever. Sealing the payloads (#1080) stops it reading the node ids of hosts that begin advertising after its removal, but it can

Read the thread · 2026-07-31 · open · 0 comments

Skill sources and their shipped plugin copies can drift silently

.agents/skills/<name>/SKILL.md and plugin/skills/<name>/SKILL.md are byte-identical copies of the same document, and nothing checks that they stay that way.

.claude/skills/ and .codex/skills/ are symlinks to the .agents source, so those three always agree. The plugin/ copy is a real file — it is what ships to users who install the plugin — and it is maintained by hand. Editing one and not the other produces no build failure, no test failure, and no lint: the repo's guidance and the

Read the thread · 2026-07-30 · closed · 0 comments

Upgrade rmcp to 3.0 (MCP 2026-07-28 support)

rmcp 3.0 is out and adds support for MCP protocol version 2026-07-28. The MCP server currently pins rmcp 2.2.

What the major bump requires

• ServerHandler::call_tool now returns the MRTR-aware CallToolResponse (Complete | InputRequired | Task) instead of CallToolResult. Every rag-rat tool answers in one round trip, so the handler always returns the Complete arm.
• Paginated results (ListToolsResult) gained result_type, ttl_ms, and cache_scope fields (SEP-2322 / SEP-254

Read the thread · 2026-07-30 · closed · 0 comments

Lens hop endpoints resolve by qualified name, collapsing overloads

Follow-up from #216.

IndexDatabase::lens_symbol_callers and lens_symbol_callees (crates/rag-rat-core/src/index/query_api/lens/hops.rs) take a qualified_name: &str and hand it straight to find_callers_with_options / trace_callees_with_options. Exact traversal expands a qname to every symbol id that carries it, so for a file with overloads sharing a qualified name, clicking one overload's CodeLens reports the callers of all of them. /api/file/graph has the same gap on the way in: ea

Read the thread · 2026-07-29 · closed · 0 comments

index: resolve locally-rooted Rust imports to their declaring scope instead of keeping the bare name

Rust receiver-type inference canonicalizes a type mention at its reference site by prefixing the enclosing inline-mod chain. That is correct for a path written out in full, but wrong for a name introduced by a use: the imported item lives at the use's path, which the lexical module chain does not describe.

#976 makes an imported bare name canonicalize to itself — dropping the module prefix rather than fabricating one. On this repo that removed 8 distinct invented owner paths across 24 cal

Read the thread · 2026-07-29 · open · 0 comments

Add first-class Cursor hook adapter

Implement a dedicated Cursor hook adapter using Cursor's native lifecycle and payload contract.

Scope:

• Normalize session start, shell execution, file read, and successful file edit events.
• Emit only bounded rag-rat context through Cursor's documented context channel.
• Trigger detached scoped reindexing after edits.
• Enable clone checks only when Cursor provides adequate pre-edit text or diff data.
• Keep malformed input and missing configuration/index/binary fail-silent and non-blocking.

Read the thread · 2026-07-28 · closed · 0 comments

sync: derive peer capability during auth and reject read-only pushes

Part of #930, toward #406.

Gap

Enrollment now persists a three-level roster role and the content fold rejects unauthorized authoring, but the transport auth phase returns only success/failure. The symmetric session therefore still accepts Entries frames from a peer enrolled as read-only and relies solely on the durable ingest/fold backstop.

Scope

• Return a per-connection peer capability from node authentication.
• Derive read/write capability from the peer's effective roster role: `

Read the thread · 2026-07-28 · closed · 0 comments

Split rag-rat-mcp server runtime, service, dispatch, and tests

Scope

Decompose crates/rag-rat-mcp/src/server.rs into a curated server/ module:

• mod.rs: RagRatService and stable exports
• service.rs: ServerHandler and dormant behavior
• dispatch.rs: call adaptation, blocking chokepoint, timeouts, inflight accounting, nudges
• runtime.rs: STDIO lifecycle
• tests.rs: existing tests

Invariants

Dormant MCP behavior, tool shapes, worker limits, inflight accounting, read timeouts, write deadlines, and panic conversion remain unchanged. No

Read the thread · 2026-07-24 · closed · 0 comments

sync: serve replicates only the account log — content needs a wire stream discriminator

`rag-rat sync serve` (from #905) replicates the account log only. The session wire (`Frame::Hello`/`Frame::Auth`) carries the account id and entry hashes but no discriminator for which stream a session moves — the store the caller passes decides the meaning (`OplogSyncStore` = account log, `OplogContentSyncStore` = /3 content). So an acceptor cannot tell account-vs-content apart from the hello, and a serve peer can replicate exactly one stream type per ALPN.

Serving content (the

Read the thread · 2026-07-24 · closed · 0 comments

distill: surface distilled records in MCP tool docs + the CLI query --json output

Part of #705 (scope item 4 — the final item). The distilled decision records now attach to retrieval + drive-by surfaces (items 1-2), but they are undocumented in the MCP tool descriptions and absent from the CLI.

• MCP tool docs: the tool descriptions (rag-rat-mcp/tools/catalog.rs) for the surfaces that now carry records — semantic_search, symbol_lookup, impact_surface, read_chunk (drive-by), and papertrail_issue_search / rationale_search (payload) — do not mention the `distilled_records

Read the thread · 2026-07-22 · closed · 0 comments

Watcher never controls WAL checkpoint cadence under churn — default passive autocheckpoint amplifies main-db write-back

Problem

maybe_checkpoint_wal only runs on quiet passes (crates/rag-rat-core/src/watch/pass.rs:468; quiet_pass at :351), and under sustained editing quiet passes almost never happen (67/107 measured passes had overlays_changed=true). Checkpoint cadence therefore falls to SQLite's default ~4 MiB passive autocheckpoint, which fires dozens of times mid-pass while ~170 MB/pass streams through the WAL. Long-lived readers keep the WAL from resetting (it plateaus around 20 MB), and each la

Read the thread · 2026-07-21 · closed · 0 comments

The remaining reports are on the project's issue tracker.