Skip to content

CHANGELOG.md

Latest commit

 

History

History
875 lines (658 loc) · 74.5 KB

CHANGELOG.md

File metadata and controls

875 lines (658 loc) · 74.5 KB

Changelog

v0.12.5

Dual-Layer Memory

  • Atomic fact extraction — same reindex --extract-facts extracts searchable facts from notes via LLM. Facts are independently searchable and boost source notes in search results. The right answer surfaces even when the fact is buried in an unrelated conversation.
  • same facts command — view, search, and manage extracted facts.
  • Fact-boosted search — hybrid search now includes fact matching as an additional signal.

Streamable HTTP MCP Transport

  • same web --mcp enables HTTP MCP endpoint — MCP clients can connect over HTTP instead of stdio. Bearer token authentication. Ready-to-copy config snippets for Claude Code and Cursor.
  • Unlocks Open WebUI and LobeChat — previously blocked by stdio-only transport.

Retrieval Quality

  • Turn-level chunking — conversational content (User/Assistant patterns) is now chunked by turn-pair instead of by heading, making individual facts in chat logs independently searchable.
  • Keyword boost in hybrid search — literal text matches get a 1.5x score boost, preventing strong keyword matches from being buried by semantically similar but wrong results.

CLI Configuration

  • same config set — set config values from CLI using dot notation: same config set ollama.url http://host.docker.internal:11434.
  • Global config — ~/.config/same/config.toml applies to all vaults. Set Ollama endpoint once.
  • same config edit --global — edit global config directly.

New MCP Tools

  • mem_restore — undo mem_forget (unsuppress a previously hidden note).
  • mem_list_suppressed — show all suppressed notes in the vault.

CLI Improvements

  • same index aliased to same reindex — one canonical command.
  • --content-type, --relationship, --direction flag aliases — clearer naming.
  • --sources flag on same add — CLI provenance tracking.
  • same display shows current mode — no-arg invocation prints state.
  • Improved search fallback messaging — actionable instructions when using keyword mode.
  • Ollama retry gating — prevents redundant retries when Ollama is unavailable.

Container Support

  • host.docker.internal allowed as Ollama endpoint — container users (Docker, OrbStack, Codespaces, devcontainers) can now point SAME at the host machine's Ollama.

Chat Model Configuration

  • SAME_CHAT_MODEL env var now works — override the chat model for consolidation, ask, and brief commands.
  • [chat] model config key — set a persistent chat model override in .same/config.toml. Precedence: env var > config > auto-detect.
  • Consolidation progress display — shows which model is being used at start, and per-group elapsed time during processing.

Claude Code Memory Import

  • same import detects Claude Code memory files — auto-scans ~/.claude/memory/ (global) and .claude/projects/*/memory/ (project-scoped). Imports with SAME frontmatter, provenance tracking, and trust_state: unknown. Skips MEMORY.md index files and de-duplicates on re-import.
  • Provenance pipeline for imported notes — provenance_source and provenance_hash frontmatter fields are now parsed by the indexer and recorded in note_sources.
  • Auto-index after import — imported files are immediately searchable via keyword index without needing a separate same reindex.

Vault UX

  • Vault feedback on every command — prints "Using vault: ()" so you always know which vault is active.
  • Ambiguity warning — when cwd has multiple vault children, shows a clear warning with options instead of silently picking one.
  • Single child auto-select — if cwd isn't a vault but has exactly one child vault, auto-selects it.
  • same config show displays sources — shows both global and vault config file paths with effective merged values.

Diagnostics & Error Handling

  • Build hash in version output — same version now shows same 0.12.5+abc1234 so different builds from the same version are distinguishable.
  • Binary shadowing detection — same doctor warns when multiple same binaries exist in PATH with different checksums.
  • Embedding errors logged — embedding provider failures now print the actual error to stderr instead of silently degrading to keyword search.
  • Graph LLM errors logged — graph extraction initialization failures now explain why regex fallback is active.
  • MCP JSON error handling — 8 json.MarshalIndent calls now return proper error responses instead of potentially sending corrupt JSON.
  • Resolved URL in config show — same config show displays the effective Ollama URL after raw config values.
  • Reindex lockfile — prevents concurrent same reindex runs from overloading Ollama. Stale locks from dead processes are automatically reclaimed.
  • Demo improvements — clearer demo flow and error handling.
  • Health grace period — avoids false positives on recently indexed vaults.
  • Doctor warning improvements — clearer messaging for common issues.

Security

  • Provenance trust boundary — provenance_source frontmatter is only trusted for notes in the imports/ directory (created by same import). MCP save_note cannot write to imports/ and cannot plant external provenance.
  • Agent ownership on mem_forget — when called with an agent parameter, only the creating agent can suppress a note. Vault owner (no agent) can still suppress anything.
  • Context usage path validation — auto-injected provenance paths from context_usage are now validated against vault boundaries before reading or hashing.
  • Import file permissions — imported files use 0600 and import directories use 0700.
  • Provenance cleanup on delete — DeleteByPath now removes note_sources rows in the same transaction, preventing stale provenance from persisting across note replacements.
  • 28 credential detection patterns — AI APIs, cloud providers, git tokens, payment keys, observability tools.
  • Guard on MCP save_note — warns on detected credentials before write.
  • Path validation hardening — URL-encoded traversal, Unicode fullwidth, and symlink escape detection.
  • MCP error sanitization — no internal paths leaked in error responses.
  • Note sources cleanup — stale provenance records cleaned up properly.
  • Consolidation sanitization — user input in consolidation is sanitized before LLM processing.

Bug Fixes

  • Stale note timing now shows source file modification time — same health and staleness hooks previously showed "just now" for all stale notes after reindex. Now correctly shows when the source file was actually modified.
  • CHANGELOG.md excluded from indexing — added to default .sameignore patterns to prevent oversized chunks.
  • Decision attribution preserved on append — save_decision no longer rewrites file-level agent frontmatter when appending.
  • Dimension mismatch handling — graceful fallback when embedding dimensions don't match stored vectors.

CI

  • Lint CI — golangci-lint integrated into CI pipeline.
  • npm-publish graceful skip — no longer fails when the version already exists on npm.
  • GitHub Actions upgraded to Node.js 24 — all actions pinned to Node.js 24-compatible versions.

v0.12.1

Search & Retrieval

  • Metadata search filters — filter search results by trust state, content type, domain, and tags via CLI flags (--trust stale, --type decision, --tag auth) and MCP parameters
  • same stale command — convenience command to list all stale notes in your vault
  • Content-type search boosting — queries mentioning "session" or "handoff" automatically boost handoff results; "decided" or "decision" boost decision results. Subtle 1.2-1.3x multiplier breaks ties without overriding strong semantic matches
  • Trust-aware search display — all CLI search results now show color-coded trust state (green validated, yellow stale, red contradicted). Hidden for unknown.
  • 93.3% Recall@5 on held-out validation set (30 blind test cases, semantic search, nomic-embed-text). Evaluation methodology documented in eval/METHODOLOGY.md

Smart File Exclusion

  • .sameignore file — gitignore-style patterns for excluding files from indexing. Auto-created on same init with smart defaults (node_modules, .git, binaries, lock files, IDE config, build artifacts)
  • same ignore command — view current patterns, same ignore add "*.log", same ignore reset
  • Integrated into both the indexer and the file watcher

Knowledge Graph

  • Tag-based graph connections — notes sharing frontmatter tags are automatically connected via entity nodes. No LLM required. If note A has tags: [auth, api] and note B has tags: [auth, security], they're connected through the shared "auth" entity. Domain field also creates connections.
  • Interactive graph visualization — Cytoscape.js force-directed graph in the web dashboard. Nodes colored by type (decision=teal, handoff=purple, entity=amber), sized by edge count. Hover highlights connections, click navigates to notes. Layout switcher (cose, circle, grid). Mini-graph on dashboard page.

Briefing

  • Trust-aware briefing — same brief now includes trust state annotations and provenance sources. Stale decisions are flagged with warnings. Validated notes get checkmarks.
  • --no-llm mode — structured data-only briefing without LLM dependency. Shows current focus, key decisions with trust tags, stale context, recent activity.
  • Graceful LLM fallback — if the LLM is unavailable, brief automatically falls back to the structured view instead of erroring.

Performance

  • Index-now-embed-later — FTS5 keyword indexing happens instantly (sub-second for 100 notes). Vector embeddings are backfilled in the background. Search works immediately via keyword fallback while embeddings process. Progress display: "Embedding: 15/35 notes (keyword search active)".
  • Ollama model unloading — after reindex completes, SAME sends keep_alive: 0 to Ollama to free the embedding model from memory. Prevents stale runner processes consuming CPU.

Web Dashboard

  • Complete visual redesign — DaisyUI + Tailwind CSS dark theme. Glassmorphism cards, gradient stat numbers, trust badges on every note.
  • Trust Overview section — 4-card grid showing validated/stale/contradicted/unknown counts with color-coded numbers.
  • Sidebar + content layout — proper side-by-side layout with fixed 220px sidebar, sticky navigation.
  • Confidence bars on note cards showing confidence level.
  • Freshness indicators and keyboard shortcut hints.

Bug Fixes

  • Safe force reindex — same reindex --force no longer deletes the existing index before re-indexing. Notes are deleted individually before their replacements are inserted. A failed reindex preserves existing data.
  • Brief empty vault message — when noteCount=0 but markdown files exist on disk, shows "X files not indexed, run same reindex" instead of the misleading "vault is empty"
  • RawToSearchResult refactor — replaced 8 manual SearchResult construction sites with a single helper function, preventing field omission bugs (7 sites were missing TrustState)

Testing

  • Three-tier evaluation methodology — internal eval (68 tuning cases), held-out eval (30 blind cases), MemoryAgentBench adapter (external benchmark). Documented in eval/METHODOLOGY.md.
  • Eval suite — 35-note curated vault, 68 test cases, bash + Go runners, Recall@5 and MRR metrics
  • MemoryAgentBench adapter — Python adapter for ICLR 2026 benchmark (146 test cases, 4 splits)
  • Graph extraction training data — 30 training + 10 validation examples for fine-tuning

CI

  • npm-publish graceful skip — no longer fails when the version already exists on npm
  • GitHub Actions upgraded to Node.js 24 — all actions (checkout, setup-go, upload-artifact, download-artifact, setup-node) pinned to Node.js 24-compatible versions
  • Smoke test updated — accepts both lite and progressive index modes

v0.12.0

Memory Integrity

  • note_sources table for provenance tracking — records what files/notes each note was derived from, with SHA256 hashes at capture time
  • trust_state field on notes: validated, stale, contradicted, unknown
  • same health now shows Trust section: validated/stale/unknown counts with specific stale sources listed
  • Health score updated to 5-factor model (added trust factor)
  • MCP save_note accepts sources parameter for explicit provenance tracking
  • Graph extraction records discovered file references as provenance sources
  • Staleness hook enhanced to detect source file divergence
  • Reindex serializes graph extraction after embeddings (fixes Ollama timeout on constrained machines)
  • Fixed NewClientWithURL timeout (10s → 120s) for production Ollama usage

Trust-Aware Retrieval

  • Trust penalty in search scoring — stale notes rank 25% lower, contradicted notes 60% lower. Validated and unknown notes are unaffected.
  • trust_state field returned on all search results (vector, keyword, hybrid, FTS5, federated) — MCP clients and agents can caveat answers based on trust
  • Context surfacing tags stale/contradicted notes visibly in injected context so agents know when retrieved knowledge may be outdated
  • Graph 1-hop expansion — top vector results are expanded through graph edges to surface related notes (decisions, references, dependencies). Max 2 supplemental results at 60% dampened score.

Kaizen

  • same kaizen command for continuous improvement tracking — log friction, bugs, and ideas as you work
  • save_kaizen MCP tool for agent-driven improvement logging with provenance tracking
  • Kaizen items shown in same health recommendations

Crash Resilience

  • PreCompact hook wired for session checkpointing — handoff notes are saved before context compaction, not just on session stop
  • Separate debounce (2min) for checkpoint vs full handoff (5min)

Added

  • same tips command with vault hygiene, security, and model selection guidance
  • same graph enable / same graph disable commands to toggle graph mode without editing config files
  • Automatic container environment detection (Docker, Kubernetes, Codespaces, Gitpod) — surfaces in status and doctor output
  • Graceful Ctrl+C cancellation during reindex and init — first press stops cleanly with partial progress, second press force-quits (Mac, Linux, Windows)
  • Human-readable error messages for common embedding failures: connection refused, timeout, auth errors, dimension mismatches
  • Graph LLM recommendation surfaced across init, demo, status, doctor, and tips when a capable model is detected but graph mode is off
  • Thinking model compatibility: stripThinkingTokens() utility strips <think>, <reasoning>, <reflection> tags from LLM responses
  • Graph extraction: [graph] model config key and SAME_GRAPH_MODEL env var for choosing a non-thinking model
  • Graph extraction: --abort-on-error flag (continue-on-error is now the default)
  • Graph extraction: Ollama structured output with JSON schema for more reliable parsing
  • Dockerfile: Multi-stage build, non-root user, OCI labels
  • Command aliases: same s (search), same st (status), same vault ls (vault list)
  • Port-in-use detection for same web
  • Platform shown in same version output
  • same consolidate command: merge related notes into structured knowledge using LLM [experimental]
  • same brief command: orientation briefing of what matters right now [experimental]
  • same health command: vault health score with actionable recommendations [experimental]
  • MCP tools: mem_consolidate, mem_brief, mem_health, mem_forget for autonomous memory management
  • Note suppression: mem_forget MCP tool marks notes as suppressed (hidden from search, not deleted)
  • Reconsolidation dynamics: frequently accessed notes rank higher in search results
  • Windows ARM64 release binary

Fixed

  • Windows self-update no longer fails when a stale .old backup file is locked from a previous update — uses fallback rename chain
  • Migration failure upgrading from v0.9.1 to v0.10.0 — entry_kind index no longer created before the column exists
  • same graph stats now reads from config.toml instead of only checking the environment variable
  • Graph LLM extraction now works with thinking/reasoning models (DeepSeek-R1, QwQ, etc.)
  • same ask, demo, and tutorial commands no longer display thinking tokens
  • Ollama and OpenAI response paths strip thinking tags at the transport layer
  • URL redirect vulnerability fixed with page whitelist in web dashboard
  • MCP SDK bumped to v1.4.0 (security fix)

Changed

  • Demo rewritten with 5 realistic sample notes and a narrative arc (search, decisions, handoff recall) — transitions into init with a single prompt
  • Init onboarding redesigned: detects project language, AI tools, and git state; adds teaching moments during directory creation; suggests seed vaults based on project type
  • install.sh messaging updated to honestly communicate Ollama's role in semantic search
  • Output consistency polish: standardized checkmarks, hint capitalization, footer formatting
  • README overhauled: conversion-focused, 270 lines, collapsible reference sections
  • Search results now factor in access frequency (subtle log-scaled boost)
  • Schema migration v8: adds suppressed column to vault_notes
  • All search paths filter suppressed notes by default

Performance

  • Batch embedding requests: Ollama switched from /api/embeddings to /api/embed, OpenAI batching added — 50 chunks per request instead of 1
  • SQLite pragmas: 64 MB page cache, 256 MB mmap, temp_store in memory
  • Covering index added for incremental reindex hash comparison
  • ReindexLite parallelized with 4-worker goroutine pool (was sequential)

v0.9.1 — Cross-Platform CI and Extraction Reliability

Fixed

  • Windows build failure (Setpgid is Unix-only) — background web process attributes now resolve via platform-specific helpers so Windows cross-compiles no longer fail
  • Graph extraction junk decision nodes from code blocks — decision extraction now ignores fenced code block content to avoid shell/example noise
  • Graph extraction treating URLs as local files — URL-like references are filtered out from file-node extraction
  • Graph extraction placeholder/template path nodes — placeholder and template-style paths are excluded so graph nodes reflect real vault references

Added

  • Cross-compile CI check for Windows and Linux — CI now runs GOOS={windows,linux} GOARCH=amd64 CGO_ENABLED=0 go build ./cmd/same on every push/PR
  • Linux CI test runner — CI now executes go test ./... -count=1 on ubuntu-latest with CGO_ENABLED=1

v0.9.0 — Knowledge Graph, Provider Flexibility, and Hardening

Added

  • Graph LLM extraction policy controls — new SAME_GRAPH_LLM=off|local-only|on setting gates optional LLM-based graph enrichment during reindex (off by default, local-only for localhost-only chat endpoints)
  • Manual CI release gate (Release Candidate workflow) — maintainers can now run baseline release checks (and optional full provider matrix) on demand via GitHub Actions workflow dispatch
  • Verified self-update downloads — same update now requires sha256sums.txt from releases and verifies the downloaded binary checksum before install
  • pi advanced profile preset (same profile use pi) — tuned memory retrieval thresholds for Raspberry Pi / low-resource setups, with CLI guidance for lightweight embedding choices
  • Provider-agnostic chat routing for same ask + graph LLM extraction — new SAME_CHAT_* controls support auto, ollama, openai, and openai-compatible providers with optional fallback queues
  • Knowledge graph command group (same graph) — new CLI for graph traversal and diagnostics:
    • same graph stats — node/edge counts by type and relationship
    • same graph query — depth-limited recursive traversal from a start node
    • same graph path — shortest path between two nodes
    • same graph rebuild — clear and fully rebuild graph links from indexed notes (references/decisions/agents)
  • Schema migration v6 (graph_nodes, graph_edges) — graph tables are now first-class in the versioned migration pipeline
  • Incremental graph sync test coverage — added regression coverage for graph cleanup on note delete/force clear and lite single-file updates
  • Web graph APIs (/api/graph/stats, /api/graph/connections/{path}) — dashboard can now read graph density/relationship stats and fetch per-note traversal paths with input validation + private-path filtering
  • Web dashboard graph UX — added a graph highlights section (nodes, edges, avg degree, top relationship types) and a note-level “Knowledge Connections” panel that renders relationship-labeled paths
  • Tutorial graph lesson (same tutorial graph) — onboarding now includes a hands-on relationship traversal lesson showing how links become graph paths
  • Example configuration files — added examples/ reference configs for Ollama, OpenAI, OpenAI-compatible, keyword-only, and Raspberry Pi setups, plus MCP integration examples for Claude Code and Cursor
  • Shell completion command (same completion) — added CLI generation for bash, zsh, and fish completion scripts via Cobra's completion support
  • Docker support artifacts — added multi-stage Dockerfile, .dockerignore, and docs/docker.md usage guide (keyword-only by default with optional external Ollama)
  • make lint target + .golangci.yml — added project lint configuration and Makefile integration for errcheck, govet, staticcheck, misspell, goimports, and related checks
  • Man page generation (same gendocs) — added hidden CLI command + make man target to generate man pages from the Cobra command tree

Fixed

  • Provider-neutral diagnostics and status UX — same status/same doctor now report embedding/chat/graph runtime state across ollama, openai, openai-compatible, and none modes instead of Ollama-only assumptions
  • same watch rename/delete consistency — watcher now removes stale indexed paths on file rename, cleans up entries when files vanish before debounce flush, and rejects out-of-vault relative-path escapes
  • same init --provider input validation — invalid provider names now fail fast with a clear error instead of falling into confusing runtime fallback paths
  • Seed manifest/cache parity checks — cached manifests now enforce the same seed name/path validation as fresh downloads (including rejection of embedded . / .. traversal segments), preventing trust gaps when falling back to local cache
  • same seed remove consistency safeguards — destructive path checks now run before registry mutation, root seed-dir deletion is explicitly refused, and delete failures trigger best-effort registry rollback
  • same seed install --force path safeguards — install now refuses dangerous destinations (filesystem root, home directory, or seed root parent), preventing accidental high-impact deletes on forced reinstall
  • Seed extraction path normalization — tar extraction now normalizes leading ./ manifest seed paths (compatibility with normalized relative paths) and surfaces directory-creation failures instead of silently skipping them
  • Seed extraction declared-size enforcement — extraction now rejects payloads larger than a tar entry’s declared size and propagates file close/write failures explicitly
  • Seed extraction containment checks — archive extraction boundary validation now uses path-relative containment checks to avoid prefix-confusion edge cases
  • Vault feed containment checks — source/destination path enforcement now uses filepath.Rel-based boundary checks instead of raw prefix matching for safer cross-platform behavior
  • Vault subpath guard strictness — SafeVaultSubpath now rejects absolute subpath inputs and uses boundary-safe containment checks before writing handoff/decision files
  • Guard allowlist path matching — file allowlist entries now require exact path matches (no basename-only fallback), preventing nested-file bypasses like docs/README.md
  • MCP hidden-path write guard — safeVaultPath now rejects dot-prefixed segments anywhere in the path (for example notes/.hidden/file.md), not only at root level
  • MCP path containment checks — safeVaultPath now uses filepath.Rel-based vault-boundary checks for absolute/symlink containment instead of separator prefix matching
  • Web API path validation parity — note/related/graph connection endpoints now share a single path safety guard that rejects traversal, hidden-dot segments, and Windows drive-prefix absolute paths
  • Filesystem write error handling — config saves, registry writes, MCP note/decision appends, handoff + decision-log generation, init .gitignore updates, tutorial/demo setup cleanup, seed config path rewrites/install rollback cleanup, verbose-log rotation/appends, and index stats persistence now handle write/cleanup failures explicitly instead of silently ignoring them
  • Self-update cleanup hardening — same update now surfaces temp-file cleanup failures explicitly and validates Windows backup-path cleanup before binary replacement
  • Web JSON response diagnostics — dashboard API JSON encode failures now emit explicit server-side warnings instead of failing silently
  • Init lock fallback diagnostics — same init now warns when lock enforcement is bypassed due lockfile/permission failures, instead of failing silently
  • Seed manifest cache write diagnostics — cache directory/file write failures are now surfaced as explicit warnings instead of being silently ignored after successful manifest fetches
  • Budget report export reliability — SaveBudgetReport now fails clearly when report directories cannot be created, with regression coverage for parent-path failure
  • Vault registry lock cleanup hardening — config lock handling now surfaces stale-lock removal and cleanup failures instead of silently swallowing them
  • Init lock stale-recovery hardening — stale init lock removal failures now return explicit errors, and lockfile cleanup paths now warn on removal failures
  • Graph consistency on deletes and force-clear — deleting notes now removes related graph nodes/edges and prunes orphan non-note graph nodes; force-clear now resets graph tables as well
  • Graph freshness during same watch — watcher now updates graph data in both semantic and keyword-only (provider = "none") modes
  • Keyword-only reindex UX — same reindex now reliably falls back to lite mode when embeddings are disabled (provider = "none"), with clear next-step messaging
  • Semantic reindex resilience — when embedding calls fail for every file (e.g., Ollama unavailable), same reindex now auto-falls back to lite indexing instead of finishing with an empty index
  • Reindex fast-fail preflight — same reindex now probes embedding availability once before full-file processing, reducing noisy per-file embed failures when the provider is down
  • --vault precedence — CLI --vault now correctly overrides env/config vault paths, preventing accidental indexing/querying of the wrong vault
  • Graph path lookup ergonomics — same graph path now resolves note/file type mismatches (for paths that exist as file nodes but were requested as note nodes, and vice versa)
  • Graph query readability — same graph query now reconstructs full edge sequences so output shows relationship-labeled paths, not node-only chains
  • Markdown cross-note linking — .md references now map to note nodes, enabling direct note-to-note traversal (for example, notes/a.md -> notes/b.md)
  • Graph reference-path normalization — extraction now rejects absolute/external-style path references (for example, Users/..., /..., .windsurf/worktrees) to reduce noisy out-of-vault graph nodes
  • Precheck scope clarity — make precheck now labels its blocklist scan as repo-scope release hygiene and explicitly notes what it does not cover (user vaults, full history, forks, mirrors); make precheck-full adds an all-tracked-files blocklist sweep
  • Config validation hardening — top_k is clamped to [1, 100], threshold is clamped to [0.0, 1.0], and openai-compatible now warns when base_url is missing
  • Lint fixes — resolved unused variables, import ordering, misspellings, and unchecked error returns flagged by golangci-lint across CLI/internal packages
  • Flaky seed cache test isolation — fixed TestSeedCmd_ListNoNetwork to avoid cross-test proxy environment caching issues

Documentation

  • README: Web dashboard section — added dedicated same web usage examples and a dashboard capability summary
  • README: Docker install option — added Docker build/run examples and linked docs/docker.md
  • README: shell completion setup — added bash, zsh, and fish completion instructions
  • README: CLI reference updates — added same web and same completion entries
  • README: comparison table update — added a web dashboard row to SAME vs. alternatives

Tests

  • CLI test coverage expansion — added command-level tests for search, doctor, index, completion, web, pin, feedback, log, model, seed, repair, display, hooks, and mcp

Privacy

  • No privacy boundary changes — graph extraction still operates only on indexed notes; _PRIVATE/ remains excluded because it is never indexed

v0.8.3 — Keyword-Only Mode & ARM Linux

Fixes keyword-only search (the biggest issue for no-Ollama environments), adds linux-arm64 pre-built binaries, and adds provider = "none" for permanent keyword-only mode.

Security

  • BUG-035: Hardened prompt injection sanitization — sanitizeContextTags (hooks) and neutralizeTags (MCP) now neutralize LLM-specific injection delimiters: Llama/Mistral [INST]/[/INST], <<SYS>>/<</SYS>>, and XML <![CDATA[ sequences. Previously only XML-like structural tags were sanitized
  • Fixed case-insensitive tag matching for IMPORTANT — the IMPORTANT tag was listed in uppercase but compared against lowercased text, so <IMPORTANT> injection payloads were not neutralized. Now correctly matched case-insensitively
  • MCP get_note now sanitizes output — previously returned raw note content to agents without neutralizing XML-like tags, allowing stored prompt injection via crafted notes. Now applies neutralizeTags before returning content
  • MCP get_session_context now sanitizes pinned notes and handoff text — pinned note text and latest handoff text were returned unsanitized, allowing injection via crafted pinned/handoff content
  • MCP recent_activity now filters _PRIVATE/ paths — defense-in-depth gap: recent notes listing did not filter private paths at the application level
  • Guard: added GitHub token (ghp_/ghs_) and Slack token (xoxb-/xoxp-) PII patterns — pre-commit scanner now detects these common credential types
  • Comprehensive security test suite — added dedicated security tests for injection sanitization, plugin validation, MCP input validation (agent normalization, rate limiting, tag neutralization), web dashboard middleware (localhost-only, security headers, private path filtering), path traversal prevention, claims path normalization, search term sanitization, and guard pattern detection

Fixed

  • Keyword-only search works in same search and same ask — added LIKE-based keyword fallback when FTS5 is unavailable. Previously, keyword-only mode indexed notes successfully but search returned "No search index available". The CLI now matches the web dashboard's 3-tier fallback chain: vector → FTS5 → LIKE-based keyword
  • Context surfacing hooks work in keyword-only mode — the user-prompt hook previously early-returned when no embedding provider was available, showing a diagnostic instead of searching. Now falls through to keyword search (FTS5 → LIKE), so context surfacing works without Ollama
  • same doctor no longer gives false positive for keyword search — the "Finding relevant notes" check now actually tests keyword search instead of just counting notes
  • same doctor no longer cascades errors when vault path is wrong — if vault path fails, database-dependent checks show "skipped" instead of cascading into confusing "permission denied" errors. Summary box shows actionable VAULT_PATH guidance
  • find_similar_notes gives clear message in keyword-only mode — instead of a confusing "not in index" error, explains that similar notes requires semantic search
  • Config persists active provider — same init now writes the actual embedding provider to config (e.g., provider = "none") instead of always writing provider = "ollama"
  • MCP and hooks portability — .mcp.json and .claude/settings.json now use same from PATH instead of hardcoded absolute binary paths. Existing users: run same setup mcp and same setup hooks to update for cross-machine portability
  • Search no-results guidance — same search now suggests --all flag when no results found. Federated search suggests same reindex
  • same --version flag — same --version now works (previously only same version worked)
  • Ollama error classification — connection errors now classified by root cause: connection_refused (not running), permission_denied (sandbox policy), timeout (model loading), dns_failure. Doctor, hooks, and retry logs all show classified reasons instead of generic "request failed"
  • Doctor: Ollama skip reason — when Ollama is unavailable, doctor shows why (e.g., "permission denied — localhost may be blocked by sandbox/runtime policy") instead of just "skipped (lite mode)"
  • Doctor: reindex upgrade prompt — when Ollama is running but index is keyword-only, doctor prompts to run same reindex to enable semantic search
  • Permission-denied errors not retried — sandbox EACCES/EPERM errors now fast-fail instead of retrying 3 times with exponential backoff
  • NULL agent scan regressions — search, note listing, pinned-note reads, and same doctor keyword checks now safely handle legacy rows where agent is NULL

Added

  • provider = "none" for keyword-only mode — explicit opt-out of embeddings. Use SAME_EMBED_PROVIDER=none, --provider none, or set provider = "none" in config. Skips Ollama check entirely during init
  • linux-arm64 pre-built binary — CI now builds for ARM Linux (Codespaces, OrbStack, Raspberry Pi, AWS Graviton). No more building from source on aarch64
  • same reindex --verbose — shows each file being processed during reindex, matching the same init -v flag
  • Doctor: portability checks — detects hardcoded absolute binary paths in .mcp.json and .claude/settings.json, suggests same setup mcp / same setup hooks to fix
  • Seed install --all hint — success message now shows same search --all as a next step for cross-vault search
  • same claim command group — advisory read/write file claims for multi-agent coordination:
    • same claim <file> --agent <name>
    • same claim --read <file> --agent <name>
    • same claim --list
    • same claim --release <file> [--agent <name>]
  • MCP session context git awareness — get_session_context now includes branch, recent commits, dirty files, and untracked files when available
  • Agent attribution on MCP writes — save_note, save_decision, and create_handoff now accept optional agent and persist it in note frontmatter/index metadata
  • Agent-aware MCP filtering — search_notes_filtered now supports filtering by agent
  • Web dashboard agent badges — note cards and note viewer now display @agent attribution when present
  • Dashboard polish pass — improved responsive breakpoints, keyboard search navigation (/, Esc, arrows, Enter), loading skeletons, friendlier empty/error states, print stylesheet, ARIA/focus improvements, and snippet term highlighting
  • same vault rename — rename a vault alias without re-registering. Updates the vault registry in-place

v0.8.2 — Handoff Quality & UX

Comprehensive upgrade to session handoffs based on real-world user feedback. Richer auto-generated handoffs, better session continuity, bug reporting infrastructure, and UX polish.

Improved

  • Rich auto-handoffs — now extract decision titles (from save_decision calls), notes created (from save_note calls), tool usage breakdown, user/assistant message split, and word-boundary truncation for topics
  • Forward-looking context — auto-handoffs now include a "Next steps" section extracted from the last assistant messages
  • Bootstrap parser — includes ALL sections from handoff files instead of silently dropping unknown ones. Known sections are prioritized, unknown sections appended in original order
  • Placeholder filtering — bootstrap strips sections containing only (see ...), (review ...), or (none) placeholder text
  • Configurable handoff age — hooks.handoff_max_age_days in config.toml (default: 2 days). Controls how far back SAME looks for handoff notes at session start
  • Clearer session loading message — stderr now shows "← previous session loaded" instead of the confusing "↩ recovered previous session"
  • File truncation indicator — when more than 20 files are changed, handoff shows "...and N more" instead of silent truncation
  • Large note warnings — reindex warns when notes exceed 30KB, which can reduce search quality
  • App attribution headers — openai-compatible provider now sends X-Title and HTTP-Referer headers for OpenRouter compatibility
  • --provider flag for init — same init --provider openai-compatible skips Ollama check and uses the specified embedding provider
  • Standardized handoff filenames — MCP create_handoff now uses the same date-based naming convention as auto-handoffs
  • Interactive seed selection during init — users with existing notes now see the full seed catalog with an option to install one directly, instead of a dim tip line
  • Embedding model chooser during init — shows available models (nomic, snowflake, mxbai, etc.) and lets users pick before indexing
  • Styled reindex and stats output — human-friendly formatted output instead of raw JSON dumps
  • Grouped help commands — all commands now organized into groups in --help output (no more loose ungrouped commands)
  • Readable search scores — verbose mode shows Relevance: 87% instead of Score: 0.873
  • Complete MCP tool list — same setup mcp now shows all 12 tools including search_across_vaults

Added

  • Web dashboard — same web launches a local read-only dashboard at 127.0.0.1:4078. Browse notes, search semantically, view vault composition, read decisions and handoffs, see related notes. Single embedded HTML file, no external dependencies. Localhost-only with security headers (CSP, X-Frame-Options DENY, DNS rebinding protection). Works without Ollama (keyword fallback). --port and --open flags available
  • Dashboard insights — contextual tips based on vault state: index staleness warnings, search mode guidance, seed install suggestions, content type coaching, pinned notes tips
  • Bug reporting — "Report a Bug" link in README and npm package, GitHub issue templates (bug report + feature request), same doctor shows issues URL when checks fail, MCP index_stats description guides agents to same doctor and GitHub Issues
  • Stop hook message control — handoff creation message shows once per session; subsequent updates are silent

Fixed

  • npm-publish CI — release workflow now checks if version already exists before publishing, preventing false failures on re-runs

v0.8.1 — Hotfix

Critical fix for Claude Code v2 transcript format. Stop hooks (decision extractor, handoff generator, feedback loop) were silently returning empty results because the transcript parser expected flat JSON but Claude Code v2 nests messages in a {"type":"...","message":{...}} envelope.

Fixed

  • Transcript parser for Claude Code v2 format — processEntry() now unwraps nested message envelopes and falls back to the top-level type field as a role identifier. Stop hooks (decision extractor, handoff generator, feedback loop) now correctly parse transcripts from Claude Code v2.x.
  • HookInput JSON tag mismatch — session_id and hook_event_name fields now use snake_case JSON tags matching Claude Code's actual input format. Previously used camelCase which caused silent deserialization failures.
  • Windows binary detection — detectBinaryPath() now appends .exe on Windows, checks %LOCALAPPDATA%\Programs\SAME\, and quotes paths with spaces on all platforms (not just Windows).
  • Cross-platform path handling — handoff note paths use filepath.Join instead of string concatenation. Tilde expansion handles both ~/ and ~\ for Windows.
  • Scanner error handling — transcript parser now checks scanner.Err() after reading, reporting partial data on I/O errors instead of silently truncating.
  • Provider-agnostic error messages — embedding failure message now says "check 'same doctor'" instead of hardcoding Ollama-specific instructions. Works correctly for OpenAI and llama.cpp users.
  • Doctor suggestion text — "run 'same init'" corrected to "run 'same setup hooks'" for missing hooks.
  • MCP setup prompt — "Connect to Cursor/Windsurf?" changed to "Set up MCP server for AI tools?" to avoid confusing Claude Code users.

Improved

  • Internal commands hidden — hook, migrate, budget, plugin, push-allow no longer appear in same --help. Still accessible for debugging.
  • Seed tab completion — same seed install, same seed info, and same seed remove now offer tab-completable seed names.
  • Cleaner CLI output — SilenceErrors and SilenceUsage prevent cobra from dumping usage text on errors. Seed install indexing messages simplified.
  • Corrected documentation — README, CHANGELOG, and AGENTS.md updated with accurate seed counts (10 seeds, 612+ notes) and diagnostic check counts (19).

v0.8.0 — Seed Installer

One-command install of pre-built knowledge vaults. same seed install claude-code-power-user downloads, extracts, indexes, and registers a seed vault — ready to search in seconds.

Added

  • same seed command group — browse and install pre-built knowledge vaults:
    • same seed list — show available seeds with note counts and descriptions. Flags: --refresh, --json
    • same seed install <name> — download, extract, index, and register a seed vault. Flags: --path, --force, --no-index
    • same seed info <name> — show detailed seed metadata (audience, tags, requirements)
    • same seed remove <name> — uninstall a seed vault. Flags: --yes, --keep-files
  • 10 official seed vaults — Claude Code Power User, AI Agent Architecture, Personal Productivity OS, Security Audit Framework, DevOps Runbooks, Indie Hacker Playbook, Open Source Launch Kit, Freelancer Business Kit, Home Chef Essentials, Fitness & Wellness. Browse at github.com/sgx-labs/seed-vaults
  • Seed manifest registry — seeds.json hosted in seed-vaults repo with schema versioning, 1-hour client-side cache, stale-cache fallback on network errors
  • same model command — show current embedding model and switch between models with same model use <name>. Lists all 10 supported models with dimensions and descriptions. Warns about reindex after switching.
  • Automatic lite mode fallback — seed install falls back to keyword-only indexing when Ollama isn't available
  • Version compatibility check — seeds declare min_same_version; install rejects incompatible seeds with upgrade guidance

Security

  • Tarball extraction hardening — path traversal prevention (null bytes, ../, absolute paths, hidden files), symlink + hardlink rejection, file extension allowlist (.md, .toml, .json, .txt, .yml, .yaml, .example, .gitkeep), per-file 10MB size limit, 500 file count limit, 50MB total tarball limit, final containment check after filepath.Join
  • Manifest validation — io.LimitReader on HTTP response (1MB max), schema version check, seed name validation (lowercase alphanumeric + hyphens only)
  • Install safety — atomic cleanup on failure (partial install removed), --force required to overwrite, seed remove only deletes under ~/same-seeds/ (safety boundary), vault registration uses existing lockfile pattern
  • Seed name validation — lowercase alphanumeric with hyphens, 64-char max, no leading/trailing hyphens

v0.7.5 — OpenAI-Compatible Embeddings

Added

  • OpenAI-compatible embedding provider — SAME now supports any server that exposes the OpenAI-compatible /v1/embeddings endpoint. Use provider = "openai-compatible" with llama.cpp, VLLM, LM Studio, or any other compatible inference engine. API key is optional for local servers. Configure via [embedding] in config or SAME_EMBED_PROVIDER / SAME_EMBED_BASE_URL / SAME_EMBED_MODEL environment variables.
  • Non-localhost security warning — when using an OpenAI-compatible provider with a remote base URL, SAME prints a warning that embedding requests will leave your machine. Local servers (localhost, 127.0.0.1, ::1) are silent.
  • --hooks-only flag for init — same init --hooks-only skips MCP setup for Claude Code-only workflows (mirrors existing --mcp-only for Cursor/Windsurf).
  • Expanded embedding model support — auto-detected dimensions for 5 new models: snowflake-arctic-embed2 (1024), embeddinggemma (768), qwen3-embedding (1024), nomic-embed-text-v2-moe (768), bge-m3 (1024). All recognized in dimension defaults and embedding-only model filter.
  • Provider-aware onboarding — same init detects openai/openai-compatible providers and skips the Ollama connectivity check, showing provider/model/endpoint info instead.
  • Pre-release verification gate — make precheck runs 9 automated checks (version consistency, build, tests, PII scan, git identity, JSON validation, CLI smoke test).

Fixed

  • Unknown config keys now warn — unrecognized keys in config.toml print a warning with suggestions (e.g. exclude_paths → skip_dirs). Previously unknown keys were silently ignored.
  • Welcome notes skip existing vaults — same init no longer creates welcome/ in vaults that already have markdown files. Governed vaults with existing structure are left untouched.
  • SAME_EMBED_BASE_URL missing from LoadConfig() — the env var was handled in EmbeddingProviderConfig() but not in the config-file loader, causing inconsistency when both paths were used.
  • OPENAI_API_KEY fallback for openai-compatible — LoadConfig() now checks the env var for both openai and openai-compatible providers.
  • BaseURL not passed to embedding provider — indexer.go and init.go now pass the configured BaseURL when constructing embedding providers, fixing silent failures when using openai-compatible with a custom endpoint.

v0.7.4 — Quality Pass

Added

  • CLI help grouping — same --help now organizes commands into 6 logical groups: Getting Started, Search & Discovery, Knowledge Management, Diagnostics, Configuration, and Advanced. Makes it easy to find what you need.
  • Doctor lite mode — same doctor no longer fails when Ollama isn't running. Embedding-dependent checks show "skipped (lite mode)" instead of red failures, and the summary shows "SAME is running in lite mode (keyword search). Install Ollama for semantic search."
  • Readable search scores — search results show human labels ("Excellent match", "Strong match", "Good match", "Weak match") instead of raw percentages.
  • MCP tool list on setup — same init and same setup mcp now print all 12 available MCP tools with descriptions after registering the server.

Fixed

  • Token budget starvation — context surfacing now continues scanning past oversized notes instead of stopping. If a large note doesn't fit in the remaining token budget, smaller high-relevance notes behind it can still be included.
  • Config clobber during init — SetDisplayMode() and SetProfile() now load from the target vault's config file instead of auto-detecting. Previously, when CWD != vaultPath, they would fall back to DefaultConfig() and overwrite the just-written config with empty/default values.
  • DefaultConfig embedding model — DefaultConfig().Embedding.Model now returns "nomic-embed-text" instead of empty string, preventing broken config files when the default is serialized.
  • Onboarding "1" input — confirm() now accepts "1" as yes, matching the mental model users have after the numbered experience-level prompt (1/2).
  • [ollama].model fallback — users who set a custom model in [ollama] without an [embedding] section now get their override applied correctly.
  • neutralizeTags dead variable — removed unused lower variable in MCP server tag sanitization.

Changed

  • CLI descriptions polished — bench, config, guard, vault, and watch commands rewritten for plain-language audience ("Test how fast search is" instead of "Run search performance benchmarks").
  • Hook error messages — 8 user-facing error messages improved: consistent same: prefix, no leaked internal errors, actionable hints pointing to same doctor or same init.
  • Glama MCP badge — added to README for MCP directory visibility.

v0.7.3 — Bootstrap & Vault UX

Fixed

  • Session-bootstrap hook not wired — same init now installs all 6 hooks including session-bootstrap. Previously only 5 of 6 were configured, so session orientation context was never delivered on SessionStart.
  • Wrong vault when CWD is a vault — vault resolution now checks CWD for vault markers before falling back to the registry default. If you're standing in a vault directory, that vault is used regardless of what the registry default is.
  • same init didn't set default vault — running same init in a new directory now always sets that vault as the registry default, not just on first-ever init.
  • same status vault display unclear — vault section now shows the active vault prominently with its resolution source (auto-detected from cwd, registry default, --vault flag). Registered vaults list uses → for active and * for default, sorted alphabetically. JSON output now includes all 6 hooks.

v0.7.0 — Cross-Vault Federation

Search across all your vaults from one place. Manage multiple vaults from the CLI. Propagate notes between vaults with privacy guards.

Added

  • Federated search — same search --all searches every registered vault in one query. same search --vaults work,personal searches specific vaults. Results merged by score with graceful degradation (semantic → FTS5 → keyword per vault). 50-vault limit.
  • MCP: search_across_vaults — federated search tool for any MCP client. Brings the total to 12 MCP tools.
  • same vault command group — same vault list, same vault add <alias> <path>, same vault remove <alias>, same vault default <alias> for managing the vault registry (~/.config/same/vaults.json).
  • same vault feed — one-way note propagation between vaults. Copy notes from a source vault into the current vault's fed/<alias>/ directory. Includes PII guard (scans for email/phone/SSN patterns), symlink rejection, 10MB file size limit, self-feed prevention, and --dry-run mode.
  • store.AllNotes() — returns all chunk_id=0 notes excluding _PRIVATE/, ordered by modified date.
  • 20 new tests — sanitizeAlias (14 cases), safeFeedPath (19 cases), FederatedSearch (empty query, too many vaults, private note exclusion, mixed vault health, graceful skip).
  • Progressive feature discovery — CLI teaches new capabilities at the moment they become relevant. vault add hints about --all when 2+ vaults registered. search hints about same related. reindex hints about same watch. status shows available vaults and same ask when a chat model is detected. doctor validates vault registry health (18 checks total).
  • Pinned notes in session bootstrap — pinned notes now survive context compaction. Previously only surfaced during per-prompt context, pinned notes are now included at session start (Priority 1, 2000-char budget) so your AI always has your most important context.

Fixed

  • Vault registry Save() merge bug — same vault remove silently failed because Save() re-read the registry from disk and merged back deleted entries. Removed merge logic so removes take effect immediately.
  • MCP works without Ollama — MCP server now starts and serves search results even when Ollama is unavailable. Search falls back gracefully: HybridSearch → FTS5 → keyword. Previously, npx @sgx-labs/same mcp refused to start without Ollama, breaking MCP client setups.
  • MCP create_handoff overwrite — multiple handoffs on the same day no longer overwrite each other; uses minute-level timestamps for uniqueness.
  • MCP registry manifests — server.json updated to official MCP registry schema ($schema, isRequired, registryBaseUrl). smithery.yaml uses npx -y instead of bare same. npm package.json adds mcpName for registry auto-discovery.
  • Hook output format — Stop and SessionStart events now use systemMessage for correct Claude Code rendering.
  • URL corrections — all references to ollama.ai updated to ollama.com across install scripts, CLI, and tutorial. Discord invite links synced across all files.
  • same status hook display — now shows all 6 hooks instead of 4.

Changed

  • Brand refresh — ASCII banner updated from 12-line "STATELESS AGENT" red gradient to compact "SAME" blue gradient with "Stateless Agent Memory Engine" subtitle. Updated across CLI, install.sh, and install.ps1.

Codebase

  • CLI decomposed — cmd/same/main.go split into 18 focused files: search_cmd.go, vault_cmd.go, ask_cmd.go, status_cmd.go, doctor_cmd.go, display_cmd.go, demo_cmd.go, tutorial_cmd.go, and more. Main.go now handles only command registration.
  • Hooks decomposed — context_surfacing.go split into search_strategies.go, term_extraction.go, text_processing.go, verbose_logging.go.
  • Test coverage expanded — new test suites for hooks runner, session recovery, session bootstrap, instance registry, indexer, memory, setup, MCP handlers, and store edge cases.
  • AGENTS.md — contributor guide for AI coding agents.

Security

  • MCP server hardening (15 fixes) — query length limits (10K chars), file size guard on get_note (1MB), write rate limiting (30 ops/min sliding window), snippet sanitization neutralizes 12 prompt-injection tag patterns before returning results to AI clients. save_decision and create_handoff use IndexSingleFile instead of full reindex to prevent DoS. Decision titles sanitized. Append-mode provenance tracking. Federated search resolves aliases from registry map directly.
  • MCP tool annotations — all 12 tools declare readOnlyHint, destructiveHint, and idempotentHint per the MCP 2025-06-18 spec. Helps MCP clients enforce least-privilege.
  • Defense-in-depth _PRIVATE/ filtering — case-insensitive UPPER(path) NOT LIKE added to VectorSearch, VectorSearchRaw, and ContentTermSearch SQL queries. Pinned notes skip _PRIVATE/ paths. Hooks use case-insensitive isPrivatePath().
  • SQL LIKE injection prevention — escapeLIKE() helper escapes %, _, and \ in user-supplied search terms across KeywordSearch and ContentTermSearch.
  • Expanded tag sanitization — context injection now neutralizes 12 tag types (added system-reminder, system, instructions, tool_result, tool_use, IMPORTANT) to block stored prompt injection.
  • Vault feed hardening — sanitizeAlias() strips path separators, traversal characters, and null bytes. safeFeedPath() blocks absolute paths, traversal, private/hidden directories, and null bytes. Federated search error messages use aliases only (no raw filesystem paths).
  • Self-closing tag injection — neutralizeTags() (MCP) and sanitizeContextTags() (hooks) now neutralize self-closing tags (<tag/>) and tags with attributes (<tag attr="...">), closing prompt injection vectors.
  • Bootstrap context injection — session bootstrap output is now sanitized before wrapping in <session-bootstrap> tags, preventing stored prompt injection via crafted handoff/decision content.
  • LIKE injection in title search — KeywordSearchTitleMatch now uses escapeLIKE() to prevent SQL wildcard injection via % and _ characters.

v0.6.1 — Hardening, Recovery & Visibility

Security hardening, crash recovery, search improvements, UX polish, and every hook now proves it's working.

Added

  • SessionStart crash recovery — 3-tier priority cascade recovers context even when terminal is closed without Stop firing: handoff (full, completeness 1.0) → instance registry (partial, 0.4) → session index (minimal, 0.3). Schema v3 adds session_recovery table for telemetry.
  • Hook status lines — every hook now prints a one-line receipt to stderr: session recovery source, decisions extracted, handoffs saved, stale notes flagged, referenced notes boosted. same display quiet silences all output.
  • HybridSearch for MCP and CLI — search_notes, search_notes_filtered, and same search now use HybridSearch (semantic + keyword + fuzzy title) instead of raw VectorSearch. Better results for partial matches and typos.
  • Command groups — same --help organizes commands into 5 groups: Essential, Search & Browse, Configuration, Advanced, Other
  • same hooks — new command showing all 6 hooks with name, event, status, and description
  • --json flag — same status --json and same doctor --json for machine-readable output
  • Star ratings — search results show ★★★★☆ 85% instead of raw scores. --verbose for raw numbers.
  • 98 new tests — security (plugin injection, path traversal, symlinks), edge cases (empty inputs, large inputs, concurrent access), store operations
  • NPM distribution — npx @sgx-labs/same mcp --vault /path for MCP clients. Zero-dependency wrapper downloads prebuilt binary from GitHub Releases at install time. Release workflow auto-publishes to npm on tag push.

Security

22 vulnerabilities fixed (3 critical, 8 high, 11 medium):

  • CRITICAL: Plugin command injection — validatePlugin() with shell metachar regex, path traversal block, exec permission check
  • CRITICAL: Hard-coded vec0 768 dims — EmbeddingDim() now dynamic per provider/model
  • CRITICAL: OpenAI gets Ollama URL — conditional BaseURL, only set for Ollama provider
  • HIGH: SSRF in init — localhost validation (127.0.0.1/::1/localhost) before HTTP
  • HIGH: Symlink escape — EvalSymlinks() + ancestor walk in safeVaultPath and config
  • HIGH: Settings destruction — return error on malformed JSON instead of silent overwrite
  • HIGH: Embedding mismatch — use resolved model name in SetEmbeddingMeta
  • HIGH: PII via os.Hostname() — SHA-256 hash → machine-a1b2c3d4 format
  • FTS5 query injection — sanitizeFTS5Term() strips *, ^, -, " and other operators
  • Case-insensitive _PRIVATE/ — UPPER(n.path) NOT LIKE across all SQL queries
  • JSON error sanitization — vault paths → directory name only, no raw hostnames in errors
  • File permissions — all MCP/config writes to 0o600

Fixed

  • Embedding pipeline — OpenAI provider URL, retry logic, all-zero vector detection, API key leak prevention, dimension validation
  • Search quality — absolute+relative scoring blend, per-note token cap (400), FTS5 OR instead of AND, composite indexes for common queries
  • same scope → same status — referenced non-existent command
  • Indexer double-read — eliminated redundant file reads during indexing
  • IncrementAccessCount — batch update instead of per-path

Changed

  • README rewritten — pain-first opening, architecture diagram, feature matrix, comparison table, updated benchmarks
  • Schema v3 — adds session_recovery table; auto-migrates from v2
  • MCP tool list updated — setup now shows all 11 tools with accurate descriptions

v0.6.0 — Reliability, Privacy & Polish

Self-diagnosing retrieval, pinned notes, keyword fallback, vault privacy structure, RAG chat, interactive demo, write-side MCP tools, security hardening, and a full polish pass.

Added

  • Write-side MCP tools — 5 new MCP tools bring the total to 11. Your AI can now save notes, log decisions, and create session handoffs — not just read:
    • save_note — create or update markdown notes (auto-indexed, dot-dir protected, 100KB limit)
    • save_decision — log structured decisions with status and date
    • create_handoff — session handoffs with summary, pending items, and blockers
    • get_session_context — one-call orientation: pinned notes + latest handoff + recent activity + stats
    • recent_activity — recently modified notes (clamped to 50)
  • same ask — ask questions, get answers FROM your notes with source citations. Uses a local Ollama LLM to synthesize answers from semantically relevant notes. Auto-detects the best available chat model. 100% local, no cloud APIs. Example: same ask "what did we decide about authentication?"
  • same demo — interactive demo that creates a temporary vault with 6 realistic sample notes, indexes them, runs search, and showcases same ask. Works without Ollama (keyword-only mode).
  • same tutorial — modular learn-by-doing system with 6 lessons: semantic search, decisions, pinning, privacy tiers, RAG chat, and session handoffs. Run all lessons (same tutorial) or jump to any topic (same tutorial search, same tutorial pin). Creates real notes and runs real commands — you learn the CLI by using it.
  • SAME Lite (keyword-only mode) — SAME now works without Ollama. When Ollama is unavailable, same init offers keyword-only mode using SQLite FTS5. All features work — search, ask, demo, tutorial — with keyword matching instead of semantic search. Install Ollama later and same reindex upgrades to full semantic mode. Zero dependencies beyond the binary.
  • Project-aware init — same init now detects existing project documentation (README.md, docs/, ARCHITECTURE.md, CLAUDE.md, .cursorrules, ADR/) and offers to index them. Zero new notes required — your project already has context.
  • same pin — pin important notes so they're always included in every session: same pin path/to/note.md, same pin list, same pin remove path/to/note.md. Pinned notes inject with maximum priority regardless of query.
  • same repair — one-command database recovery: backs up same.db, force-rebuilds the index, and confirms. The go-to command when something breaks.
  • same feedback — manual thumbs-up/down for notes: same feedback "path" up boosts retrieval confidence; same feedback "path" down penalizes. Supports glob-style paths.
  • Vault seed structure — same init now creates a three-tier privacy directory structure: sessions/ (handoffs), _PRIVATE/ (never indexed, never committed), plus a .gitignore template enforcing privacy boundaries
  • FTS5 keyword fallback — when Ollama is down or slow, context surfacing falls back to SQLite FTS5 full-text search instead of failing silently
  • Doctor retrieval diagnostics — 8 new same doctor checks: embedding config mismatch, SQLite PRAGMA integrity, retrieval utilization rate, config file validity, hook installation, DB integrity, index freshness, log file size
  • Schema migration system — schema_meta table with version-gated migrations; GetMeta()/SetMeta() for metadata storage; auto-migrates between schema versions
  • Embedding mismatch guard — detects when embedding provider/model/dimensions change without reindexing; surfaces clear guidance; Provider interface gains Model() method
  • Hook execution timeout — 10-second timeout prevents hung Ollama from blocking prompts; returns <same-diagnostic> on timeout
  • AI-facing diagnostics — when hooks fail (DB missing, Ollama down), the AI sees <same-diagnostic> blocks with suggested user actions instead of silent failure
  • Ollama retry with backoff — 3 attempts with exponential backoff (0/2/4s) for 5xx and network errors
  • Usage data pruning — records older than 90 days pruned during reindex
  • Configurable noise filtering — [vault] noise_paths in config.toml or SAME_NOISE_PATHS env var
  • MCP directory manifests — server.json (official MCP registry), smithery.yaml (Smithery.ai) for directory submissions
  • GitHub Sponsors — .github/FUNDING.yml configuration
  • MCP server test coverage — 22 tests for safeVaultPath, filterPrivatePaths, clampTopK, and helpers
  • 45+ new tests — store, search, indexer, config, and MCP packages

Security

11 fixes from 6 rounds of pre-release security auditing:

  • Dot-path blocking in MCP — save_note can no longer overwrite .same/config.toml, .git/, .gitignore
  • DB path PII fix — index_stats returns same.db not the full filesystem path
  • MCP error sanitization — all MCP error messages changed to static strings; no internal paths leak to AI
  • find_similar_notes path validation — now validates through safeVaultPath
  • Write size limits — 100KB max on save_decision and create_handoff content
  • <plugin-context> tag sanitization — opening tag now stripped (was only stripping closing tag)
  • Config file permissions — all config writes changed from 0o644 to 0o600 (5 occurrences)
  • Backup file permissions — same repair backup changed to 0o600
  • OLLAMA_URL scheme validation — blocks file://, ftp://; only http/https allowed
  • Empty input validation — same search, same ask, same feedback reject empty input
  • Plugin timeout safety — cmd.Process nil check before Kill()

Fixed

  • Replaced all panics with errors — OllamaURL() and validateLocalhostOnly() now return errors instead of crashing
  • TOML skip_dirs now applied — LoadConfig() applies [vault] skip_dirs to the global SkipDirs map
  • Verbose log permissions — changed from 0o644 to 0o600 (owner-only)
  • Noise path filter — uses HasPrefix instead of Contains to prevent false matches

Changed

  • Go 1.25 — standardized across go.mod, CI, release workflow, install scripts, README
  • Schema version 2 — adds FTS5 virtual table for keyword fallback; auto-migrates from v1
  • Context surfacing resilience — embedding failures trigger keyword fallback instead of returning errors
  • CLI descriptions rewritten — all user-facing commands use outcome language (e.g. "Scan your notes and rebuild the search index" instead of "Index vault into SQLite")
  • README restructured — same demo above the fold, MCP tools table promoted, numbers section, SAME Lite callout, eval methodology
  • MCP tool descriptions improved — all 11 tools with agent-oriented "when to use" guidance
  • Error messages friendlier — "escapes vault boundary" → "outside your notes folder"; timeouts and connection failures include actionable guidance
  • Box is now default display — full mode shows the cyan Unicode box automatically
  • Noise filtering off by default — add noise_paths to config if you want path-based filtering
  • Intel Mac install — install.sh uses ARM binary + Rosetta instead of non-existent darwin-amd64

v0.5.4 — Windows Installer Overhaul

Fixed critical Windows installation issues.

Fixed

  • PowerShell 5.1 compatibility — ANSI escape codes now work in Windows PowerShell (not just PS7)
  • TLS 1.2 enforcement — Installer works on older Windows systems
  • PATH works immediately — No need to restart terminal after install
  • Better Ollama detection — Checks process and API, not just PATH
  • Windows Defender guidance — Clear instructions when antivirus blocks the binary
  • Unblock downloaded file — Removes "downloaded from internet" security flag

Added

  • PowerShell version display during install
  • Corporate proxy detection hint in error messages
  • Execution policy bypass instructions on website
  • Windows added to site structured data (SEO)

v0.5.3 — Push Protection & Display Fixes

Safety rails for multi-agent workflows.

Added

  • Push protection — Prevents accidental pushes to wrong repos when running multiple agent instances
    • same push-allow [repo] creates one-time push ticket
    • same guard settings set push-protect on enables with auto-hook install
    • same guard settings set push-timeout N configures ticket expiry (10-300s)
    • Works across multiple Claude instances sharing same machine
  • Visual feedback box — Unicode box output showing surfaced notes, match terms, and token counts
  • CI setup for vibe coders — same ci init creates GitHub Actions workflow
    • Auto-detects project type (Go, Node, Python)
    • same ci explain teaches what CI is
    • Educational output guides users through next steps

Changed

  • Context surfacing output uses the visual feedback box for full mode
  • Guard settings now show push protection status and hook installation state

v0.5.2 — Self-Update

One-command updates, no more curl.

Added

  • same update — Check for and install the latest version from GitHub releases
    • Detects platform (darwin-arm64, linux-amd64, windows-amd64)
    • Downloads correct binary
    • Replaces itself atomically
    • --force flag to reinstall even if on latest
  • Handles dev builds gracefully (warns instead of failing)

Changed

  • Version check now suggests same update instead of curl command

v0.5.1 — Onboarding & UX Polish

Better first-run experience and vibe-coder friendly commands.

Added

  • Welcome notes — 3 example notes copied to .same/welcome/ during init, demonstrating recommended format and providing searchable onboarding content
  • Profile system — same profile use precise|balanced|broad to adjust precision vs coverage tradeoffs, with token usage warnings
  • Display modes — same display full|compact|quiet to control output verbosity
  • Experience level question — Setup asks if you're new to coding or experienced, sets appropriate defaults
  • Cloud sync warning — Detects Dropbox, iCloud, OneDrive, Google Drive and warns about database conflicts
  • Large vault time estimates — Shows estimated indexing time for 500+ note vaults
  • Dependency checks — Verifies Go 1.25+ and CGO with platform-specific install instructions
  • ASCII art banner — STATELESS AGENT logo with red gradient in installer

Changed

  • Default display mode is now "full" (verbose box) instead of compact
  • Installer has friendlier messaging and visual polish

v0.5.0 — Public Launch

Landing page, branded CLI, multi-provider embeddings, and feedback loop.

Added

  • Multi-provider embedding support — pluggable embedding backend with Ollama (default) and OpenAI providers. Configure via [embedding] config section or SAME_EMBED_PROVIDER / SAME_EMBED_MODEL / SAME_EMBED_API_KEY env vars
  • Feedback loop — notes surfaced during a session that the agent actually references get an access count boost, improving future retrieval confidence
  • Landing page at statelessagent.com — dark terminal aesthetic, install-first design
  • Branded CLI output — STATELESS AGENT ASCII art with red gradient, section headers, boxed summaries, and footer across same init, same status, and same doctor
  • Post-init explanation — completion message now explains what SAME does: context surfacing, decision extraction, handoffs, feedback loop, staleness checks
  • Donations — Buy Me a Coffee + GitHub Sponsors links in README and landing page

Changed

  • Embedding architecture — embedding.Client replaced with embedding.Provider interface; all call sites updated
  • README overhauled — sell first, document second; collapsed <details> sections for CLI reference, configuration, and MCP; streamlined FAQ
  • install.sh now also available at statelessagent.com/install.sh

v0.4.0 — Public Release Polish

Eval-driven optimization, security hardening, CLI improvements.

Added

  • Composite scoring: semantic + recency + confidence signal blending
  • Distance threshold and composite threshold tuning via config
  • Eval harness for measuring retrieval quality
  • Security: prompt injection pattern scanning in context snippets
  • same budget command for context utilization stats
  • Config file support (.same/config.toml) with same config show/edit

Changed

  • Default distance threshold tuned from 15.0 to 16.2 based on eval results
  • Hook output formatting improvements

v0.3.0 — Standalone Release

SAME is now a standalone Go project, decoupled from any specific vault infrastructure.

Breaking Changes

  • Data directory moved: .scripts/same/data/ → .same/data/. Run same reindex --force after updating.
  • Plugins path moved: .scripts/same/plugins.json → .same/plugins.json.
  • Go module renamed: now github.com/sgx-labs/statelessagent.
  • Default handoff directory: Now sessions. Override with SAME_HANDOFF_DIR.
  • Default decision log: Now decisions.md. Override with SAME_DECISION_LOG.

Added

  • Multi-tool vault detection: recognizes .same, .obsidian, .logseq, .foam, .dendron markers
  • SAME_DATA_DIR env var to override data directory location
  • SAME_HANDOFF_DIR env var to override handoff directory
  • SAME_DECISION_LOG env var to override decision log path
  • SAME_SKIP_DIRS env var to add custom skip directories
  • Security: _PRIVATE/ exclusion from indexing and context surfacing
  • Security: Ollama localhost-only validation
  • Security: Prompt injection detection in context surfacing snippets
  • Security: same doctor checks for private content leaks and Ollama binding
  • MCP server name changed from vault-search to same

Removed

  • Obsidian-specific vault detection fallback
  • Personal path defaults
  • Node.js/Python infrastructure (package.json, vault-search Python server)
  • Raycast scripts, eval harness, docs (deferred to separate repos)

v0.2.0

  • Initial Go rewrite of SAME
  • Vector search with sqlite-vec
  • Claude Code hooks (context surfacing, decision extraction, handoff generation, staleness check)
  • MCP server with 6 tools
  • Composite scoring (semantic + recency + confidence)
  • Vault registry for multi-vault support
  • File watcher for auto-reindex
  • Budget tracking for context utilization