@superbased/observer 1.4.5 → 1.4.6

package/README.md

@@ -2,13 +2,50 @@
 
 [![npm](https://img.shields.io/npm/v/@superbased/observer.svg)](https://www.npmjs.com/package/@superbased/observer)
 [![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://www.apache.org/licenses/LICENSE-2.0)
+[![Platforms: Linux • macOS • Windows](https://img.shields.io/badge/platforms-linux%20%7C%20macos%20%7C%20windows-blue.svg)](https://github.com/marmutapp/superbased-observer)
 
-SuperBased Observer — capture, normalize, compress, and analyze AI
-coding tool activity across **Claude Code**, **Codex**, **Cursor**,
-**Cline/Roo Code**, and **GitHub Copilot**.
+**Capture, normalize, compress, and analyze every AI coding tool call you
+run** — across Claude Code, Codex, Cursor, Cline / Roo Code, and GitHub
+Copilot — in one local single-binary tool. No telemetry, no cloud, no
+data leaves your machine.
+
+```
+  ┌──────────────┐    ┌──────────────────────────────────┐
+  │  AI clients  │ ─▶ │ observer proxy   :8820 (local)   │ ─▶ Anthropic / OpenAI
+  │  (Claude     │    │   • intercepts every API call    │
+  │   Code,      │    │   • compresses conversations     │
+  │   Cursor,    │    │   • records api_turns + tokens   │
+  │   Codex …)   │    └──────────────┬───────────────────┘
+  │              │                   ▼
+  │              │ ◀── observer mcp (stdio JSON-RPC) ◀── 12 cross-client tools
+  └──────┬───────┘
+         │ JSONL session logs
+         ▼
+  ┌──────────────────────────────────────────────────────────────┐
+  │  observer watch    SQLite store    :8081 dashboard           │
+  │   • normalizes tool calls            • cost over time        │
+  │   • freshness, redundancy detection  • compression savings   │
+  │   • patterns, suggestions            • discovery / waste     │
+  └──────────────────────────────────────────────────────────────┘
+```
+
+# Table of contents
+
+- [Install](#install)
+- [Five-minute quickstart](#five-minute-quickstart)
+- [Per-AI-client setup](#per-ai-client-setup)
+- [Architecture in detail](#architecture-in-detail)
+- [Dashboard tour](#dashboard-tour)
+- [MCP tools reference](#mcp-tools-reference)
+- [Compression mechanisms](#compression-mechanisms)
+- [Cost and token math](#cost-and-token-math)
+- [Terminology and glossary](#terminology-and-glossary)
+- [CLI reference](#cli-reference)
+- [Configuration](#configuration)
+- [Troubleshooting](#troubleshooting)
+- [Security and privacy](#security-and-privacy)
+- [Source, contributing, license](#source-contributing-license)
 
-Single binary. Local-only (no network calls from the observer
-itself). Pure-Go, no CGO.
 
 ## Install
 
@@ -17,66 +54,699 @@ npm install -g @superbased/observer
 observer --version
 ```
 
-The package ships pre-built binaries for:
+Pre-built binaries ship for:
 
 | Platform              | Architecture |
-| --------------------- | ------------ |
+|-----------------------|--------------|
 | Linux                 | x64, arm64   |
 | macOS (Intel)         | x64          |
 | macOS (Apple Silicon) | arm64        |
 | Windows               | x64          |
 
-If you're on a platform that isn't listed, file an issue or build the
-binary from source — see the
-[main repo](https://github.com/marmutapp/superbased-observer).
+The package uses the `optionalDependencies`-per-platform pattern (same
+shape as `esbuild` / `swc` / `@biomejs/biome`) — only the binary
+matching your machine downloads. No postinstall network calls, no
+compile step.
 
-## Quick start
+If your platform isn't listed, build from source — instructions in
+the [main repo](https://github.com/marmutapp/superbased-observer).
+
+
+## Five-minute quickstart
 
 ```bash
-# 1. Install the package + register hooks with the AI clients you use.
+# 1) Install + register hooks with whichever AI clients you use.
 npm install -g @superbased/observer
-observer init
+observer init                 # interactive: pick clients
+
+# 2) Start the long-running services (watcher + dashboard + proxy).
+observer start &              # all three in one process
+                              # (or run them separately, see CLI ref)
+
+# 3) Engage the proxy by pointing your AI client at the local URL.
+#    See "Per-AI-client setup" for the matching env var.
+export ANTHROPIC_BASE_URL=http://127.0.0.1:8820     # Claude Code
+export OPENAI_BASE_URL=http://127.0.0.1:8820/v1     # Codex / OpenAI
+
+# 4) Open the dashboard.
+open http://127.0.0.1:8081/   # macOS
+xdg-open http://127.0.0.1:8081/   # Linux
+start http://127.0.0.1:8081/   # Windows
+```
+
+After ten minutes of normal AI-coding usage, the dashboard will be
+populated with cost over time, per-tool activity, compression
+savings, and stale-reread waste signals.
+
+
+## Per-AI-client setup
+
+Different clients send to different upstreams. The local proxy on
+`127.0.0.1:8820` handles all of them — routes by URL path. Set the
+env var that matches your client(s); both can coexist on one machine.
+
+| AI client                       | Env var                                               |
+|---------------------------------|-------------------------------------------------------|
+| **Claude Code**                 | `ANTHROPIC_BASE_URL=http://127.0.0.1:8820`            |
+| **Cursor** (Anthropic mode)     | `ANTHROPIC_BASE_URL=http://127.0.0.1:8820`            |
+| **Codex**                       | `OPENAI_BASE_URL=http://127.0.0.1:8820/v1` (note `/v1`) |
+| **Cursor** (OpenAI mode)        | `OPENAI_BASE_URL=http://127.0.0.1:8820/v1`            |
+| **Cline / Roo Code**            | `ANTHROPIC_BASE_URL=...` or `OPENAI_BASE_URL=...` per provider |
+| **GitHub Copilot**              | not yet — JSONL adapter only (no proxy support)       |
+
+### Persistent setups
+
+**Claude Code** (`~/.claude/settings.json`):
+
+```json
+{
+  "env": { "ANTHROPIC_BASE_URL": "http://127.0.0.1:8820" },
+  "hooks": { /* `observer init` writes these */ }
+}
+```
 
-# 2. Start the long-running watcher + dashboard + proxy in one process.
-observer start &
+**Codex** (`~/.codex/config.toml`):
 
-# 3. Engage the proxy by pointing your AI client at the local URL.
-#    Claude Code / Cursor (Anthropic mode):
+```toml
+[env]
+OPENAI_BASE_URL = "http://127.0.0.1:8820/v1"
+```
+
+**Shell rc** (`~/.bashrc` / `~/.zshrc`) — affects every program:
+
+```bash
 export ANTHROPIC_BASE_URL=http://127.0.0.1:8820
-#    Codex / Cursor (OpenAI mode):
 export OPENAI_BASE_URL=http://127.0.0.1:8820/v1
+```
+
+
+## Architecture in detail
+
+Five components running side by side:
+
+### 1. JSONL adapters (passive ingest)
+
+Watch `~/.claude/projects/`, `~/.codex/sessions/`, etc. for new
+session log lines. Normalize per-client tool names to a shared
+taxonomy (`read_file`, `run_command`, `spawn_subagent`, …) and write
+them into the actions table. **Always-on**, no setup required —
+`observer init` registers the hooks once.
+
+What this gets you: every tool call you've ever run, queryable.
+
+### 2. API reverse proxy (active capture)
+
+A localhost HTTP server (`127.0.0.1:8820`) you point your AI client
+at via `ANTHROPIC_BASE_URL` / `OPENAI_BASE_URL`. Intercepts every
+request before it hits Anthropic / OpenAI and:
+
+- Records exact token usage from the upstream `usage` envelope (the
+  most accurate cost source — proxy beats JSONL parsing here).
+- Runs the **conversation compression pipeline** to trim large
+  `tool_result` blocks and drop low-importance messages before
+  forwarding upstream.
+- Captures the `cost_usd` the upstream reports (when present).
+
+What this gets you: ground-truth cost numbers and conversation
+compression savings you can measure.
+
+### 3. SQLite store
+
+A single file at `~/.observer/observer.db`. Tables include:
+
+- `projects`, `sessions`, `actions` — the taxonomy
+- `api_turns` — one row per proxy-intercepted upstream request
+- `token_usage` — JSONL-derived token-row events (deduped via spec §A1)
+- `file_state` — content hashes for freshness classification
+- `compression_events` — per-event compression detail (post-migration 010)
+- `project_patterns` — derived patterns from `observer patterns`
+- `failure_context`, `action_excerpts` — diagnostic data
+
+Pure-Go via `modernc.org/sqlite`, no CGO. WAL mode by default.
+
+### 4. Local dashboard (`:8081`)
+
+Eight tabs covering: Overview, Cost, Sessions, Actions, Tools,
+Compression, Discovery, Patterns. See [Dashboard tour](#dashboard-tour).
+
+Static HTML + Chart.js. No analytics, no external requests.
+
+### 5. MCP server (stdio)
+
+12 read-only tools the AI client itself can call mid-conversation —
+`check_file_freshness`, `get_last_test_result`, `search_past_outputs`,
+etc. Powers cross-client tool sharing: if Claude Code ran `go test`,
+Cursor's MCP query for the latest test result will return Claude
+Code's run. See [MCP tools reference](#mcp-tools-reference).
+
+
+## Dashboard tour
+
+Open `http://127.0.0.1:8081/` after `observer start`. Eight tabs.
+
+### Overview tab
+
+High-level snapshot of the selected window:
+
+- **KPI tiles**: Sessions count, API turns (proxy-captured), Token
+  rows (JSONL-recovered), Failures (24h)
+- **Cost over time** chart — daily token volume, split into the four
+  billable buckets (net input / cache read / cache write / output)
+- **Actions over time** chart — total actions vs failures
+- **Top models (by tokens)** chart — top-8 models stacked by net
+  input / cache read / output
+- **Top tools (actions over time)** — per-AI-client stacked-area
+  showing when each client is active
+
+### Cost tab
+
+Per-model breakdown over the selected window. Tokens split into the
+four billing buckets, with computed dollar cost and a reliability
+flag (`high` = upstream-reported, `medium` = computed, `low` =
+estimated, `unreliable` = no pricing entry). Daily stacked-bar chart
+of token volume per day.
+
+Hover any column header for tooltip; click for the full definition
+in the help drawer.
+
+### Sessions tab
+
+One row per AI-coding session. Each session has a stable ID, a tool
+(claude-code / cursor / codex / cline / copilot), a working-directory
+project, action count, sub-agent action count (when the session
+spawned sub-agents via the `Agent` tool), and — if `observer score`
+has run — quality / errors / redundancy ratios.
+
+Click a row to open the session-detail modal: action breakdown,
+token buckets, cost, recent actions.
+
+### Actions tab
+
+The flat firehose: every recorded tool call, normalized across
+adapters. Filter by action type (`read_file`, `write_file`,
+`run_command`, `spawn_subagent`, `todo_update`, `mcp_call`, …).
+Pagination caps at 50 rows per page; total count is shown next to
+the heading.
+
+### Tools tab
+
+Per-AI-client (the *client* — claude-code / cursor / codex / etc., not
+the per-tool name) aggregates plus three views:
+
+- **KPI tiles**: Total actions, Distinct tools, Overall success rate,
+  Busiest tool
+- **Activity over time** stacked-area showing per-tool action volume
+  per day
+- **Action-type mix per tool** horizontal stacked bar — what each
+  tool actually does (read_file vs edit_file vs run_command vs
+  search_text vs spawn_subagent)
+- The full per-tool aggregate table with first/last seen
+
+### Compression tab
+
+How many tokens and dollars the conversation-compression pipeline
+saved by trimming requests before forwarding upstream:
+
+- **KPI tiles**: Tokens saved (est.), Dollars saved (est.), Bytes
+  saved, Turns compressed
+- **Savings per day** chart — daily tokens-saved (left axis) and
+  bytes-saved (right axis)
+- **Savings by mechanism** stacked bar — segments per mechanism
+  (json / code / logs / text / diff / html / drop). **Toggle the
+  y-axis between tokens and bytes** with the chart-header switch.
+- **Per-model breakdown table** — tokens saved ~, $ saved ~, bytes
+  saved, saved %, turns, tool-results compressed, dropped, markers
+- **Recent compression events** — paginated per-event detail with
+  mechanism, original / compressed / saved bytes, message slot, importance
+  score (for drops), and a **Source** column showing whether the
+  event came from a main-thread or sub-agent runtime call
+
+### Discovery tab
+
+Wasted-effort signals:
+
+- **Stale rereads** — files re-read after they changed inside the
+  same session. KPI tiles show count, ~Tokens wasted, ~$ wasted (at
+  your blended input rate), affected files. The **CROSS-THREAD**
+  column flags re-reads that crossed the parent ↔ sub-agent
+  boundary — these are the "pass content via Agent's prompt
+  parameter" candidates.
+- **Repeated commands** — commands run multiple times with no
+  relevant inputs changed in between (e.g. you ran `go test` three
+  times without editing anything between runs).
+- **Cross-tool overlap** — files touched by ≥2 AI clients in the
+  window (e.g. claude-code AND cursor both edited `auth.ts`). This
+  is the visible side of cross-platform tool-call sharing via the
+  MCP server.
+
+### Patterns tab
+
+Repeatable behaviours the observer noticed across your sessions —
+"after running `go test`, you almost always run `go vet`", "when
+working on `auth.ts`, you also touch `login.tsx`", etc. Each pattern
+has a confidence score (decay-weighted: more observations + recent
+observations push it higher) and an observation count.
+
+`observer patterns` derives them; `observer suggest` writes the
+high-confidence ones into `CLAUDE.md` / `AGENTS.md` / `.cursorrules`
+so new sessions inherit your habits.
+
+### Help drawer
+
+Press `?` anywhere on the dashboard or click the **? Help** button in
+the topbar. Every column header, KPI tile, chart label, and filter
+control on every tab is annotated — hover any element to see a
+one-liner tooltip; click to open the drawer at the matching glossary
+entry.
+
+The drawer has full descriptions, formulas, data sources, examples,
+"why it matters", "what to do", and cross-links. Search at the top.
+Deep-linkable via URL fragment — `#help/metric.stale_count` opens
+the drawer at that entry.
+
+Each compression mechanism (json / code / logs / text / diff / html
+/ drop) has a "Full methodology · see more" expandable section
+explaining the actual algorithm.
+
+
+## MCP tools reference
+
+Every AI client running through `observer init` gets these 12 tools
+registered as an MCP server. They're **read-only** queries against
+the unified database, so any agent can read any other agent's
+recorded work — true cross-platform tool-call sharing.
+
+| Tool                              | Purpose |
+|-----------------------------------|---------|
+| `check_file_freshness`            | Has this file been read in the current session? Has it changed since? |
+| `get_file_history`                | Full read/edit history of a file across all sessions and clients |
+| `get_session_summary`             | Roll-up stats for a session: action count, success rate, cost, token buckets |
+| `search_past_outputs`             | Full-text search across recorded tool outputs (FTS5-indexed) |
+| `get_last_test_result`            | Most recent `go test` / `npm test` / `pytest` etc. output |
+| `get_failure_context`             | Recent failures: which command, which file, which session |
+| `get_action_details`              | One specific action's full record (target, args, output excerpt) |
+| `check_command_freshness`         | Has this command been run in the current session? With what result? |
+| `get_session_recovery_context`    | Recent activity for resuming a paused session |
+| `get_project_patterns`            | High-confidence patterns derived from this project's history |
+| `get_cost_summary`                | Daily / per-model / per-session cost rollups |
+| `get_redundancy_report`           | Stale rereads, repeated commands, cross-tool overlap for the project |
+
+**Cross-tool sharing**: when `observer init` registers the MCP server
+with Claude Code AND Cursor (and Codex…), all of them call the same
+12 tools against the same database. Cursor's `get_last_test_result`
+returns Claude Code's last test run; Codex's `check_file_freshness`
+reflects edits made by Cursor.
+
 
-# 4. Open the dashboard.
-open http://127.0.0.1:8081/
+## Compression mechanisms
+
+The conversation-compression pipeline runs inside the proxy on every
+upstream request. It tries to fit the request body within
+`target_ratio × original_bytes` (default 0.85) without breaking the
+conversation's referential integrity.
+
+Two passes, in order:
+
+### Pass 1 — per-content-type compression
+
+Each `tool_result` block is sniffed for content type, then routed to
+a content-aware compressor. Six compressors:
+
+| Mechanism | What it does | When it fires |
+|-----------|--------------|---------------|
+| **json**  | Replaces every scalar value with a type sentinel (`"<string>"`, `"<number>"`, …) preserving structure (keys, arrays, nesting). Arrays of length > 1 collapse to one element with `_len: N`. | API responses, structured logs, telemetry exports. JSON tool_results are usually the biggest savings target. |
+| **code**  | Heuristic skeleton: keeps top-of-file imports + signature lines (function / method / class / struct / interface / type), drops bodies. | Source code files. **Opt-in only** (excluded from default `compress_types`) because the model often needs to read full code. |
+| **logs**  | Two-pass: collapses adjacent identical lines to `<line> [×N]`, then head+tail-truncates to 200 lines if still long. | Log-shaped output — `go test ./...`, `npm run build`, polling/retry loops. Lossless on distinct lines; only the truncate pass is lossy. |
+| **text**  | Catch-all for content not classified as code/json/logs/diff/html. Head+tail truncation: keeps 40 + 40 lines on inputs over 80 lines. | Markdown bodies, README excerpts, narrative descriptions. |
+| **diff**  | Strips unified-diff context beyond ±1 line of each change. Keeps every header, every `+`/`-` line, drops the rest with elision markers. | `git diff`, patch tool outputs. Lossless on changes; lossy only on the cheap-to-rebuild context. |
+| **html**  | Three regex passes: strips `<script>`, `<style>`, and HTML comments. Tag attributes + visible text + structural elements survive. | `web_fetch` results pulling whole HTML pages — usually 80%+ scripts/styles. |
+
+### Pass 2 — drop with marker
+
+If Pass 1's compressed body is still over budget, the budget enforcer
+ranks remaining messages by **importance score** (a deterministic
+weighted sum) and drops the lowest-scored non-preserved ones until
+the budget is met. Each dropped message is replaced by a single
+marker block (a placeholder text) so the conversation flow stays
+intact for the model.
+
+**Importance score** = `0.4 × recency + 0.3 × reference + 0.15 × density + 0.15 × role`
+
+- **Recency** = `(i+1) / n` — newest message scores 1.0, oldest scores 1/n
+- **Reference** = `1.0` if any of the message's `tool_use_ids` is
+  cited by a later `tool_result`, OR any of its `referenced_ids` points
+  to a live tool_use; `0.0` otherwise. Tool-pair-live messages always
+  get full weight regardless of position.
+- **Density** = fraction of non-whitespace runes (whitespace-padded
+  outputs get dropped first)
+- **Role** = `system 1.0`, `user 0.9`, `assistant 0.7`, `tool 0.5`
+  (tool outputs are the most-compressible by policy)
+
+**Preserved messages** (never droppable):
+- The last `PreserveLastN` messages (default 4)
+- Any `system` role message
+- **Tool-pair-live messages**: any message whose `tool_use_id` is
+  referenced by a later tool_result (parent side), AND any
+  `tool_result` message whose `referenced_id` points to a live
+  tool_use (consumer side)
+
+Tool-pair preservation is symmetric — dropping either side leaves an
+orphan that Anthropic rejects with 400.
+
+### Per-event detail
+
+Every drop and every per-type compression is recorded as a row in
+the `compression_events` table (post migration 010). The Compression
+tab's "Recent compression events" view surfaces these with
+mechanism, original / compressed / saved bytes, message slot, and
+importance score (for drops).
+
+### Tuning
+
+`observer config` settings (in `~/.observer/config.toml`):
+
+```toml
+[compression.conversation]
+enabled = true
+target_ratio = 0.85
+mode = "token"             # or "cache" — restricts drops to tail half
+preserve_last_n = 4
+prefix_bytes = 8192
+compress_types = ["json", "logs", "text"]   # add "code", "diff", "html" to opt in
+min_bytes_to_compress = 1024
+
+[compression.conversation.weights]
+recency = 0.4
+reference = 0.3
+density = 0.15
+role = 0.15
+```
+
+High importance scores on dropped events (≥0.5) suggest the threshold
+is too aggressive — raise `target_ratio` (e.g. 0.9 or 0.95).
+
+
+## Cost and token math
+
+Anthropic's usage envelope reports four token buckets per request,
+each at a different rate:
+
+| Bucket             | What it is                                                                  | Bills at                    |
+|--------------------|------------------------------------------------------------------------------|-----------------------------|
+| `net_input`        | Fresh prompt tokens not served from cache                                   | model's standard input rate |
+| `cache_read`       | Prompt tokens served from Anthropic's ephemeral cache                       | ~10% of input rate          |
+| `cache_creation`   | Tokens written to ephemeral cache                                            | 1.25× input (5m tier) or 2× input (1h tier) |
+| `output`           | What the model generated                                                     | typically 5× input rate     |
+
+`prompt_context = net_input + cache_read + cache_creation`
+`total_tokens  = prompt_context + output`
+
+The cost engine (`internal/intelligence/cost`) computes USD via:
+
+```
+cost_usd = (net_input × p.input + cache_read × p.cache_read +
+            cache_creation_5m × p.cache_creation +
+            cache_creation_1h × p.cache_creation_1h +
+            output × p.output) ÷ 1,000,000
 ```
 
-## What you get
+When the upstream API returns `cost_usd` in the response envelope
+(proxy-sourced rows), that value is preferred over the computed one —
+ground truth, reliability=high.
+
+### Blended input rate
+
+The Discovery tab's "~$ wasted" tile uses a **blended input rate**
+computed from your last-30d api_turns mix: each model's input rate
+weighted by the prompt-token volume it consumed. Example: if you
+spent 70% of prompt tokens on opus-4-7 ($15/1M) and 30% on
+haiku-4-5 ($1/1M), the blended rate is `0.7 × 15 + 0.3 × 1 = $10.80/1M`.
+
+Falls back to $3/1M (claude-sonnet-4 input) on fresh installs with
+no proxy data.
 
-- **Real-time dashboard** at `http://127.0.0.1:8081/` — cost over
-  time, tool activity, compression savings, stale rereads, patterns,
-  per-tool action breakdowns, and full-workbook Excel export.
+### JSONL dedup
+
+When the proxy isn't engaged, observer falls back to parsing the AI
+client's on-disk session log. Clients echo the same cumulative usage
+on every content block of a multi-block response, so naive parsing
+counts one API call 2-4×. Two layers of dedup catch this:
+
+1. **Adapter-level**: dedupes on Anthropic `message.id` at write time
+2. **Cost-engine-level**: dedupes on `(source_file, model, timestamp-bucketed-to-minute, tokens)` at read time
+
+Migration 007 ran a one-time pass collapsing pre-fix duplicates.
+
+
+## Terminology and glossary
+
+Quick reference; the in-platform help drawer (press `?` on the
+dashboard) has the full versions with cross-links.
+
+- **Action** — one normalized tool call recorded by an adapter. Action
+  types are taxonomic and cross-client: `read_file`, `write_file`,
+  `edit_file`, `run_command`, `search_text`, `search_files`,
+  `web_search`, `web_fetch`, `mcp_call`, `spawn_subagent`,
+  `todo_update`, `ask_user`, `task_complete`, `user_prompt`,
+  `unknown`.
+- **API turn** — one HTTP request captured by the local proxy. Records
+  one row in `api_turns` per request, with the upstream usage envelope
+  intact.
+- **Cache 5m vs 1h tier** — Anthropic's prompt cache has two TTLs.
+  Default is 5 minutes; `cache_control: {type: ephemeral, ttl: 3600}`
+  extends to 1 hour at 2× the write cost. Reads bill the same rate
+  regardless of tier.
+- **Compression event** — one individual compression decision (one
+  per-type compress, or one drop) recorded post migration 010.
 - **Conversation compression** — pre-forward trimming of API request
-  bodies. Per-mechanism breakdown (json / code / logs / text / diff /
-  html / drop) so you can see exactly what saved how many tokens.
-- **MCP server** with 12 tools every connected AI client can use to
-  query each other's recorded work — `get_last_test_result`,
-  `check_file_freshness`, `get_failure_context`, etc.
-- **Patterns engine** — `observer patterns` derives repeatable
-  behaviours from session history; `observer suggest` writes them
-  into `CLAUDE.md` / `AGENTS.md` / `.cursorrules` so new sessions
-  inherit them.
-
-Full feature reference + walkthrough:
-[main repo README](https://github.com/marmutapp/superbased-observer).
-
-## How the npm package works
-
-This package is a thin Node.js shim that locates the right pre-built
-binary (shipped via `optionalDependencies`) and spawns it. Same shape
-as `esbuild` / `swc` / `@biomejs/biome`. No postinstall network calls,
-no compile step at install time.
-
-## License
-
-Apache 2.0 — see
-[LICENSE](https://github.com/marmutapp/superbased-observer/blob/main/LICENSE).
+  bodies. Pass 1 = per-content-type compression, Pass 2 = drop with
+  marker. See [Compression mechanisms](#compression-mechanisms).
+- **Cross-platform tool calling** — every AI client connected via
+  `observer init` can call the 12 MCP tools against the unified
+  database. So Cursor's `get_last_test_result` can return a `go test`
+  Claude Code ran an hour earlier.
+- **Cross-thread reread** — the parent thread re-reads a file the
+  sub-agent already saw (or vice versa) within the same session. Fix:
+  pass content via the Agent tool's `prompt` parameter rather than
+  letting the child re-read.
+- **Freshness state** — per-read tag from the freshness engine: `fresh`
+  (first read in this session, OR re-read with same content),
+  `stale` (re-read after change in same session), `missing` (file no
+  longer exists), `modified-elsewhere` (file changed by something
+  other than an observable AI action).
+- **Mechanism** — one of `json`, `code`, `logs`, `text`, `diff`,
+  `html` (per-content-type compressor) or `drop` (low-importance
+  message replaced by a marker).
+- **Pattern** — a derived behaviour: `command_pair` (X often
+  followed by Y), `cross_tool_file` (file touched by multiple
+  clients), `knowledge_snippet` (consistent topic-specific habit),
+  `failure_correlation` (X often precedes a failure of Y),
+  `session_summary`. Each has a decay-weighted confidence score 0-1.
+- **Project** — working-directory root that owns sessions and
+  actions. Derived from cwd at session start; `/.git/worktrees/...`
+  paths fold back to the working-tree root.
+- **Proxy vs JSONL** — proxy intercepts upstream HTTP calls (ground
+  truth, reliability=high). JSONL parses the AI client's on-disk
+  session log (works without configuring a base URL, but client
+  echoes cumulative usage on every block, requiring dedup —
+  reliability=unreliable for token counts on Claude Code).
+- **Reliability** — cost-engine confidence: `high` (upstream-
+  reported), `medium` (computed from known pricing), `low` (some
+  buckets estimated), `unreliable` (no pricing entry).
+- **Session** — one continuous AI-coding conversation in a single
+  tool, scoped to one working directory. Has a stable ID (Claude
+  Code's UUID, Codex's rollout ID, …).
+- **Sidechain** — actions emitted inside a sub-agent runtime spawned
+  via the parent's `Agent` tool. Sub-agents share the parent's
+  session_id; the `is_sidechain` column distinguishes them. The
+  Discovery tab's CROSS-THREAD column counts stale rereads that
+  crossed this boundary.
+- **Stale reread** — same-session re-read of a file whose content
+  changed between reads. Cross-session reads are excluded (a fresh
+  session has no memory of a prior session's read).
+- **Tool** — in this dashboard, "tool" means the *AI client*
+  (claude-code, cursor, codex, cline, copilot), not the per-tool
+  name (`read_file`, `run_command`). The latter is "Tool name" on
+  the Actions tab.
+- **Tool-pair integrity** — Anthropic requires every `tool_result`
+  block to have a corresponding `tool_use` block in a preceding
+  message. The compression pipeline preserves both sides of every
+  live pair to satisfy this constraint.
+
+
+## CLI reference
+
+Every command supports `--help` for the full surface.
+
+| Subcommand                         | Purpose |
+|------------------------------------|---------|
+| `observer init`                    | Register hooks + MCP server with installed AI clients |
+| `observer uninstall`               | Reverse `observer init` |
+| `observer start`                   | Run watcher + dashboard + proxy in one process (recommended) |
+| `observer watch`                   | Long-running JSONL watcher only |
+| `observer dashboard --addr ADDR`   | HTTP dashboard only |
+| `observer proxy start`             | Reverse proxy only |
+| `observer scan`                    | One-shot ingest of existing JSONL files (catch-up after install) |
+| `observer status`                  | DB stats + recent activity |
+| `observer doctor`                  | Diagnostic — checks paths, schemas, hook registration |
+| `observer tail`                    | Live tail of incoming events |
+| `observer cost`                    | Per-model cost summary CLI |
+| `observer score`                   | Compute quality_score / error_rate / redundancy_ratio for sessions |
+| `observer discover`                | Stale rereads + repeated commands report (CLI version of the Discovery tab) |
+| `observer patterns`                | Derive patterns from session history |
+| `observer learn`                   | Adapter for ingesting external JSONL exports |
+| `observer suggest`                 | Write high-confidence patterns into CLAUDE.md / AGENTS.md / .cursorrules |
+| `observer summarize`               | Roll-up summary across sessions |
+| `observer export`                  | Export DB to xlsx / json |
+| `observer prune`                   | Manual retention pass (delete old data) |
+| `observer backfill --is-sidechain` | Re-walk JSONL to populate is_sidechain on pre-migration rows |
+| `observer metrics`                 | Prometheus-format metrics endpoint |
+| `observer serve`                   | MCP server (stdio JSON-RPC) — usually invoked by `observer init` registration |
+| `observer tail`                    | Live event stream |
+
+
+## Configuration
+
+`~/.observer/config.toml` — created with defaults on first run.
+
+```toml
+[paths]
+db_path = "~/.observer/observer.db"
+log_dir = "~/.observer/logs"
+
+[proxy]
+listen_addr = "127.0.0.1"
+port = 8820
+anthropic_upstream = "https://api.anthropic.com"
+openai_upstream = "https://api.openai.com"
+
+[dashboard]
+listen_addr = "127.0.0.1"
+port = 8081
+
+[compression.conversation]
+enabled = true
+target_ratio = 0.85
+mode = "token"
+preserve_last_n = 4
+prefix_bytes = 8192
+compress_types = ["json", "logs", "text"]
+min_bytes_to_compress = 1024
+
+[compression.conversation.weights]
+recency = 0.4
+reference = 0.3
+density = 0.15
+role = 0.15
+
+[compression.shell]
+enabled = true
+# per-command filters configured under [compression.shell.filters]
+
+[retention]
+prune_on_startup = true
+max_actions = 5_000_000
+max_age_days = 365
+
+[pricing]
+# Per-model overrides if the baked-in pricing is wrong for you.
+# [pricing.models."claude-opus-4-7"]
+# input = 15
+# output = 75
+# cache_read = 1.5
+# cache_creation = 18.75
+```
+
+
+## Troubleshooting
+
+### Empty dashboard / "No proxy traffic"
+
+The JSONL adapter populates passively after `observer init`, but
+ground-truth cost / compression numbers require the proxy. Set
+`ANTHROPIC_BASE_URL=http://127.0.0.1:8820` (Claude Code) or
+`OPENAI_BASE_URL=http://127.0.0.1:8820/v1` (Codex) in the shell
+that launches your AI client.
+
+Verify with `observer status | grep api_turns` — count should
+climb during AI-client activity.
+
+### `observer --version` says `dev`
+
+You're on a non-released build. Reinstall with `npm install -g @superbased/observer` or rebuild with the workflow's `-X main.version=$VERSION` ldflag.
+
+### `tool_result block must have a corresponding tool_use block`
+
+Anthropic 400. Means the conversation-compression pipeline dropped
+a `tool_use` while keeping its matching `tool_result`. Versions
+prior to 1.3.2 had this bug; upgrade. If you're on 1.3.2+ and still
+see it, file an issue with the conversation prefix.
+
+### `tool use concurrency issues`
+
+Anthropic 400 surfaced in Claude Code as this message. Means the
+parallel-tool-use case (multiple `tool_use` blocks in one assistant
+message) isn't paired correctly with the multi-block tool_result
+that follows. Versions prior to 1.3.2 had this bug; upgrade.
+
+### Cross-thread numbers are 0
+
+Pre-migration data was ingested without the `is_sidechain` flag.
+Run `observer backfill --is-sidechain` once to re-walk JSONL and
+populate the flag on existing rows.
+
+### Migration error: `duplicate column name`
+
+Race condition between concurrent daemon startups, fixed in 1.4.1.
+Upgrade. If you still see it, run daemons serially: `observer
+watch`, wait, then `observer dashboard`, then `observer proxy
+start` (or just use `observer start` which runs all three in one
+process).
+
+### Dashboard port already in use
+
+```bash
+observer dashboard --addr 127.0.0.1:8082    # pick a different port
+# or
+[dashboard]
+port = 8082                                  # in config.toml
+```
+
+
+## Security and privacy
+
+- **No telemetry**. The observer never makes outbound network calls
+  except: (a) the proxy forwarding to the upstream API on your
+  behalf, (b) optional codegraph-MCP integration if you've configured
+  it. Both are explicit.
+- **Local-only HTTP**. The proxy and dashboard bind to
+  `127.0.0.1` by default. Don't bind to `0.0.0.0` unless you've
+  thought about it — there's no auth.
+- **Secrets scrubbing**. Tool inputs and outputs are passed through
+  a configurable scrubber (`internal/scrub`) that redacts API keys,
+  bearer tokens, and other obvious secrets before persistence. Not a
+  guarantee — review the rules in `~/.observer/config.toml` for your
+  own secret formats.
+- **Database**. `~/.observer/observer.db` is a SQLite file. Owns
+  the same security as your `~/.claude/` and `~/.codex/` session
+  logs (which already contain the same content). Encrypt the disk
+  if your threat model needs that.
+- **Logs**. `~/.observer/logs/` carries operational logs (start /
+  stop, errors, retention). Not request bodies — those are in the
+  DB only.
+
+
+## Source, contributing, license
+
+- **Source**: https://github.com/marmutapp/superbased-observer
+- **Specification**: `superbased-final-spec-v2.md` in the repo
+- **Issues**: https://github.com/marmutapp/superbased-observer/issues
+- **License**: [Apache 2.0](https://github.com/marmutapp/superbased-observer/blob/main/LICENSE)
+- **Author**: Santosh Kathira <contact@marmut.app>
+
+This npm package is a thin Node.js shim that resolves the right
+pre-built binary at runtime and spawns it. Same shape as `esbuild` /
+`swc` / `@biomejs/biome`. The Go source lives in the main repo;
+binaries are cross-compiled per release tag via GitHub Actions and
+published as `@superbased/observer-<platform>-<arch>` per-platform
+packages.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@superbased/observer",
-  "version": "1.4.5",
+  "version": "1.4.6",
   "description": "SuperBased Observer — capture, normalize, compress, and analyze AI coding tool activity across Claude Code, Codex, Cursor, Cline/Roo, and Copilot.",
   "keywords": [
     "ai",
@@ -36,11 +36,11 @@
     "LICENSE"
   ],
   "optionalDependencies": {
-    "@superbased/observer-linux-x64": "1.4.5",
-    "@superbased/observer-linux-arm64": "1.4.5",
-    "@superbased/observer-darwin-x64": "1.4.5",
-    "@superbased/observer-darwin-arm64": "1.4.5",
-    "@superbased/observer-win32-x64": "1.4.5"
+    "@superbased/observer-linux-x64": "1.4.6",
+    "@superbased/observer-linux-arm64": "1.4.6",
+    "@superbased/observer-darwin-x64": "1.4.6",
+    "@superbased/observer-darwin-arm64": "1.4.6",
+    "@superbased/observer-win32-x64": "1.4.6"
   },
   "scripts": {
     "test": "node bin/observer.js --version"