driftguard-tokepi 0.5.0__tar.gz

driftguard_tokepi-0.5.0/CHANGELOG.md

@@ -0,0 +1,451 @@
+# Changelog
+
+All notable changes to DriftGuard are recorded here. The format follows
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and the project uses
+[semantic versioning](https://semver.org/) (pre-1.0: minor versions may break).
+
+## [Unreleased]
+
+## [0.5.0] - 2026-11-11
+
+First release published to PyPI. It contains the 0.4.0 foundation and
+everything built on top of it: the activity log, the signal box dashboard
+redesign, the analytics (baseline comparison, spread, latency), the MCP gateway,
+and the framework adapters.
+
+### Changed
+- **`driftguard-server` fails gracefully without the server extra.** Running it
+  from a base `pip install driftguard-tokepi` now prints `pip install "driftguard-tokepi[server]"`
+  instead of a `ModuleNotFoundError` traceback. Importing `driftguard.server` is
+  lazy, so the entry point and the `driftguard-mcp-proxy` command never pull in
+  FastAPI.
+
+### Added
+- **Framework adapters for LangChain, the OpenAI Agents SDK and the Claude Agent
+  SDK.** `DriftGuardCallbackHandler` (a LangChain/LangGraph callback handler),
+  `DriftGuardRunHooks` (OpenAI Agents SDK run hooks) and
+  `DriftGuardClaudeAgentHooks` (Claude Agent SDK tool hooks) each record one
+  agent event per tool call — with its duration and outcome — and attribute each
+  model turn's tokens to the next tool, reusing `AgentContext`. Each plugs into
+  the framework's own hook mechanism in one or two lines. None of the frameworks
+  is a dependency: each adapter subclasses its framework's base when installed
+  and still works for testing without it. 10 new tests.
+- **Latency as a first-class metric.** Telemetry accepts `latency_ms`, a new
+  `latency_limit_ms` policy rule (default 10,000 ms) flags requests slower than
+  the limit, and the reading flows through the summary (`averages.latency_ms`,
+  `violation_rates.latency_limit_ms`), the Overview's All readings, the Drift
+  Risk Breakdown, the Policy page and its impact preview, and the Test Telemetry
+  form. A slow request adds 0.15 to its risk score — enough to tip a borderline
+  event into `warning`, never enough to flag a fast one, and an event that
+  reports no latency is never flagged. The SDK already forwards it
+  (`capture_metrics(latency_ms=...)`). Existing databases gain the two columns on
+  startup. 9 new backend tests and a frontend preview test.
+- **MCP gateway (`driftguard-mcp-proxy`).** A stdio proxy that sits between any
+  MCP client and a downstream MCP server, forwards every JSON-RPC message
+  unchanged, and records each `tools/call` as a DriftGuard agent event with its
+  duration and outcome — so closed agents that speak MCP can be observed with no
+  code change. One session's calls group under one task, so a retried failing
+  tool is flagged as blocked. Config comes from flags, `DRIFTGUARD_*` env vars or
+  a JSON file (the same keys as the Claude Code integration). It logs only to
+  stderr, passes a message it can't parse through untouched, and resolves the
+  server command on `PATH` (including `.cmd` wrappers on Windows). New
+  `driftguard.gateway` package, 13 tests.
+- **View filters in the address.** Alerts keeps its tab and severity in the
+  address (`#/alerts?tab=resolved&severity=critical`) and Logs keeps its search,
+  kind, outcome and task (`#/logs?q=pytest&outcome=failed`), so a filtered list
+  can be linked. Defaults are left out and the values go when you leave the view.
+  Back and forward move between them, and a link to another view carries only
+  the global filters.
+- **Plate panels.** The Overview's timeline, Drift Risk Breakdown, Top Root
+  Causes and Open Alerts are named with engraved plate labels over a sleeper
+  rule, the same idiom as the instruments, instead of a plain title over a
+  hairline. `Panel` gains `variant="plate"`.
+- **Bulk resolve and reopen in one request.** `PATCH /projects/{id}/alerts` takes
+  `{"ids": [...], "resolved": bool}` for up to 500 alerts and returns the alerts
+  now in that state plus any ids that weren't the project's. Ids from other
+  projects are never touched. Each changed alert is still pushed on the
+  WebSocket. The Alerts view's selection actions use it, so resolving a hundred
+  alerts is one round trip. 3 new tests.
+- **Frontend unit tests.** The dashboard gains Vitest with jsdom: 33 tests over
+  hash routing and shareable links, the keyboard shortcuts (including the `g`
+  timeout, typing, modifiers and dialogs), input validation, which scores a
+  source reports, the baseline wording, and the Policy impact preview, whose
+  maths moved from the view into `lib/policyPreview.ts`. `npm test` runs in CI
+  before the build.
+- **Demo slides.** `docs/demo.md` links a nine-card deck that follows the running
+  order, built from the current screenshots.
+- **`startup.bat`** for Windows development: creates the environment on first
+  run, starts the API on port 8000 against the local SQLite file and the Vite dev
+  server on port 5173 with hot reload. `demo.bat` stays the one-click demo.
+- **Claude Code templates** (`integrations/claude_code/templates/`): a
+  `driftguard.json` and a `settings.local.json` with all six hooks, ready to copy
+  into another repository's `.claude/` folder. Three placeholders to fill in; no
+  key is shipped.
+- **Signal box redesign of the dashboard.** The Overview now leads with one
+  status plate, then a track diagram of the most urgent task, then four
+  instruments.
+  - The status plate shows one lamp and one word (`Critical`, `Warning`,
+    `Stable` or `No data`), the diagnosis sentence, the recommendation and a
+    single action.
+  - The track draws each tool a task used as a block section, with the
+    blocking tool held at a red signal. It also appears on Agent Diagnosis and
+    in the task drawer.
+  - Open alerts, mean risk score, response quality and tool failure rate sit in
+    one strip. Gauges draw each reading against its policy floor or limit.
+    Everything else is under a collapsed **All readings**.
+  - The timeline draws the prompt-token limit and both score floors as dashed
+    lines, hatches the span below the floors, and offers its data as a table.
+- **Baseline comparison and sample-size guard.** `GET /projects/{id}/summary`
+  gains `baseline`: the window's mean retrieval score, response quality and risk
+  score against the same length of time just before it (at least 24 hours).
+  - A change counts only when it is at least 3 standard errors wide (a Welch
+    two-sample z-test) and at least 0.05 in size.
+  - With under 10 events on either side it says "too few events to compare"
+    instead of giving a verdict. "All time" has no comparison.
+  - The Overview shows it under Mean risk score, Response quality and Mean
+    retrieval score, and adds a "Worse than usual" line to the status plate. It is
+    advice only: it raises no alert and doesn't change the status.
+  - `driftguard/baseline.py` holds the maths, with 14 new tests.
+- **Response quality for Claude Code, from outcome checks.** When a turn runs tests,
+  a linter, a type check or a build, the hook scores the turn's last model call 1.0
+  if the last check passed and 0.0 if it failed. Turns that run no check get no
+  score, so the average is over checked turns only. Pass or fail also reads the
+  output for what failing runs print, because a pipe into `tail` hides the exit code,
+  and a command that never started is ignored. The basis is in the event's metadata.
+  `"quality_from_checks": false` turns it off. There is deliberately **no retrieval
+  score** for Claude Code: it has no retrieval step, and scoring empty searches would
+  raise false degradation warnings. 54 new tests.
+- **Scores a source never sends are left off the dashboard.** The Overview drops the
+  Response quality instrument (three instruments instead of four), the Mean retrieval
+  score reading, the matching Drift Risk Breakdown row, and the timeline's line, axis
+  and legend entry, and All readings says what was left out. A score counts as sent if
+  it appears in the window or the period before. Scores sent for only some events are
+  drawn as dots, not joined by a line. The Policy preview counts only events that
+  carry the reading ("8 of 25") and says when none do.
+- **Spread behind each average.** `GET /projects/{id}/summary` gains `spread`:
+  the median and the 10th and 90th percentiles of each metric, from the latest
+  5,000 readings in the window. The Overview shows them under Mean risk score,
+  Response quality and Mean retrieval score, with the share of events under the
+  floor, so an average can't hide a split. It appears from 10 events. 6 new tests.
+- **Restore defaults on Policy.** A button fills every field with the thresholds a
+  new project starts with, for you to review and save. It is backed by a new
+  `GET /projects/{id}/policy/defaults` (session or key), so the defaults live only
+  on the server. 2 new tests.
+- **Attempt timeline in the task drawer.** One row per tool call on a shared time
+  axis, with each bar ending when the attempt was recorded and running back by its
+  duration. A retry loop shows as a bunch of red bars. It shows the latest 40
+  attempts and has a text summary.
+- **Delete account from the dashboard.** Project Settings has a "Your account"
+  panel. Deleting lists the projects that will go, and asks for your email and
+  password. A wrong password deletes nothing. It uses the existing
+  `DELETE /auth/me`.
+- **Keyboard shortcuts.** `g` then `o` `d` `l` `e` `a` `n` `p` `s` `t` jumps
+  between views, `/` focuses the search box, `r` refreshes and `?` lists them.
+  They are off while typing or with a dialog open.
+- **Shareable links.** The address carries the view, the project and both
+  filters (`#/alerts?env=prod&range=7d&project=showcase`). Opening a link
+  applies them. **Reset filters** appears when a filter is not at its default.
+- **Bulk resolve and reopen** on Alerts, with select-all. The Alerts tabs now
+  respond to arrow keys, Home and End, and the new-alert form opens from a
+  **New alert** button.
+- **Policy impact preview.** Each threshold says how many of the loaded events
+  or tasks it flags, before and after your edit. Fields validate as you leave
+  them. The save bar lists every change and warns when coverage drops.
+- **Logs view** in the dashboard: the full history of agent activity, newest
+  first.
+  - Server-side search, plus kind and outcome filters.
+  - Live updates, with a pause button.
+  - A detail drawer, including "show only this task".
+  - CSV and JSON export.
+- **Agent event kinds.** `kind` accepts `tool_call` (the default), `llm_call`,
+  `prompt`, `response`, `session_start` and `session_end`. Non-tool kinds don't
+  need `tool_name` or `status`, and are ignored by the diagnosis and agent alerts.
+- **Opt-in content storage.**
+  - Agent events accept `input` and `output` previews.
+  - The server stores them only when the project's new `capture_content`
+    setting is on. It's off by default.
+  - Stored content is scrubbed, cut to 2,000 characters, and pruned by retention.
+- `PATCH /projects/{id}` changes a project's name or `capture_content`.
+- **Overview agent activity.** `GET /projects/{id}/summary` gains
+  `agent_activity`, computed from the same rows as Logs: tool calls, failures,
+  failure rate, tasks, finished turns, tokens used and average tool duration. The
+  Overview shows these as a row of cards. Retrieval and quality cards say "Not
+  reported by this source" when the telemetry has no such score.
+- `GET /projects/{id}/agent-events` gains `kind`, `tool_name`, `agent_name`,
+  `outcome`, `q` (search) and `before_id` (paging).
+- `GET /projects/{id}/agent-events/export?format=csv|json` downloads up to 10,000
+  filtered events.
+- The WebSocket sends `{"type": "activity", "count": n}` after agent events are
+  ingested. Agent Diagnosis and Logs use it to refresh live.
+- **SDK.**
+  - `DriftGuardClient(capture_content=True)` sends content previews. Without it,
+    `input` and `output` are removed before events are queued.
+  - With it on, `@driftguard_tool` and the MCP adapter send tool arguments and
+    results.
+  - `AgentContext.record_activity(kind, ...)` logs prompts, model calls and
+    other non-tool activity.
+
+- `examples/seed_showcase.py` fills a project with a week of varied sample data
+  for every dashboard view, using only its API key.
+- **Claude Code integration** (`integrations/claude_code/`): a hook script
+  that reports a Claude Code session to DriftGuard with no code changes.
+  - Each prompt becomes a task, and each tool call a `tool_call` event.
+  - Failures come from `PostToolUseFailure`.
+  - Prompts, turn endings and session start/end appear in Logs.
+  - Tool calls get their duration, model and tokens from the session
+    transcript. Cached context isn't counted, and each model call is counted
+    once. A later tool call from the same model call shows 0.
+  - Each turn is reported once, when Claude stops, as one `response`. It holds
+    the prompt, the answer, the turn's duration and the tokens of all its model
+    calls.
+  - Session start and end are opt-in (`"session_events": true`).
+  - Each finished turn sends LLM telemetry, one event per model call, with new
+    input tokens and the full context size. Turn it off with
+    `"send_telemetry": false`.
+  - A turn whose start isn't in the hook state falls back to the last prompt you
+    typed. It skips messages Claude Code inserts, and ignores starts older than
+    6 hours.
+  - It uses only the standard library, gives up after 2 seconds, never fails
+    Claude Code, and logs problems to `.claude/driftguard-hook.log`.
+
+### Changed
+- **New look.** The dashboard moves from a pure-black graphite console to a dark
+  black enamel panel with ivory type. Colour now appears only where
+  something deviates (red, amber, green lamps, each with an icon and a word).
+  Fonts are Barlow, Barlow Condensed and JetBrains Mono, replacing Space
+  Grotesk and Plus Jakarta Sans. The logo is the text "DriftGuard" alone, with no
+  symbol, and the browser-tab icon is a plain "D". The decorative background grid
+  is removed.
+- **True-black canvas and brighter colours.** The page background is pure black
+  with panels lifted a shade, ivory text and the route colour are brighter, and
+  the red, amber, green and steel lamp colours are more luminous.
+- The Logs kind filter now offers only Tool call, Prompt and Response. Model
+  call, Session start and Session end are no longer listed there, though rows of
+  those kinds still appear under All activity.
+- **Themed dropdowns.** Every select uses a themed control. In Chrome and Edge
+  the open list is a black panel with a hairline border, a highlighted row and a
+  check mark on the chosen item. Other browsers keep their own list in the same
+  colours.
+- **Layout fixes from the design review.** The track diagram's sections share the
+  full width. Agent Diagnosis gains a gauge of the longest run of failures in a
+  row against the block threshold. Top Root Causes moves under Drift Risk
+  Breakdown and Open Alerts runs full width, so there are no empty gaps. On a
+  phone the chart legend is one scrolling line, Test Telemetry is an icon, and the
+  filter row wraps.
+- The baseline line says **No clear change** instead of "Steady", and every verdict
+  shows the events it rests on, for example "(89 events now, 93 before)". Under 30
+  events on a side it adds that a small drop could be missed. The research study
+  found a real 0.09 drop is detected only about a third of the time at 10 events,
+  so "steady" over-claimed.
+- Disabled navigation, placeholder and footer text now meets 4.5:1 contrast.
+  The header selects have visible labels (**Env**, **Range**), and the
+  **Updated** time shows from tablet width up. The sidebar groups its nine
+  views into Monitor and Configure.
+- The Overview moves Total events, Tokens at stake, retrieval score and the
+  agent counts into a collapsed **All readings** list. The "Active Alerts"
+  panel is now **Open Alerts**, and the Alerts panel title follows the tab.
+- Agent Diagnosis shows three headline instruments, with the other counts
+  under **All counts**.
+- The dashboard's Refresh button also reloads the project list, and the list
+  reloads when you return to the tab. A project created through the API or a
+  script no longer needs a page reload to appear.
+- Existing databases gain `projects.capture_content` and
+  `agent_events.kind`/`input_text`/`output_text` automatically on startup.
+  Existing agent events become `tool_call` events.
+
+### Documentation
+- **Research: the baseline study** (`research/baseline_study.py`,
+  `research/BASELINE_STUDY.md`). A seeded simulation of the product's own rules,
+  fixed floors against the baseline comparison, with known truth. It reports
+  detection rates, false alarms and delays for one decision per window, for
+  noisier data, for checking after every event, and for slow decline, with 11
+  tests. Findings: the two rules fail in opposite places so both stay; the
+  10-event guard works but "Steady" over-claims at low counts; and a 3-sigma
+  check repeated at every event raises false alarms in 5 percent of streams, which
+  needs a rule of several checks in a row before anything alerts. The research
+  README now explains what each part can and can't show.
+- **New guide: Measuring retrieval score and response quality**
+  (`docs/scoring-guide.md`). How to calculate each score, what to do when a source
+  has no evaluator (and why a proxy from telemetry is only a rough signal),
+  choosing the floors from your own data, how the scores become alerts, and what
+  the numbers can't tell you. The Logs Details column and its troubleshooting row
+  now explain that a dash means content storage is off.
+- **Documentation refresh.** The architecture diagram is restyled to the new
+  look (SVG and PNG) and notes the dashboard's shortcuts and shared links. The
+  dashboard guide, demo outline, setup guides and architecture page use the new
+  names (**Env**, **Range**, instruments, track), and every screenshot is
+  retaken with a full week of data, plus new ones for Policy and the shortcut
+  sheet. `docs/development.md` lists the design files.
+- **`PRODUCT.md`** records the product's users, purpose, positioning and
+  constraints for design work.
+- **Design system** (`docs/design-system.md`) rewritten for the signal box
+  look. The **dashboard guide** covers the new Overview, shortcuts, shared
+  links, bulk actions and the policy preview, and its screenshots are
+  retaken.
+- **`SETUP.md`**: a step-by-step guide for teammates, from a fresh laptop to
+  a running DriftGuard with sample data, on Windows, macOS and Linux.
+- **Getting started** (`docs/quickstart.md`) rewritten: three ways to run it
+  (`demo.bat`, from source, Docker), four ways to send data (sample data, live
+  demo, Claude Code, SDK), and a troubleshooting table.
+- New **How DriftGuard works** (`docs/concepts.md`): the problem, the two kinds
+  of events, a glossary, scoring and diagnosis with worked examples, and a
+  60-second summary.
+- **Architecture diagram** as an image (`docs/images/architecture.svg`, plus a
+  PNG for slides), and new diagrams of the dashboard and of the telemetry, agent
+  event and Claude Code hook flows. Also a server module table and a
+  view-to-endpoint table.
+- **Dashboard guide** rewritten with screenshots and a table for every card and
+  column, plus a map of which source fills which view.
+- The demo guide gains a 15-minute presentation outline. The README gains a
+  "Start here" table, the diagram and a screenshot.
+
+## [0.4.0] - unreleased
+
+A rebuild of the foundation. v0.3 described several features as done that
+didn't work: authentication, PostgreSQL, live alerts, and retention. This
+release implements them, with tests.
+
+### Security
+- **Removed the unauthenticated `default` account.** v0.3 let any request with
+  no token act as a shared account, so anyone could list, read, and change
+  those projects.
+- **Real authentication.** Email and password accounts, with passwords hashed by
+  scrypt. `POST /auth/register`, `/auth/login`, `/auth/logout`, `GET /auth/me`,
+  and `DELETE /auth/me`. v0.3's `POST /auth/session?account_id=` handed out a
+  token to anyone who knew an account ID.
+- Sessions are revocable. Each JWT's `jti` is stored, so logout and account
+  deletion invalidate it.
+- **Project API keys are hashed**, shown once, and can be rotated
+  (`POST /projects/{id}/api-key`). Existing plaintext keys are hashed on first
+  start. Keys now have a `dg_live_` prefix.
+- **Fixed project takeover.** v0.3 used `INSERT OR REPLACE`, so creating a
+  project with another account's ID overwrote it. Creating an existing ID now
+  returns `409`.
+- A project API key can no longer change policy. Policy changes, deletion, key
+  rotation, and alert changes need the owner's session.
+- The WebSocket requires a session token or API key, sent as a subprotocol so
+  it stays out of URL access logs.
+- Production refuses to start without a strong `DRIFTGUARD_SECRET_KEY`.
+- Rate limits: separate limits for login (per IP), the API (per session or
+  key, 300/min), and ingestion (per key). The
+  Redis limiter is atomic and uses wall-clock time; v0.3 used
+  `time.monotonic()`, which isn't comparable across processes.
+- Scrubbing now covers API keys and tokens (OpenAI, Anthropic, GitHub, AWS,
+  Google, Slack, bearer, private keys). Card numbers need a valid Luhn checksum,
+  so ordinary long numbers are no longer redacted.
+- Security headers on every response, and HSTS in production.
+- Upgraded to FastAPI 0.14x and Starlette 1.x. Starlette 0.38 had published
+  advisories.
+
+### Added
+- `driftguard.server`: an app factory (`create_app`) and a working
+  `driftguard-server` command. The old entry point pointed at an ASGI object and
+  couldn't run.
+- **Automatic alerts.**
+  - Critical telemetry raises a `drift` alert, at most one per root cause per
+    cooldown.
+  - A blocked agent task raises one `agent` alert. Its wasted-token figure keeps
+    updating, and the alert resolves itself when the task recovers.
+  - Alerts have `id`, `source`, `task_id`, `root_cause`, and
+    `resolved`/`resolved_at`.
+  - `PATCH /projects/{id}/alerts/{alert_id}` resolves or reopens an alert.
+- **Server-side scoring.** Each ingested event stores `risk_score`, `severity`,
+  `root_cause`, and `recommendation`.
+- **Batch ingestion:** `POST /events/{id}/batch` and
+  `POST /agent-events/{id}/batch`, up to 500 events each.
+- Events accept `occurred_at` and `metadata`. Agent events accept `input_hash`.
+- `GET /projects/{id}/summary` rewritten: SQL aggregates, `violation_rates`,
+  `events_by_severity`, `top_root_causes`, and open-alert totals.
+- `GET /projects/{id}/agent-events` with a `task_id` filter. Event lists page
+  with `before_id`.
+- Realtime alerts at `WS /ws/projects/{id}` with an `action` field (`created`,
+  `resolved`, `reopened`), fanned out across workers through Redis.
+- Slack and generic webhook notifications, delivered on a background thread.
+- A scheduled retention job (`DRIFTGUARD_RETENTION_DAYS`). It prunes telemetry,
+  agent events, resolved alerts, and expired sessions. v0.3 had the code but
+  never ran it.
+- **Agent diagnosis v2.**
+  - Per-task results in `tasks`.
+  - Recovery detection: a later success of the same tool.
+  - Policy settings `blocked_after_failures` and `retry_window_minutes`.
+  - Redundant-call detection.
+  - Recommendations based on the error type.
+- **Adapters.**
+  - `@driftguard_tool` supports async functions. `AgentContext` uses
+    contextvars and supports `async with` and `auto_sync`.
+  - `record_llm_usage()` attributes model tokens to tool calls.
+  - `usage_from_response()` reads OpenAI, Anthropic, and Gemini responses.
+  - `MCPMiddleware.instrument_session()` wraps an MCP `ClientSession`.
+  - Inputs are fingerprinted automatically.
+- **SDK.**
+  - Batched sync; failed batches stay queued in order (v0.3 could drop or
+    duplicate events).
+  - A bounded queue and client-side `occurred_at` timestamps.
+  - An optional `project_id` default, plus `flush()` and
+    `sync_agent_events_async()`.
+- **Dashboard** moved into `frontend/` from the separate `Design-` repo,
+  rebranded to DriftGuard, and rebuilt against the real API.
+- Schema setup is safe when several workers start at once (a PostgreSQL
+  advisory lock; retried on SQLite).
+- The server closes its database connection pool on shutdown.
+- Docker: a multi-stage, non-root image that bundles the dashboard, with a
+  healthcheck. Compose uses healthchecks, and Postgres and Redis are no longer
+  exposed on the host.
+- CI: ruff, pytest on Python 3.10–3.13, the full server suite on PostgreSQL 16
+  and Redis 7, the dashboard build, and security scans that fail on findings.
+  Releases check that the wheel bundles the dashboard.
+- `examples/live_demo.py` and [docs/demo.md](docs/demo.md): a step-by-step live
+  demo covering drift, a blocked agent, and automatic recovery.
+  `demo.bat` starts the whole demo on Windows in one step.
+- Docs: new API reference, architecture, configuration, deployment,
+  development, and roadmap pages. `DESIGN.md` moved to `docs/design-system.md`. All existing docs were rewritten to match the
+  code. A `LICENSE` file was added (MIT, as `pyproject.toml` already declared).
+
+### Changed
+- **Package extras.** `pip install driftguard` now installs only `httpx`. The
+  server needs `driftguard[server]`; add `[postgres]` and `[redis]` as needed,
+  or `[all]`. NumPy and pandas are no longer dependencies.
+- **Persistence rewritten** on SQLAlchemy Core. PostgreSQL works; v0.3 used
+  SQLite-only SQL and silently fell back to a local SQLite file. There are no
+  in-memory caches, so it's safe with several workers. Filtering happens in SQL.
+- `recommend_mitigation` and `check_drift` use the project policy's thresholds;
+  v0.3 compared against hardcoded values.
+- Root cause text: "trajectory quality degradation" is now "response quality
+  degradation".
+- The policy returns `blocked_after_failures` as an integer (`3`, not `3.0`),
+  matching what `PUT /projects/{id}/policy` accepts.
+- `.env.example` lists every server variable, including the `driftguard-server`
+  defaults `DRIFTGUARD_HOST`, `PORT` and `DRIFTGUARD_WORKERS`.
+- Configuration uses `DRIFTGUARD_*` names. `JWT_SECRET`, `API_SECRET`, and
+  `RATE_LIMIT_PER_MINUTE` still work but are deprecated.
+- The research prototypes (`data`, `detector`, `benchmark`, `experiment`,
+  `reporting`, and the demo API) moved to `research/`.
+
+### Removed / breaking
+- These routes are gone:
+  - `POST /accounts` → use `/auth/register`
+  - `POST /auth/session` → use `/auth/login`
+  - `GET /accounts/me` → use `/auth/me`
+  - `POST /alerts` → use `POST /projects/{id}/alerts`
+  - `GET /alerts/{id}` → use `GET /projects/{id}/alerts`
+  - `GET /events/{id}` → use `GET /projects/{id}/events`
+  - `/projects/{id}/metrics` and `/projects/{id}/dashboard` → use
+    `/projects/{id}/summary`
+  - `WS /ws/alerts/{id}` → use `WS /ws/projects/{id}`
+- `driftguard.routes`, `driftguard.service`, `driftguard.middleware`,
+  `driftguard.config`, `driftguard.workers`, and `driftguard.notifications` moved
+  under `driftguard.server`. `driftguard.analytics` and
+  `driftguard.dashboard_data` were removed; `/summary` replaces them.
+- `driftguard.compute_drift_scores` was removed from the public API. It was the
+  synthetic research detector and now lives in `research.detector`. Use
+  `driftguard.evaluate_drift` for LLM telemetry.
+- `capture_metrics` no longer adds a `project` key to the payload.
+- `requirements.txt` was removed; `pyproject.toml` is the single source of
+  dependencies.
+- The legacy static dashboard (`static/dashboard.js`, `styles.css`) and the
+  1,200-line embedded HTML dashboard in `api.py` were removed.
+
+## [0.3.0] - 2026-08
+
+- The SDK client, FastAPI server, project API keys, agent event ingestion, and
+  the first blocking-tool diagnosis.

driftguard_tokepi-0.5.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Sean Pereira
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

driftguard_tokepi-0.5.0/MANIFEST.in

@@ -0,0 +1,5 @@
+include README.md CHANGELOG.md LICENSE
+recursive-include driftguard/server/static *
+prune frontend
+prune research
+prune tests