detectkit 0.44.0__tar.gz → 0.45.0__tar.gz

{detectkit-0.44.0/detectkit.egg-info → detectkit-0.45.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: detectkit
-Version: 0.44.0
+Version: 0.45.0
 Summary: Metric monitoring with automatic anomaly detection
 Author: detectkit team
 License: MIT

{detectkit-0.44.0 → detectkit-0.45.0}/detectkit/__init__.py

@@ -4,7 +4,7 @@ detectk - Anomaly Detection for Time-Series Metrics
 A Python library for data analysts and engineers to monitor metrics with automatic anomaly detection.
 """
 
-__version__ = "0.44.0"
+__version__ = "0.45.0"
 
 from detectkit.core.interval import Interval
 from detectkit.core.models import ColumnDefinition, TableModel

{detectkit-0.44.0 → detectkit-0.45.0}/detectkit/cli/assets/claude/CLAUDE.section.md

@@ -48,15 +48,25 @@ version — **read the relevant one on demand** instead of guessing:
   values that need your real connection details) and, optionally, a first alert
   channel.
 - **A new metric** — use the **`dtk-new-metric`** skill; it walks the config out
-  to a YAML file that validates and is ready to run.
-- **Tune / pick a detector, or build an alert from scratch and tune it** — use
-  the **`dtk-autotune`** skill; it gathers seasonality + incidents, runs
-  `dtk autotune`, and explains the chosen, annotated config (an automatic,
-  cross-validated search). For **hands-on** tuning instead, `dtk tune` opens a
-  browser view where you drag the detector's knobs against the real series and
-  write the result back in place — and its **Autotune** mode can run that same
-  search server-side and re-seed the knobs without leaving the cockpit (see
-  `cli.md`).
+  to a YAML file that validates and is ready to run. A robust starter is enough —
+  it's the **sandbox** the user then refines with one of the tuning skills below.
+- **Tune a detector (hands-on, recommended)** — use the **`dtk-tune`** skill: it
+  loads some history and opens an interactive browser **cockpit** where the user
+  turns the detector's knobs on their **real** series and watches the band,
+  anomalies and would-fire alerts recompute live, then **Apply**s the result back
+  into the metric YAML in place. It is the **umbrella** for tuning: its **Autotune**
+  mode runs the full `dtk autotune` engine **server-side** and re-seeds the knobs,
+  and its **Label / Review** modes mark or confirm incidents — so the user can
+  auto-search, hand-tune and label in one place. **Prefer this whenever the user
+  wants to be in the loop** (see, judge, adjust). This is also how you build an
+  alert from scratch and dial it in: scaffold the metric (`dtk-new-metric`), then
+  open the cockpit.
+- **Tune a detector (fully automatic)** — use the **`dtk-autotune`** skill: a
+  cross-validated **command-line** search (seasonality → detector → params →
+  window) that writes an annotated `__tuned_<id>.yml` without touching the
+  original. Best for CI or "just pick one for me, I won't look." It runs the same
+  engine `dtk-tune`'s Autotune mode does — pick the skill by whether the user wants
+  hands-on or hands-off.
 - **Hit a detectkit bug, or have feedback** — once you've ruled out a local
   config fix (see the gotchas below), use the **`dtk-feedback`** skill to file a
   redacted bug report, feature request, or comment as a GitHub issue on the

{detectkit-0.44.0 → detectkit-0.45.0}/detectkit/cli/assets/claude/rules/alerting.md

@@ -164,8 +164,7 @@ config needed. Control it project-wide with `alert_help_url` in
 - **a URL string** → your own runbook/wiki page instead
 - **`false`** → hide the link entirely
 
-Per-channel rendering (defaults only; resolved by
-`ProjectConfig.resolve_alert_help_url`):
+Per-channel rendering (defaults only; the resolved help URL is rendered per channel as follows):
 
 - **Slack / Mattermost / generic webhook** — a clickable `How to read this alert`
   label in the compact `Links` field (alongside `Dashboard` + any extra links),
@@ -185,7 +184,7 @@ mirrors `{dashboard_url}` / `{dashboard_line}`. See the template table below.
 
 With no custom `template`, each channel renders a structured, branded message
 (alert-centric: the rule that fired leads, the anomaly value is evidence). The
-shared value computation lives in one place (`BaseAlertChannel.build_context`),
+shared value computation lives in one place,
 so templates and native rendering stay consistent. Every alert title/headline
 leads with a colored **status circle** — 🔴 anomaly, 🟢 recovery, 🟡 no-data,
 🔵 pipeline error — so the status reads from color alone. It also leads with the
@@ -241,8 +240,7 @@ Latest reading** fields bound the span. Labels are self-describing so the onset
 isn't misread as the alert-fire moment: **Anomaly began** is the resolved onset,
 **not** when the alert fired. Recovery shows the fuller **began → fired →
 recovered** timeline (`Incident lasted …`), where **Alert fired** =
-`onset + (consecutive_required − 1) × interval` (computed in `build_context`,
-exposed as `{fired_display}`, omitted when the run is capped). The true
+`onset + (consecutive_required − 1) × interval` (exposed as `{fired_display}`, omitted when the run is capped). The true
 streak/onset is resolved only when an alert fires/clears (a bounded lookback over
 the detection history; a run older than the window shows `over …`), so the hot
 no-alert path stays cheap. Exposed to templates as `{anomaly_lead}` /

{detectkit-0.44.0 → detectkit-0.45.0}/detectkit/cli/assets/claude/rules/overview.md

@@ -114,6 +114,5 @@ series, watch the band recompute live, then write the config back into the metri
 
 ## Authoritative sources
 
-The shipped detectkit docs (`docs/` in the repo, or
-<https://dtk.pipelab.dev>) and the `CHANGELOG.md` are authoritative for
+The detectkit docs at <https://dtk.pipelab.dev> and the project changelog are authoritative for
 behavior. These rule files summarize them for the installed version.

{detectkit-0.44.0 → detectkit-0.45.0}/detectkit/cli/assets/claude/rules/project.md

@@ -45,8 +45,7 @@ overrides it; unset → a built-in `0.5`. Tuning-only — it never touches the p
 ### `alert_help_url` — "How to read this alert" link
 
 Every default-rendered alert on every channel carries a `How to read this alert`
-link for non-operator stakeholders. Tri-state, resolved by
-`ProjectConfig.resolve_alert_help_url`:
+link for non-operator stakeholders. Tri-state:
 
 - **unset / null** (default) → the official detectkit guide
   (`https://dtk.pipelab.dev/guides/reading-alerts/`).

{detectkit-0.44.0 → detectkit-0.45.0}/detectkit/cli/assets/claude/skills/dtk-autotune/SKILL.md

@@ -20,6 +20,14 @@ have them, otherwise an unsupervised objective), and writes a ready-to-run
 config named `<metric>__tuned_<id>` whose comment header explains every
 decision. Each run is recorded in the `_dtk_autotune_runs` table.
 
+**This skill is the fully-automatic, command-line path** — best for CI or when the
+user just wants a strong config picked for them without watching. If the user wants
+to be **in the loop** — see the series, judge the alerts, turn the knobs, mark
+incidents by eye — prefer the **`dtk-tune`** skill: it opens an interactive browser
+cockpit that runs this **same engine** in its **Autotune** mode (server-side,
+re-seeding the knobs) *and* lets them hand-tune and **Apply** in one place. The two
+share the same search; pick by whether the user wants hands-on or hands-off.
+
 Work the steps in order. Do not invent SQL, incident times, or channel names —
 gather them. This skill is the procedure; for field detail read the matching
 file under `.claude/rules/detectkit/` (`autotune.md`, `detectors.md`,
@@ -79,8 +87,9 @@ generalize is rejected.
 Labels turn tuning from a good unsupervised default into a config optimised
 against *your* real incidents (it then optimises MCC and also tunes the alert
 window `consecutive_anomalies`). **The easiest, most reliable way to produce them
-is to mark incidents in `dtk tune` — offer this first**, before asking the user to
-recall timestamps:
+is to mark incidents in `dtk tune` — offer this first**: open the cockpit and
+**guide the user through marking incidents on the chart** (see the **`dtk-tune`**
+skill for driving it), rather than asking them to recall timestamps:
 
 ```bash
 dtk tune --select <name>

{detectkit-0.44.0 → detectkit-0.45.0}/detectkit/cli/assets/claude/skills/dtk-new-metric/SKILL.md

@@ -73,8 +73,12 @@ Pick with the user (see `detectors.md` for the decision table):
 - **Unsure** → `mad` (robust default), `threshold: 3.0`.
 
 If you don't yet know the right detector or parameters, scaffold a robust starter
-(`mad`, `threshold: 3.0`) and let the **`dtk-autotune`** skill pick the detector
-and tune it on real data / labeled incidents after this metric loads.
+(`mad`, `threshold: 3.0`) — treat it as a **sandbox** the user will refine on the
+real series after this metric loads. Two complementary ways to refine it (Step 7):
+the **`dtk-tune`** skill opens an interactive browser cockpit where they turn the
+knobs by hand and watch the band recompute live (with autotune built in), or the
+**`dtk-autotune`** skill searches for the best detector/params automatically. You
+don't need the perfect detector now — a sound starter is enough to open the cockpit.
 
 Set `window_size` to fit the data cadence (non-seasonal 100–500 points; seasonal
 several full cycles) and `min_samples` ≈ 10–30% of it. For a metric with a
@@ -163,6 +167,17 @@ dtk run --select <name> --steps detect    # confirm the detector runs
 dtk test-alert <name>                      # confirm channels work (if alerting)
 ```
 
-Report the created file path and the commands to run it for real. If the user
-wants the detector chosen or tuned automatically (rather than the starter
-config), hand off to the **`dtk-autotune`** skill now.
+Report the created file path and the commands to run it for real. Then offer to
+**refine the starter on the real series** — the metric so far is a sandbox, and
+this is where most of the value is. Pick the path with the user:
+
+- **Hands-on / "let me see and turn the knobs"** → the **`dtk-tune`** skill: it
+  loads some history and opens an interactive browser cockpit where they adjust
+  the detector live and **Apply** the result back. Its **Autotune** mode runs the
+  automatic search *in place* too, so this one path covers both styles — prefer it
+  whenever the user wants to be in the loop.
+- **Fully automatic / "just pick one for me"** → the **`dtk-autotune`** skill: a
+  cross-validated command-line search that writes an annotated tuned config
+  (optionally against incidents you label first).
+
+If they're happy with the robust starter for now, they can just run it and tune later.

detectkit-0.45.0/detectkit/cli/assets/claude/skills/dtk-tune/SKILL.md

@@ -0,0 +1,184 @@
+---
+name: dtk-tune
+description: >-
+  Interactively tune a detectkit metric's detector in the browser cockpit: stand
+  up a sandbox config, open `dtk tune`, and guide the user to turn the detector's
+  knobs on their real series and watch the confidence band recompute live —
+  **with the autotune engine built in** (an Autotune mode that searches
+  server-side and re-seeds the knobs), plus Label/Review to mark or confirm
+  incidents — then write the chosen config back into the metric YAML. Use when
+  the user wants to tune a detector by hand / interactively / in the browser, see
+  and adjust the band or alerts live, set up a tuning sandbox to "turn the knobs"
+  themselves, reduce false alerts hands-on, or visually mark/confirm incidents.
+  This is the primary, hands-on entry point for dialing in a metric; it contains
+  autotune, so prefer it whenever the user wants to be in the loop.
+---
+
+# Tune a detectkit metric interactively (the cockpit)
+
+`dtk tune` opens a localhost **browser cockpit** over a metric's **real**
+persisted series. The user turns the detector's knobs and the confidence band,
+flagged anomalies and would-fire alerts **recompute live** — no DB round-trip,
+nothing leaves the machine — then clicks **Apply** to write the chosen config
+**back into the metric YAML in place** (safely: validated, with the previous
+version archived first).
+
+It is the **hands-on umbrella** for tuning. The cockpit has four modes and one of
+them runs the **real `dtk autotune` engine server-side** and re-seeds every knob
+with the winner — so the user can auto-search *and* hand-tune *and* label
+incidents without leaving the page. **Prefer this skill whenever the user wants
+to be in the loop** (see, judge, adjust). Use the **`dtk-autotune`** skill
+instead only for a fully hands-off, command-line search (e.g. CI, "just pick one
+for me, I won't look").
+
+Your job is to **set up the sandbox, open it, and guide the user through it** —
+they drive the chart; you prepare it, explain each control in plain language, and
+handle the follow-up (`dtk run` / `dtk clean`). This skill is the procedure; for
+field detail read the matching file under `.claude/rules/detectkit/` (`cli.md`
+has the full `dtk tune` reference, plus `detectors.md`, `autotune.md`,
+`alerting.md`).
+
+## Step 0 — A metric with loaded data (the sandbox)
+
+The cockpit charts a metric's **already-loaded** `_dtk_datapoints`, so you need a
+metric that exists and has some history loaded.
+
+- **A project root** contains `detectkit_project.yml`. If `profiles.yml` is still
+  the `dtk init` placeholder, set up the DB first with **`dtk-setup-project`**.
+- **No metric yet?** This is the "stand up a sandbox to turn the knobs" path: hand
+  off to **`dtk-new-metric`** to design the query + a robust **starter** detector
+  (e.g. `mad`, `threshold: 3.0`) — that starter is exactly the sandbox the user
+  will refine here. Keep query design in that skill; never fabricate SQL here.
+- **Load some history** (the more, the better the picture):
+
+  ```bash
+  dtk run --select <name> --steps load                 # incremental
+  dtk run --select <name> --steps load --from 2026-01-01   # backfill more history
+  ```
+
+  If the chart later looks empty or too short, load more with `--from`.
+
+## Step 1 — Open the cockpit
+
+```bash
+dtk tune --select <name>
+```
+
+The selector must resolve to a **single** metric. This starts a `127.0.0.1`
+server and opens the browser. Useful flags:
+
+- **Remote / headless machine** — add `--no-open` and share the printed URL (the
+  server is localhost-only; the user opens it via a tunnel/port-forward).
+- **Focus a window** — `--from <date>` / `--to <date>` bound what's loaded.
+- **No write-back / share a snapshot** — `--no-serve` writes a static read-only
+  HTML preview instead: the sliders still recompute, but there is **no Apply**
+  (and **no Autotune** — that needs the live server); **Save incidents**
+  downloads the labels file instead of writing it.
+
+Tell the user plainly: *"I've opened an interactive chart of your metric — turn
+the knobs and watch the band; tell me what you see and I'll help you decide."*
+
+## Step 2 — Orient the user to the cockpit
+
+One chart fills the screen (the **windshield**); the live quality metrics ride
+**pinned over it**; every control lives in an **always-visible side rail** that is
+**mode-aware** (it shows only the current mode's controls). A **mode switch** above
+the chart picks the job. Navigate a dense series by **scrolling to zoom**,
+**dragging to pan**, double-clicking to reset, and dragging the **navigator
+strip** below the chart. The **Points shown** slider trims to the most-recent N
+points so recompute (and the read) is faster — view-only, never written.
+
+These controls stay visible in **every** mode (they shape the band and the
+alerts): **Points shown**, the alert rule (**direction** + **consecutive
+anomalies**), and the **y = 0** reference-line toggle.
+
+## Step 3 — The four modes (autotune is one of them)
+
+Walk the user to the mode that fits what they want to do:
+
+- **Tune** — *turn the knobs.* The band leads. Adjust **detector type** (MAD /
+  Z-Score / IQR, or **Manual** = fixed lower/upper bounds), **threshold**,
+  **window size**, **recency weighting + half-life**, **detrend**, **smoothing**,
+  and **seasonality groups**; the band + anomalies + would-fire alerts recompute
+  on every change. The window-size / half-life readouts show the equivalent
+  wall-clock span. This is the core "turn the knobs yourself" loop.
+
+- **Autotune** — *let the engine search, then refine.* Click **Run autotune** and
+  the **same engine `dtk autotune` uses** (seasonality → detector → grid → window,
+  cross-validated) runs **server-side over the window currently shown** (the
+  **Points shown** trim — the exact series on screen, not the full history), using
+  any marked/confirmed incidents as ground truth. It then **re-seeds every knob**
+  with the winner and shows the score + decision log. It's **advisory** — it
+  computes and re-seeds only, persists nothing (no run record, no `__tuned_<id>`
+  file, no detections), so the session stays lock-free; the user reviews the band
+  and **Applies**. Each run also streams a structured `LABELS → … → RESULT` log to
+  the terminal you launched from. (This is why `dtk tune` is the umbrella: the user
+  gets autotune **and** hand-tuning in one place.)
+
+- **Label** — *mark the real incidents* (ground truth that sharpens both the live
+  quality metrics and a supervised Autotune). **Drag a span** over each incident;
+  or **Lasso anomalies** (loop a cloud of anomaly dots into one span per run); or
+  **Threshold capture** (grab every span past a horizontal line). **Save
+  incidents** writes a versioned `incidents/<metric>/*.yml` — the same store
+  `dtk autotune` reads.
+
+- **Review** — *confirm the fired alerts.* When a config already looks good, click
+  each alert marker to cycle its verdict un-reviewed → **valid** (green) → **false
+  alarm** (slate). **Confirming an alert valid marks it as a ground-truth
+  incident**, and **Confirm all unreviewed valid** does the lot — so a clean
+  metric is validated in a few clicks **without hand-drawing spans**.
+
+**Suggesting a path.** If the user *can recognise* their incidents on the chart,
+start in **Label** (or **Review** to confirm good alerts) to give the engine
+ground truth, then **Autotune** for a strong config, then **Tune** by eye, then
+**Apply**. If they just want a quick strong default, go straight to **Autotune**
+(unsupervised) and refine. If they want full manual control, stay in **Tune**.
+
+## Step 4 — Read the live quality, then Apply
+
+As they tune, the metrics bar shows **incident catch rate (recall)**,
+**false-alert rate** ("≈1 in N false"), and **reviewed N/M** — recomputed against
+the marked + confirmed incidents (only those inside the loaded window are scored).
+An optional **false-alert budget** (`false_alert_budget`, a fraction in `(0, 1]`,
+on the metric then project, default `0.5`) gently flags the chip when the
+false-alert rate exceeds it (tuning-only; labeling stays optional).
+
+When the user is happy, they click **Apply to metric**. detectkit then, in order:
+**validates** the config (a bad/untunable config is rejected and **nothing is
+written**), **archives** the current YAML verbatim under
+`metrics/.history/<metric>/`, and **re-emits** the metric in place with the tuned
+detector (and updates the first alerting block's `consecutive_anomalies` if it has
+one). Applying ends the session; saving incidents does not.
+
+## Step 5 — Recompute under the new config
+
+`dtk tune` takes **no pipeline lock** and the live preview is a faithful
+*approximation* — the **next `dtk run` is the source of truth**. Because the
+detector params changed, the detector's identity changed, so detections recompute
+under the new id and the old rows are orphaned:
+
+```bash
+dtk run --select <name>                    # load → detect → alert under the new config
+dtk run --select <name> --report           # optional: a self-contained HTML report to confirm behavior
+dtk clean --select <name> --execute        # prune the orphaned old detector rows
+dtk test-alert <name>                      # if alerting is configured
+```
+
+## When to hand off instead
+
+- **Fully automatic, no browser** (CI, "just pick one for me") → the
+  **`dtk-autotune`** skill (it runs the same engine from the command line and
+  writes an annotated `__tuned_<id>.yml` without touching the original).
+- **No metric yet** → **`dtk-new-metric`** to scaffold the sandbox, then come back.
+- **DB not connected** → **`dtk-setup-project`**.
+
+## Final checklist — verify before declaring done
+
+- [ ] Metric exists, resolves to a **single** selector, and has loaded datapoints
+      (ran `--steps load`, backfilled with `--from` if the chart was thin).
+- [ ] You opened `dtk tune` (with `--no-open` + shared URL on a remote machine)
+      and told the user it's an interactive chart they drive.
+- [ ] You explained the relevant mode(s) — including that **Autotune is built in**
+      and runs over the window shown — and which controls stay visible everywhere.
+- [ ] After **Apply**, you ran `dtk run` to recompute and `dtk clean --execute`
+      to prune the orphaned old detector rows (and `dtk test-alert` if alerting).

{detectkit-0.44.0 → detectkit-0.45.0}/detectkit/cli/commands/init_claude.py

@@ -12,7 +12,7 @@ It writes three things into the target directory:
 - ``.claude/rules/detectkit/`` — the reference docs the assistant reads on
   demand (overview, cli, project, metrics, detectors, alerting, autotune).
 - ``.claude/skills/`` — user-facing skills (``dtk-setup-project``,
-  ``dtk-new-metric``, ``dtk-autotune``, ``dtk-feedback``).
+  ``dtk-new-metric``, ``dtk-tune``, ``dtk-autotune``, ``dtk-feedback``).
 
 The source of truth for all of the above lives in ``detectkit/cli/assets/claude``
 and ships with the package, so re-running this command after upgrading detectkit

{detectkit-0.44.0 → detectkit-0.45.0/detectkit.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: detectkit
-Version: 0.44.0
+Version: 0.45.0
 Summary: Metric monitoring with automatic anomaly detection
 Author: detectkit team
 License: MIT

{detectkit-0.44.0 → detectkit-0.45.0}/detectkit.egg-info/SOURCES.txt

@@ -61,6 +61,7 @@ detectkit/cli/assets/claude/skills/dtk-autotune/SKILL.md
 detectkit/cli/assets/claude/skills/dtk-feedback/SKILL.md
 detectkit/cli/assets/claude/skills/dtk-new-metric/SKILL.md
 detectkit/cli/assets/claude/skills/dtk-setup-project/SKILL.md
+detectkit/cli/assets/claude/skills/dtk-tune/SKILL.md
 detectkit/cli/commands/__init__.py
 detectkit/cli/commands/autotune.py
 detectkit/cli/commands/clean.py