docs: design the ambient drift coach (always-on sentinel)

The first Control loop / scheduler component (architecture §3): an
always-on sentinel beside the one-mode harness that judges activity
against the ~/owc frame when no session is declared, and surfaces drift
to the status bar, a notify-send toast, and a web-UI banner.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
2026-06-05 21:03:10 -04:00
parent 682eb603fa
commit 9d7786c3e6
@@ -0,0 +1,298 @@
# Ambient Drift Coach — Design
**Date:** 2026-06-05
**Status:** Approved, ready for implementation planning
**Scope:** Give Keel an always-on presence. Today Keel only coaches *inside* a
declared focus session; the moment you are just working, the harness goes idle,
drops every window snapshot, and the status bar reads `idle`. This slice adds an
**ambient drift coach**: a sentinel that watches the window stream continuously
(even when no mode is active), judges your recent activity against the broad
`~/owc` frame (goals, values, life-domains) plus today's Marvin tasks, and —
when you have clearly slid off what matters — surfaces a single coaching line to
the status bar, a `notify-send` toast, and a web-UI banner you can act on.
This builds the **Control loop / scheduler** component from
`docs/keel-architecture.md` §3, which today has zero implementation, and is the
first thing that makes Keel *show up unprompted* (SOUL: "when my partner drifts,
I name it").
## Problem
The focus mode is already a capable proactive coach — local on-task matching, an
LLM drift judge, a semantic "nudge" over recent window titles, surfaced to the
status bar and web UI, all sensibly debounced. But every bit of it hangs off a
**declared `Commitment`**. You must open the UI and start a session first. With
no session:
- `harness.RecordWindow` finds no active mode and silently drops the snapshot.
- `statusfile.Render` of an empty envelope returns `idle`.
- No brain call ever happens; Keel is blind and mute.
So the common case — Felix working without ceremonially declaring a task — gets
nothing. The fix is an observer that runs regardless of the active mode and is
grounded in the standing frame instead of a per-session commitment.
## Approach (and the one real fork)
The ambient coach is **not a mode**. It is an always-on **sentinel** that runs
*beside* the one-mode harness. Two harness invariants make "ambient as a mode"
the wrong shape:
1. The harness hosts exactly one mode and drops evidence when idle — an ambient
watcher must run continuously, including when idle.
2. `harness.Start` returns `ErrBusy` if a different mode is active — an ambient
mode would either block you from starting focus, or need a special
auto-yield carve-out that erodes the single-mode contract.
A sentinel sidesteps both and matches §3, where the control loop is a distinct
component beside the harness. **Coexistence rule:** when any harness mode is
active (`h.State().ActiveMode != ""`), the sentinel stays silent — focus mode
already owns coaching during a declared session, so there is never double
coaching.
## Components
Each unit has one responsibility, a defined interface, and is testable alone.
### 1. `internal/notify` — notification effector (new)
The OS-level "interrupt me" primitive, mirroring `enforce.Guard`.
- **Interface:** `Notifier interface { Notify(ctx context.Context, title, body string) error }`
- **Linux adapter:** shells `notify-send <title> <body>`. Best-effort: returns
an error for diagnostics; callers never block on it.
- **Constructor:** `NewNotifier() Notifier` returns the real adapter when the
`notify-send` binary is found on `PATH` (via `exec.LookPath`), else a nop. On
non-Linux it is a nop. This means a missing binary degrades to "no toast,"
never a failure.
- **Nop:** `Notify` returns nil, does nothing.
- **Depends on:** `os/exec` only. Leaf package.
### 2. `internal/frame` — shared frame assembler (new; DRY extraction)
The "what matters to Felix" brief, today duplicated inside off-screen's
`collect.go`. Extracting it removes the duplication and gives the sentinel the
same grounding off-screen uses.
- **Interface:** `Assemble(ctx context.Context, know knowledge.Source, tasks tasks.Provider) string`
returns a bounded markdown brief with three sections: **Today's tasks**
(from `tasks.Today`), **Goals** (`~/owc/goals-2026.md` via `knowledge.Load`),
and **Life domains** (every `~/owc/resources/bug-*.md` via `knowledge.Load`).
- **Best-effort:** never panics, never errors; tolerates nil ports and missing
files, degrading to a short, mostly-empty brief. Same truncation cap
(`maxBriefBytes`) and rune-safe clipping off-screen uses today.
- **Refactor:** off-screen's `briefTasks` / `briefGoals` / `briefLifeDomains`
are replaced by a call to `frame.Assemble`; off-screen keeps its own
proposal-history section (`briefHistory`) prepended around it. Off-screen's
existing tests must stay green.
- **Depends on:** `knowledge`, `tasks`. Imports nothing from `ai`/`mode`.
### 3. `internal/ai` — `AmbientDrift` method (new; mirrors `Nudge`)
The brain call grounded in the frame, not a commitment.
- **Method:** `func (s *Service) AmbientDrift(ctx context.Context, frame string, recentTitles []string) (string, error)`
returns `""` when the activity plausibly serves the frame (on-track), and a
one-sentence coaching line when it is a slide.
- **Prompt:** an "ambient coach" prompt stating the user has **not** declared a
task; it is handed the frame and the recent window-title sequence and asked
whether the activity serves **any** of what matters, **forgivingly**
legitimate breaks, rest, and unplanned-but-useful work are on-track. Output is
`ONLY` JSON: `{"drifting": <bool>, "message": "<one sentence; REQUIRED when drifting>"}`.
- **Parse (`parseAmbientDrift`):** reuses `extractJSON` and the shared
`ErrEmptyResponse` / `ErrNoJSON` sentinels. `drifting:false``""`.
`drifting:true` with a message → the message. `drifting:true` with no message →
`""` (silence) — the ambient signal is advisory, so the safe degenerate is to
say nothing, exactly like `parseNudge`.
- **Depends on:** the existing `Backend`. `ai` stays a leaf (primitives only).
### 4. `internal/ambient` — the Sentinel (new; the meat)
Always-on observer: consume evidence → tick on cadence → judge → surface.
**Construction:** `New(deps)` where deps carry `*ai.Service`, `knowledge.Source`,
`tasks.Provider`, `notify.Notifier`, `memory.Store`, `clock func() time.Time`,
and `activeMode func() string` (reads `h.State().ActiveMode` for the coexistence
rule). Config: `cadence time.Duration` and `mode` (`off` | `status` | `notify`).
**State (under one mutex):** `recentTitles []string` (consecutive-dedup ring,
capped like focus's `recentTitlesMax`), `lastWindow evidence.WindowSnapshot`,
`line string` (current drift line; `""` when on-track), `lastEvalTitles`
(snapshot of the ring at the last brain call), `wasDrifting bool`,
`episodeNotified bool`, `snoozeUntil time.Time`, `gen int`, and `onChange []func()`.
**Methods:**
- `OnWindow(snap evidence.WindowSnapshot)` — called from the evidence fan-out.
Updates `lastWindow`, appends the title to the ring (skip consecutive dup,
cap). Cheap and lock-bounded; fires no brain call.
- `Run(ctx context.Context)` — cadence loop: a ticker every `cadence` calls
`evaluate()`; returns on ctx cancel.
- `evaluate()` — the decision (see Trigger discipline below).
- `Line() string` — current ambient line for the surfaces (`""` when on-track,
quiet, or snoozed).
- `Snooze(d time.Duration)` — sets `snoozeUntil = clock()+d`; clears the line and
episode flags; fires `onChange`. While snoozed, `evaluate()` returns early.
- `SetConfig(cadence time.Duration, mode string)` — live settings update.
- `AddOnChange(f func())` — register a surface refresh (status wake, web
broadcast). Fired whenever the line changes.
**Depends on:** `ai`, `frame`, `notify`, `memory`, `evidence`, and a
`activeMode` closure. It does **not** import `harness` (closure injection keeps
it decoupled and testable).
### 5. Wiring — `cmd/keeld/main.go`
- **Build** the notifier (`notify.NewNotifier()`) and the sentinel from settings
(cadence, mode) and the shared ports (the same `ai.Service`, `knowledge`,
`tasks`, `memory`, clock used to build `harness.Services`), with
`activeMode: func() string { return h.State().ActiveMode }`.
- **Fan the evidence stream to both** the harness and the sentinel:
```go
go src.Watch(ctx, func(w evidence.WindowSnapshot) {
h.RecordWindow(w)
sentinel.OnWindow(w)
})
```
- **Run** the sentinel: `go sentinel.Run(ctx)`.
- **Surface refresh:** register both the status writer's `Wake` and the web
broadcast with `sentinel.AddOnChange(...)`, so a new ambient line re-renders
the bar and pushes over SSE promptly.
- **Live settings:** `applyFn` (the existing POST `/settings` re-wire) also calls
`sentinel.SetConfig(...)` from the new settings, so the dial takes effect
without a restart.
## Trigger discipline (aggressive, but not trigger-happy)
A `notify-send` toast on a false positive is far more irritating than a stale
status line, so `evaluate()` is conservative and brain-call-thrifty:
1. If `mode == off` → return.
2. If snoozed (`clock() < snoozeUntil`) → return.
3. If `activeMode() != ""` → a mode is coaching; clear the line (if set), fire
`onChange`, return. (Coexistence rule.)
4. **Cheap gate:** if the ring is unchanged since `lastEvalTitles` **and** we are
currently on-track (`line == ""`) → skip the brain call. This naturally
pauses calls during steady deep work *and* while AFK (no new titles arrive in
either case), while still re-checking an unresolved drift.
5. Assemble the frame (`frame.Assemble`) and call
`ai.AmbientDrift(ctx, frame, recentTitles)` under a timeout.
6. Apply the result:
- **On-track (`""`):** if a line was set, clear it and reset
`wasDrifting=false`, `episodeNotified=false`; fire `onChange`. Otherwise no-op.
- **Drifting (`msg`):**
- **New episode** (`!wasDrifting`): set `line=msg`, `wasDrifting=true`,
`episodeNotified=false`; fire `onChange`. The **status bar / banner show
immediately**, but **no toast yet**.
- **Sustained** (already `wasDrifting` on this consecutive eval): record an
`ambient_nudge` event to `memory` (`{ "message": msg, "title":
lastWindow.Title, "class": lastWindow.Class }`), and — if `mode == notify`
and `!episodeNotified` — fire exactly one `notify-send` toast and set
`episodeNotified=true`. Further sustained ticks do not re-toast.
- On a brain **error**, leave the prior line intact and log — never fabricate
drift (same discipline as `Nudge`).
7. Update `lastEvalTitles` to the current ring.
Net effect with the default ~5-minute cadence: a 40-second glance at Twitter
never pops a toast (it resolves before the second tick); a sustained ~10-minute
slide shows on the bar at the first tick and pops one toast at the second.
Memory records only *sustained* drifts, independent of the surface mode.
## Surfaces
- **Status bar (`internal/statusfile`):** `Render` gains the current ambient line
(passed in by the `Writer`, which reads `sentinel.Line`). When the envelope is
idle (`ActiveMode == ""`): a non-empty line renders `⚠ <line>`; an empty line
renders `idle` as today. Non-idle envelopes are unchanged. The `Writer` adds an
`ambientLine func() string` source and `sentinel.AddOnChange(writer.Wake)`
drives prompt re-render; the existing minute ticker still applies.
- **Web UI (`internal/web` + `static/app.js`):** the served state gains an
`"ambient": "<line>"` field (read from `sentinel.Line`), pushed over the
existing SSE channel via `sentinel.AddOnChange(broadcast)`. `app.js` renders a
**banner** when the field is non-empty: the coaching line plus two controls —
**Snooze 1h** (`POST /ambient/snooze`) and **Start focus** (reuses the existing
start-session affordance so a nudge converts into a declared session). The
banner clears automatically when the line goes empty.
- **New endpoint:** `POST /ambient/snooze` → `sentinel.Snooze(1 * time.Hour)`.
Returns 204. No body needed in v1.
## Config (the agency-vs-cage dial, §8)
`settings.Settings` gains two fields (env-seedable, mirroring the existing ones):
- `ambient_mode` (`KEEL_AMBIENT_MODE`): `off` | `status` | `notify`. Default
`notify`. Validated in the settings applier (like `ai_backend`); an unknown
value is rejected with a wrapped sentinel so POST `/settings` can map it to 400.
- `ambient_cadence_secs` (`KEEL_AMBIENT_CADENCE_SECS`): integer seconds between
evaluations. Default `300`. A non-positive value falls back to the default.
The web settings form gains a select (`ambient_mode`) and a number input
(`ambient_cadence_secs`); `applyFn` pushes both into `sentinel.SetConfig`.
## Error handling & graceful degradation
- **No `notify-send` binary / non-Linux:** nop notifier; status line and banner
still work.
- **Brain error / timeout:** `AmbientDrift` returns an error; the sentinel logs
and leaves the prior line intact. Never fabricates drift.
- **AW down:** `memory` is already the nop store; `ambient_nudge` writes are
dropped silently.
- **`knowledge` / `tasks` errors:** `frame.Assemble` degrades to a partial brief
(best-effort), never errors.
- **Headless / no X11:** the evidence source is a no-op, so `OnWindow` is never
called, the ring stays empty, the cheap gate skips every brain call, and the
sentinel is silent. No new X11 requirement.
- **Nothing in the sentinel can crash the daemon** — every external call is
best-effort and lock-bounded.
## Testing
- **`internal/ambient`** (fake clock, fake `ai` with scriptable `AmbientDrift`
returns and a call counter, fake `notify.Notifier` recording calls, fake
`tasks`/`knowledge`, `activeMode` stub, `memory.Fake`):
- on-track → no line, no toast, no memory.
- single drift tick → line set, **no** toast, no memory yet.
- two consecutive drift ticks → line set, **exactly one** toast, one
`ambient_nudge` event.
- third consecutive drift tick → still one toast (once per episode).
- back on-track → line cleared, flags reset.
- `activeMode() != ""` → no eval, line cleared.
- `mode == status` → drift sets the line but never toasts (memory still records
the sustained drift).
- `mode == off` → nothing happens, no brain call.
- snooze → no eval / no line during the window.
- cheap gate → unchanged ring + on-track → assert the fake `ai` call count does
**not** increase.
- **`internal/notify`:** `NewNotifier` with `notify-send` absent → nop; `Notify`
returns nil, no panic. (The real shell path is not unit-tested.)
- **`internal/ai`:** `parseAmbientDrift` — on-track→`""`, drift+msg→msg, drift
with no msg→`""`, empty→`ErrEmptyResponse`, no-JSON→`ErrNoJSON`.
- **`internal/frame`:** `Assemble` with fake ports — sections present; nil ports
and missing files degrade without error; truncation honored. Off-screen's
existing tests stay green after the refactor.
- **`internal/statusfile`:** idle + ambient line → `⚠ …`; idle + empty line →
`idle`; non-idle envelopes unaffected.
- **`internal/web`:** served state includes `ambient`; `POST /ambient/snooze`
routes to `Snooze` and returns 204.
- **`internal/settings`:** `ambient_mode` validation (good values accepted,
unknown rejected with the sentinel); cadence default on non-positive.
## Out of scope (deliberate, for a later slice)
- **AFK precision via the AW afk bucket.** v1 uses the cheap-gate heuristic
(no new titles ⇒ no call), which covers both AFK and deep-focus well enough.
Reading AW's input-derived afk bucket to *prove* presence is a future collector.
- **Ambient enforcement** (window-minimize). The sentinel is advisory only;
minimizing windows with no declared session would be far too cagey. Enforcement
stays a focus-session decision.
- **Trend / spiral detection** over the accumulating `ambient_nudge` history
(§7). This slice *produces* that history; reading it for cross-time patterns is
the next increment.
- **Richer banner actions** beyond snooze + start-focus (e.g. "I'm on it",
per-domain mute).
## Open questions
- **Default cadence.** 300 s is a starting guess; may want tuning after living
with it. Configurable, so cheap to change.
- **Snooze duration.** Fixed 1 h in v1; a picker is a later refinement.