Files
zzg 8a775b71fb Add fathom insights and a Stop hook, so findings come to you
A finding you have to remember to go and look at is a finding you do not see.
This closes the loop: `fathom insights` reports findings for other programs,
a Claude Code Stop hook runs it when a turn ends, and /insights shows the
detail on request.

The CLI reports as text, --json, or --jsonl, and --new returns only what has
not been reported before. The server owns the reported-findings state, because
two sessions can end at the same moment and a read-then-write would have both
report the same thing.

Two rules follow from running on every turn:

- It must be fast. --hook asks a running server for an analysis already
  computed rather than doing one, and never falls back to analysing directly,
  which could take seconds. Measured at ~20ms whether or not a server is up.
- It must never disrupt the session it observes. Every path exits 0 and prints
  at most one line. A Stop hook exiting 2 tells Claude Code to keep going
  rather than stop, so a crash here would hijack the session it was supposed to
  quietly watch.

Surfacing is a higher bar than appearing in the tab, and biased towards
silence. A tab is consulted; a notification arrives whether or not you wanted
it, and the costs are asymmetric -- a missed finding waits patiently, a wrong
nudge trains you to ignore all of them. Medium severity and $0.50 minimum, not
already reported unless it has recurred since, and at most one nudge every two
hours.

A local model, if one is running, can veto a finding as not worth interrupting
for. It can never promote one that did not clear the bar, is never consulted
about findings below it, and a model failure keeps the finding rather than
silencing it. Findings it rejects are recorded, so it is not asked the same
question every turn.

Also fixes a latent hazard this would have made routine. OpenShard runs tail
recovery, which truncates, so a CLI opening a store the server is mid-append to
could cut off the chunk being written -- the reader destroying the writer's
data. Added a read-only open that skips recovery and opens O_RDONLY, and
pointed every read-only command at it, including `fathom query`, which had the
same problem already.

The nudge's cost figure counted each finding separately, so a set of them could
claim more money than the window ever spent: $12.52 against a window costing
$7.94. Now counted once per trace.

claude/ ships the skill and the hook snippet with an installer. The snippet is
printed rather than applied, because a script should not blindly edit a
settings file that may already have hooks in it, and a mangled settings.json
silently disables everything in it.
2026-09-16 00:26:11 -07:00

300 lines
8.6 KiB
Go

package insights
import (
"context"
"encoding/json"
"fmt"
"os"
"sort"
"strings"
"time"
)
// Surfacing decides what is worth interrupting someone about.
//
// This is a different and much higher bar than appearing in the UI. A tab is
// consulted; a notification arrives whether or not you wanted it, at the end of
// a turn, while you are thinking about something else. The cost of a wrong call
// is asymmetric — a missed finding waits patiently in the tab, while a wrong
// nudge trains you to ignore all of them — so everything here is biased towards
// staying quiet.
// SurfaceOptions tune the bar.
type SurfaceOptions struct {
// MinSeverity is the floor. Anything below is never surfaced, whatever a
// model thinks of it.
MinSeverity Severity
// MinCostUSD is a second floor, because severity alone rounds too coarsely
// at the bottom.
MinCostUSD float64
// Cooldown is the minimum gap between nudges, regardless of how many new
// findings appear. Several notifications in a row is how a feature like
// this gets turned off.
Cooldown time.Duration
}
func DefaultSurfaceOptions() SurfaceOptions {
return SurfaceOptions{
MinSeverity: SevMedium,
MinCostUSD: 0.50,
Cooldown: 2 * time.Hour,
}
}
// Surfaced is the record of what has already been reported.
type surfacedState struct {
// Seen maps insight id to the LastSeen it had when reported. A pattern that
// recurs after being reported becomes new again, which is the behaviour
// worth having: a habit you were told about and then kept doing is a
// stronger signal than the first occurrence was.
Seen map[string]int64 `json:"seen"`
// LastNudge is when the reader was last interrupted, for the cooldown.
LastNudge int64 `json:"last_nudge"`
}
// Nudge is what the hook reports.
type Nudge struct {
// Message is the line shown to the reader. Empty means say nothing.
Message string `json:"message,omitempty"`
// Insights are the findings that cleared the bar, most important first.
Insights []Insight `json:"insights,omitempty"`
// Suppressed explains why nothing is being shown, for debugging a hook
// that appears to do nothing. It is not shown to the reader.
Suppressed string `json:"suppressed,omitempty"`
}
func severityRank(s Severity) int {
switch s {
case SevHigh:
return 3
case SevMedium:
return 2
case SevLow:
return 1
}
return 0
}
// SurfaceNew returns findings not yet reported, and records them.
//
// Returning and recording in one call is deliberate. Two sessions can end at
// the same moment, and a read-then-write would have both report the same
// finding.
func (a *Analyzer) SurfaceNew(opts SurfaceOptions) Nudge {
res, ready := a.Latest()
if !ready {
return Nudge{Suppressed: "no analysis has completed yet"}
}
a.mu.Lock()
defer a.mu.Unlock()
state := a.loadSurfaced()
now := a.now()
if state.LastNudge > 0 && opts.Cooldown > 0 {
since := now.Sub(time.UnixMicro(state.LastNudge))
if since < opts.Cooldown {
return Nudge{Suppressed: fmt.Sprintf(
"within the %s cooldown (last nudge %s ago)",
opts.Cooldown.Round(time.Minute), since.Round(time.Minute))}
}
}
var candidates []Insight
var belowBar int
for _, th := range res.Themes {
for _, in := range th.Insights {
if severityRank(in.Severity) < severityRank(opts.MinSeverity) || in.CostUSD < opts.MinCostUSD {
belowBar++
continue
}
if when, ok := state.Seen[in.ID]; ok && in.LastSeen <= when {
continue // already reported, and it has not happened again since
}
candidates = append(candidates, in)
}
}
if len(candidates) == 0 {
return Nudge{Suppressed: fmt.Sprintf(
"nothing new above the bar (%d finding(s) below it)", belowBar)}
}
sort.SliceStable(candidates, func(i, j int) bool {
if severityRank(candidates[i].Severity) != severityRank(candidates[j].Severity) {
return severityRank(candidates[i].Severity) > severityRank(candidates[j].Severity)
}
return candidates[i].CostUSD > candidates[j].CostUSD
})
// A local model gets a veto, never a promotion. It can decide a finding
// that cleared the numeric bar is still not worth an interruption; it can
// never surface one that did not clear it. Detection and thresholds stay
// where they can be reasoned about.
if a.status.Enabled && a.judge != nil {
kept := candidates[:0]
for _, in := range candidates {
worth, reason := a.judgeWorth(in)
if worth {
kept = append(kept, in)
continue
}
_ = reason
}
if len(kept) == 0 {
// Record them anyway. The model said these are not worth an
// interruption; asking it again about the same findings on every
// turn would burn time to reach the same answer.
a.recordSurfaced(state, candidates, now, false)
return Nudge{Suppressed: "the local model judged these not worth interrupting for"}
}
candidates = kept
}
a.recordSurfaced(state, candidates, now, true)
return Nudge{
Message: nudgeMessage(candidates),
Insights: candidates,
}
}
// nudgeMessage writes the line the reader sees.
//
// Deliberately templated rather than model-written. It is the one piece of text
// that arrives unbidden, so it should read the same way every time and say
// exactly what it can support: what was found and where to look.
func nudgeMessage(ins []Insight) string {
const call = "Run /insights to learn more."
if len(ins) == 1 {
what := ins[0].Label
if what == "" {
what = strings.TrimSuffix(ins[0].Title, ".")
}
return fmt.Sprintf("You've got insights on your Claude usage — %s (%s). %s",
what, money(ins[0].CostUSD), call)
}
return fmt.Sprintf("You've got insights on your Claude usage — %d new findings covering %s. %s",
len(ins), money(coveredCost(ins)), call)
}
// coveredCost is the spend the findings actually speak to, counted once per
// trace.
//
// Summing each finding's cost double-counts, because several findings routinely
// name the same traces — one run can be both a repeated sequence and a repeated
// ask. Naively summed, a set of findings can claim more money than was spent in
// the whole window, which is exactly what the first version of this message
// did: "$12.52" against a window that cost $7.94.
func coveredCost(ins []Insight) float64 {
perTrace := map[string]float64{}
for _, in := range ins {
for _, e := range in.Evidence {
if e.CostUSD > perTrace[e.TraceID] {
perTrace[e.TraceID] = e.CostUSD
}
}
}
var total float64
for _, c := range perTrace {
total += c
}
// A finding with no evidence cannot be attributed to a trace; fall back to
// its own figure rather than reporting nothing.
if total == 0 {
for _, in := range ins {
total += in.CostUSD
}
}
return total
}
// judgeWorth asks the local model whether a finding justifies an interruption.
func (a *Analyzer) judgeWorth(in Insight) (bool, string) {
ctx, cancel := context.WithTimeout(context.Background(), 12*time.Second)
defer cancel()
worth, reason, err := a.judge.Worth(ctx, JudgeInput{
Title: in.Title,
Body: in.Body,
Occurrences: in.Occurrences,
CostUSD: in.CostUSD,
})
if err != nil {
// A model that is slow, missing, or confused must not silence a finding
// that already cleared every numeric bar. Failure keeps it.
return true, "model unavailable"
}
return worth, reason
}
// JudgeInput is what the model is shown. It gets the finding's own words and
// numbers and nothing else — no traces, no prompts, no file contents.
type JudgeInput struct {
Title string
Body string
Occurrences int
CostUSD float64
}
// Judge decides whether a finding is worth interrupting for.
type Judge interface {
Worth(ctx context.Context, in JudgeInput) (bool, string, error)
}
func (a *Analyzer) loadSurfaced() *surfacedState {
st := &surfacedState{Seen: map[string]int64{}}
if a.surfacedPath == "" {
if a.surfaced != nil {
return a.surfaced
}
a.surfaced = st
return st
}
b, err := os.ReadFile(a.surfacedPath)
if err != nil {
return st
}
if err := json.Unmarshal(b, st); err != nil || st.Seen == nil {
return &surfacedState{Seen: map[string]int64{}}
}
return st
}
func (a *Analyzer) recordSurfaced(state *surfacedState, ins []Insight, now time.Time, nudged bool) {
for _, in := range ins {
state.Seen[in.ID] = in.LastSeen
}
if nudged {
state.LastNudge = now.UnixMicro()
}
a.surfaced = state
if a.surfacedPath == "" {
return
}
b, err := json.MarshalIndent(state, "", " ")
if err != nil {
return
}
tmp := a.surfacedPath + ".tmp"
if err := os.WriteFile(tmp, b, 0o644); err != nil {
return
}
os.Rename(tmp, a.surfacedPath)
}
// ResetSurfaced forgets what has been reported, so everything is new again.
func (a *Analyzer) ResetSurfaced() {
a.mu.Lock()
defer a.mu.Unlock()
a.recordSurfaced(&surfacedState{Seen: map[string]int64{}}, nil, a.now(), false)
}