@titan-design/session-analytics 0.4.1 → 0.5.0

package/README.md

@@ -29,10 +29,33 @@ priceRequest(
 - `classifySession(facts)` — `agent_spawned`, `human_interactive`, `headless_sdk` or
   `other_headless`, plus `coordinator` or `adhoc` for human sessions.
 - `CONTEXT_BANDS`, `GAP_BANDS`, `bandOf`, `contextBand`, `gapBand`.
-- `costReport(db, { since, until, days, top, transcriptsDiscovered, facetVersion })` — the
-  standing cost report as one JSON object, and `costReportSchema`, its zod schema.
+- `costReport(db, { since, until, days, top, transcriptsDiscovered, facetVersion, actionRules,
+  mechanicalClasses, episodeRoles, noActionClasses, brokerLogLines, handoff })` — the standing cost report as one JSON
+  object, and `costReportSchema`, its zod schema. `byAction` gives each role's cost by action
+  class, and `mechanicalShare` the cost of the `mechanicalClasses` (default
+  `DEFAULT_MECHANICAL_CLASSES`) over the window total, with each role's share over that role's
+  cost. `wakeEpisodes` cuts the wakes of the `episodeRoles` (default `DEFAULT_EPISODE_ROLES`:
+  `coordinator` and `worker:coordinator`) into episodes; see "Wake episodes" below.
+  `handoffThreshold` fits the handoff model per role and model; see "Handoff threshold" below.
+- `buildWakeEpisodes`, `summarizeWakeEpisodes`, `episodeNames`, `wakeEpisodesSchema`,
+  `WAKE_FROM_KINDS`, `DEFAULT_NO_ACTION_CLASSES` — the pure pieces behind `wakeEpisodes`.
+- `handoffThreshold(rows, teleports, agentSessions, options?)`, `sweepK`, `costPerRequest`,
+  `isBootAction`, `parseTeleportEvents`, `handoffThresholdSchema` — the pure pieces behind
+  `handoffThreshold`.
+- `ACTION_CLASSES`, `DEFAULT_ACTION_RULES`, `DEFAULT_MECHANICAL_CLASSES`,
+  `classifyRequest(calls, rules?)` — one action class per request from its tool calls: the
+  first rule in list order that any call matches, `text-only` with no calls, `other` with no
+  match. Rules match a tool name, a Bash command head, or a read or written path.
+- `readRequestToolCalls(db, window)` — each request's tool calls as `ActionCall`s, with the
+  `command_heads`, `file_read` and `file_write` signals session-read extracted for them.
 - `renderCostReportText(report)` — the same report as plain-text tables, ending with
   `LIST_PRICE_CAVEAT` and the price-table and coverage footer.
+- `renderCostReportSections(report, sections)`, `COST_REPORT_SECTIONS` — chosen sections of that
+  text under the same header, with the caveat and footer last; one question's answer, not the
+  whole report.
+- `scope` on `costReport` and `cacheTtlReport` options (`ReportScope`: `sessionIds`,
+  `agentPrefix` on agent-chat names, `roles`), and `scopeFilter(db, scope)` behind it — narrows
+  every request-keyed field to some sessions. Compactions and coverage stay window-wide.
 - `roleFromProfile`, `workerRole(facts)`, `sessionRole(classification, facts)` — worker-v1
   roles, including the standing-peer overlay.
 - `buildEpisodes(input, "worker-v1" | "coordinator-v1")` (pure), `readEpisodeInput`,
@@ -49,6 +72,89 @@ const report = costReport(openDatabase(graphPath, { readonly: true }), { days: 7
 process.stdout.write(renderCostReportText(report));
 ```
 
+## Wake episodes
+
+An episode is one arrival that wakes a session and the requests after it, up to the next
+arrival in the same transcript. An arrival is a turn-start record or a mid-loop delivery, the
+`queued_command` that reaches a busy seat inside a tool loop. A tool result is never an arrival.
+Mid-loop deliveries are most of a busy coordinator's events, so leaving them out would show a
+handful of expensive wakes and hide the rest.
+
+`wakeEpisodes.byCause` gives, per cause, episodes, mid-loop episodes, requests, cost,
+`requestsPerEpisode`, `costPerEpisode`, and the no-action count and cost. Causes are
+session-read's, except that `channel_system` is reported as `agent_lifecycle`: agent-chat's
+notice that an agent exited or changed state. Requests in the window whose arrival came before
+it are counted under `unattributed`.
+
+**No-action rule.** An episode is no-action when every one of its requests has an action class
+in `noActionClasses`, by default `read-investigate`, `text-only` and `other`. An episode with no
+requests, such as the second report of a burst, is no-action. Anything else, including a
+journal write, a PR check or a message, counts as action. The rule reads the TP-501 facet
+rather than regexes over tool input, so the default action rules and a seat's own `actionRules`
+decide it the same way they decide `byAction`.
+
+**From.** Each episode has a sender kind. `broker` is agent-chat itself, the sender of the
+lifecycle notice. `broadcast` is a message whose `msg_id` reached more than one session. `seat`
+is a name carried by a top-level session or by a session spawned with a coordinator profile.
+`agent` is any other named sender, and `none` is an arrival with no sender. `pairs` is the
+sender-by-receiver matrix. A seat sender is named, and every other sender collapses to its kind,
+because spawned agents have one-off names.
+
+## Handoff threshold
+
+When should a session hand over to a fresh one? Each session's terms come from its own requests,
+main thread only, priced from `PRICE_TABLE` rather than the graph's stored price table:
+
+- **Boot cost B**: the cost of every request up to and including the first action, a dispatch,
+  a send (`BOOT_TOOL`) or a file write.
+- **Boot fill f0**: the context tokens at that request.
+- **Growth g**: the mean rise in fill per request after boot. Only rises count, so a
+  compaction's drop does not cancel the growth before it.
+- **Read price p**: the model's cache-read rate.
+
+A cycle runs from f0 to the threshold K in n = (K - f0) / g requests, so a request costs
+B / n + p x (f0 + K) / 2. `cohorts` averages the terms over each role and model, sweeps K over
+`DEFAULT_K_SWEEP` and reports the best K, the extra cost per request at each `configuredK`
+(the charter's `teleport_k` and `retire_k`), and the same sweep in `halfBoot`, where only half
+the boot is overhead. `sessions` carries the per-session terms and best K. A session with no
+boot action, or on an unpriced model, counts in `unbootedSessions`.
+
+`teleports` gives the exit fill of each handover: the fill of the outgoing session's last
+request at or before the broker's `teleport_started` line. The package does not open the broker
+log. Pass its lines as `brokerLogLines`; the `from` agent id maps to sessions through
+`session_origin.agent_id`.
+
+`reviewers` compares `reviewerPrs` reviews (default 10). A fresh reviewer per PR pays its boot
+and its own reads each time. A standing reviewer, priced from the `standingRole` cohort, boots
+once and then reads a context that every earlier PR grew. A row's requests per PR come from the
+reviewer sessions on its own model; a standing model with no reviewers of its own takes the
+mean of the newest reviewer cohort, the model whose latest session ends last; `requestsFrom`
+names that model. With no reviewer session at all it is 0 and `requestsFrom` is `"pooled"`.
+A session with no request after boot does not count toward its cohort's growth. A reviewer's
+review often sits inside its boot (its first write or send comes late), so requests per PR
+count only what follows it and can understate a reviewer that does its work before it writes.
+
+The report reads a window, so a session that started before it has its boot cut short.
+
+## Cache TTL what-if
+
+`cacheTtlReport(db, { since, until, days })` asks what `CLAUDE_CODE_PROMPT_CACHE_TTL=5m` would
+have saved against the 1h TTL. `cacheTtlWhatIf(rows, prices?)` is the pure core over
+`TtlRequestRow`s, which `readTtlRows(db, window)` reads with the cost report's roles and each
+session's spawn profile. `cacheTtlWhatIfSchema` is its zod schema and `renderCacheTtlText` its
+text form.
+
+- **Reprice.** Every 1h write is priced at the 5m write rate instead, from `findPrice`.
+- **Rebuild.** A request whose gap falls in `REBUILD_GAP_BANDS` (the `gapBand`s from 5 minutes
+  up) finds a 5m cache gone. Its cache read is charged again at the 5m write rate, less the read
+  it no longer pays. Past an hour the 1h cache had expired too, so the read is already near zero
+  and only the reprice applies.
+- **Net.** Per role and per profile: reprice saving less rebuild cost, and the same per session.
+  `lossRoles` lists the roles whose net is negative, typically seats that wait on CI.
+
+The gap is the request's `gap_ms` when the miner stored one. Otherwise it is the time since the
+session's previous request on the same thread, which may fall before the window.
+
 ## Things that will bite you
 
 Fable's cache read is **0.025** of its input rate, not the 0.1 every other model uses.
@@ -59,6 +165,27 @@ because defaulting bills a new model at an old model's rate without saying so. T
 report lists such models under `unpricedModels`.
 
 The cost report prices through the graph's `price` table, not through `PRICE_TABLE`. A graph
-whose price rows were never synced reports every request as unpriced.
+whose price rows were never synced reports every request as unpriced, and one synced from an
+older `PRICE_TABLE` keeps pricing at the old rates. The cost report takes a caller-opened,
+read-only graph and never writes it. Before reporting, the caller must run session-graph's
+`reconcilePrices(graph, PRICE_TABLE, { tableVersion: PRICE_TABLE_VERSION, source: "session-analytics" })`
+through a writable connection; `titan-miner` does so on every open, other openers do not.
+
+The default action rules are generic. Rules that name a seat's own journal files or scorer
+scripts belong in the caller's config, passed as `actionRules`, never in this package. The
+defaults read session-read's `command_heads` signal, which keeps only a path's shape, so
+`gh api -X PUT repos/o/r/pulls/5/merge` reaches the classifier as `gh api PUT pulls/merge`. No
+default rule matches that head yet, so it is not a merge.
+
+A tool call belongs to the latest request at or before it in its transcript, the request that
+issued it. The `context_contribution` view maps the other way, to the request a block feeds.
+
+session-read keeps `from` and `msg_id` from a channel tag but not its `broadcast` attribute, so
+a broadcast is inferred from its `msg_id` reaching more than one session. A multicast to named
+recipients counts as a broadcast too.
+
+Coordinator seats spawned with the `opus-coordinator` profile report as `worker:coordinator`.
+Before that profile was mapped they fell into `worker:unknown`, and the `coordinator` role held
+only human-driven seats.
 
 Full reference: `site/reference/session-analytics.md`.