Details dialog, compaction, and the same-engine sub-agent fix (0.4.0)

CHANGELOG.md

@@ -1,3 +1,34 @@
+## [Unreleased]
+### Added
+- A **details dialog**, opened by `details ›` under the sidebar boxes,
+  `ctrl+shift+d` or `/headsup`, and closed the same way or with `esc`. Last
+  turn on the left, session on the right; one at a time below 110 columns,
+  switched with `tab`. The sidebar is unchanged.
+  - Last turn: where the time went, in seconds and shares that add up to the
+    total; every step with its tool calls and their times; tokens in all
+    five kinds and context used; the engine's own figures only, marked `◆`,
+    including MTPLX acceptance at every depth and per-step engine rates;
+    retry reasons in full.
+  - Session: speed and time-to-first-token spread (min, median, p90, max),
+    time split in seconds, tools by time, retries by reason, and coverage:
+    how many turns had engine figures and why the others did not.
+- Compaction is named. OpenCode compacting the conversation mid-turn gets
+  its own share of the turn's time, and the step it delayed says so. On an
+  engine that publishes cumulative counters (Splash, llama.cpp, llamafile,
+  the vLLM family), the compaction's own request is read before and after
+  and taken out of the turn, so the turn keeps its engine figures; where it
+  can't be, the reason reads `compaction ran this turn`.
+
+### Fixed
+- A sub-agent on the same counter engine as its parent no longer makes the
+  parent's turn lose its engine figures. The sub-agent's own report moved
+  the engine-wide starting reading, so the parent's window held only its
+  last step and was declined as `overlapping requests`. Each turn now takes
+  its own starting reading when it starts.
+- When engine figures were declined, the dialog says the engine's figures
+  were left out and why, rather than reading as though the engine reported
+  nothing.
+
 ## [0.3.3] – 2026-09-25
 ### Fixed
 - Generation speed counts the time a model spends writing a tool call's

README.md

@@ -26,7 +26,9 @@ sub-agent   191 tok
 ```
 
 Two boxes, each opened and closed by clicking its heading: the last turn,
-and the session so far.
+and the session so far. `details ›` under them opens the full picture:
+where each turn's time went, every step and tool call, and the session's
+spread and coverage (see [Details](#details)).
 
 Requires [**OpenCode 2**](https://opencode.ai/v2/docs). For the v1 line
 (OpenCode 1.18.x), see
@@ -36,6 +38,7 @@ Requires [**OpenCode 2**](https://opencode.ai/v2/docs). For the v1 line
 
 - [Install](#install)
 - [Keys](#keys)
+- [Details](#details)
 - [Configuration](#configuration)
 - [Supported Engines](#supported-engines)
 - [Engine Details](#engine-details)
@@ -68,10 +71,11 @@ Equivalent, if you keep your config in version control:
 | --- | --- |
 | `ctrl+shift+m` | Collapse/expand the last-turn box. Clicking its heading does the same. |
 | `ctrl+shift+h` | Open/close the per-turn history panel. |
+| `ctrl+shift+d` | Open/close the details dialog. So do `/headsup` and clicking `details ›`. |
 
-Both are registered with stable command ids (`headsup.toggle`,
-`headsup.panel`), so they can be remapped from your own OpenCode keybind
-config and are reachable from the command palette.
+All three are registered with stable command ids (`headsup.toggle`,
+`headsup.panel`, `headsup.details`), so they can be remapped from your own
+OpenCode keybind config and are reachable from the command palette.
 
 The Session box has no key; click its heading. Collapsed, each box keeps
 one figure rather than becoming a bare label:
@@ -80,6 +84,39 @@ one figure rather than becoming a bare label:
 ▸ MTPLX · last turn  34.4 tok/s
 ```
 
+## Details
+
+A dialog with the last turn on the left and the session on the right. On a
+terminal narrower than 110 columns it shows one at a time; `tab` switches.
+It scrolls with the wheel, `↑` `↓` and page up/down; `esc` closes it.
+
+**Last turn**
+- **Where the time went**, in seconds and as shares that add up to the
+  turn's total: waiting for the first token, generating, tools, sub-agents,
+  compaction, and the rest. A moment is counted once, so a tool running
+  beside a sub-agent is not counted twice.
+- **Steps**: tokens, tok/s and time to first token per step, each tool call
+  and how long it ran, retries, and a step that waited on a compaction.
+- **Tokens**: output, reasoning, fresh input, cache read and cache write;
+  context used against the model's limit.
+- **◆ Engine**: only what the engine itself measured -- MTPLX's acceptance
+  at every depth and its verify passes, prefill and decode as tokens and
+  seconds, cache reuse, draft acceptance, and per-step rates where the
+  engine is read per step. When its figures were not used, it says which
+  and why.
+- Retry and error reasons in full, and sub-agent totals.
+
+**Session** (the current model's turns)
+- Speed as an average and a spread (min, median, p90, max), a trend, and
+  time to first token (median, p90, max).
+- Where the time went, in seconds, across the session.
+- Tools by time, retries by reason, tokens in all five kinds.
+- **Coverage**: how many turns had the engine's own figures, and why the
+  rest did not (first turn, compaction, overlapping requests, no engine
+  telemetry, ...).
+
+Everything unmarked is OpenCode's own data; `◆` marks the engine's.
+
 ## Configuration
 
 Two keys, because two things are genuinely preferences. Everything else

adapters/koboldcpp.ts

@@ -17,7 +17,7 @@
 
 import { httpJson, type HttpOptions } from "../http"
 import { nn, ni } from "../format"
-import { rowsOf, timeRows, viewText, nt, type Row, type TurnView } from "../rows"
+import { rowsOf, timeRows, viewText, nt, phase, type Row, type TurnView } from "../rows"
 /**
  * The fields of /api/extra/perf this plugin reads. The endpoint returns more
  * (image/TTS/transcription counters, horde bookkeeping, seeds) that describe
@@ -252,7 +252,19 @@ export function koboldView(
     t.generationsInWindow !== undefined && t.generationsInWindow > 1
       ? [`${ni(t.generationsInWindow)} generations this turn`, "(last shown only)"]
       : []
-  return { engine: "KoboldCpp", rows, notes, key: t.decodeTokS !== undefined ? `${nn(t.decodeTokS)} tok/s` : undefined }
+  const detail: Row[] = [
+    ...rowsOf("speed", [t.decodeTokS !== undefined ? `${nn(t.decodeTokS)} tok/s` : ""]),
+    ...rowsOf("prefill", [t.prefillTokS !== undefined ? `${ni(t.prefillTokS)} tok/s` : "", phase(t.promptTokens, t.prefillS)]),
+    ["decode", phase(t.completionTokens, t.decodeS)],
+    ...rowsOf("draft", [t.draftAcceptRate !== undefined ? `${ni(t.draftAcceptRate * 100)}% accepted` : ""]),
+  ]
+  return {
+    engine: "KoboldCpp",
+    rows,
+    notes,
+    key: t.decodeTokS !== undefined ? `${nn(t.decodeTokS)} tok/s` : undefined,
+    detail,
+  }
 }
 
 /** The view as text; kept for tests that look for a figure. */

adapters/llamacpp.ts

@@ -22,7 +22,7 @@
 import { sumLabeledMetric } from "../prometheus-text"
 import { httpText, type HttpOptions } from "../http"
 import { nn, ni } from "../format"
-import { rowsOf, timeRows, viewText, nt, type Row, type TurnView } from "../rows"
+import { rowsOf, timeRows, viewText, nt, phase, type Row, type TurnView } from "../rows"
 
 export interface LlamaCppCounters {
   promptTokens: number
@@ -142,7 +142,12 @@ export function llamaCppView(
     ...rowsOf("tokens", [nt(t.completionTokens), host.includesSubagents ? "incl. sub-agents" : ""]),
     ...timeRows(host.total ?? t.decodeS + t.prefillS, host.retries),
   ]
-  return { engine: label, rows, notes: [], key: t.decodeTokS !== undefined ? `${nn(t.decodeTokS)} tok/s` : undefined }
+  const detail: Row[] = [
+    ...rowsOf("speed", [t.decodeTokS !== undefined ? `${nn(t.decodeTokS)} tok/s` : ""]),
+    ...rowsOf("prefill", [t.prefillTokS !== undefined ? `${ni(t.prefillTokS)} tok/s` : "", phase(t.promptTokens, t.prefillS)]),
+    ["decode", phase(t.completionTokens, t.decodeS)],
+  ]
+  return { engine: label, rows, notes: [], key: t.decodeTokS !== undefined ? `${nn(t.decodeTokS)} tok/s` : undefined, detail }
 }
 
 /** The view as text; kept for tests that look for a figure. */

adapters/mlxserve.ts

@@ -23,7 +23,7 @@
 
 import { httpJson, type HttpOptions } from "../http"
 import { nn, ni } from "../format"
-import { rowsOf, timeRows, viewText, nt, type Row, type TurnView } from "../rows"
+import { rowsOf, timeRows, viewText, nt, phase, type Row, type TurnView } from "../rows"
 
 /** One record from /v1/metrics/requests, as the server names its fields. */
 export interface MlxServeRequest {
@@ -295,7 +295,21 @@ export function mlxServeView(
   // as a tenfold slowdown rather than a one-off load.
   if (t.coldStart) notes.push("cold start (model loaded)")
   if (t.requests > 1 && t.steps === undefined) notes.push(`${ni(t.requests)} requests this turn`)
-  return { engine: "mlx-serve", rows, notes, key: t.decodeTokS !== undefined ? `${nn(t.decodeTokS)} tok/s` : undefined }
+  const detail: Row[] = [
+    ...rowsOf("speed", [t.decodeTokS !== undefined ? `${nn(t.decodeTokS)} tok/s` : ""]),
+    ...rowsOf("ttft", [t.decodeTokS !== undefined && t.ttft !== undefined ? `${nn(t.ttft, 2)}s` : ""]),
+    ["tokens", nt(t.completionTokens)],
+    ...rowsOf("prompt", [t.promptTokens !== undefined ? `${nt(t.promptTokens)} tok` : ""]),
+    ["request", `${nn(t.totalS, 2)}s`],
+    ...rowsOf("requests", [t.requests > 1 ? nt(t.requests) : ""]),
+  ]
+  return {
+    engine: "mlx-serve",
+    rows,
+    notes,
+    key: t.decodeTokS !== undefined ? `${nn(t.decodeTokS)} tok/s` : undefined,
+    detail,
+  }
 }
 
 /** The view as text; kept for tests that look for a figure. */

adapters/mtplx.ts

@@ -140,7 +140,24 @@ export function mtplxView(
       rows.push(["accepted", `${l.mean_accept_probability_by_depth.map((p) => Math.round(p * 100)).join("/")}%`])
     }
   }
-  return { engine: "MTPLX", rows, notes: [], key: decode !== undefined ? `${nn(decode)} tok/s` : undefined }
+  // The dialog's engine section: MTPLX's own figures only, with the verify
+  // passes behind the MTP rate and acceptance at every depth.
+  const depths = Array.isArray(l.mean_accept_probability_by_depth) ? l.mean_accept_probability_by_depth : []
+  const detail: Row[] = [
+    ...rowsOf("speed", [decode !== undefined ? `${nn(decode)} tok/s` : ""]),
+    ...rowsOf("ttft", [ttft !== undefined ? `${nn(ttft, 2)}s` : ""]),
+    ...rowsOf("prefill", [prefill !== undefined ? `${ni(prefill)} tok/s` : ""]),
+    ...rowsOf("tokens", [completion !== undefined ? nt(completion) : ""]),
+    ...rowsOf("request", [num(l.request_elapsed_s) !== undefined ? `${nn(num(l.request_elapsed_s) as number, 2)}s` : ""]),
+    ...(verify !== undefined && verify > 0 && completion !== undefined
+      ? rowsOf("MTP", [`${nn(completion / verify, 2)}x`, `${nt(verify)} verify passes`])
+      : []),
+    ...rowsOf(
+      "accepted",
+      depths.map((p, i) => `${Math.round(p * 100)}% at depth ${i + 1}`)
+    ),
+  ]
+  return { engine: "MTPLX", rows, notes: [], key: decode !== undefined ? `${nn(decode)} tok/s` : undefined, detail }
 }
 
 /** The view as text; kept for tests that look for a figure. */

adapters/omlx.ts

@@ -23,7 +23,7 @@
 
 import { httpJson, type HttpOptions } from "../http"
 import { nn, ni } from "../format"
-import { rowsOf, timeRows, viewText, nt, type Row, type TurnView } from "../rows"
+import { rowsOf, timeRows, viewText, nt, phase, type Row, type TurnView } from "../rows"
 
 /** Cumulative counters as this plugin reads them. */
 export interface OmlxSample {
@@ -132,6 +132,10 @@ export function omlxView(
       ],
       notes: [],
       key: `${nn(now.avgGen)} tok/s avg`,
+      detail: [
+        ["speed", `${nn(now.avgGen)} tok/s (server avg)`],
+        ["prefill", `${ni(now.avgPrefill)} tok/s (server avg)`],
+      ],
     }
   }
 
@@ -173,6 +177,16 @@ export function omlxView(
     ],
     notes: [],
     key: `${nn(decode)} tok/s${decodeLabel}`,
+    // The host's rate stands in when oMLX's can't be recovered; it is not
+    // the engine's, so the engine section leaves it out.
+    detail: [
+      ...rowsOf("speed", [useHostRate ? "" : `${nn(decode)} tok/s${decodeLabel}`]),
+      ["prefill", `${ni(prefill)} tok/s${prefillLabel}`],
+      ["tokens", nt(completion)],
+      ["prompt", `${nt(promptTokens)} tok`],
+      ...rowsOf("cached", [cached > 0 ? `${nt(cached)} tok` : ""]),
+      ["requests", nt(now.requests - prev.requests)],
+    ],
   }
 }
 

adapters/prometheus.ts

@@ -9,7 +9,7 @@
 import { httpText, type HttpOptions } from "../http"
 import { sumLabeledMetric } from "../prometheus-text"
 import { nn, ni } from "../format"
-import { rowsOf, timeRows, viewText, nt, type Row, type TurnView } from "../rows"
+import { rowsOf, timeRows, viewText, nt, phase, type Row, type TurnView } from "../rows"
 export interface PromSpec {
   prefix: string
   promptTokens: string
@@ -346,7 +346,25 @@ export function promView(
     ["prompt", nt(diff.promptTokens)],
     ...rowsOf("cached", [diff.cachedTokens > 0 ? nt(diff.cachedTokens) : ""]),
   ]
-  return { engine: label, rows, notes: [], key: decodeTokS !== undefined ? `${nn(decodeTokS)} tok/s${overall}` : undefined }
+  // The engine's own figures only: its rate and TTFT for a single request
+  // (over several, the counters give only means, and the host's stand in).
+  const detail: Row[] = [
+    ...rowsOf("speed", [single && diff.decodeTokS !== undefined ? `${nn(diff.decodeTokS)} tok/s` : ""]),
+    ...rowsOf("ttft", [single && diff.ttft !== undefined ? `${nn(diff.ttft, 2)}s` : ""]),
+    ...rowsOf("prefill", [single && diff.prefillTokS !== undefined ? `${ni(diff.prefillTokS)} tok/s` : ""]),
+    ["tokens", nt(diff.completionTokens)],
+    ["prompt", `${nt(diff.promptTokens)} tok`],
+    ...rowsOf("cached", [diff.cachedTokens > 0 ? `${nt(diff.cachedTokens)} tok` : ""]),
+    ...rowsOf("request", [single && diff.durationS !== undefined ? `${nn(diff.durationS, 2)}s` : ""]),
+    ["requests", nt(ttftCount)],
+  ]
+  return {
+    engine: label,
+    rows,
+    notes: [],
+    key: decodeTokS !== undefined ? `${nn(decodeTokS)} tok/s${overall}` : undefined,
+    detail,
+  }
 }
 
 /** The view as text, or null when declined; kept for tests that look for a figure. */

adapters/splash.ts

@@ -18,7 +18,7 @@
 import { sumLabeledMetric } from "../prometheus-text"
 import { httpText, type HttpOptions } from "../http"
 import { nn, ni } from "../format"
-import { rowsOf, timeRows, viewText, nt, type Row, type TurnView } from "../rows"
+import { rowsOf, timeRows, viewText, nt, phase, type Row, type TurnView } from "../rows"
 
 /**
  * Names verified against the server's own metrics.py (Splash 1.0), which maps
@@ -219,7 +219,15 @@ export function splashView(
   // Only when a turn spanned several requests nobody accounted for, so the
   // figures above read as sums rather than as one reply.
   const notes = t.requests > 1 && host.steps === undefined ? [`${ni(t.requests)} requests this turn`] : []
-  return { engine: "Splash", rows, notes, key: t.decodeTokS !== undefined ? `${nn(t.decodeTokS)} tok/s` : undefined }
+  const detail: Row[] = [
+    ...rowsOf("speed", [t.decodeTokS !== undefined ? `${nn(t.decodeTokS)} tok/s` : ""]),
+    ...rowsOf("prefill", [t.prefillTokS !== undefined ? `${ni(t.prefillTokS)} tok/s` : "", phase(t.promptTokens, t.prefillS)]),
+    ["decode", phase(t.completionTokens, t.decodeS)],
+    ...rowsOf("cached", [t.cachedTokens > 0 ? `${nt(t.cachedTokens)} tok reused` : ""]),
+    ...rowsOf("draft", [t.draftAcceptRate !== undefined ? `${ni(t.draftAcceptRate * 100)}% accepted` : ""]),
+    ["requests", nt(t.requests)],
+  ]
+  return { engine: "Splash", rows, notes, key: t.decodeTokS !== undefined ? `${nn(t.decodeTokS)} tok/s` : undefined, detail }
 }
 
 /** The view as text; kept for tests that look for a figure. */

counters.ts

@@ -0,0 +1,39 @@
+// Taking a known request out of a counter window.
+//
+// Pure, like the other shared modules. Engines that publish cumulative
+// counters -- Splash, llama.cpp, llamafile, the Prometheus engines -- give a
+// turn's figures as the difference between a reading before it and one after.
+// A request of OpenCode's own inside that window (a compaction, measured on
+// Splash) lands in the difference too, and the turn is declined because its
+// tokens no longer match. When that request was bracketed by readings of its
+// own, its difference can be taken back out.
+
+/** Field-wise `after - before`, for every numeric field. */
+export function counterDelta<T extends object>(before: T, after: T): Partial<Record<keyof T, number>> {
+  const out: Partial<Record<keyof T, number>> = {}
+  for (const k of Object.keys(after) as Array<keyof T>) {
+    const a = after[k]
+    const b = before[k]
+    if (typeof a === "number" && typeof b === "number") out[k] = a - b
+  }
+  return out
+}
+
+/**
+ * The baseline moved forward by each bracketed request's own difference, so
+ * that `now - shifted` holds everything in the window except those requests.
+ * Non-numeric fields keep the baseline's. A bracket that went backwards (a
+ * restarted engine) can't be trusted, and leaves the baseline as it was.
+ */
+export function shiftBaseline<T extends object>(prev: T, brackets: ReadonlyArray<{ before: T; after: T }>): T {
+  const shifted = { ...prev } as Record<string, unknown>
+  for (const { before, after } of brackets) {
+    const d = counterDelta(before, after) as Record<string, number | undefined>
+    if (Object.values(d).some((v) => v !== undefined && v < 0)) continue
+    for (const [k, v] of Object.entries(d)) {
+      const cur = shifted[k]
+      if (typeof cur === "number" && v !== undefined) shifted[k] = cur + v
+    }
+  }
+  return shifted as T
+}