test: address dual-review findings for #371

- add nudge breakdown line test asserting rendered reasoning category
- fix 2 pre-existing vacuous acp_status tests (partial mocks dropped by
  filterMessages → passed with zero visible messages)
- tighten protected-reasoning assertion to exact value
- system prompt breakdown example percentages now sum to 100%

1086/1086 tests passing

Author: ework-agent

devlog/2026-09-08_reasoning-in-context-estimate/WORKLOG.md

@@ -26,16 +26,32 @@
 
 ### Tests
 - `tests/inject-utils-pure.test.ts`: +4 — reasoning counted in `reasoningTokens`/`total`; total formula includes reasoning; mixed message (msgTotal vs messageTokens + largestRanges footprint); no-reasoning regression guard.
-- `tests/protection-aware-stats.test.ts`: +1 — reasoning on a protected message counted in `protectedTokens`.
+- `tests/protection-aware-stats.test.ts`: +1 — reasoning on a protected message counted in `protectedTokens` (exact `reasoningTokens === 200`).
 - `tests/acp-status.test.ts`: +3 — overview reasoning category (100 text/33% + 200 reasoning/67% of 300); reasoning-only message visible (overview 100% + drilldown line); drilldown per-message footprint includes reasoning.
+- `tests/inject.test.ts`: +1 — rendered nudge breakdown line shows `2.0K reasoning (Q%)` (added per test review).
+- `tests/acp-status.test.ts` (pre-existing fixes): 2 vacuous tests (partial mocks dropped by `filterMessages`) given complete mocks + real listing-line assertions (per both reviews).
+
+## Dual-agent review (AGENTS.md §5.3 + §5.6)
+
+Both reviewers: **APPROVE**. Findings addressed in this branch:
+
+- **Test reviewer finding (minor)**: rendered nudge breakdown line was untested → added `tests/inject.test.ts` "E2E: nudge breakdown line shows reasoning category with token count (#371)" (asserts `2.0K reasoning (\d+%)` in the injected nudge).
+- **Both reviewers (minor)**: pre-existing vacuous tests `tests/acp-status.test.ts:286-305` / `:307-328` used partial mocks (missing `info.sessionID`/`info.time.created`) that `filterMessages` drops → tests passed with ZERO visible messages. Fixed: complete mocks + assertions on the actual `m00001 (...) text|bash` listing lines.
+- **Test reviewer (nit)**: tightened `comp.reasoningTokens >= 200` → `assert.equal(comp.reasoningTokens, 200)`.
+- **Code reviewer (nit)**: system prompt example percentages now sum to 100% (were 121% pre-existing, 131% after first edit).
+
+Not addressed (documented, non-blocking):
+- Composition-vs-range divergence (code reviewer minor): `buildCompressibleRanges` range tokens still exclude reasoning (intentional — ranges = compressible amounts; the pipeline's min-size check `countMessageCharacters` also excludes reasoning, so adding it there risks phantom "Range too small" rejections per #37). Consequence: "Effective compressible: ~X" (nudge) and overview totals now include reasoning while per-range lines don't. Documented in PR description; candidate follow-up issue (source-tagged).
+- Reasoning-only messages render with `text` label in the drilldown (`toolName || "text"`); `classifyMessageType` would say `reasoning` — label-semantics change, out of scope.
+- Per-message `dcp-message-id` token annotation (`countMessageCharacters`, token-utils.ts) still excludes reasoning — pre-existing, out of scope, candidate follow-up.
 
 ## Verification
 
 - `npm run typecheck` — clean.
-- `npm run test` — **1085/1085 pass** (was 1077 on master; +8 new).
+- `npm run test` — **1086/1086 pass** (was 1077 on master; +9 new).
 - `npm run build` — clean.
 - `npm run format:check` — repo-wide pre-existing Prettier drift (423 files fail on clean master, incl. all 7 touched files); CI does not run format checks; no reformat to keep the diff minimal.
-- Test-input fidelity note: `filterMessages` (`lib/messages/shape.ts:14-24`) drops messages lacking `info.sessionID`/`info.time.created` — the new acp_status tests use complete mocks. (Pre-existing tests at `tests/acp-status.test.ts:286/:307` use partial mocks and only assert section headers, so they pass with zero visible messages — flagged for the review, not fixed here.)
+- Test-input fidelity note: `filterMessages` (`lib/messages/shape.ts:14-24`) drops messages lacking `info.sessionID`/`info.time.created` — all acp_status tests (new + 2 pre-existing fixed per review) use complete mocks.
 
 ## Open items / known interactions
 

lib/prompts/system.ts

@@ -73,7 +73,7 @@ CONTEXT BREAKDOWN
 
 When context usage passes a threshold, the system appends a breakdown showing where your context tokens are spent:
 
-Breakdown: 5.2K system (21%) | 12.3K tool (40%) | 3.1K summaries (10%) | 8.5K code (28%) | 6.5K text (22%) | 2.4K reasoning (10%)
+Breakdown: 4.2K system (21%) | 8.0K tool (40%) | 2.0K summaries (10%) | 2.6K code (13%) | 2.2K text (11%) | 1.0K reasoning (5%)
 
 - "system" = system prompt tokens (AGENTS.md, tool definitions — not compressible)
 - "tool" = tool call outputs (largest category — compress first when consumed)

tests/acp-status.test.ts

@@ -285,7 +285,10 @@ test("acp_status: scope=uncompressed defaults to ranges view", async () => {
 
 test("acp_status: scope=uncompressed view=messages shows per-message listing", async () => {
     const mockMsgs = [
-        { info: { id: "raw-1", role: "assistant" }, parts: [{ type: "text", text: "hello world" }] },
+        {
+            info: { id: "raw-1", role: "assistant", sessionID: SID, time: { created: 1 } },
+            parts: [{ type: "text", text: "hello world" }],
+        },
     ]
     const mockClient = makeMockClient(mockMsgs)
     const state = makeState([], new Map())
@@ -302,12 +305,13 @@ test("acp_status: scope=uncompressed view=messages shows per-message listing", a
 
     assert.match(result, /UNCOMPRESSED/)
     assert.match(result, /Sorted by/)
+    assert.match(result, /m00001 \(\d+\) text/, "per-message listing must include the visible message")
 })
 
 test("acp_status: scope=uncompressed view=messages with tool filter shows filter in header", async () => {
     const mockMsgs = [
         {
-            info: { id: "raw-1", role: "assistant" },
+            info: { id: "raw-1", role: "assistant", sessionID: SID, time: { created: 1 } },
             parts: [{ type: "tool", tool: "bash", state: { input: { command: "ls" } } }],
         },
     ]
@@ -325,6 +329,7 @@ test("acp_status: scope=uncompressed view=messages with tool filter shows filter
     const result = await statusTool.execute({ scope: "uncompressed", view: "messages", tool: "bash" } as any, { sessionID: SID } as any)
 
     assert.match(result, /UNCOMPRESSED — bash:/)
+    assert.match(result, /m00001 \(\d+\) bash/, "filtered listing must include the bash message")
 })
 
 test("acp_status: invalid scope falls back to overview", async () => {

tests/inject.test.ts

@@ -646,6 +646,37 @@ test("E2E: nudge recommendation content includes composition breakdown and compr
     )
 })
 
+test("E2E: nudge breakdown line shows reasoning category with token count (#371)", () => {
+    const state = createSessionState()
+    state.modelContextLimit = 1_000_000
+    state.nudges.lastPerMessageNudgeTokens = 200_000
+    const config = buildConfig()
+    config.compress.maxContextLimit = 800_000
+    config.compress.minContextLimit = 200_000
+
+    const messages: WithParts[] = [
+        userMsg("u1", "hello"),
+        {
+            info: {
+                id: "a1", role: "assistant", sessionID: SID, agent: "a", time: { created: 2 },
+                tokens: { input: 200_000, output: 55_000 },
+            } as WithParts["info"],
+            parts: [
+                { id: "a1-r", messageID: "a1", sessionID: SID, type: "reasoning" as const, text: "z".repeat(8_000) },
+                textPart("a1", "done"),
+            ],
+        },
+    ]
+    injectCompressNudges(state, config, logger, messages, {} as any)
+
+    assert.equal(state.nudges.shouldInjectThisTurn, true, "should nudge (55K growth >= 50K threshold)")
+
+    const injected = suffixText(messages)
+    assert.ok(injected.includes("Breakdown:"), "nudge must include composition breakdown")
+    // 8_000 chars of reasoning / 4 = 2_000 tokens
+    assert.match(injected, /2\.0K reasoning \(\d+%\)/, "breakdown must show reasoning category with its token count")
+})
+
 test("growth floor: nudge suppressed when growth below floor (issue #27 anti-thrashing)", () => {
     // 1M model: growthFloor = max(5000, 0.45×50000) = 22500
     // Growth of 5K < 22500 → no nudge output at all

tests/protection-aware-stats.test.ts

@@ -220,7 +220,7 @@ test("estimateContextComposition: reasoning on protected message counted in prot
     setupRefs(state, messages)
 
     const comp = estimateContextComposition(messages, state, ["skill"])
-    assert.ok(comp.reasoningTokens >= 200, "reasoning tokens counted")
+    assert.equal(comp.reasoningTokens, 200, "reasoning tokens counted exactly (800 chars / 4)")
     assert.ok(
         comp.protectedTokens >= comp.reasoningTokens,
         "protected tokens include the protected message's reasoning",