tickmarkr 2.8.3 → 2.8.4

package/dist/adapters/claude-code.js

@@ -238,6 +238,9 @@ export const claudeCode = {
     // The empty -p argument selects print mode while stdin carries the prompt, keeping its nonce out
     // of process argv. The redirect path is shell-quoted independently from the model.
     headlessCommand: (promptFile, model, effort) => `claude -p '' --model ${shq(model)}${claudeEffortFlag(effort)} --permission-mode bypassPermissions --strict-mcp-config --mcp-config '{"mcpServers":{}}' --output-format text < ${shq(promptFile)}`,
+    // Queue row 184: `--output-format text` prints nothing until the turn ends, so a working pane reviewer reads zero seat bytes at
+    // the first-liveness beat; the beat must not re-route it (2.8.3 killed every claude pane reviewer at 30 s).
+    paneReviewBuffersOutput: true,
     // HYG-03 / OBS-137: the residual first-entry dialog is workspace trust, not MCP config loading.
     // Claude's only store is global last-writer-wins ~/.claude.json, so tickmarkr still does not seed it;
     // the daemon safely answers only the exact adapter-declared dialog once per slot.

package/dist/adapters/kimi.js

@@ -1,3 +1,4 @@
+import { spawnSync } from "node:child_process";
 import { readdirSync, readFileSync, realpathSync, statSync } from "node:fs";
 import { homedir } from "node:os";
 import { join } from "node:path";
@@ -171,7 +172,32 @@ export function confirmKimiSeedBanner(banner, assignedModel) {
     }
     return { ok: true, status: "confirmed", ...(sessionId === undefined ? {} : { sessionId }) };
 }
-const kimiTuiLaunch = (model) => `kimi -y -m ${shq(model)}`;
+const KIMI_HELP_TIMEOUT_MS = 10_000;
+function kimiHasNeverAskOption(help) {
+    let inOptions = false;
+    for (const line of help.split(/\r?\n/)) {
+        if (line === "Options:") {
+            inOptions = true;
+            continue;
+        }
+        if (inOptions && /^[^\s].*:$/.test(line))
+            break;
+        if (inOptions && /^  --auto[ \t]{2,}Start in Never Ask mode\b/.test(line))
+            return true;
+    }
+    return false;
+}
+const kimiTuiLaunch = (model) => {
+    let neverAsk = false;
+    try {
+        const help = spawnSync("kimi", ["--help"], {
+            encoding: "utf8", timeout: KIMI_HELP_TIMEOUT_MS, stdio: ["ignore", "pipe", "ignore"],
+        });
+        neverAsk = !help.error && help.status === 0 && kimiHasNeverAskOption(help.stdout);
+    }
+    catch { /* unreadable help keeps the existing launch form */ }
+    return `kimi ${neverAsk ? "--auto" : "-y"} -m ${shq(model)}`;
+};
 const KIMI_ANSI_SGR_RE = /\u001B\[[0-9;]*m/g;
 const KIMI_EDITOR_TOP_RE = /^╭─+╮$/;
 const KIMI_EDITOR_INPUT_RE = /^│ > .*│$/;
@@ -215,7 +241,10 @@ export const KIMI_INPUT_BOX = declareInputBox("kimi", {
     fingerprint: "│ > ",
     match: (paneText) => matchesKimiEditorBox(paneText, false),
     emptyMatch: (paneText) => matchesKimiEditorBox(paneText, true),
-    launchCommand: (command) => command.startsWith("kimi -y -m "),
+    launchCommand: (command) => {
+        const match = /^kimi (?:--auto|-y) -m '([^']+)'$/.exec(command);
+        return match !== null && MODEL_ID_RE.test(match[1]);
+    },
     readinessTimeoutMs: 15_000,
 });
 // Shared launch-then-seed surface (T6) + banner confirm (T7/T2). One definition so the adapter

package/dist/adapters/types.d.ts

@@ -195,6 +195,7 @@ export interface WorkerAdapter {
     channels(cfg: TickmarkrConfig): BillingChannel[];
     headlessCommand(promptFile: string, model: string, effort?: Effort): string;
     harnessBannerRows?: readonly string[];
+    paneReviewBuffersOutput?: boolean;
     interactiveCommand(promptFile: string, model: string, effort?: Effort): string | null;
     interactiveSeed?: InteractiveSeed;
     resumeCommand?(sessionId: string, promptFile: string, model: string, effort?: Effort): string;

package/dist/gates/acceptance.js

@@ -1,11 +1,13 @@
-import { readdirSync } from "node:fs";
-import { join } from "node:path";
+import { randomBytes } from "node:crypto";
+import { lstatSync, mkdtempSync, readdirSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { isAbsolute, join, relative, resolve } from "node:path";
 import { z } from "zod";
 import { channelKey, shq } from "../adapters/types.js";
 import { compileNative } from "../compile/native.js";
 import { DEFAULT_DIFF_CAP } from "../config/config.js";
 import { renderAcceptanceItem } from "../graph/schema.js";
-import { sh } from "../run/git.js";
+import { DEFAULT_SHELL_TIMEOUT_MS, sh, shell } from "../run/git.js";
 import { checkTaskDiffCaps, fetchTaskDiff, isProtectedEvidence, setAsideReceiptPath } from "./review.js";
 import { appendAnchoredReview, COMPLETION_FAKING_CHECKLIST, extractVerdictJson, generateVerdictNonce, runLlm, SeatLaunchError, verdictNonceLine } from "./llm.js";
 import { classifyVerdictCause } from "./verdict-cause.js";
@@ -13,6 +15,8 @@ import { reviewableLogicDiff } from "./artifact-manifest.js";
 import { diffSidePath } from "./git-paths.js";
 import { classifyFailureOutput } from "./baseline.js";
 import { operationFactsAtHead, readOperationFacts } from "./operation-facts.js";
+import { isVitestTestCommand, manifestChildEnv, readTestReport, singleForkRetryCommand, toManifestPath } from "./test-manifest.js";
+import { TEST_REPORTER_SOURCE } from "./test-reporter.js";
 // Fable F4: acceptance judge shares review's 900s timeout — 300s default killed frontier judges on cap-sized diffs.
 const JUDGE_TIMEOUT_MS = 900_000;
 const CitationSchema = z.object({ path: z.string(), line: z.number().int() });
@@ -98,6 +102,26 @@ export function testFiltered(testCmd, name) {
     const fwd = wrapped ? "-- " : "";
     return `${testCmd} ${fwd}-t ${shq(pattern)}`;
 }
+// Queue row 179: the criterion's suite file as the runner's file argument AHEAD of the title filter, appended to a
+// base that already carries its forwarding separator (npmForwarded, under either runner). testFiltered keeps its
+// two-argument form: a criterion with no suite file, and the eval harness.
+const fileFiltered = (base, file, name) => `${base} ${shq(file)} -t ${shq(testFilterPattern(name))}`;
+// The forwarding separator, decided on shell WORDS — the test gate's tokenizer (test-manifest.ts shellTokens/words,
+// unexported) beside its unquoting — never on the raw text: behind any `NAME=value` assignments (whose value, even
+// ` -- `, is not a word of the command), an npm manager word, bare, quoted or escaped, gets ONE `--` unless a later
+// word already is `--` (bare, quoted or escaped). npx, pnpm, yarn and a direct runner forward their arguments as they
+// are. Both runners' branches append through it; the Vitest branch asks the gate's parser only whether the reporter
+// can be forwarded at all (V5), since that parser's own separator reads the raw text.
+const SHELL_TOKEN = /(?:[^\s'"\\&|;<>]+|'[^']*'|"(?:\\.|[^"\\])*"|\\.)+|[&|;<>]+/g;
+const unquoted = (token) => token.replace(/'([^']*)'|"((?:\\.|[^"\\])*)"|\\(.)/g, (_all, single, double, escaped) => single ?? (double !== undefined ? double.replace(/\\(["\\$`])/g, "$1") : escaped));
+function npmForwarded(testCmd) {
+    const raw = testCmd.match(SHELL_TOKEN) ?? [];
+    let manager = 0;
+    while (/^[A-Za-z_]\w*=/.test(raw[manager] ?? ""))
+        manager++;
+    const argv = raw.slice(manager).map(unquoted);
+    return argv[0] === "npm" && !argv.includes("--") ? `${testCmd} --` : testCmd;
+}
 const VitestListedTestsSchema = z.array(z.object({
     name: z.string(),
     file: z.string(),
@@ -238,6 +262,99 @@ function oracleExecutionFailure(label, code, stdout, stderr) {
         details: `oracle failed: ${label} (exit ${code})${tail(stderr || stdout)}`,
     };
 }
+// today's rule for a test oracle's run: exit 0 AND a positive printed run count (OBS-55: a zero-match run is vacuous).
+function countedTestRun(label, filter, r) {
+    if (r.code !== 0)
+        return oracleExecutionFailure(label, r.code, r.stdout, r.stderr);
+    const out = (r.stderr || "") + "\n" + (r.stdout || "");
+    const ran = testsRan(out);
+    if (ran === null || ran < 1) {
+        return { gate: "acceptance", pass: false,
+            details: `oracle failed: ${label} — name filter matched zero tests (filter: ${filter})${tail(out)}` };
+    }
+    return undefined;
+}
+/**
+ * Queue row 179 (D-2018, D-2019): a test criterion that declares its suite file L is satisfied by L alone — run
+ * 1248's oracles named one file and a same-titled test in another satisfied them. Closed table:
+ * L1 L absent or not a regular file (a directory included) in the worktree: a work verdict naming L, under either
+ *    runner, and the test command never starts.
+ * Vitest: L alone, file argument then title filter, under the test gate's nonce-bound per-file reporter, because
+ *    Vitest's file filter selects EVERY path containing L. PASS only when that report, of this nonce and ending in
+ *    its run-end certificate with exit 0, shows a passed body in module L and none run in any other module.
+ *    V2 L without the title (another file holding it): no passed body in L. V3 a deeper path ending with L holding
+ *    it, whether or not L does: a body run in another module. V4 report missing, another nonce or no certificate,
+ *    and V5 a script body the reporter cannot be forwarded to: fail closed, unverified.
+ * Other runner: L as its file argument before the title filter, today's run-count rule. Declared residual: such a
+ *    runner gives no per-file evidence, so a test of the title in a path containing L can still satisfy it there.
+ */
+async function landedTestOracle(testCmd, title, landing, worktree) {
+    const label = `test "${title}" in ${landing}`;
+    const passed = { gate: "acceptance", pass: true, details: `✓ test: ${title} (exit 0) in ${landing}` };
+    const work = (why) => ({ gate: "acceptance", pass: false, details: `oracle failed: ${label} — ${why}` });
+    const unverified = (why) => ({ gate: "acceptance", pass: false,
+        details: `oracle failed: ${label} — ${why}; unverified, failing closed`,
+        meta: { cause: "oracle-execution", classification: "infra", infra: true, retryable: false } });
+    const path = resolve(worktree, landing);
+    const within = relative(worktree, path);
+    let regular = false;
+    try {
+        regular = !within.startsWith("..") && !isAbsolute(within) && lstatSync(path).isFile();
+    }
+    catch { /* absent */ }
+    if (!regular)
+        return work(`suite file ${landing} is absent from the worktree or is not a regular file; the test command never started`);
+    if (!isVitestTestCommand(testCmd, worktree)) {
+        return countedTestRun(label, title, await sh(fileFiltered(npmForwarded(testCmd), landing, title), worktree)) ?? passed;
+    }
+    // V5: the test gate's own parser, through its one pure export, throws on a compound or non-direct script body,
+    // where an appended --reporter would reach some other command. Only its refusal is read: its separator reads the
+    // raw text, so a quoted `"--"` or an assignment's ` -- ` value would misplace the file, filter and reporter.
+    try {
+        singleForkRetryCommand(testCmd, worktree, [], []);
+    }
+    catch (error) {
+        return unverified(`the per-file reporter cannot be forwarded to this test command (${error instanceof Error ? error.message : String(error)})`);
+    }
+    const cmd = fileFiltered(npmForwarded(testCmd), landing, title);
+    const dir = mkdtempSync(join(tmpdir(), "tickmarkr-oracle-"));
+    try {
+        const nonce = randomBytes(16).toString("hex");
+        const reporterPath = join(dir, `test-reporter-${nonce}.mjs`);
+        const reportPath = join(dir, `test-report-${nonce}.json`);
+        writeFileSync(reporterPath, TEST_REPORTER_SOURCE);
+        const r = await shell(`${cmd} --reporter=${shq(reporterPath)} --outputFile=${shq(reportPath)}`, worktree, DEFAULT_SHELL_TIMEOUT_MS, false, { env: { ...manifestChildEnv(worktree), TICKMARKR_TEST_REPORT: reportPath, TICKMARKR_TEST_NONCE: nonce } });
+        const raw = readTestReport(reportPath);
+        const report = raw?.nonce === nonce ? raw : undefined;
+        if (r.code !== 0) {
+            // the reporter replaces the runner's own output: its failure identities are the evidence of a red.
+            const named = report ? [...Object.values(report.completed).flatMap((c) => c.failures ?? []), ...(report.certificate?.diagnostics ?? [])] : [];
+            return oracleExecutionFailure(label, r.code, r.stdout, [r.stderr, ...named].filter(Boolean).join("\n"));
+        }
+        if (!raw)
+            return unverified("the runner exited 0 but wrote no per-file report");
+        if (!report)
+            return unverified(`the per-file report carries nonce "${raw.nonce}", not this invocation's`);
+        if (report.certificate?.exitCode !== 0)
+            return unverified("the per-file report never reached its run-end certificate with exit 0");
+        const key = toManifestPath(within, worktree);
+        // a module completed more than once (one path in several projects) keeps only its LAST counters, so its record
+        // cannot show that no earlier completion ran a body: it counts as run.
+        const repeated = new Set(report.duplicateCompletions ?? []);
+        const elsewhere = [...new Set([...Object.keys(report.started), ...Object.keys(report.completed)])].filter((file) => {
+            const t = report.completed[file]?.tests;
+            return file !== key && (repeated.has(file) || !t || t.passed + t.failed > 0);
+        }).sort();
+        if (elsewhere.length)
+            return work(`the runner's per-file report shows a test body run outside ${landing}, in ${elsewhere.join(", ")} — only ${landing} can satisfy this criterion`);
+        if (!((report.completed[key]?.tests?.passed ?? 0) > 0))
+            return work(`the runner's per-file report shows no passing test of that title in ${landing}`);
+        return passed;
+    }
+    finally {
+        rmSync(dir, { recursive: true, force: true });
+    }
+}
 // v1.70 / OBS-129: the new-file line numbers each changed HUNK spans, per path — parsed from the
 // unified-diff so a citation is validated against real changed regions, not a substring of the whole
 // diff text (which also matches the diff's own +++/@@ headers). A hunk's span is every new-file line it
@@ -355,17 +472,18 @@ judge, via, opts = {}) {
                 return { gate: "acceptance", pass: false,
                     details: `oracle failed: test "${a.test}" — no test command configured to run it (failing closed)` };
             }
-            const r = await sh(testFiltered(opts.testCmd, a.test), worktree);
-            const out = (r.stderr || "") + "\n" + (r.stdout || "");
-            if (r.code !== 0) {
-                return oracleExecutionFailure(`test "${a.test}"`, r.code, r.stdout, r.stderr);
+            // queue row 179: a declared suite file is the only file that can satisfy its criterion.
+            if (a.landing !== undefined) {
+                const landed = await landedTestOracle(opts.testCmd, a.test, a.landing, worktree);
+                if (!landed.pass)
+                    return landed;
+                passedDet.push(landed.details);
+                continue;
             }
             // OBS-55: exit 0 alone is vacuous when the name filter matched zero tests — fail closed.
-            const ran = testsRan(out);
-            if (ran === null || ran < 1) {
-                return { gate: "acceptance", pass: false,
-                    details: `oracle failed: test "${a.test}" — name filter matched zero tests (filter: ${a.test})${tail(out)}` };
-            }
+            const red = countedTestRun(`test "${a.test}"`, a.test, await sh(testFiltered(opts.testCmd, a.test), worktree));
+            if (red)
+                return red;
             passedDet.push(`✓ test: ${a.test} (exit 0)`);
         }
     }

package/dist/gates/llm.js

@@ -494,15 +494,20 @@ async function runViaDriverDetailed(adapter, model, prompt, cwd, via, timeoutMs
                 // OBS-1177: the beat is armed for an INTERACTIVE (pane) driver only. On the subprocess driver
                 // a headless `claude -p` buffers every byte until it exits, so 0 seat bytes at 30 s is not a
                 // dead launch; that seat keeps its full ceiling, which still bounds it (fail-closed by timeout).
-                if (reviewing && via.driver.interactive && !firstLivenessObserved && now - startedAt >= REVIEW_FIRST_LIVENESS_MS) {
+                // Queue row 184: an adapter whose pane review command buffers its whole output until exit
+                // (claude-code's `claude -p`) has written nothing of its own until its verdict, working or
+                // dead, so the beat cannot tell the two apart and does not arm for it; its ceiling bounds it.
+                if (reviewing && via.driver.interactive && adapter.paneReviewBuffersOutput !== true && !firstLivenessObserved && now - startedAt >= REVIEW_FIRST_LIVENESS_MS) {
                     firstLivenessObserved = true;
                     // RS-2: the beat reads seat-authored bytes ALONE. CPU evidence never holds a preamble-only
                     // capture open to the ceiling; a seat that has not written one byte of its own is re-routed.
-                    // OBS-944 is why a buffering runner earns no exemption here: the claude-code seat's 901 s
-                    // capture was byte-identical across two legs and ended at the pane-identity line — that
-                    // seat never started, it was not quietly working. RULING-229-06 puts the beat on the PANE
-                    // path only; a headless `-p` runner (runHeadlessDetailed, no pane, no beat) buffers every
-                    // byte until completion and keeps its full ceiling.
+                    // OBS-944: the claude-code seat's 901 s capture was byte-identical across two legs and ended
+                    // at the pane-identity line — that seat never started. Until 2.8.3 the orca checkout frame
+                    // counted as seat bytes and kept every claude pane seat past this beat; once it stopped
+                    // counting (row 135(a)) the beat killed working claude reviewers too, hence the declared
+                    // exemption above (row 184), which restores that pre-2.8.3 bound for buffering adapters only.
+                    // RULING-229-06 puts the beat on the PANE path only; a headless `-p` runner
+                    // (runHeadlessDetailed, no pane, no beat) buffers every byte until completion and keeps its full ceiling.
                     if (seatAuthoredBytes === 0) {
                         launchNeverStarted = true;
                         forceClose = true;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tickmarkr",
-  "version": "2.8.3",
+  "version": "2.8.4",
   "description": "Spec in, verified work out.",
   "type": "module",
   "license": "MIT",