tickmarkr 2.8.3 → 2.8.5

package/dist/cli/commands/fleet.js

@@ -7,7 +7,7 @@ import { allAdapters, discoverChannels, doctorAgeMs, initDoctorReuse, modelAuthE
 import { catalogModelAdvisory, catalogTierRanking, declaredModelWindow, fleetUnclassifiedModels } from "../../adapters/model-lints.js";
 import { CATALOG_REFRESH_TIMEOUT_MS, formatCatalogRefreshLegs, readCachedCatalog, refreshCatalogCommand } from "../../adapters/catalog-remote.js";
 import { CLAUDE_ALIAS_IDENTITY_STAMPS, readClaudeAliasIdentity } from "../../adapters/claude-code.js";
-import { ConfigError, DEFAULT_CONFIG, fleetEditableFromConfig, fleetEditableEquals, formatFleetPrint, globalConfigDir, DENY_SCOPES, initialDenyPropOf, loadConfigWithMode, lowerLayerModelOverrides, overlayBytesLoadError, readOverlayFile, stagedDenyKeyOf, renderFleetOverlayWrite, repoOverlayPath, repoShadowedUserKeys, ROUTING_MODES, TIER_RANK, unifiedYamlDiff, universeEntryMatches, } from "../../config/config.js";
+import { ConfigError, DEFAULT_CONFIG, fleetEditableFromConfig, fleetEditableEquals, formatFleetPrint, globalConfigDir, DENY_SCOPES, initialDenyPropOf, loadConfigWithMode, lowerLayerModelOverrides, overlayBytesLoadError, ownRepoOverlayPath, readOverlayFile, stagedDenyKeyOf, renderFleetOverlayWrite, repoOverlayPath, repoShadowedUserKeys, ROUTING_MODES, TIER_RANK, unifiedYamlDiff, universeEntryMatches, } from "../../config/config.js";
 import { exclusionReason, projectFleetWhy, renderFleetWhy } from "../../config/fleet-why.js";
 import { overlayHoldingLine, overlayHoldsMap } from "../../config/fleet-overlay.js";
 import { SHAPES, TIERS } from "../../graph/schema.js";
@@ -355,7 +355,10 @@ export async function assembleFleetEditor(cwd, adapters, io, opts) {
     // them (destinationEdits below) — the repository overlay when it declares the key or masks it
     // with null at it or above, else the user overlay.
     const userPath = join(globalDir, "config.yaml");
-    const repoPath = repoOverlayPath(cwd);
+    // Queue row 195 / D-2102: the save destination is the checkout's OWN file even when the reader
+    // fallback (repoOverlayPath) has this linked worktree reading its main checkout's overlay — a
+    // first save here creates the worktree's own overlay, which then wins wholesale.
+    const repoPath = ownRepoOverlayPath(cwd);
     const currentUserText = () => (existsSync(userPath) ? readFileSync(userPath, "utf8") : "");
     const pathOf = (destination) => (destination === "repo" ? repoPath : userPath);
     // the candidate bytes stand in for the destination layer in the one loader path
@@ -1243,7 +1246,12 @@ export async function assembleFleetEditor(cwd, adapters, io, opts) {
     // T10: both overlays are guarded whichever one is written, so a holder that moved after the review
     // (the destination itself) refuses instead of retargeting. Null = still current.
     const stalePreview = (pending, reviewed, prior) => {
-        const destinationNow = prior ?? (pending.destination === "repo" ? currentRepoOverlayText(cwd) : currentUserText());
+        // Queue row 195 / D-2102: a repo-destination review renders against the EFFECTIVE repository
+        // overlay — in a linked worktree with no overlay of its own that is the main checkout's file,
+        // while the write lands in the worktree's own (ownRepoOverlayPath). The guard therefore
+        // re-reads the effective layer, not the destination's prior bytes; either file moving after
+        // the review moves that read (an own file appearing wins wholesale), so the refusal stands.
+        const destinationNow = pending.destination === "repo" ? currentRepoOverlayText(cwd) : (prior ?? currentUserText());
         const repoNow = pending.destination === "repo" ? destinationNow : currentRepoOverlayText(cwd);
         const userNow = pending.destination === "user" ? destinationNow : currentUserText();
         if (repoNow !== pending.repo) {

package/dist/cli/commands/init.js

@@ -24,7 +24,7 @@ const ENVIRONMENTS_FOOTER = [
     "environments:",
     "  herdr — the full cockpit — every worker, judge, and consult is a visible pane you can watch and unblock · https://herdr.dev · auto picks it first when HERDR_ENV=1",
     "  orca — visible terminals in the Orca app — auto picks it inside an Orca terminal after herdr and before subprocess; set driver: orca to force it · https://onorca.dev",
-    "  claude code — tickmarkr init --agent installs the /tkr skills + AGENTS.md so Claude Code (or any agent CLI) drives the loop natively",
+    "  claude code — tickmarkr init --agent installs the /tkr skills + AGENTS.md so Claude Code (or any agent CLI) drives the loop natively; in Claude Desktop's Code tab (neither herdr nor Orca) one single-tier session runs the loop on the headless subprocess driver, watched with tickmarkr status --watch --plain",
     "  anywhere — no herdr or Orca terminal? same fail-closed gates, headless subprocess driver (legacy: no herdr? same fail-closed gates, headless subprocess driver when Orca markers are absent)",
 ].join("\n");
 /** Current run truth, lock first; falls back to an abandoned journal. */
@@ -115,7 +115,7 @@ A run is green only when ALL of these hold: the run-end event exists in the jour
 
 ### Verified handoffs
 
-When relaying missions between agents, never use bare send-text (\`herdr agent send\` / pane send-text) — it omits Enter. Use \`herdr pane run <pane> "<message>"\` or \`herdr notification show "<message>"\`. Confirm delivery by reading the TARGET composer afterward; if the sent text still sits unsubmitted after ~15 s, retry the Enter only when that composer read shows the staged text and no permission prompt or numbered choice holds focus — an Enter sent onto an active prompt approves it or picks a choice instead of submitting the message (OBS-1119). Under that guard, send-keys enter to that pane once and read again (OBS-975); when a prompt or choice holds focus instead, resolve it first and re-read before retrying. Never report "relayed" without read-back.
+When relaying missions between agents, never use bare send-text (\`herdr agent send\` / pane send-text) — it omits Enter. Use \`herdr pane run <pane> "<message>"\` or \`herdr notification show "<message>"\`. Confirm delivery by reading the TARGET composer afterward; if the sent text still sits unsubmitted after ~15 s, retry the Enter only when that composer read shows the staged text and no permission prompt or numbered choice holds focus — an Enter sent onto an active prompt approves it or picks a choice instead of submitting the message (OBS-1119). Under that guard, send-keys enter to that pane once and read again (OBS-975); when a prompt or choice holds focus instead, resolve it first and re-read before retrying. Never report "relayed" without read-back. On a plain host (neither herdr nor Orca) there is no pane to read back: hand off through the host's cross-session message tool, and count the handoff confirmed only by the peer's reply.
 
 ### Orient before you act — this block may be the ONLY guidance your host loaded
 

package/dist/cli/commands/verify.js

@@ -14,7 +14,7 @@ import { getTask, loadGraph } from "../../graph/graph.js";
 import { GATE_NAMES } from "../../graph/schema.js";
 import { executionSignal } from "../../run/execution-budget.js";
 import { addDetachedWorktree, linkNodeModules, reflogWitness, removeWorktree, shGit, shGitOk } from "../../run/git.js";
-import { foldOwedChecks, integrationMapped, owedCriteria, Journal, OWED_DISCHARGE_EVENT } from "../../run/journal.js";
+import { foldOwedChecks, integrationMapped, owedCriteria, Journal, OWED_DISCHARGE_EVENT, parseRunId } from "../../run/journal.js";
 import { describeLeaseWait, withRepositoryLease } from "../../run/lease.js";
 import { isPidLive } from "../../run/lock.js";
 import { LAUNCHSERVICES_PORT_STOP, probeLaunchServicesPorts, recordLaunchServices } from "../../run/launchservices-check.js";
@@ -195,8 +195,17 @@ export async function verify(argv, cwd = process.cwd(), options = {}) {
     const stateRoot = await verifyStateRoot(cwd);
     const fileRoot = (file) => existsSync(join(cwd, ".tickmarkr", file)) ? cwd : stateRoot;
     if (resolve(stateRoot) !== resolve(cwd)) {
-        console.error(`verify: state files resolved read-only from ${join(stateRoot, ".tickmarkr")} (linked worktree; per-file origins: `
-            + STATE_FILES.map(file => `${file}: ${join(fileRoot(file), ".tickmarkr", file)} (${fileRoot(file) === cwd ? "local" : "common root"})`).join("; ") + ")");
+        // Queue row 195 / D-2102: the fallback is no longer "read-only" — the one line names every
+        // state file with the path it was read from and its origin, AND every path this invocation
+        // writes: the shared verdict store always, doctor.json when present, the --record journal when given.
+        const reads = STATE_FILES.map(file => `${file}: ${join(fileRoot(file), ".tickmarkr", file)} (${fileRoot(file) === cwd ? "local" : "common root"})`).join("; ");
+        const doctorPath = join(fileRoot("doctor.json"), ".tickmarkr", "doctor.json");
+        const writes = [
+            `verdict store ${getVerdictStore(join(stateRoot, ".tickmarkr")).dir}`,
+            ...(existsSync(doctorPath) ? [`doctor.json ${doctorPath}`] : []),
+            ...(values.record ? [`record journal ${join(stateRoot, ".tickmarkr", "runs", parseRunId(values.record), "journal.jsonl")}`] : []),
+        ].join("; ");
+        console.error(`verify: linked worktree — state files resolve from ${join(stateRoot, ".tickmarkr")} (reads: ${reads}; writes: ${writes})`);
     }
     const cfg = loadConfig(fileRoot("config.yaml"));
     const head = (await shGitOk("git rev-parse HEAD", cwd)).trim();

package/dist/compile/index.js

@@ -1,6 +1,7 @@
 import { existsSync, readFileSync, statSync } from "node:fs";
 import { join } from "node:path";
 import { parse } from "yaml";
+import { repoOverlayPath } from "../config/config.js";
 import { validateGraph } from "../graph/schema.js";
 import { taskUnitContractErrors } from "./collateral.js";
 import { blocksCompile, ownershipFindings, renderOwnershipFinding } from "./ownership.js";
@@ -37,7 +38,9 @@ function detect(src) {
 function repoOverlayMode(repoRoot) {
     if (!repoRoot)
         return undefined;
-    const path = join(repoRoot, ".tickmarkr", "config.yaml");
+    // Queue row 195 / D-2102: the ONE repo-layer path rule — a linked worktree without an overlay of
+    // its own proves against the main checkout's file, never against defaults.
+    const path = repoOverlayPath(repoRoot);
     if (!existsSync(path))
         return undefined;
     try {

package/dist/config/config.d.ts

@@ -459,6 +459,21 @@ export type FleetEditable = Partial<Record<DenyScopeKey, string[]>> & {
     map: Record<string, MapEntry>;
     floors: Record<string, Tier>;
 };
+/** The checkout's OWN repo-layer overlay — every writer's destination, and the reader's answer in
+ *  every case but the one fallback below. */
+export declare function ownRepoOverlayPath(repoRoot: string): string;
+/**
+ * Queue row 195 / D-2102: .tickmarkr/ is gitignored, so a fresh session worktree has no overlay of
+ * its own and every command there would silently run on defaults plus the global layer. ONE rule,
+ * and every repo-layer reader goes through it: a checkout whose own state-dir config.yaml exists
+ * reads it; otherwise a LINKED worktree — a gitfile checkout whose trusted common dir
+ * (src/run/git-trust.ts trustedCommonDir) lies outside it — reads
+ * dirname(common dir)/<state dir>/config.yaml when that file exists; every other case (a main
+ * checkout or its subdirectory, no enclosing checkout, a git trust refusal, an absent main file)
+ * keeps the checkout's own path. WRITERS never follow this fallback: they use ownRepoOverlayPath.
+ * The trust check includes missing, redirected or symlinked commondir metadata and config.worktree
+ * under extensions.worktreeConfig; none can authorize a shared overlay.
+ */
 export declare function repoOverlayPath(repoRoot: string): string;
 export declare function readOverlayFile(path: string): Record<string, unknown>;
 /** A discovered fleet universe row: the served models of one adapter and, per model, the resolved

package/dist/config/config.js

@@ -1,11 +1,12 @@
-import { existsSync, readFileSync } from "node:fs";
+import { existsSync, lstatSync, readFileSync, realpathSync } from "node:fs";
 import { homedir } from "node:os";
-import { join } from "node:path";
+import { dirname, join, sep } from "node:path";
 import { filesGlob } from "../graph/files-glob.js";
 import { Document, isScalar, parse } from "yaml";
 import { z } from "zod";
 import { CITED_MODEL_WINDOWS } from "../adapters/model-windows.js";
 import { stateDirName } from "../graph/graph.js";
+import { assertGitTrust, trustedCommonDir } from "../run/git-trust.js";
 import { EFFORTS, SHAPES, TIERS } from "../graph/schema.js";
 const TierEnum = z.enum(TIERS, {
     error: (iss) => `Invalid option: expected one of "cheap"|"mid"|"frontier" (got ${JSON.stringify(iss.input)})`,
@@ -744,7 +745,7 @@ export function globalConfigDir() {
 }
 export function overlayPreferShapes(repoRoot, opts = {}) {
     const shapes = new Set();
-    for (const layer of [readYaml(join(opts.globalDir ?? globalConfigDir(), "config.yaml")), readYaml(join(repoRoot, stateDirName(repoRoot), "config.yaml"))]) {
+    for (const layer of [readYaml(join(opts.globalDir ?? globalConfigDir(), "config.yaml")), readYaml(repoOverlayPath(repoRoot))]) {
         const map = layer?.routing?.map;
         if (!map)
             continue;
@@ -835,7 +836,7 @@ export function loadConfigWithMode(repoRoot, opts = {}) {
     // v1.52 T2: repoOverlayText substitutes candidate bytes for the on-disk repo layer — the fleet
     // write guard validates EXACTLY what it is about to write through this one loader path.
     const repoCfg = opts.repoOverlayText === undefined
-        ? readYaml(join(repoRoot, stateDirName(repoRoot), "config.yaml"))
+        ? readYaml(repoOverlayPath(repoRoot))
         : parse(opts.repoOverlayText);
     const merged = deepMerge(deepMerge(structuredClone(DEFAULT_CONFIG), globalCfg), repoCfg);
     const r = TickmarkrConfigSchema.safeParse(merged);
@@ -1017,9 +1018,70 @@ export function configTemplate(overlay) {
 }
 // Fleet overlay mutation, legacy serialization, and diff rendering live together in fleet-overlay.ts.
 export { FLEET_OVERLAY_KEYS, fleetEditableEquals, fleetRepoOverlayFromDelta, renderFleetOverlayWrite, repoOverlayYaml, serializeFleetOverlay, unifiedYamlDiff, } from "./fleet-overlay.js";
-export function repoOverlayPath(repoRoot) {
+/** The checkout's OWN repo-layer overlay — every writer's destination, and the reader's answer in
+ *  every case but the one fallback below. */
+export function ownRepoOverlayPath(repoRoot) {
     return join(repoRoot, stateDirName(repoRoot), "config.yaml");
 }
+/**
+ * Queue row 195 / D-2102: .tickmarkr/ is gitignored, so a fresh session worktree has no overlay of
+ * its own and every command there would silently run on defaults plus the global layer. ONE rule,
+ * and every repo-layer reader goes through it: a checkout whose own state-dir config.yaml exists
+ * reads it; otherwise a LINKED worktree — a gitfile checkout whose trusted common dir
+ * (src/run/git-trust.ts trustedCommonDir) lies outside it — reads
+ * dirname(common dir)/<state dir>/config.yaml when that file exists; every other case (a main
+ * checkout or its subdirectory, no enclosing checkout, a git trust refusal, an absent main file)
+ * keeps the checkout's own path. WRITERS never follow this fallback: they use ownRepoOverlayPath.
+ * The trust check includes missing, redirected or symlinked commondir metadata and config.worktree
+ * under extensions.worktreeConfig; none can authorize a shared overlay.
+ */
+export function repoOverlayPath(repoRoot) {
+    const own = ownRepoOverlayPath(repoRoot);
+    if (existsSync(own))
+        return own;
+    let dotGit;
+    try {
+        for (let dir = realpathSync(repoRoot);;) {
+            const candidate = join(dir, ".git");
+            try {
+                lstatSync(candidate);
+                dotGit = candidate;
+                break;
+            }
+            catch { /* keep walking */ }
+            const up = dirname(dir);
+            if (up === dir)
+                break;
+            dir = up;
+        }
+    }
+    catch {
+        return own;
+    }
+    // A directory .git is a main checkout (or one of its subdirectories); only a gitfile can be linked.
+    if (!dotGit)
+        return own;
+    try {
+        if (!lstatSync(dotGit).isFile())
+            return own;
+    }
+    catch {
+        return own;
+    }
+    let common;
+    try {
+        assertGitTrust(repoRoot);
+        common = trustedCommonDir(repoRoot);
+    }
+    catch {
+        return own;
+    }
+    const checkoutRoot = dirname(dotGit);
+    if (common === checkoutRoot || common.startsWith(`${checkoutRoot}${sep}`))
+        return own; // not linked
+    const shared = ownRepoOverlayPath(dirname(common));
+    return existsSync(shared) ? shared : own;
+}
 export function readOverlayFile(path) {
     const raw = readYaml(path);
     if (raw === undefined || raw === null)

package/dist/drivers/host-env.d.ts

@@ -0,0 +1,3 @@
+export declare const HOST_SESSION_VARS: readonly ["CLAUDECODE", "CLAUDE_CODE_ENTRYPOINT", "CLAUDE_CODE_SESSION_ID", "CLAUDE_CODE_HOST_SESSION_ID", "CLAUDE_CODE_CHILD_SESSION", "CLAUDE_CODE_MESSAGING_SOCKET", "CLAUDE_CODE_MESSAGING_TOKEN", "CLAUDE_CODE_SDK_HAS_HOST_AUTH_REFRESH", "CLAUDE_CODE_SESSION_ATTENDED", "CLAUDE_CODE_EXECPATH", "CLAUDE_CODE_DESKTOP_APP_VERSION", "CLAUDE_CODE_TERMINAL_MCP_TOOLS", "CLAUDE_CODE_ENABLE_ASK_USER_QUESTION_TOOL", "CLAUDE_CODE_EMIT_TOOL_USE_SUMMARIES", "CLAUDE_CODE_ENABLE_SDK_FILE_CHECKPOINTING", "CLAUDE_CODE_REPORT_FINDINGS", "CLAUDE_CODE_STARTUP_FAILURE_RESULTS", "CLAUDE_CODE_EAGER_FLUSH", "CLAUDE_CODE_DISABLE_CRON", "CLAUDE_CODE_SIMPLE", "CLAUDE_CODE_DISABLE_TERMINAL_TITLE", "CLAUDE_CODE_ACCOUNT_UUID", "CLAUDE_CODE_ORGANIZATION_UUID", "CLAUDE_CODE_USER_EMAIL", "CLAUDE_CODE_OAUTH_SCOPES", "CLAUDE_AGENT_SDK_VERSION", "CLAUDE_AGENT_SDK_MCP_NO_PREFIX", "CLAUDE_PID", "CLAUDE_EFFORT", "CLAUDE_PREVIEW_CLASSIFIER_FLOOR", "CLAUDE_CODE_BRIDGE_SESSION_ID"];
+/** A copy of `env` without the host-session names; the input (the daemon's own process.env) is never mutated. */
+export declare function withoutHostSession(env?: NodeJS.ProcessEnv): NodeJS.ProcessEnv;

package/dist/drivers/host-env.js

@@ -0,0 +1,47 @@
+// Queue row 193 (D-2102 add.1): a daemon launched from an agent session (the Claude Desktop Code tab, or a
+// claude CLI seat on Orca or herdr) carries that session's host variables, and an agent CLI that inherits
+// them gains the host's tools and effort. This CLOSED denylist is stripped at every choke point an agent
+// CLI or gate command passes: the shell seam, the subprocess driver's seal, the pane seeds and the bare
+// probes. Exact names only — never a prefix: an unlisted CLAUDE_CODE variable and every user auth or
+// provider variable reach children byte-equal. Declared residual: a host variable a future host adds
+// reaches children until it is listed here. Leaf module: imports nothing from src/.
+export const HOST_SESSION_VARS = [
+    "CLAUDECODE",
+    "CLAUDE_CODE_ENTRYPOINT",
+    "CLAUDE_CODE_SESSION_ID",
+    "CLAUDE_CODE_HOST_SESSION_ID",
+    "CLAUDE_CODE_CHILD_SESSION",
+    "CLAUDE_CODE_MESSAGING_SOCKET",
+    "CLAUDE_CODE_MESSAGING_TOKEN",
+    "CLAUDE_CODE_SDK_HAS_HOST_AUTH_REFRESH",
+    "CLAUDE_CODE_SESSION_ATTENDED",
+    "CLAUDE_CODE_EXECPATH",
+    "CLAUDE_CODE_DESKTOP_APP_VERSION",
+    "CLAUDE_CODE_TERMINAL_MCP_TOOLS",
+    "CLAUDE_CODE_ENABLE_ASK_USER_QUESTION_TOOL",
+    "CLAUDE_CODE_EMIT_TOOL_USE_SUMMARIES",
+    "CLAUDE_CODE_ENABLE_SDK_FILE_CHECKPOINTING",
+    "CLAUDE_CODE_REPORT_FINDINGS",
+    "CLAUDE_CODE_STARTUP_FAILURE_RESULTS",
+    "CLAUDE_CODE_EAGER_FLUSH",
+    "CLAUDE_CODE_DISABLE_CRON",
+    "CLAUDE_CODE_SIMPLE",
+    "CLAUDE_CODE_DISABLE_TERMINAL_TITLE",
+    "CLAUDE_CODE_ACCOUNT_UUID",
+    "CLAUDE_CODE_ORGANIZATION_UUID",
+    "CLAUDE_CODE_USER_EMAIL",
+    "CLAUDE_CODE_OAUTH_SCOPES",
+    "CLAUDE_AGENT_SDK_VERSION",
+    "CLAUDE_AGENT_SDK_MCP_NO_PREFIX",
+    "CLAUDE_PID",
+    "CLAUDE_EFFORT",
+    "CLAUDE_PREVIEW_CLASSIFIER_FLOOR",
+    "CLAUDE_CODE_BRIDGE_SESSION_ID",
+];
+/** A copy of `env` without the host-session names; the input (the daemon's own process.env) is never mutated. */
+export function withoutHostSession(env = process.env) {
+    const out = { ...env };
+    for (const k of HOST_SESSION_VARS)
+        delete out[k];
+    return out;
+}

package/dist/drivers/orca.js

@@ -7,6 +7,7 @@ import { createWorktree, FORK_CAP_ENV, resolvedForkCap, sh } from "../run/git.js
 import { Journal, parseRunId } from "../run/journal.js";
 import { stateDirName } from "../graph/graph.js";
 import { readWatchBoard, requestWatchBoardStop, stopWatchBoard, supervisionPresencePath, WATCH_OWNER_ENV } from "../run/supervision.js";
+import { HOST_SESSION_VARS } from "./host-env.js";
 import { MAX_BUF } from "./subprocess.js";
 import { formatOwnedName, panesToClose, parseOwnedName } from "./types.js";
 // Orca (onorca.dev) as a third execution surface beside herdr and subprocess. tickmarkr keeps
@@ -966,7 +967,8 @@ export class OrcaDriver {
         // UI's active worktree (`active`/`current`) nor the daemon's cwd — and the command itself moves
         // into the checkout, because Orca cannot be asked to place a terminal in a path it does not
         // track (OBS-1004). Shell startup can swallow that prefix; prove it from scrollback below.
-        const payload = `export ${FORK_CAP_ENV}=${shq(process.env[FORK_CAP_ENV] ?? resolvedForkCap())}; ${cmd}`;
+        // Queue row 193: the fork cap stays first; the agent session's host variables never reach the worker.
+        const payload = `export ${FORK_CAP_ENV}=${shq(process.env[FORK_CAP_ENV] ?? resolvedForkCap())}; unset ${HOST_SESSION_VARS.join(" ")}; ${cmd}`;
         launchGuards.getStore()?.(); // v2.6.8 T1: eligibility is re-read here, after every awaited lookup above
         const env = await this.call("create", [
             "terminal", "create", "--worktree", `path:${tracked}`, "--title", st.title, "--command", inCheckout(st.cwd, payload),

package/dist/drivers/subprocess.d.ts

@@ -2,7 +2,7 @@ import type { ExecutorDriver, NotifyOpts, Slot, TransportState } from "./types.j
 export declare const MAX_BUF: number;
 export declare const HERDR_CONTROL_VARS: readonly ["HERDR_ENV", "HERDR_SOCKET_PATH", "ORCA_TERMINAL_HANDLE", "ORCA_PANE_KEY", "ORCA_TAB_ID"];
 /**
- * Copy of worker env with the fork cap applied and host control-plane vars stripped.
+ * Copy of worker env with the fork cap applied and host control-plane and host-session vars stripped.
  * Workers retain the run's concurrency-derived cap (resolvedForkCap), independently of
  * verification's suite-window budget. The operator's own export still wins.
  */

package/dist/drivers/subprocess.js

@@ -1,6 +1,7 @@
 import { spawn } from "node:child_process";
 import { shq } from "../adapters/types.js";
 import { createWorktree, FORK_CAP_ENV, resolvedForkCap } from "../run/git.js";
+import { HOST_SESSION_VARS, withoutHostSession } from "./host-env.js";
 // HARD-03: cap retained worker output so a chatty worker can't grow tickmarkr's heap unbounded.
 // 2MB is safe against BOTH consumers: the largest read() call site asks for 1000 lines
 // (daemon.ts:217/247), and every marker (TICKMARKR_RESULT, TICKMARKR_EXIT_<nonce>) is emitted at the
@@ -21,12 +22,12 @@ export const HERDR_CONTROL_VARS = [
     "ORCA_TAB_ID",
 ];
 /**
- * Copy of worker env with the fork cap applied and host control-plane vars stripped.
+ * Copy of worker env with the fork cap applied and host control-plane and host-session vars stripped.
  * Workers retain the run's concurrency-derived cap (resolvedForkCap), independently of
  * verification's suite-window budget. The operator's own export still wins.
  */
 export function sealHerdrEnv(env = process.env) {
-    const out = { ...env };
+    const out = withoutHostSession(env); // queue row 193
     if (!(FORK_CAP_ENV in out))
         out[FORK_CAP_ENV] = resolvedForkCap();
     for (const k of HERDR_CONTROL_VARS)
@@ -37,7 +38,7 @@ export function sealHerdrEnv(env = process.env) {
 export function herdrSealShellPrefix(env = process.env) {
     const forkCap = sealHerdrEnv(env)[FORK_CAP_ENV] ?? resolvedForkCap();
     return `export ${FORK_CAP_ENV}=${shq(forkCap)}; ` +
-        HERDR_CONTROL_VARS.map((k) => `unset ${k}`).join("; ") + "; ";
+        HERDR_CONTROL_VARS.map((k) => `unset ${k}`).join("; ") + `; unset ${HOST_SESSION_VARS.join(" ")}; `;
 }
 export class SubprocessDriver {
     id = "subprocess";

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/baseline.d.ts

@@ -60,7 +60,7 @@ export declare function beginGateEvidence(cwd: string, gate: string, command: st
             subjectCommit: string | null;
         };
         termination: {
-            kind: "unknown" | "timeout" | "signal" | "not-started" | "exit";
+            kind: "unknown" | "signal" | "timeout" | "not-started" | "exit";
             exitCode: number | null;
             signal: string | null;
             timedOut: boolean | null;

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;