sortie-dogs 0.12.16 → 0.12.17

package/README.md

@@ -87,19 +87,29 @@ goal; preserving quality while reducing unnecessary expensive work is.
 
 ### SWE-bench evaluation
 
-A frozen Sortie-dogs v0.10.14 build was evaluated on 23 fixed tasks.
-
-- Official verifier PASS: 6 / 23
-- Scored FAIL: 17 / 23
-- Infrastructure blocked: 0 / 23
-- Total estimated model cost: $15.75
-- Median agent runtime: 15.2 min
-
-The same Sortie-dogs candidate, model routes, budgets, tools, and verification
-procedure were held fixed for all tasks. This is one frozen evaluation, not a
-general success-rate claim.
-
-Full methodology and per-task results: [benchmark details](docs/benchmark-v0.10.14-dev23.md)
+Official SWE-bench Lite `dev` results on the same 23 public instances:
+
+| Candidate / benchmark adapter | Official resolved | Empty patches | Run conditions | Details |
+| --- | ---: | ---: | --- | --- |
+| v0.10.6 candidate (`859c396`) | 4 / 23 (17.4%) | 5 | Four inference slots | [Handoff](docs/swebench-handoff-2026-09-21.md) |
+| Frozen v0.10.14 build | 6 / 23 (26.1%) | — | Four slots; $1.50/task | [Per-task results](docs/benchmark-v0.10.14-dev23.md) |
+| v0.12.8 (`e0f8cef` adapter) | 5 / 23 (21.7%) | 9 | Four slots; budget amended across two batches; 30-minute timeout | [Campaign](docs/swebench-v0128-dev23-2026-09-26.md) |
+| v0.12.8 (`84ccdf1` main adapter) | 6 / 23 (26.1%) | 2 | Fresh 23-task run; four slots; 30-minute timeout | [Main-integrated run](docs/swebench-main-84ccdf1-dev23-2026-09-26.md) |
+| v0.12.15 (`0c9690d`) | 5 / 23 (21.7%) | 3 | Fresh 23-task ext4 retry; four slots; 40-minute timeout | [Campaign](docs/swebench-v01215-dev23-2026-09-27.md) |
+| v0.12.16 (`9b05a34` release; `b1a6c0e` runner) | 4 / 23 (17.4%) | 5 | Fresh 23-task run; eight slots; 40-minute timeout | [Campaign](docs/swebench-v01216-dev23-2026-09-27.md) |
+
+Every row has 23 submitted official predictions; an empty patch counts against
+the score, not as a missing evaluation. The v0.10.14 report does not separately
+summarize empty patches. The v0.12.15 row is the separately approved fresh
+run after an initial `/tmp` quota failure, not an additional score for that
+failed attempt. Inference completion is **not** official resolution.
+Budgets, runtime/adapter versions, and execution conditions changed between
+campaigns, so this table is a history of observed results, not a controlled
+head-to-head comparison or a general success-rate claim. The v0.10.14 run
+estimated $15.75 in model cost and a 15.2-minute median agent runtime.
+The v0.12.16 run is the first eight-slot inference score in this table;
+five runners timed out, and this does not establish a model-quality regression
+against runs with different concurrency and conditions.
 
 Historical qualification references remain in [benchmark reference](docs/benchmark-reference.md).
 

package/dist/asset-version.d.ts

@@ -3,5 +3,5 @@
  * installed project marker without importing every asset body.
  */
 export declare const RUNTIME_ASSET_VERSION = "0.3.89-completion-proof-v1";
-export declare const V010_RUNTIME_ASSET_VERSION = "0.12.16-operation-efficiency-v1";
+export declare const V010_RUNTIME_ASSET_VERSION = "0.12.17-review-cost-paths-v1";
 export type RuntimeAssetVersion = typeof RUNTIME_ASSET_VERSION | typeof V010_RUNTIME_ASSET_VERSION;

package/dist/asset-version.js

@@ -3,4 +3,4 @@
  * installed project marker without importing every asset body.
  */
 export const RUNTIME_ASSET_VERSION = "0.3.89-completion-proof-v1";
-export const V010_RUNTIME_ASSET_VERSION = "0.12.16-operation-efficiency-v1";
+export const V010_RUNTIME_ASSET_VERSION = "0.12.17-review-cost-paths-v1";

package/dist/core/operator-mission.d.ts

@@ -53,6 +53,8 @@ export interface OperatorMission {
     callID: string | null;
     dispatchOpen: boolean;
     runID: string | null;
+    /** Git HEAD before this mission's first implementation unit, retained across replans and commits. */
+    reviewBaseline?: string;
     /** A cancelled run replaced by a later real user turn; never reuse its acceptance or evidence. */
     supersededRunID?: string;
     plans: number;

package/dist/core/operator-mission.js

@@ -283,8 +283,10 @@ export class OperatorMissionRuntime {
         return this.update(root, state => {
             const expected = this.task(state);
             const resume = state.coordinator !== null && args.task_id === state.coordinator && typeof args.prompt === "string" && args.prompt.trim();
+            // The description is a presentation label. The opaque prompt binds the original request and
+            // requirements; changing only the label must not force a second Coordinator dispatch.
             if (["cancelled", "completed"].includes(state.phase) || state.dispatchOpen || args.subagent_type !== expected.subagent_type ||
-                (!resume && (args.prompt !== expected.prompt || args.description !== expected.description || args.task_id))) {
+                (!resume && (args.prompt !== expected.prompt || args.task_id))) {
                 const reason = ["cancelled", "completed"].includes(state.phase) ? `mission is ${state.phase}`
                     : state.dispatchOpen ? "Coordinator Task is still admitted; read operator_status to reconcile its native completion"
                         : args.task_id && args.task_id !== state.coordinator ? `this location owns Coordinator ${state.coordinator ?? "not yet claimed"}, not ${String(args.task_id)}`

package/dist/core/operator-runtime.d.ts

@@ -343,6 +343,7 @@ export declare class OperatorRuntime {
     /** Reconstruct the admitted child's instructions without trusting a lossy conversation summary. */
     workerContext(root: string, child: string): Promise<string | undefined>;
     continuationCheckpoint(root: string): Promise<string | undefined>;
+    private controlReference;
     private verifyControls;
     /** Follow only the host's authorized parent-link repair; keep the durable control pin in sync. */
     acknowledgeHostHandoffRepair(root: string, taskID: string, path: string, original: string, repaired: string): Promise<void>;

package/dist/core/operator-runtime.js

@@ -910,9 +910,11 @@ export class OperatorRuntime {
             const contents = [JSON.stringify(handoff), JSON.stringify(manifest)];
             controls.push({ path: handoffPath, content: contents[0] }, { path: manifestPath, content: contents[1] });
             const prompt = ["role: implementation", `task_id: ${taskID}`, `project_root: ${this.projectRoot}`,
-                `source_manifest: ${(unit.write.length ? unit.write : unit.read).join(", ") || "none"}`, `operation_manifest: ${manifestRelative}`, `handoff_path: ${handoffPath}`,
+                `source_manifest: ${(unit.write.length ? unit.write : unit.read).join(", ") || "none"}`, `operation_manifest: ${manifestRelative}`,
+                `handoff_path: ${handoffPath}`,
                 `goal_declaration_path: ${declarationPath}`, "acceptance:", ...plan.acceptance.map(value => `  - ${value}`),
                 "validation:", ...unit.validation.map(value => `  - ${value}`),
+                ...(mission ? ["Read handoff_path and other repository files using their project-relative paths as supplied. The native working directory is project_root; do not prepend or reconstruct its absolute path for read/search/shell. Copy project_root only when binding the write gate. Preserve explicitly declared external paths."] : []),
                 mission ? "Read-only investigation commands are unrestricted. Use shell to reproduce and diagnose without asking for command registration. Keep all writes, including generated/transient outputs and cleanup, inside unit.write. Run formal validation exactly as listed, in order and in separate calls, so the host records its real result. If a write scope or formal check must change, return the precise change to your Coordinator; it can extend/redeclare immediately within the original requirements. Diagnostic success is not formal acceptance evidence."
                     : "Execute validation in its declared order. Earlier entries may be approved generator, build, formatter, or exact cleanup commands required before canonical criterion tests. Every persistent or transient generator output must be declared in unit.write. Cleanup may remove only declared unit.write outputs and must be an explicit ordered command after generation and before post-commit or canonical validation; never add an ignore rule or remove an undeclared path. If any necessary command, input, output, or cleanup is missing, do not run an undeclared command or variant and do not use resume evidence tooling to invent permission; return a contract-repair decision.",
                 ...(mission ? [] : ["Preserve existing public API success and error return semantics unless acceptance explicitly changes them, and cover those compatibility boundaries in the declared validation."]),
@@ -1930,12 +1932,12 @@ export class OperatorRuntime {
         // Keep this prefix stable across reads/edits: changing counters belong in status/tool results.
         return `SORTIE_WORKER_CONTEXT\n${JSON.stringify({ root_session_id: root, run_id: state.runID,
             unit_id: unit.unit.id, task_id: /^task_id: (.+)$/m.exec(unit.task.prompt)?.[1],
-            project_root: this.projectRoot, handoff_path: unit.handoffPath, operation_manifest: unit.manifestPath,
+            project_root: this.projectRoot, handoff_path: this.controlReference(unit.handoffPath), operation_manifest: this.controlReference(unit.manifestPath),
             objective: unit.unit.objective, read: unit.unit.read, write: unit.unit.write,
             validation: unit.unit.validation, acceptance: state.acceptance,
             ...(unit.repairValidation === null ? {} : { validation_only: true }),
         })}\nThis is the durable assignment for this same Worker, including after compaction or reload. ` +
-            "A missing path in a summary does not mean the handoff is missing. Read the retained handoff/manifest and inspect the actual diff and outputs before claiming no edits or replaying work. " +
+            "A missing path in a summary does not mean the handoff is missing. Read the retained handoff/manifest using these project-relative paths without prepending project_root, and inspect the actual diff and outputs before claiming no edits or replaying work. " +
             `Use ${this.profile.toolPrefix}operator_status for observed validation and remaining work; do not repeat successful checks for an unchanged candidate. ` +
             "This context is an assignment, not proof of execution or acceptance.";
     }
@@ -1961,6 +1963,11 @@ export class OperatorRuntime {
                                 ? `call ${this.profile.toolPrefix}operator_status and execute its prepare_operator next_action in this turn; do not stop after reporting status`
                                 : `call ${this.profile.toolPrefix}operator_status and execute its exact next_action in this turn; do not stop after reporting status` });
     }
+    controlReference(path) {
+        const local = relative(this.projectRoot, path);
+        return local === ".." || local.startsWith(`..${sep}`) || isAbsolute(local)
+            ? path : local.replaceAll("\\", "/");
+    }
     async verifyControls(unit) {
         if (!isAbsolute(unit.handoffPath) || !isAbsolute(unit.manifestPath))
             throw new Error("operator-control-path-invalid");

package/dist/plugin/mission-review.d.ts

@@ -1,6 +1,7 @@
 import type { OperatorState } from "../core/operator-runtime.js";
 import { type MissionEvidenceExcerpt, type OperatorMission } from "../core/operator-mission.js";
 import { type RuntimeProfile } from "../core/runtime-profile.js";
+export declare function missionReviewBaseline(directory: string): Promise<string | undefined>;
 /** A child ID alone cannot establish the initial phase after a restart or failed dispatch. */
 export declare function initialMissionReviewPrompt(mission: OperatorMission, prompts: readonly string[]): string | undefined;
 /** Recover from native completion records, never from the pending review's inherited child ID. */
@@ -9,7 +10,7 @@ export declare function completedMissionReviewPrompts(mission: OperatorMission |
     messages(id: string): Promise<readonly Record<string, unknown>[]>;
 }): Promise<string[]>;
 /** Pin all scoped tracked/untracked source bytes, including deletions; display a bounded excerpt only. */
-export declare function missionReviewSource(directory: string, run: OperatorState, evidence?: readonly MissionEvidenceExcerpt[]): Promise<{
+export declare function missionReviewSource(directory: string, run: OperatorState, evidence?: readonly MissionEvidenceExcerpt[], baseline?: string): Promise<{
     fingerprint: string;
     excerpt: string;
 }>;

package/dist/plugin/mission-review.js

@@ -13,6 +13,16 @@ import { normalizeManifestScope } from "../core/path.js";
 import { declaredArtifacts } from "./declared-artifacts.js";
 const exec = promisify(execFile);
 const record = (value) => value !== null && typeof value === "object" && !Array.isArray(value);
+export async function missionReviewBaseline(directory) {
+    try {
+        const { stdout } = await exec("git", ["rev-parse", "--verify", "HEAD^{commit}"], { cwd: directory });
+        const oid = stdout.trim();
+        return /^[a-f0-9]{40,64}$/u.test(oid) ? oid : undefined;
+    }
+    catch {
+        return undefined;
+    } // A mission can start before Git is initialized.
+}
 /** A child ID alone cannot establish the initial phase after a restart or failed dispatch. */
 export function initialMissionReviewPrompt(mission, prompts) {
     return [mission.review?.initialPrompt, ...prompts].find(prompt => typeof prompt === "string" &&
@@ -86,8 +96,8 @@ export async function completedMissionReviewPrompts(mission, profile, root, requ
     return prompts;
 }
 /** Pin all scoped tracked/untracked source bytes, including deletions; display a bounded excerpt only. */
-export async function missionReviewSource(directory, run, evidence = []) {
-    const hash = createHash("sha256").update(JSON.stringify(run.units.map(unit => ({ unit: unit.unit, hashes: unit.hashes }))));
+export async function missionReviewSource(directory, run, evidence = [], baseline) {
+    const hash = createHash("sha256").update(JSON.stringify({ baseline, units: run.units.map(unit => ({ unit: unit.unit, hashes: unit.hashes })) }));
     let selected = "";
     for (const entry of evidence) {
         const absolute = resolve(directory, entry.path);
@@ -136,11 +146,40 @@ export async function missionReviewSource(directory, run, evidence = []) {
     const ignored = local.length ? await git(["ls-files", "-z", "--others", "--ignored", "--exclude-standard", "--", ...scopes]) : "";
     for (const path of ignored.split("\0").filter(Boolean))
         untracked.add(path);
-    const diff = local.length ? await git(["diff", "--no-ext-diff", "--no-textconv", "HEAD", "--", ...scopes]).catch(() => git(["diff", "--no-ext-diff", "--no-textconv", "--", ...scopes])) : "";
+    const omitted = [];
+    let diff = "";
+    if (local.length && baseline) {
+        // HEAD-only diffs are empty once the Worker commits. Show bounded changes from the
+        // mission's original HEAD across all replans, not arbitrary alphabetical repository files.
+        const changed = (await git(["diff", "--name-only", "-z", "--no-ext-diff", baseline, "--", ...scopes]))
+            .split("\0").filter(Boolean);
+        const shown = changed.slice(0, 16);
+        const heading = changed.length ? `Changed since mission baseline (${baseline}; ${changed.length} paths):\n` : "";
+        const headers = shown.map(path => `\n--- changed: ${path} ---\n`);
+        const available = Math.max(0, 24_000 - Buffer.byteLength(selected + heading) -
+            headers.reduce((size, title) => size + Buffer.byteLength(title), 0) - 500);
+        const allowance = shown.length ? Math.min(3_000, Math.floor(available / shown.length)) : 0;
+        diff = heading;
+        for (const [index, path] of shown.entries()) {
+            const patch = await git(["diff", "--no-ext-diff", "--no-textconv", baseline, "--", path]);
+            const bytes = Buffer.from(patch);
+            if (allowance <= 0) {
+                omitted.push(path);
+                continue;
+            }
+            diff += `${headers[index]}${bytes.subarray(0, allowance).toString("utf8")}`;
+            if (bytes.length > allowance)
+                omitted.push(path);
+        }
+        omitted.push(...changed.slice(shown.length));
+    }
+    else if (local.length) {
+        diff = await git(["diff", "--no-ext-diff", "--no-textconv", "HEAD", "--", ...scopes])
+            .catch(() => git(["diff", "--no-ext-diff", "--no-textconv", "--", ...scopes]));
+    }
     const unchanged = diff.length === 0;
     let excerpt = selected + diff;
     const paths = [...new Set([...names.split("\0"), ...ignored.split("\0")].filter(Boolean))].sort();
-    const omitted = [];
     for (const path of paths) {
         hash.update(JSON.stringify(path));
         try {

package/dist/plugin/profiled.js

@@ -4,7 +4,7 @@ import { DEFAULT_OPERATOR_PROPOSAL_BUDGET, OPERATOR_APPROVAL_CONTRACT, OPERATOR_
 import { CANONICAL_AGENT_ROLES, canonicalAgent, profileAgent, profileTool, V010_RUNTIME_PROFILE } from "../core/runtime-profile.js";
 import { SortieDogsPlugin as canonicalPlugin } from "./index.js";
 import { taskChildSessionID } from "./task-result-repair.js";
-import { relative, resolve, sep } from "node:path";
+import { isAbsolute, relative, resolve, sep } from "node:path";
 import { readFile, realpath } from "node:fs/promises";
 import { BUILT_IN_MODEL_CATALOG } from "./model-routing.js";
 import { goalFingerprint } from "../core/goal-bound.js";
@@ -14,7 +14,7 @@ import { normalizeCommand } from "./gate.js";
 import { normalizeRelativePath } from "../core/path.js";
 import { MISSION_EVIDENCE_GAP_REVIEW_LIMIT, OperatorMissionRuntime, missionPacket, missionPlan, missionReviewAccepted, missionReviewTask, missionCommandOutcome, missionConversationContext, missionExecutionStatus, missionValidationCommand, missionReviewTraces, missionReviewVerdict } from "../core/operator-mission.js";
 import { publishMissionProgress } from "./mission-progress.js";
-import { completedMissionReviewPrompts, initialMissionReviewPrompt, missionReviewSource } from "./mission-review.js";
+import { completedMissionReviewPrompts, initialMissionReviewPrompt, missionReviewBaseline, missionReviewSource } from "./mission-review.js";
 import { missionLocations, missionLocationPacket } from "./mission-location.js";
 import { SOURCE_REVIEW_RISK_TAGS } from "../core/consultation.js";
 import { createHash } from "node:crypto";
@@ -280,6 +280,21 @@ export function createProfiledPlugin(profile, assetVersion) {
                 return role;
             return CANONICAL_AGENT_ROLES.includes(value) ? `foreign/${value}` : value;
         }
+        function workerControlPrompt(text, outward) {
+            // Project only the generated header, never acceptance, commands or the user's objective.
+            // Durable Tasks and the shared engine retain their canonical absolute identities.
+            if (!/^role: implementation\ntask_id: operator-/u.test(text))
+                return text;
+            const end = text.indexOf("\nacceptance:\n");
+            if (end < 0)
+                return text;
+            return text.slice(0, end).replace(/^(handoff_path|goal_declaration_path): (.+)$/gm, (_line, name, path) => {
+                const absolute = resolve(input.directory, path);
+                const local = relative(input.directory, absolute);
+                const inside = local !== ".." && !local.startsWith(`..${sep}`) && !isAbsolute(local);
+                return `${name}: ${outward && inside ? local.replaceAll("\\", "/") : absolute}`;
+            }) + text.slice(end);
+        }
         function translate(value, outward) {
             if (Array.isArray(value))
                 return value.map(item => translate(item, outward));
@@ -315,7 +330,23 @@ export function createProfiledPlugin(profile, assetVersion) {
                             if ((method === "prompt" || method === "promptAsync") && record(request) && record(request.path) &&
                                 typeof request.path.id === "string" && retired.has(request.path.id))
                                 throw new Error("runtime-profile-revoked");
-                            return translate(await operation.apply(group, args.map(argument => translate(argument, true))), false);
+                            const result = translate(await operation.apply(group, args.map(argument => translate(argument, true))), false);
+                            // Native Worker history stores short references. Only its generated user
+                            // header is projected back for core recovery; root requests and tool evidence
+                            // keep their original text, including any path-like lines in task data.
+                            if (method === "messages" || method === "message") {
+                                const data = payload(result);
+                                for (const message of Array.isArray(data) ? data : [data]) {
+                                    if (!record(message) || !record(message.info) || message.info.role !== "user" ||
+                                        message.info.agent !== "dog-worker" || !Array.isArray(message.parts))
+                                        continue;
+                                    for (const part of message.parts)
+                                        if (record(part) && part.type === "text" && typeof part.text === "string") {
+                                            part.text = workerControlPrompt(part.text, false);
+                                        }
+                                }
+                            }
+                            return result;
                         };
                     },
                 });
@@ -1324,10 +1355,15 @@ export function createProfiledPlugin(profile, assetVersion) {
                         if (relocated)
                             return JSON.stringify({ ...relocated, budget: await control.currentBudget(context.sessionID) });
                     }
-                    const mission = await retainCancelledMissionAcceptance(context.sessionID, await missions.start(context.sessionID, args.requirements, args.intent === "replace" || args.intent === "new", {
+                    let mission = await retainCancelledMissionAcceptance(context.sessionID, await missions.start(context.sessionID, args.requirements, args.intent === "replace" || args.intent === "new", {
                         kind: args.kind === "operation" ? "operation" : "implementation",
                         context: missionConversationContext(await messages(context.sessionID)),
                     }));
+                    if (!mission.reviewBaseline) {
+                        const baseline = await missionReviewBaseline(input.directory);
+                        if (baseline)
+                            mission = await missions.update(context.sessionID, state => { state.reviewBaseline = baseline; });
+                    }
                     return JSON.stringify({ ...locationObservation(context.sessionID), ...(mission.dispatchOpen
                             ? missionDispatchPacket(mission, await operators.read(context.sessionID))
                             : { mission_id: mission.id, requirements: mission.requirements, task: missions.task(mission),
@@ -1365,6 +1401,11 @@ export function createProfiledPlugin(profile, assetVersion) {
             if (budget && budget.remaining_units < plan.units.length && !same)
                 throw new Error(`mission-budget-exhausted: plan needs ${plan.units.length} units; ${budget.remaining_units} remain. ` +
                     "The current run is retained. Correct the plan within the remaining budget, or report a necessary cumulative extension to Operator.");
+            if (!mission.reviewBaseline) {
+                const baseline = await missionReviewBaseline(input.directory);
+                if (baseline)
+                    mission = await missions.update(root, state => { state.reviewBaseline ??= baseline; });
+            }
             // Cancellation marks the durable units before interrupting their native sessions. A cancelled
             // status alone therefore cannot prove the old Worker stopped or its reservation settled.
             const cancelledPredecessor = previous?.phase === "cancelled" &&
@@ -1453,7 +1494,7 @@ export function createProfiledPlugin(profile, assetVersion) {
                 const evidence = args.evidence;
                 if (evidence !== undefined && (!Array.isArray(evidence) || evidence.length > 6))
                     throw new Error("mission-review-evidence: select at most six focused excerpts");
-                const source = await missionReviewSource(input.directory, run, evidence);
+                const source = await missionReviewSource(input.directory, run, evidence, mission.reviewBaseline);
                 const requestFingerprint = goalFingerprint({ run: run.runID, source: source.fingerprint, risk, traces });
                 if (mission.review?.requestFingerprint === requestFingerprint && mission.review.verdict !== "pending") {
                     return JSON.stringify({ ...missionPacket(mission, run), status: "review-recorded",
@@ -1494,7 +1535,7 @@ export function createProfiledPlugin(profile, assetVersion) {
         async function assertMissionReview(root, mission) {
             const run = await operators.required(root), review = mission.review;
             if (!review || review.runID !== run.runID || !missionReviewAccepted(review) ||
-                review.source !== (await missionReviewSource(input.directory, run, review.evidence)).fingerprint)
+                review.source !== (await missionReviewSource(input.directory, run, review.evidence, mission.reviewBaseline)).fingerprint)
                 throw new Error("mission-review-required-or-stale");
         }
         tools[submitMission] = { description: "Coordinator: return only a completion candidate, a user-only decision, or a proven external/scope/budget blocker. Continue ordinary investigation, scope extensions and corrections yourself. Unit progress is published automatically without waking Operator.",
@@ -1746,6 +1787,11 @@ export function createProfiledPlugin(profile, assetVersion) {
                 const mapped = translate(output, false);
                 await core["chat.message"]?.(translate(chat, false), mapped);
                 Object.assign(output, translate(mapped, true));
+                if (role === "dog-worker")
+                    for (const part of output.parts) {
+                        if (record(part) && part.type === "text" && typeof part.text === "string")
+                            part.text = workerControlPrompt(part.text, true);
+                    }
                 if (role === "dog-worker" && explicitWorkerSelection) {
                     if (explicitWorkerSelection.model) {
                         const split = explicitWorkerSelection.model.indexOf("/");
@@ -1795,9 +1841,9 @@ export function createProfiledPlugin(profile, assetVersion) {
                 if (repairAccess !== null) {
                     const bind = profileTool(profile, "sortie_bind_write_gate"), release = profileTool(profile, "sortie_release_write_gate");
                     const allowedRead = request.tool.toLowerCase() === "read" && typeof args.filePath === "string" &&
-                        resolve(args.filePath) === resolve(repairAccess.handoff_path);
+                        resolve(input.directory, args.filePath) === resolve(repairAccess.handoff_path);
                     const allowedBind = request.tool === bind && args.project_root === input.directory &&
-                        resolve(String(args.manifest_path ?? "")) === resolve(repairAccess.manifest_path);
+                        resolve(input.directory, String(args.manifest_path ?? "")) === resolve(repairAccess.manifest_path);
                     const allowedRelease = request.tool === release;
                     const allowedValidation = ["bash", "shell"].includes(request.tool.toLowerCase()) && typeof args.command === "string" &&
                         repairAccess.expected_command !== null && normalizeCommand(args.command) === repairAccess.expected_command;

package/dist/plugin/sortie-debrief.d.ts

@@ -56,6 +56,12 @@ export interface Debrief {
             readonly cacheReadTokens: number;
             readonly cacheWriteTokens: number;
         };
+        readonly estimatedCost?: {
+            readonly usd: number;
+            readonly pricedRequests: number;
+            readonly unpricedRequests: number;
+            readonly complete: boolean;
+        };
         readonly hostCostZero?: boolean;
     }[] | null;
     /** Shares use observed tokens; incomplete history must not erase known model usage. */

package/dist/plugin/sortie-debrief.js

@@ -293,6 +293,9 @@ export function buildDebrief(receipt, contract, observation, records) {
             return { model, tokens, percent: tokens / total * 100,
                 ...(usage?.complete ? { inputCache: { uncachedInputTokens: usage.uncachedInputTokens,
                         cacheReadTokens: usage.cacheReadTokens, cacheWriteTokens: usage.cacheWriteTokens } } : {}),
+                ...(usage ? { estimatedCost: { usd: usage.estimatedCost, pricedRequests: usage.pricedRequests,
+                        unpricedRequests: usage.unpricedRequests,
+                        complete: usage.complete && usage.pricingComplete && usage.unpricedRequests === 0 && usage.pricedRequests > 0 } } : {}),
                 ...(usage?.complete && usage.costAvailable && tokens > 0 && usage.cost === 0 ? { hostCostZero: true } : {}) };
         }) : null,
         mixCoverage: { complete: observation.complete && usageComplete, observedTokens: total },
@@ -310,6 +313,7 @@ export function buildDebrief(receipt, contract, observation, records) {
 const label = (text) => text.replace(/[\r\n\t]/gu, " ").replace(/[\\`*_{}\[\]()<>!|]/gu, "").slice(0, 120);
 const milliseconds = (value) => value === null ? "未記録"
     : value < 1000 ? `${value}ms` : `${(value / 1000).toFixed(2)}s`;
+const dollars = (usd) => `$${usd.toFixed(usd > 0 && usd < 0.00005 ? 6 : 4)}`;
 const gauge = (percent) => {
     const eighths = Math.max(0, Math.min(80, Math.round(percent * 0.8)));
     const whole = Math.floor(eighths / 8), remainder = eighths % 8;
@@ -324,13 +328,21 @@ export function renderDebrief(debrief) {
     const packVisible = pack?.slice(0, 4) ?? [];
     if (pack !== null && pack.length > 4)
         packVisible.push({ model: "その他", count: pack.slice(4).reduce((sum, entry) => sum + entry.count, 0) });
-    const mix = debrief?.mix == null ? null : [...debrief.mix].sort((a, b) => b.tokens - a.tokens || a.model.localeCompare(b.model));
+    const mix = debrief?.mix == null ? null : [...debrief.mix].sort((a, b) => (b.estimatedCost?.pricedRequests ? 1 : 0) - (a.estimatedCost?.pricedRequests ? 1 : 0) ||
+        (b.estimatedCost?.usd ?? 0) - (a.estimatedCost?.usd ?? 0) || b.tokens - a.tokens || a.model.localeCompare(b.model));
+    const pricedUsd = mix?.reduce((sum, entry) => sum + (entry.estimatedCost?.usd ?? 0), 0) ?? 0;
+    const hasPriced = mix?.some(entry => (entry.estimatedCost?.pricedRequests ?? 0) > 0);
+    const partiallyPriced = debrief?.mixCoverage?.complete === false || mix?.some(entry => !entry.estimatedCost?.complete);
     const visible = mix?.slice(0, 4) ?? [];
     if (mix !== null && mix.length > 4) {
         const tail = mix.slice(4);
         const cacheKnown = tail.every((entry) => entry.inputCache !== undefined);
         visible.push({ model: "その他", tokens: tail.reduce((sum, entry) => sum + entry.tokens, 0),
             percent: tail.reduce((sum, entry) => sum + entry.percent, 0),
+            estimatedCost: { usd: tail.reduce((sum, entry) => sum + (entry.estimatedCost?.usd ?? 0), 0),
+                pricedRequests: tail.reduce((sum, entry) => sum + (entry.estimatedCost?.pricedRequests ?? 0), 0),
+                unpricedRequests: tail.reduce((sum, entry) => sum + (entry.estimatedCost?.unpricedRequests ?? 0), 0),
+                complete: tail.every(entry => entry.estimatedCost?.complete) },
             ...(cacheKnown ? { inputCache: {
                     uncachedInputTokens: tail.reduce((sum, entry) => sum + entry.inputCache.uncachedInputTokens, 0),
                     cacheReadTokens: tail.reduce((sum, entry) => sum + entry.inputCache.cacheReadTokens, 0),
@@ -342,14 +354,22 @@ export function renderDebrief(debrief) {
         : value === "WAIVED" ? "免除" : "未記録";
     const counts = new Map(packVisible.map((entry) => [entry.model, entry.count]));
     return [
-        `モデル内訳    ${mix === null ? "usage未取得" : debrief?.mixCoverage?.complete === false
-            ? "観測済みtoken比率（一部未取得）" : "token比率"}`,
+        `モデル内訳    ${mix === null ? "usage未取得" : !hasPriced
+            ? "予測費用未換算（モデル別token数）" : pricedUsd === 0
+            ? "予測費用比率（合計$0・算出不可）" : partiallyPriced
+            ? "換算済み予測費用の比率（一部未換算）" : "予測費用比率"}`,
         ...visible.map((entry) => {
             const count = counts.get(entry.model);
             const input = entry.inputCache === undefined ? 0 : entry.inputCache.uncachedInputTokens +
                 entry.inputCache.cacheReadTokens + entry.inputCache.cacheWriteTokens;
             const cache = entry.inputCache === undefined || input === 0 ? "未取得" : `${(entry.inputCache.cacheReadTokens / input * 100).toFixed(0)}%`;
-            return `🐕 ${label(entry.model)} ${gauge(entry.percent)} ${entry.percent.toFixed(1)}% ${entry.tokens.toLocaleString("ja-JP")} tokens${count === undefined ? "" : ` ×${count}`} ↺${cache}${entry.hostCostZero ? "†" : ""}`;
+            const cost = entry.estimatedCost;
+            const share = cost?.pricedRequests
+                ? pricedUsd > 0
+                    ? `${gauge(cost.usd / pricedUsd * 100)} ${(cost.usd / pricedUsd * 100).toFixed(1)}% · ${dollars(cost.usd)}${cost.complete ? "" : "（一部未換算）"}`
+                    : `比率算出不可 · ${dollars(cost.usd)}`
+                : "費用未換算";
+            return `🐕 ${label(entry.model)} ${share} · ${entry.tokens.toLocaleString("ja-JP")} tokens${count === undefined ? "" : ` ×${count}`} ↺${cache}${entry.hostCostZero ? "†" : ""}`;
         }),
         ...(visible.some((entry) => entry.hostCostZero) ? ["   †host費用0計上あり（無料・全体価格の評価ではありません）"] : []),
         `⚡ 実行重複率 ${debrief?.overlap !== undefined && debrief.overlap.wallMilliseconds > 0

package/dist/runtime-mission-assets.js

@@ -124,6 +124,8 @@ ${controls(profile, ["plan_units", "operator_next", "operator_status", "expand_u
 You are Coordinator. Own most of the practical work and dispatch within the saved original request.
 Read/search and confirmation shell commands are available; source edits belong to Worker. Do not edit
 through shell. There is no proposal/approval/contract-repair round trip in this route.
+Use project-relative paths for repository read/search/shell; the native working directory is project_root.
+Do not reconstruct or prepend the absolute workspace path. Preserve explicitly requested external paths.
 
 For release/benchmark work, retain the user's selected package/environment and record its receipt/hash.
 A local repack and a published tarball can have different hashes; that alone does not prohibit a requested
@@ -215,6 +217,10 @@ mode: subagent
 Implement the assigned unit promptly. Use the user's language in its handoff. Read the exact handoff_path
 once, then call ${profile.toolPrefix}bind_write_gate with project_root and operation_manifest before writes.
 The host generates these documents; do not rewrite or reconstruct them. Copy opaque paths exactly.
+Use handoff_path and operation_manifest as supplied: new local references are relative to the native
+project working directory; a retained older absolute reference can be used without rewriting it.
+Keep repository read/search/shell paths relative; do not prepend project_root or copy an absolute root from
+tool output. project_root is the exact binding identity, not a prefix to rebuild. Retain explicit external paths.
 The binding remains valid throughout this Task until return or a control/source authorization change.
 Use ${profile.toolPrefix}operator_status when the task needs native runtime identity, mission state or
 remaining budget. It is a read-only observation available to Worker; do not request a separate unit or

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "sortie-dogs",
-  "version": "0.12.16",
+  "version": "0.12.17",
   "description": "Bounded agent harness and validated orchestration loop plugin for OpenCode",
   "keywords": [
     "opencode",