sortie-dogs 0.10.3 → 0.10.5

package/README.md

@@ -74,7 +74,7 @@ bounded implementation, canonical validation, and evidence-backed completion.
 
 Guides: [日本語](docs/guide-ja.md) · [简体中文](docs/guide-zh-CN.md) · [テスト実行](docs/testing.md) · [CLI testing](docs/cli-testing.md)
 
-Release: [v0.10.3](https://github.com/zufall-upon/Sortie-dogs/releases/tag/v0.10.3)
+Release: [v0.10.5](https://github.com/zufall-upon/Sortie-dogs/releases/tag/v0.10.5)
 
 ## Latest local benchmark case study
 
@@ -92,20 +92,20 @@ Run configuration was fixed per product configuration:
   children on `openai/gpt-5.6-sol` / `medium`, with the pinned Sortie package and runtime assets.
   No Luna, Astra, or Opus messages were observed in these trials.
 
-| Metric · one frozen task | Bare OpenCode | Sortie v0.9.12 | Sortie v0.10.1 RC qualification |
-| --- | ---: | ---: | ---: |
-| Attempts needed | 3 | 5 | 1 |
-| Completed runs compared | 3 | 3 | 1 |
-| Verified PASS | 0/3 | 0/3 | 0/1 |
-| Task checks · F2P | 11.1% · 3/27 | 85.2% · 23/27 | 88.9% · 8/9 |
-| Retained checks · P2P | 282/282 | 282/282 | 94/94 |
-| Median agent wall | 24.5 min | 25.7 min | 23.2 min · n=1 |
-| Median model steps | 43 | 39 | 41 · CLI stream only |
-| Implementation child sessions · total | 0 | 13 | 3 |
-| Estimated API-equivalent cost · median completed run | $3.53 | $2.85 | **$4.48** · n=1 |
-| Estimated cost · completed runs | $10.69 | $9.74 | **$4.48** |
-| Additional interrupted-attempt cost · same fixed snapshot | $0 | $6.20 | $0 |
-| Total cost to acquire completed runs · same fixed snapshot | $10.69 | $15.94 | **$4.48** |
+| Metric · one frozen task | Bare OpenCode | Sortie v0.9.12 | Sortie v0.10.1 RC qualification | Sortie v0.10.3 qualification · Terra/xhigh | Sortie v0.10.4 qualification · Terra/xhigh | Sortie v0.10.5 candidate qualification · Terra/xhigh + Luna/max |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Attempts needed | 3 | 5 | 1 | 1 | 2 | 1 recorded success |
+| Completed runs compared | 3 | 3 | 1 | 1 | 2 | 1 recorded success |
+| Verified PASS | 0/3 | 0/3 | 0/1 | 0/1 | 0/2 | **1/1** |
+| Task checks · F2P | 11.1% · 3/27 | 85.2% · 23/27 | 88.9% · 8/9 | 77.8% · 7/9 | 72.2% · 5/9 and 8/9 | **100% · 9/9** |
+| Retained checks · P2P | 282/282 | 282/282 | 94/94 | 94/94 | 94/94 | **94/94** |
+| Median agent wall | 24.5 min | 25.7 min | 23.2 min · n=1 | 22.7 min · n=1 | 31.1 min · n=2 | 43.1 min · n=1 |
+| Median model steps | 43 | 39 | 41 · CLI stream only | 36 · CLI stream only | 44 · CLI stream only | 25 · CLI stream only |
+| Implementation child sessions · total | 0 | 13 | 3 | 5 | 10 | 3 |
+| Estimated API-equivalent cost · median completed run | $3.53 | $2.85 | **$4.48** · n=1 | **$3.79** · n=1 | **$3.82** · n=2 | **$1.72** · n=1 |
+| Estimated cost · completed runs | $10.69 | $9.74 | **$4.48** | **$3.79** | **$7.64** | **$1.72** |
+| Additional interrupted-attempt cost · same fixed snapshot | $0 | $6.20 | $0 | $0 | $0 | Not reported |
+| Total cost to acquire completed runs · same fixed snapshot | $10.69 | $15.94 | **$4.48** | **$3.79** | **$7.64** | Not reported |
 
 All three Bare runs passed 1/9 task checks. The three completed Sortie runs passed 7/9, 8/9,
 and 8/9. Every compared candidate retained 94/94 prior checks, but every official verifier still
@@ -128,12 +128,52 @@ into this RC's same-snapshot acquisition cost. Observed spend across both distin
 **$8.755191**, which is development spend rather than this RC's acquisition cost. Historical Bare and
 v0.9.12 costs retain their 2026-07-30 schedule; their cost cells are not same-rate comparisons with
 this RC estimate.
-
-![Latest local case study: Bare completed 11.1 percent of task checks at a median estimated API-equivalent cost of $3.53; Sortie completed 85.2 percent at $2.85. Sortie needed five attempts and $15.94 to collect three completed runs. Neither configuration achieved a Verified PASS.](docs/assets/quality-cost-reference.svg)
+
+The released **v0.10.3** Terra/xhigh treatment is a separate one-shot qualification, not a matched
+comparison or leaderboard result. It reached `DONE` and the localized Docker-free official verifier,
+which returned reward **0**, F2P **7/9**, and P2P **94/94**. Its full root-plus-eight-descendant audit
+prices 93 assistant requests and 4,809,706 tokens at **$3.789190** API-equivalent cost: **$2.186540**
+for Terra/xhigh and **$1.602650** for Sol/low. Pricing coverage is 100%, with no unpriced or pending
+requests. This uses the same 2026-09-14 per-request schedule; the observed host cost of $0 is not used
+as cost evidence, and the estimate is not an invoice or subscription charge.
+
+The released **v0.10.4** Terra/xhigh treatment repeats that qualification twice on the same frozen
+task, pins, and verifier. Both runs reached `DONE` with confirmed cleanup, and the localized
+Docker-free official verifier returned reward **0** with P2P **94/94** each time. F2P was **5/9** and
+**8/9**, so the single v0.10.3 observation of 7/9 sits inside this spread: neither release solves the
+task, and the per-run difference is run-to-run variance at n=1 and n=2, not a measured quality change.
+
+The **v0.10.5 candidate** switches the v0.10 implementation worker default from Sol/low to
+Luna/max. One separate qualification on the same frozen task, pins, and localized Docker-free
+official verifier reached `DONE` and returned reward **1**, F2P **9/9**, and P2P **94/94**. Agent
+wall was **2,583,374 ms** (43.1 min), the CLI stream recorded 25 model steps, and the operation
+created three implementation child sessions. Its full root-plus-five-descendant audit prices 106
+assistant requests and 8,830,131 tokens at **$1.719999** API-equivalent cost: **$1.436725** for
+Terra/xhigh and **$0.283274** for Luna/max. Pricing coverage was 100%, with no unpriced or pending
+requests. The candidate package was built before the release version bump and reported package
+metadata `0.10.4`; this is a source-snapshot reference, not a released-package measurement.
+
+The v0.10.5 column deliberately reports this one verified success only. It is not a success-rate or
+acquisition-cost aggregate, and its `Not reported` cells do not imply zero cost or zero failed
+attempts outside this recorded success.
+
+The release payload this qualification exists to check is a prompt-cache correctness fix, and that
+effect is separable from task quality. Before the fix the proposal investigation child re-sent its
+whole prompt uncached on every request, because a consumed-read counter sat in the system block and a
+system element is an absolute prompt prefix. The audit now reports per-session prefix reuse, which
+isolates exactly that failure: the v0.10.3 proposal child reused a median **0.065** of the previous
+prompt and was the only flagged session in its run, while both v0.10.4 runs reuse **0.957** and
+**0.986** and flag none. Its uncached input fell from **406,202** tokens to **213,115** and
+**285,060**, and its estimated cost fell from **$0.943508** to **$0.574788** and **$0.802139**. Whole-run
+cost is not a clean measure of the same fix, because each run performed a different amount of work:
+the two v0.10.4 runs priced 97 requests at **$3.268415** and 119 requests at **$4.368122**, both at
+100% pricing coverage on the 2026-09-14 per-request schedule.
+
+![Historical local case study: Bare completed 11.1 percent of task checks at a median estimated API-equivalent cost of $3.53; Sortie v0.9.12 completed 85.2 percent at $2.85. The one-shot v0.10.3 Terra/xhigh qualification completed 77.8 percent at $3.79. Neither configuration achieved a Verified PASS.](docs/assets/quality-cost-reference.svg)
 
 Cost audits use deduplicated root and descendant session tokens, grouped by the model that produced
-each message. Historical values retain their frozen 2026-07-30 short-context schedule; the RC uses
-the product's 2026-09-14 per-request schedule described above. Completed-run cost shows execution
+each message. Historical values retain their frozen 2026-07-30 short-context schedule; the v0.10
+qualifications use the product's 2026-09-14 per-request schedule described above. Completed-run cost shows execution
 efficiency; same-snapshot acquisition cost adds interrupted attempts. These are API-equivalent
 estimates, not invoices.
 

package/dist/core/operator-runtime.js

@@ -1625,7 +1625,9 @@ export class OperatorRuntime {
                 committed_head: state.gitLifecycle.committedHead,
                 inherited: state.gitLifecycle.commitProvenance === "inherited-parent",
                 carried_uncommitted_paths: (state.gitLifecycle.carriedPaths ?? []).slice(0, 32) },
-            requirements: state.acceptance.map((criterion, index) => ({ criterion,
+            // Index-aligned with this packet's own `acceptance`. Restating each criterion here duplicated the
+            // whole acceptance text inside a packet the root re-reads on every later turn of its session.
+            requirements: state.acceptance.map((_criterion, index) => ({ index,
                 proof_ids: state.acceptanceProof[index], status: state.acceptanceProof[index].every(id => proved.has(id)) ? "observed-pass" : "unproven" })),
             source_refs: state.sourceRefs, decision: state.decision,
             contract_repair: state.contractRepair === null ? null : { code: state.contractRepair.code, unit_id: state.contractRepair.unit_id,

package/dist/plugin/profiled.js

@@ -16,7 +16,7 @@ const SERIAL_CAPABILITIES = new Set([
     "sortie_bind_write_gate", "sortie_release_write_gate", "sortie_check_contract",
     "sortie_compact_and_continue", "sortie_enable_backlog_drain",
 ]);
-const PREVIEW_WORKER_ROUTE = Object.freeze({ model: "openai/gpt-5.6-sol", variant: "low" });
+const PREVIEW_WORKER_ROUTE = Object.freeze({ model: "openai/gpt-5.6-luna", variant: "max" });
 const PREVIEW_SCOUT_ROUTE = Object.freeze({ model: "openai/gpt-5.6-luna", variant: "xhigh" });
 const PREVIEW_OPERATIONS_ROUTE = Object.freeze({ model: "openai/gpt-5.6-terra", variant: "xhigh" });
 const PREVIEW_PRIMARY_ROUTE = Object.freeze({ model: "openai/gpt-5.6-sol", variant: "low" });
@@ -137,6 +137,27 @@ export function createProfiledPlugin(profile, assetVersion) {
             }
             return parentRoot;
         }
+        /**
+         * Resolve the owning root of a proposal investigation child for prompt assembly only.
+         *
+          * `rootFor` stops resolving this child as soon as the proposal leaves `investigating`, which is the
+         * correct authorization answer: the child must not run another tool. It is the wrong answer for the
+         * system prefix, because losing the root also drops every profile element and changes the absolute
+         * prompt prefix, so the child's post-submit turn re-sent its whole investigation uncached. This
+         * resolver grants no tool authority and is never consulted on an execute path.
+         */
+        async function proposalPromptRoot(id) {
+            if (retired.has(id))
+                return undefined;
+            const who = await identity(id);
+            if (who.role !== "dog-operator" || !who.parent)
+                return undefined;
+            const parentRoot = await rootFor(who.parent, 1);
+            if (!parentRoot)
+                return undefined;
+            const proposal = await proposals.read(parentRoot);
+            return proposal?.phase === "submitted" && proposal.proposal_session_id === id ? parentRoot : undefined;
+        }
         function mapAgent(value, outward) {
             if (outward) {
                 if (value.startsWith("foreign/"))
@@ -436,6 +457,19 @@ export function createProfiledPlugin(profile, assetVersion) {
             }
             return { ...packet, budget };
         }
+        /**
+         * Proposal accounting without the submitted packet body.
+         *
+         * The full proposal carries the plan, every unit, and the acceptance text a further time. Once the
+         * root has compared it, approval froze it into the execution run, whose own packet already reports
+         * acceptance, units, and scopes. Re-emitting it on approval and on every later status call appended
+         * a redundant copy to the one session that re-reads its whole context on every turn. `content_hash`
+         * stays, so the exact submitted revision is still identifiable.
+         */
+        function proposalIdentity(state) {
+            const { proposal: _packet, ...identity } = proposals.packet(state);
+            return identity;
+        }
         tools[status] = { description: "Read the durable root-owned operator outcome and host budget counters (max_units, consumed_units, reserved_units, remaining_units) without claiming acceptance or retrying work. An investigating proposal returns its exact short Task reference only before a Task has been admitted; an existing admission never yields a redispatch Task.",
             args: {}, execute: async (_args, context) => {
                 await requireRoot(context.sessionID);
@@ -452,7 +486,7 @@ export function createProfiledPlugin(profile, assetVersion) {
                         ? "compare and approve the exact proposal; do not call operator_next before approval prepares a run"
                         : "approved proposal has no operator run; do not call operator_next; reconcile the approval or preparation failure";
                 return JSON.stringify(state ? { ...await operatorPacket(state), ...(draft ? { pending_draft: draft } : {}),
-                    ...(proposal ? { proposal: proposals.packet(proposal) } : {}) }
+                    ...(proposal ? { proposal: proposalIdentity(proposal) } : {}) }
                     : draft ? { ...draft, ...(proposal ? { proposal: proposals.packet(proposal) } : {}) }
                         : proposal ? { profile: profile.id, proposal: proposals.packet(proposal),
                             ...(proposal.phase === "investigating" && proposal.proposal_call_id === null ? { task: proposals.referenceTask(proposal),
@@ -701,7 +735,7 @@ export function createProfiledPlugin(profile, assetVersion) {
                         throw new Error("operator-proposal-approved-plan-invalid");
                     await registerPreparedGoal(context.sessionID, prepared.state);
                     const committed = await proposals.approve(context.sessionID, approval);
-                    return JSON.stringify({ ...proposals.packet(committed), execution: JSON.parse(preparedTask(prepared.state)) });
+                    return JSON.stringify({ ...proposalIdentity(committed), execution: JSON.parse(preparedTask(prepared.state)) });
                 });
             } };
         const ownTools = new Set([prepare, repair, next, status, cancel, complete, resume, resolveContractRepair,
@@ -1109,6 +1143,15 @@ export function createProfiledPlugin(profile, assetVersion) {
                         throw new Error("operator-contract-repair-validation-command-missing");
                     await operators.recordRepairValidationResult((await rootFor(id)), id, repairAccess.expected_command, exit);
                 }
+                if (!ownership && request.tool.toLowerCase() === "read") {
+                    const proposal = await proposals.read((await rootFor(id)));
+                    if (proposal?.phase === "investigating" && proposal.proposal_session_id === id) {
+                        output.output = `${output.output ?? ""}\n\nSORTIE_PROPOSAL_BUDGET actual_reads=${proposal.read_count}; ` +
+                            `remaining_reads=${proposal.intent.proposal_budget.max_reads - proposal.read_count}; ` +
+                            `submissions=${proposal.submission_count}; ` +
+                            `remaining_submissions=${proposal.intent.proposal_budget.max_submissions - proposal.submission_count}.`;
+                    }
+                }
                 if (ownership) {
                     const child = taskChildSessionID(output);
                     if (child)
@@ -1156,7 +1199,7 @@ export function createProfiledPlugin(profile, assetVersion) {
                 await core["permission.ask"]?.(request, output);
             },
             "experimental.chat.system.transform": async (request, output) => {
-                const root = await rootFor(request.sessionID);
+                const root = await rootFor(request.sessionID) ?? await proposalPromptRoot(request.sessionID);
                 if (!root)
                     return;
                 await core["experimental.chat.system.transform"]?.(request, output);
@@ -1164,12 +1207,22 @@ export function createProfiledPlugin(profile, assetVersion) {
                     `Shared MkII protocol role names are logical: ${protocolMap}. Use only ${profile.toolPrefix} tools for this profile. ` +
                     "Never rewrite user acceptance or evidence to rename protocol roles. Final acceptance belongs only to the root coordinator.");
                 const proposal = await proposals.read(root);
-                if (proposal?.phase === "investigating" && proposal.proposal_session_id === request.sessionID) {
+                if (proposal?.phase !== "approved" && proposal?.proposal_session_id === request.sessionID) {
+                    // Only immutable identity and the frozen budget caps belong here. Consumed counters move with
+                    // every accounted read, and a system element is an absolute prompt prefix: restating them here
+                    // invalidated the whole cached prefix on every later request of the same investigation, so the
+                    // Task prompt and all accumulated reads were re-billed uncached. They are reported on the read
+                    // result instead, which is appended after the stable prefix.
+                    //
+                    // The phase is deliberately not part of this condition. Removing an element is the same absolute
+                    // prefix change as rewriting one: gating on `investigating` dropped this block the moment a
+                    // submission succeeded, so the child's final turn re-sent the entire accumulated investigation
+                    // uncached. The terminal submit result is appended after this prefix and is more recent.
                     (output.system ??= []).push(`SORTIE_PROPOSAL_PHASE investigating; intent=${proposal.intent_id}; root=${root}; child=${request.sessionID}. ` +
                         `This durable phase remains authoritative after compaction even when the latest message is a generic continuation. ` +
                         `Continue the admitted read-only investigation and submit through ${submitProposal}. Do not call ${next} or dispatch workers: no execution run exists yet. ` +
-                        `actual_reads=${proposal.read_count}; remaining_reads=${proposal.intent.proposal_budget.max_reads - proposal.read_count}; ` +
-                        `submissions=${proposal.submission_count}; remaining_submissions=${proposal.intent.proposal_budget.max_submissions - proposal.submission_count}.`);
+                        `max_reads=${proposal.intent.proposal_budget.max_reads}; max_submissions=${proposal.intent.proposal_budget.max_submissions}. ` +
+                        `Consumed budget is reported as SORTIE_PROPOSAL_BUDGET on each read result and by ${submitProposal}; never infer it from this element.`);
                 }
             },
             "experimental.text.complete": async (request, output) => {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "sortie-dogs",
-  "version": "0.10.3",
+  "version": "0.10.5",
   "description": "Bounded agent harness and validated orchestration loop plugin for OpenCode",
   "keywords": [
     "opencode",