akm-cli 0.9.25-alpha.1 → 0.9.25-alpha.2

package/dist/commands/improve/reflect-noise.js

@@ -9,8 +9,13 @@
  * cosmetic edit through costs a review, suppressing a real fix loses work — so
  * anything uncertain is `substantive`: code (fenced or indented) compares
  * verbatim, and headings, tables and breaks never absorb the next line.
+ *
+ * `findReflectDefect` is its counterpart for an edit that is not noise but a
+ * defect no judge needs to weigh: placeholder text, talk about the edit itself,
+ * or frontmatter copied into the body.
  */
 import { parse as yamlParse } from "yaml";
+import { parseFrontmatter } from "../../core/asset/frontmatter.js";
 /**
  * `noop` (identical up to trailing whitespace) and `cosmetic` (identical after
  * normalizing frontmatter as parsed YAML and prose as unwrapped text) never
@@ -36,6 +41,126 @@ export function classifyReflectChange(sourceContent, candidateContent) {
     }
     return "substantive";
 }
+/** Not TODO, TBD or FIXME: an asset may carry those on purpose, and the owner does not want them refused. */
+const DEFAULT_PLACEHOLDERS = [
+    "please confirm",
+    "please verify",
+    "to be confirmed",
+    "to be determined",
+    "to be verified",
+];
+const DEFAULT_META_COMMENTARY = [
+    "feedback signal",
+    "feedback signals",
+    "feedback indicate",
+    "feedback indicates",
+    "feedback suggest",
+    "feedback suggests",
+    "feedback ask",
+    "feedback asks",
+    "feedback says",
+    "feedback report",
+    "feedback reports",
+    "feedback request",
+    "feedback requests",
+    "this revision",
+    "the source asset",
+    "the source note",
+    "the source memory",
+    "the original asset",
+    "the original note",
+    "the original memory",
+    "the original version of this",
+    "quality gate rejected",
+    "proposal rejected",
+];
+const DEFAULT_FRONTMATTER_KEYS = [
+    "sources",
+    "updated",
+    "inferenceProcessed",
+    "captureMode",
+    "beliefState",
+    "xrefs",
+    "contradictedBy",
+    "outcomeData",
+    "orderedActions",
+    "generated",
+    "verified",
+    "description",
+    "when_to_use",
+    "tags",
+    "searchHints",
+    "quality",
+    "salience",
+    "salienceInputs",
+    "lint_skip",
+    "type",
+];
+/**
+ * The first defect the candidate has and its source lacks, or `undefined`. Each
+ * rule counts only what the revision adds, so text the asset already carried is
+ * not held against it. On 396 labelled reflect edits (83 good, 313 bad) the
+ * default lists hit 22 bad edits and no good one, so a hit is refused unjudged.
+ */
+export function findReflectDefect(sourceContent, candidateContent, filter = {}) {
+    if (gainsPhrases(filter.placeholders ?? DEFAULT_PLACEHOLDERS, sourceContent, candidateContent)) {
+        return "placeholder_added";
+    }
+    if (gainsPhrases(filter.metaCommentary ?? DEFAULT_META_COMMENTARY, sourceContent, candidateContent)) {
+        return "meta_commentary_added";
+    }
+    if (frontmatterCopiedIntoBody(sourceContent, candidateContent, filter.frontmatterKeys ?? DEFAULT_FRONTMATTER_KEYS)) {
+        return "frontmatter_copied_into_body";
+    }
+    return undefined;
+}
+/** Frontmatter fields that name other assets; one of their values newly in the body is provenance copied over. */
+const PROVENANCE_KEYS = ["sources", "xrefs", "contradictedBy"];
+/** Shorter values (a bare name or id) say too little to find in a body. */
+const PROVENANCE_VALUE_MIN_CHARS = 12;
+function escapeRegExp(text) {
+    return text.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
+/** Whether the candidate holds more of the phrases than the source: whole words, any case, any whitespace between words. */
+function gainsPhrases(phrases, source, candidate) {
+    const alternatives = phrases
+        .map((phrase) => phrase.trim().split(/\s+/).map(escapeRegExp).join("\\s+"))
+        .filter((alternative) => alternative !== "");
+    if (alternatives.length === 0)
+        return false;
+    const pattern = new RegExp(`(?<!\\w)(?:${alternatives.join("|")})(?!\\w)`, "gi");
+    return (candidate.match(pattern)?.length ?? 0) > (source.match(pattern)?.length ?? 0);
+}
+/** The body gains a line that starts with one of the frontmatter keys outside code, or a provenance value from either asset's frontmatter. */
+function frontmatterCopiedIntoBody(source, candidate, keys) {
+    if (keys.length === 0)
+        return false;
+    const keyLine = new RegExp(`^(?:${keys.map(escapeRegExp).join("|")}):(?:[ \\t].*)?$`);
+    const keyLines = (text) => parseLowValueSections(text).proseLines.filter((line) => keyLine.test(line)).length;
+    if (keyLines(candidate) > keyLines(source))
+        return true;
+    const sourceBody = lettersAndDigits(splitFrontmatter(source).body);
+    const candidateBody = lettersAndDigits(splitFrontmatter(candidate).body);
+    return [...provenanceValues(source), ...provenanceValues(candidate)].some((value) => {
+        const v = lettersAndDigits(value);
+        return v.length >= PROVENANCE_VALUE_MIN_CHARS && candidateBody.includes(v) && !sourceBody.includes(v);
+    });
+}
+/** Every string under the provenance keys of a frontmatter block: a scalar, or the items of a list. */
+function provenanceValues(content) {
+    const { data } = parseFrontmatter(content);
+    return PROVENANCE_KEYS.flatMap((key) => {
+        const value = data[key];
+        return (Array.isArray(value) ? value : [value]).filter((item) => typeof item === "string");
+    });
+}
+/** Lower-case letters and digits, each run of anything else a single space. */
+function lettersAndDigits(text) {
+    return text
+        .toLowerCase()
+        .replace(/[^a-z0-9]+/g, " ")
+        .trim();
+}
 /** A `---` frontmatter block and the rest (`fmText: null` when there is none). */
 export function splitFrontmatter(raw) {
     const m = raw.match(/^---\r?\n([\s\S]*?)\r?\n---\r?\n?([\s\S]*)$/);

package/dist/commands/improve/reflect.js

@@ -26,9 +26,8 @@ import { parseEmbeddedJsonResponse } from "../../core/parse.js";
 import { redactSensitiveText } from "../../core/redaction.js";
 import { resolveStandardsContext } from "../../core/standards/resolve-standards-context.js";
 import { warn, warnOnce } from "../../core/warn.js";
-import { MODEL_WORK_TOOLS } from "../../execution/source.js";
 import { lookup } from "../../indexer/indexer.js";
-import { DEFAULT_MODEL_WORK_TIMEOUT_MS } from "../../integrations/agent/config.js";
+import { DEFAULT_LLM_TIMEOUT_MS } from "../../integrations/agent/config.js";
 import { fallbackAnnouncement, NO_ENGINE_MESSAGE_SUFFIX, NO_ENGINE_REMEDY, withEngineFallback, } from "../../integrations/agent/engine-fallback.js";
 import { buildExecution, resolveExecution } from "../../integrations/agent/execution.js";
 import { buildReflectOutputRepairPrompt, buildReflectPrompt, parseAgentProposalPayload, REFLECT_CONTENT_CAP, REFLECT_TRUNCATION_MARKER, } from "../../integrations/agent/prompts.js";
@@ -42,7 +41,7 @@ import { deriveLessonRef } from "./distill.js";
 import { findAssetFilePath } from "./eligibility.js";
 import { resolveImproveExecution } from "./execution.js";
 import { recordLedgerAttempt } from "./ledger.js";
-import { classifyReflectChange, splitFrontmatter } from "./reflect-noise.js";
+import { classifyReflectChange, findReflectDefect, splitFrontmatter } from "./reflect-noise.js";
 import { loadRetrievalQueries, runRetrievalRegressionGate } from "./retrieval-gate.js";
 import { callStageOnce, errMessage, mintProposal, noticeSet, rejectedProposalContext, resolveQualityGateJudge, runReflectQualityJudge, } from "./stage.js";
 const MAX_FEEDBACK_LINES = 10;
@@ -494,7 +493,7 @@ export async function runReflectIteration(opts) {
         ? (opts.timeoutMs ?? null)
         : Object.hasOwn(opts.runner, "timeoutMs")
             ? (opts.runner.timeoutMs ?? null)
-            : DEFAULT_MODEL_WORK_TIMEOUT_MS;
+            : DEFAULT_LLM_TIMEOUT_MS;
     const deadline = typeof configuredTimeout === "number" ? start + configuredTimeout : undefined;
     const messages = [{ role: "user", content: opts.prompt ?? "" }];
     if (opts.priorDraft !== undefined && opts.iteration > 0) {
@@ -516,16 +515,8 @@ export async function runReflectIteration(opts) {
             parsed: { outputMode: opts.outputMode, repairAttempts },
         };
     };
-    // A reply that broke the contract. An agent or SDK engine's failure names the engine; an LLM
-    // reply keeps its message, because improve feeds a failed reflect's `error` into the next
-    // prompts as a pattern to avoid, and rewording it would change those requests.
-    const invalidReply = (err, reply) => {
-        if (runnerIsLlm(opts.runner))
-            return failure(err, "parse_error", reply, 0);
-        const attempts = repairAttempts + 1;
-        const message = `Engine "${opts.runner.engine}" reply was not a valid reflect proposal after ${attempts} attempt${attempts === 1 ? "" : "s"}: ${errMessage(err)}`;
-        return failure(new Error(message), "parse_error", reply, 0);
-    };
+    // A reply that broke the contract; its message is the parser's, on every engine kind.
+    const invalidReply = (err, reply) => failure(err, "parse_error", reply, 0);
     // The result of the dispatch that failed, kept for an agent or SDK engine.
     let dispatched;
     // Reflect parses and repairs its own reply (the repair turn below), so one dispatch, unvalidated.
@@ -756,7 +747,7 @@ function preflightReflectDispatch(runnerSpec, onNotices) {
     const prepared = resolveExecution({
         content: "Validate reflect operation transport before dispatch.",
         runner: runnerSpec,
-        current: { tools: MODEL_WORK_TOOLS },
+        modelWork: true,
     });
     const lowered = buildExecution(prepared.request, prepared.runner);
     onNotices(lowered.notices);
@@ -907,7 +898,8 @@ const NOISE_SUBREASONS = {
  * exact content that would be persisted, then mint. A judge pass is staged only
  * when the body is unchanged: a body edit the judge passes, or one made with the
  * gate off, waits for review. Size-flagged or truncation-leaking content skips
- * the judge and waits for review.
+ * the judge and waits for review. A revision with a deterministic defect is
+ * refused before the judge runs, whether or not the gate is on.
  */
 async function finalizeReflectProposal(args) {
     const { run, assetContent, result, judge, feedback } = args;
@@ -953,11 +945,16 @@ async function finalizeReflectProposal(args) {
         }, options.eventsCtx);
         return reflectFailure(run, result, "quality_rejected", message, false);
     };
+    // A defect no judge needs to weigh is refused before any judge call, whether or not the gate is on.
+    const defect = assetContent === undefined ? undefined : findReflectDefect(assetContent, payload.content, options.defectFilter);
+    if (defect)
+        return refuse(defect, { reflectDefect: defect }, `Reflect proposal refused before the judge: ${defect}`);
     let verdict;
     let judgeFailed = false;
     if (judged) {
         verdict = await runReflectQualityJudge(run.config, payload.content, assetContent ?? "", feedback, options.chat, {
             runnerSelectionFrozen: true,
+            ref: payload.ref,
             ...(judge.runner ? { llmRunner: judge.runner } : {}),
             ...(Object.hasOwn(options, "timeoutMs") ? { timeoutMs: options.timeoutMs } : {}),
             ...(options.signal ? { signal: options.signal } : {}),

package/dist/commands/improve/retrieval-gate.js

@@ -32,6 +32,10 @@ const GRADE_SCHEMA = {
     additionalProperties: false,
     properties: { grade: { type: "integer", minimum: 0, maximum: 3 }, reason: { type: "string" } },
 };
+function parseGrade(raw) {
+    const grade = parseEmbeddedJsonResponse(raw)?.grade;
+    return typeof grade === "number" && Number.isInteger(grade) && grade >= 0 && grade <= 3 ? grade : undefined;
+}
 /** Up to five distinct task queries, in the given order, whitespace collapsed. */
 export function usableRetrievalQueries(raw) {
     const out = [];
@@ -100,10 +104,11 @@ export async function runRetrievalRegressionGate(args) {
                     ...(args.signal ? { signal: args.signal } : {}),
                     ...(args.chat ? { chat: args.chat } : {}),
                 },
+                parse: parseGrade,
                 ...(args.onNotices ? { onNotices: args.onNotices } : {}),
             });
-            const grade = outcome.ok ? parseEmbeddedJsonResponse(outcome.raw)?.grade : undefined;
-            if (typeof grade !== "number" || !Number.isInteger(grade) || grade < 0 || grade > 3) {
+            const grade = outcome.ok ? parseGrade(outcome.raw) : undefined;
+            if (grade === undefined) {
                 return {
                     pass: false,
                     queries: args.queries.length,

package/dist/commands/improve/stage.js

@@ -3,9 +3,8 @@
 // file, You can obtain one at https://mozilla.org/MPL/2.0/.
 import { getImproveProcessConfig } from "../../core/config/config.js";
 import { ConfigError } from "../../core/errors.js";
-import { validateJsonSchemaSubset } from "../../core/json-schema.js";
 import { parseEmbeddedJsonResponse } from "../../core/parse.js";
-import { runStructured } from "../../core/structured.js";
+import { defaultFeedback } from "../../core/structured.js";
 import { warn } from "../../core/warn.js";
 import { runnerLlmConnection } from "../../integrations/agent/runner.js";
 import { callStructured, dispatchFailureReason, dispatchFailureResult, } from "../../llm/structured-call.js";
@@ -45,43 +44,23 @@ export function stageRunner(frozen, config, profile, processName, onNotices) {
         onNotices?.(resolved.notices);
     return resolved?.runner;
 }
-const TRANSPORT_FAILED = Symbol("stage-transport-failed");
 /**
  * One model call. Provider trouble (transport error, timeout, abort, a
  * disabled feature) comes back as `{ ok: false }`; only a configuration
- * failure throws. A reply to a call with `request.responseSchema` that fails
- * the schema gets one corrective retry. The caller's own parse still decides
- * what it accepts, so the last reply comes back even when it fails the schema.
+ * failure throws. A reply to a call with `request.responseSchema` that the
+ * stage's own `parse` rejects gets one corrective retry; the last reply comes
+ * back either way, and the caller parses it again.
  */
 export async function callStage(call) {
-    const schema = call.request?.responseSchema;
-    if (!schema)
-        return callStageOnce(call);
-    let reply = undefined;
-    let failure = undefined;
-    try {
-        await runStructured({
-            dispatch: async (feedback) => {
-                const outcome = await callStageOnce(feedback ? { ...call, prompt: `${call.prompt}\n\n${feedback}` } : call);
-                if (!outcome.ok) {
-                    failure = outcome;
-                    throw TRANSPORT_FAILED;
-                }
-                reply = outcome;
-                return outcome.raw;
-            },
-            validate: (candidate) => {
-                const errors = validateJsonSchemaSubset(candidate, schema);
-                return errors.length === 0 ? { ok: true, value: candidate } : { ok: false, errors };
-            },
-        });
-    }
-    catch (err) {
-        if (err !== TRANSPORT_FAILED)
-            throw err;
-    }
+    const reply = await callStageOnce(call);
+    if (!reply.ok || !call.request?.responseSchema)
+        return reply;
+    if ((call.parse ?? parseEmbeddedJsonResponse)(reply.raw) !== undefined)
+        return reply;
+    const feedback = defaultFeedback({ reason: "parse_error", errors: [] });
+    const retry = await callStageOnce({ ...call, prompt: `${call.prompt}\n\n${feedback}` });
     // A retry that fails in transport keeps the first reply, which the caller may still accept.
-    return reply ?? failure ?? { ok: false, reason: "error" };
+    return retry.ok ? retry : reply;
 }
 /** Timeout and abort come from the dispatch's own reason, whatever the runner's kind. */
 function failureReason(err) {
@@ -283,15 +262,24 @@ function buildChangedRegion(sourceContent, candidateContent) {
     const added = candidate.slice(prefix, candidate.length - suffix).join("\n");
     return boundedDocument(`Removed or replaced:\n${removed || "(none)"}\n\nAdded or replacement:\n${added || "(none)"}`);
 }
-/** Judge prompt for an in-place revision. */
-export function buildReflectJudgePrompt(candidateContent, sourceContent, feedback) {
+/**
+ * What the judge may do with tools when it runs on an agent engine: verify a
+ * fact the revision adds or alters, and nothing else. The plain judge's prompt
+ * is unchanged (its rubric is tuned and measured without this paragraph).
+ */
+function reflectJudgeToolRules(ref) {
+    const asset = ref ? `The asset is \`${ref}\`: read it with akm_show, ` : "Read an asset with akm_show ";
+    return `Tools: ${asset}or an asset the changed region names, only to verify a fact the revision adds or alters; the text above already shows every change. Do not search, do not read anything else, and do not use a tool to judge structure or wording. One or two reads at most. Before scoring, check three lists: (1) every statement the revision adds: find each in the asset, or as a fact the feedback states about the subject, and score QUALITY 1-2 if any is in neither; a statement is found only when the asset or the feedback says it, in any words: a new step, cause, consequence or detail that merely seems to follow is not found; feedback says what to fix and is not content, so an added statement about how the asset was used, found or verified is unsupported; (2) every fact, caveat and field of the source: find each in the revision, and score PRESERVATION 1-3 if any is missing; (3) every point the feedback makes: find the text it is about changed in the revision, and score NEED 2-3 if any is not; a note that restates the feedback does not address it. A read that finds nothing wrong raises no score above what these lists support. Then reply with the JSON.`;
+}
+/** Judge prompt for an in-place revision. `tools` is set when the judge runs on an agent engine. */
+export function buildReflectJudgePrompt(candidateContent, sourceContent, feedback, tools) {
     return [
         "You are evaluating a proposed revision to an existing akm asset.",
         "",
         "Score this revision on each criterion from 1 (poor) to 5 (excellent):",
-        "1. NEED: Does the revision fix a concrete problem in the source? Concrete problems are: something the feedback reports as wrong or missing; a factual error; or broken, garbled, truncated or missing text, including frontmatter fields such as description or when_to_use. Score 4-5 when it fixes one, even a small one. Score 1-2 when the source was already correct and the revision only rewords, restates, reformats, or adds headings, an introduction or a table of contents.",
-        "2. PRESERVATION: Does it keep every concrete fact, identifier, command, path, number and example from the source, without truncation?",
-        "3. QUALITY: Is it coherent and accurate, with no claims, steps or details that the source or the feedback does not support?",
+        "1. NEED: Does the revision fix a concrete problem in the source? Concrete problems are: something the feedback reports as wrong or missing; a factual error; broken, garbled, truncated or missing text; and a missing or broken title, description or when_to_use field. Compare the source's description with the revision's: a description with a sentence split in its middle by a stray period or line-wrap artifact (as in 'calls. asset writes'), an unbalanced or escaped quote, or a truncated ending is broken, and repairing it is a concrete problem fixed even when the rest of the revision only adds stamps or reformats; adding or removing a trailing period, or rewording a readable description, repairs nothing. A missing title, description or when_to_use is a concrete problem whether or not the feedback mentions it: empty or positive feedback does not mean the source was complete, and the frontmatter is complete only when it has all three. A when_to_use is missing when the frontmatter has none, even if the body has a 'when to use' section; moving or copying that text into the field is the fix. A missing type field or a provenance stamp such as generated or verified is not a concrete problem. Score 4-5 when the revision fixes one, even a small one, whatever else it also reformats. Score 1-2 when the source was already complete and correct and the revision only rewords, restates, reformats, adds a type field or a stamp, or adds headings, an introduction or a table of contents.",
+        "2. PRESERVATION: Does it keep every concrete fact, identifier, command, path, number, example, caveat and frontmatter field from the source, without truncation? Check the changed region line by line. Score 1-3 when any of them is dropped or weakened, even when the revision also fixes something. Trimming narrative that states no fact, or removing what the feedback asks to remove or rescope, is not a drop.",
+        "3. QUALITY: Is it coherent and accurate, with no claims, steps or details that the source or the feedback does not support? Check each added or changed statement against the source and the feedback: a statement that follows from either counts as supported, and frontmatter stamps such as type, generated, verified or quality, whatever their values, and a restatement of existing content are not claims. Score 1-2 when the revision adds a claim neither supports, turns a draft or proposal into a decision, strengthens a statement beyond the source (a preference into a requirement, a possibility into a fact, a pending fix into a done one), adds a hedge such as 'may be outdated', or a placeholder such as 'TODO' or 'verify'; a fix elsewhere in the revision does not raise this score.",
         "",
         "Feedback:",
         "```",
@@ -313,6 +301,7 @@ export function buildReflectJudgePrompt(candidateContent, sourceContent, feedbac
         buildChangedRegion(sourceContent, candidateContent),
         "```",
         "",
+        ...(tools ? [reflectJudgeToolRules(tools.ref), ""] : []),
         'Return ONLY valid JSON, no prose: {"scores": {"need": <1-5 integer>, "preservation": <1-5 integer>, "quality": <1-5 integer>}, "reason": "<one sentence>"}',
     ].join("\n");
 }
@@ -421,6 +410,7 @@ async function runQualityJudge(feature, config, prompt, keys, chat, options) {
             ...(options.signal ? { signal: options.signal } : {}),
             ...(chat ? { chat } : {}),
         },
+        parse: (raw) => parseJudgeResponse(raw, keys),
         ...(options.onNotices ? { onNotices: options.onNotices } : {}),
     });
     if (!outcome.ok) {
@@ -462,6 +452,8 @@ export function runLessonQualityJudge(config, lessonContent, sourceContent, chat
 }
 /** Judge an in-place reflect revision without new-lesson novelty criteria. */
 export function runReflectQualityJudge(config, candidateContent, sourceContent, feedback, chat, options = {}) {
-    const prompt = buildReflectJudgePrompt(candidateContent, sourceContent, feedback);
+    // A judge on an agent engine gets the tool rules; the runner is the frozen one or none.
+    const tools = options.llmRunner && options.llmRunner.kind !== "llm" ? { ref: options.ref } : undefined;
+    const prompt = buildReflectJudgePrompt(candidateContent, sourceContent, feedback, tools);
     return runQualityJudge("proposal_quality_gate", config, prompt, REFLECT_JUDGE_CRITERIA, chat, options);
 }

package/dist/commands/proposal/drain.js

@@ -25,8 +25,7 @@ import { ConfigError } from "../../core/errors.js";
 import { appendEvent } from "../../core/events.js";
 import { escapeJsonStringControls, stripCodeFences, stripThinkBlocks } from "../../core/parse.js";
 import { info, warn } from "../../core/warn.js";
-import { MODEL_WORK_TOOLS } from "../../execution/source.js";
-import { DEFAULT_MODEL_WORK_TIMEOUT_MS } from "../../integrations/agent/config.js";
+import { DEFAULT_LLM_TIMEOUT_MS } from "../../integrations/agent/config.js";
 import { buildExecution, resolveExecution } from "../../integrations/agent/execution.js";
 import { assertRunnerCredentials, runExecution, } from "../../integrations/agent/runner-dispatch.js";
 import { errMessage, noticeSet } from "../improve/stage.js";
@@ -177,10 +176,8 @@ async function dispatchJudgment(runner, prompt, seams) {
         const prepared = resolveExecution({
             content: prompt,
             runner,
-            current: {
-                ...(Object.hasOwn(runner, "timeoutMs") ? {} : { timeout: DEFAULT_MODEL_WORK_TIMEOUT_MS }),
-                tools: MODEL_WORK_TOOLS,
-            },
+            current: Object.hasOwn(runner, "timeoutMs") ? {} : { timeout: DEFAULT_LLM_TIMEOUT_MS },
+            modelWork: true,
         });
         const lowered = buildExecution(prepared.request, prepared.runner);
         notices = lowered.notices;
@@ -345,7 +342,7 @@ export async function drainProposals(opts, promoteFn = akmProposalAccept, reject
         const prepared = resolveExecution({
             content: "Validate the selected proposal judgment runner before mutation.",
             runner: opts.judgment,
-            current: { tools: MODEL_WORK_TOOLS },
+            modelWork: true,
         });
         assertRunnerCredentials(buildExecution(prepared.request, prepared.runner).runner);
     }

package/dist/commands/proposal/propose.js

@@ -27,8 +27,7 @@ import { deriveEntryProvenance } from "../../indexer/installations.js";
 import { fallbackAnnouncement } from "../../integrations/agent/engine-fallback.js";
 import { buildExecution, resolveExecution } from "../../integrations/agent/execution.js";
 import { buildProposePrompt, PROPOSAL_JSON_SCHEMA, validateProposalPayload, } from "../../integrations/agent/prompts.js";
-import { assertRunnerCredentials, collectDispatchSensitiveValues, runExecution, } from "../../integrations/agent/runner-dispatch.js";
-import { getHarness } from "../../integrations/harnesses/index.js";
+import { assertRunnerCredentials, collectDispatchSensitiveValues, runExecution, unwrapHarnessReply, } from "../../integrations/agent/runner-dispatch.js";
 import { baseFailureFields, enoentHintMessage, isEnoentFailure } from "../agent/agent-support.js";
 import { createProposal, resolveProposalQueueTarget, } from "./repository.js";
 function failureEnvelope(result, type, name, engine, notices, fallbackReason = "non_zero_exit") {
@@ -45,14 +44,6 @@ function noticeFields(notices) {
     return notices.length > 0 ? { notices } : {};
 }
 const DISPATCH_FAILED = Symbol("proposal-dispatch-failed");
-/** A reply's text, unwrapped from its harness's framing (claude's `--output-format json` envelope). */
-function replyText(execution, result) {
-    const runner = execution.runner;
-    if (runner.kind !== "agent")
-        return result.stdout;
-    const extractor = getHarness(runner.profile.platform ?? runner.profile.name)?.resultExtractor;
-    return extractor ? extractor(result).text : result.stdout;
-}
 /**
  * Resolve, lower, and dispatch the already-rendered proposal prompt with the
  * proposal's JSON Schema as its output schema, and capture the reply. A reply
@@ -92,7 +83,7 @@ async function dispatchProposalPrompt(prompt, config, options, onDispatchReady)
                 results.push(result);
                 if (!result.ok)
                     throw DISPATCH_FAILED;
-                return replyText(execution, result);
+                return unwrapHarnessReply(execution.runner, result).text;
             },
             validate: validateProposalPayload,
         });

package/dist/commands/proposal/validators/proposal-quality-validators.js

@@ -9,7 +9,8 @@
  * growing past min(max(250% of the source, 2500 bytes), 25000 bytes) suggests
  * speculation. The absolute bounds keep small assets (a p25 source is ~780
  * bytes) from tripping on one good paragraph, and 25000 (below p99) still
- * catches runaway expansion. Sources under 200 bytes are too noisy to judge.
+ * catches runaway expansion. Sources under 200 bytes are too noisy to judge. A
+ * body that does not grow is never expansion, however long its source.
  */
 import { parseFrontmatter } from "../../../core/asset/frontmatter.js";
 import { parseRefInput } from "../../../core/asset/resolve-ref.js";
@@ -159,7 +160,8 @@ export function checkReflectSize(sourceBody, proposedBody) {
         return { ok: false, code: "EXCESSIVE_SHRINKAGE", ratio };
     }
     const expandCeiling = Math.min(Math.max(REFLECT_EXPAND_RATIO_MAX * sourceLen, REFLECT_ABSOLUTE_CEILING_BYTES), REFLECT_ABSOLUTE_MAX_BYTES);
-    if (proposedLen > expandCeiling)
+    // A body that does not grow is never expansion, even where the cap sits below the source's own length.
+    if (proposedLen > sourceLen && proposedLen > expandCeiling)
         return { ok: false, code: "EXCESSIVE_EXPANSION", ratio };
     return { ok: true };
 }

package/dist/commands/read/search-cli.js

@@ -81,20 +81,6 @@ export const searchCommand = defineJsonCommand({
             description: "Include session assets (excluded from default search results via config.search.defaultExcludeTypes).",
             default: false,
         },
-        // Declared as the POSITIVE name with `default: true` so citty's native
-        // `--no-<name>` negation (it strips a leading `--no-` from ANY token and
-        // negates the remainder BEFORE consulting the declared-args table — see
-        // node_modules/citty/dist/index.mjs) does the work, the same pattern
-        // `sync --push/--no-push` uses. A flag DECLARED as `no-track-usage` could
-        // never be negated: `--no-track-usage` parses as "negate `track-usage`",
-        // a name nothing declared, leaving the real key at its default forever
-        // (F1/A1).
-        "track-usage": {
-            type: "boolean",
-            default: true,
-            description: "A successful search records usage-events telemetry. Default: on. Use --no-track-usage to run a " +
-                "search that records nothing.",
-        },
     },
     async run({ args }) {
         rejectRetiredSourceFlag();
@@ -108,7 +94,6 @@ export const searchCommand = defineJsonCommand({
         const filters = parseScopeFilterFlags(filterTokens, "--filter");
         const includeProposed = args["include-proposed"] === true;
         const belief = parseBeliefFilterMode(typeof args.belief === "string" ? args.belief : undefined);
-        const skipLogging = args["track-usage"] === false;
         const includeSessions = args["include-sessions"];
         const assets = args.assets === true;
         const outputMode = getOutputMode();
@@ -121,7 +106,6 @@ export const searchCommand = defineJsonCommand({
             includeProposed,
             belief,
             includeSessions,
-            skipLogging,
             assets,
             eventSource: resolveUsageEventSource(),
             attributionProjection: outputMode.shape === "agent" ? "agent" : outputMode.detail,
@@ -154,15 +138,6 @@ export const curateCommand = defineJsonCommand({
                 "with a workflow asset's own `budget` field (a run-cost cap) — this is a context-size target for this " +
                 "one curate call.",
         },
-        // Declared as the POSITIVE name with `default: true` — see the
-        // `track-usage` comment on `searchCommand` above for why a flag NAME
-        // must never start with `no-`.
-        "track-usage": {
-            type: "boolean",
-            default: true,
-            description: "A successful curate records usage-events telemetry for the curated items. Default: on. Use " +
-                "--no-track-usage to run a curate that records nothing.",
-        },
     },
     async run({ args }) {
         rejectRetiredSourceFlag();
@@ -173,7 +148,6 @@ export const curateCommand = defineJsonCommand({
         const limitParsed = parsePositiveIntFlag(args.limit ?? undefined);
         const limit = limitParsed && limitParsed > 0 ? limitParsed : 4;
         const source = parseSearchSource(args.from ?? "local");
-        const skipLogging = args["track-usage"] === false;
         const outputMode = getOutputMode();
         const packBudget = parsePositiveIntFlag(args.pack ?? undefined, "--pack");
         const curated = await akmCurate({
@@ -181,7 +155,6 @@ export const curateCommand = defineJsonCommand({
             type,
             limit,
             source,
-            skipLogging,
             eventSource: resolveUsageEventSource(),
             attributionProjection: outputMode.shape === "agent" ? "agent" : outputMode.detail,
         });
@@ -285,15 +258,6 @@ export const showCommand = defineJsonCommand({
             type: "string",
             description: "Exact context budget in characters. Requires --context lead; mutually exclusive with --max-tokens.",
         },
-        // Declared as the POSITIVE name with `default: true` — see the
-        // `track-usage` comment on `searchCommand` above for why a flag NAME
-        // must never start with `no-`.
-        "track-usage": {
-            type: "boolean",
-            default: true,
-            description: "A successful show records usage-events telemetry, including the search-selection linkage when this " +
-                "show follows a recent search. Default: on. Use --no-track-usage to run a show that records nothing.",
-        },
     },
     async run({ args }) {
         // `[origin//]meta[:name]` targets the stash `.meta/` convention, which is
@@ -342,14 +306,12 @@ export const showCommand = defineJsonCommand({
         if (maxContextChars !== undefined && !Number.isSafeInteger(maxContextChars)) {
             throw new UsageError("Fragment context budget is too large.", "INVALID_FLAG_VALUE");
         }
-        const skipLogging = args["track-usage"] === false;
         const result = await akmShowUnified({
             ref: args.ref,
             detail: showDetail,
             contextMode,
             maxContextChars,
             scope,
-            skipLogging,
             eventSource: resolveUsageEventSource(),
         });
         output("show", result);

package/dist/core/config/schema/engines.js

@@ -13,7 +13,7 @@ import { z } from "zod";
 // `config-types`, which type-derives from this barrel via
 // `typeof import("./config-schema")` — routing through config-types would mint
 // a config-schema ↔ config-types type cycle that collapses inference.
-import { HARNESS_AGENT_DISPATCH_IDS, harnessInferenceKeys, VALID_HARNESS_IDS, } from "../../../integrations/harnesses/ids.js";
+import { HARNESS_AGENT_DISPATCH_IDS, VALID_HARNESS_IDS } from "../../../integrations/harnesses/ids.js";
 import { WORKFLOW_MAX_TIMEOUT_MS } from "../../../workflows/resource-limits.js";
 import { chatCompletionsEndpoint, ExtraParamsSchema, engineName, nonEmptyString, positiveInt, symbolicOrWarnApiKey, } from "./primitives.js";
 /**
@@ -105,20 +105,6 @@ const LlmEngineSchema = z
         });
     }
 });
-/**
- * The inference fields an agent engine may set: the ones its platform
- * translates (`harnesses/ids.ts`). An asset's or a caller's inference reaches
- * every engine and reports what the engine does not translate as a lowering
- * notice; an engine the operator configures for a platform names only what
- * that platform carries, so the rest is an error here.
- */
-const AGENT_INFERENCE_KEYS = [
-    "temperature",
-    "maxTokens",
-    "contextLength",
-    "enableThinking",
-    "reasoningEffort",
-];
 const AgentEngineSchema = z
     .object({
     kind: z.literal("agent"),
@@ -131,30 +117,26 @@ const AgentEngineSchema = z
     model: nonEmptyString.optional(),
     timeoutMs: timeoutMsField,
     llmEngine: engineName.optional(),
-    temperature: z.number().finite().optional(),
-    maxTokens: positiveInt.optional(),
-    contextLength: positiveInt.optional(),
-    enableThinking: z.boolean().optional(),
-    reasoningEffort: nonEmptyString.optional(),
 })
     .passthrough()
     .superRefine((value, ctx) => {
-    for (const key of ["provider", "endpoint", "apiKey", "apiKeyFile", "concurrency", "extraParams", "modelAliases"]) {
+    for (const key of [
+        "provider",
+        "endpoint",
+        "apiKey",
+        "apiKeyFile",
+        "temperature",
+        "maxTokens",
+        "concurrency",
+        "extraParams",
+        "contextLength",
+        "enableThinking",
+        "reasoningEffort",
+        "modelAliases",
+    ]) {
         if (key in value)
             ctx.addIssue({ code: z.ZodIssueCode.custom, path: [key], message: `${key} is not valid on an agent engine` });
     }
-    const translated = harnessInferenceKeys(value.platform);
-    for (const key of AGENT_INFERENCE_KEYS) {
-        if (key in value && !translated.includes(key)) {
-            ctx.addIssue({
-                code: z.ZodIssueCode.custom,
-                path: [key],
-                message: `${key} is not valid on a ${value.platform} engine: ${translated.length > 0
-                    ? `the platform translates only ${translated.join(", ")}`
-                    : "the platform translates no inference fields"}`,
-            });
-        }
-    }
     if (value.platform !== "opencode-sdk" && value.llmEngine !== undefined) {
         ctx.addIssue({
             code: z.ZodIssueCode.custom,