akm-cli 0.9.25-alpha.1 → 0.9.25-alpha.3

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/session-asset.js

@@ -99,6 +99,12 @@ export function buildSessionAccessInstructions(harness, logPath, sessionId) {
             `Parse messages: jq -r 'select(.type=="message") | .message.content[]? | select(.type=="text") | .text' ${logPath}`,
         ].join("\n");
     }
+    if (harness === "codex") {
+        return [
+            `Read with: cat ${logPath}`,
+            `Parse messages: jq -r 'select(.type=="response_item" and .payload.type=="message") | .payload.content[]? | .text // empty' ${logPath}`,
+        ].join("\n");
+    }
     if (harness === "opencode") {
         return [
             `Open the SQLite database at ${JSON.stringify(logPath)} in read-only mode.`,

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 };
 }
@@ -331,7 +333,11 @@ const redactedContentValidator = {
         ];
     },
 };
-/** Defense in depth when a reflect proposal bypasses creation-time sanitization (#963). */
+/**
+ * The run-only "Avoid These Patterns" section (#963) in a reflect proposal: one
+ * made before reflect kept the body, or a body that already had it. Reflect
+ * keeps a body as it is, so a person accepting has no way to remove it.
+ */
 const reflectPromptScaffoldingValidator = {
     name: "reflect-prompt-scaffolding",
     appliesTo(proposal) {
@@ -359,15 +365,15 @@ function advisory(validator) {
         validate: (proposal, ctx) => validator.validate(proposal, ctx).map((finding) => ({ ...finding, severity: "warn" })),
     };
 }
-/** The quality validators `validateProposal` runs; the last three protect durable content and block. */
+/** The quality validators `validateProposal` runs; the last two protect durable content and block. */
 export const defaultProposalQualityValidators = [
     ...[
         descriptionQualityValidator,
         lessonContentQualityValidator,
         sourceNotSupersededValidator,
         reflectSizeGuardValidator,
+        reflectPromptScaffoldingValidator,
     ].map(advisory),
     reflectTruncationMarkerValidator,
     redactedContentValidator,
-    reflectPromptScaffoldingValidator,
 ];

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

@@ -131,14 +131,13 @@ export const defaultProposalValidators = [
  * and was previously safe to run in full there because every quality
  * validator was advisory (`advisory()` downgrades findings to `severity:
  * "warn"`, which {@link runProposalValidators}'s `ok` never treats as
- * failing). Blocking durable-content validators (the #952 truncation marker,
- * #962 redaction marker, and #963 reflected prompt scaffolding) deliberately
+ * failing). Blocking durable-content validators (the #952 truncation marker
+ * and #962 redaction marker) deliberately
  * remain outside this subset, so running the full
  * {@link defaultProposalValidators} list at mint time could throw
  * `invalid_canonical_structure` for any lesson/task/workflow reflect
- * proposal whose body leaks the truncation marker — instead of letting
- * `sanitizeReflectPayload` mint the proposal and defer it with
- * `reflect-truncation-leak`, per the #952 design. Quality validators (prose
+ * proposal whose body carries the truncation marker — instead of minting it
+ * for a person to review. Quality validators (prose
  * shape, reflect size ratio, and durable-content guards) belong at
  * `proposal accept` / drain-promotion time, which already calls
  * {@link validateProposal} (the full list) via `preflightProposalPromotion`

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/asset/asset-serialize.js

@@ -95,7 +95,7 @@ export function assembleAsset(frontmatter, body) {
  *   - exactly one `\n` terminates the file
  *
  * This helper is the single point of truth for the fence-and-body template.
- * Three command surfaces (`reflect`, `distill`, `consolidate`) call it
+ * Two command surfaces (`distill`, `consolidate`) call it
  * directly because their inputs are pre-validated LLM payloads where the
  * full `yamlStringify` may emit shapes (`|`-block scalars, anchors) that
  * the project's hand-rolled `parseFrontmatter` subset parser cannot read.

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,

package/dist/core/config/schema/improve-processes.js

@@ -103,6 +103,21 @@ const fidelityCheckField = z.object({ enabled: z.boolean().optional() }).passthr
  * byte-identical behaviour). Reflect process only.
  */
 const lowValueFilterField = z.object({ enabled: z.boolean().optional() }).passthrough().optional();
+/**
+ * The wording lists of reflect's pre-judge defect filter (`findReflectDefect`).
+ * Each list is optional: one that is set replaces that rule's default list, and
+ * an empty one turns the rule off. Phrases (`placeholders`, `metaCommentary`)
+ * match as whole words in any case; `frontmatterKeys` are exact key names.
+ * Reflect process only.
+ */
+const defectFilterField = z
+    .object({
+    placeholders: z.array(nonEmptyString).optional(),
+    metaCommentary: z.array(nonEmptyString).optional(),
+    frontmatterKeys: z.array(nonEmptyString).optional(),
+})
+    .passthrough()
+    .optional();
 /**
  * #626 — extract process: pre-LLM heuristic triage gate. When enabled, a
  * deterministic scorer decides BEFORE the extraction LLM call whether a
@@ -159,6 +174,7 @@ const REFLECT_PROCESS_FIELDS = {
     limit: processLimitField,
     qualityGate: qualityGateField,
     lowValueFilter: lowValueFilterField,
+    defectFilter: defectFilterField,
 };
 const DISTILL_PROCESS_FIELDS = {
     allowedTypes: allowedTypesField,

package/dist/core/content-safety.js

@@ -6,36 +6,12 @@ export const REDACTED_CONTENT_MARKER = "[REDACTED]";
 /** Reflect prompt section that contains run diagnostics, not proposed asset content. */
 export const REFLECT_AVOID_PATTERNS_HEADING = "Avoid These Patterns";
 const REFLECT_AVOID_PATTERNS_RE = /^##[ \t]+Avoid These Patterns[ \t]*$/i;
-const SECTION_BOUNDARY_RE = /^#{1,2}(?:[ \t]+|$)/;
 export function containsRedactedContent(content) {
     return content.includes(REDACTED_CONTENT_MARKER);
 }
 export function containsReflectPromptScaffolding(content) {
     return content.split(/\r?\n/).some((line) => REFLECT_AVOID_PATTERNS_RE.test(line));
 }
-/**
- * Remove every echoed run-only "Avoid These Patterns" section while preserving
- * the next peer/top-level section. Reflect alone calls this sanitizer; authored
- * source assets are never rewritten by this helper.
- */
-export function stripReflectPromptScaffolding(content) {
-    const newline = content.includes("\r\n") ? "\r\n" : "\n";
-    const lines = content.split(/\r?\n/);
-    const kept = [];
-    let stripped = false;
-    for (let index = 0; index < lines.length;) {
-        if (!REFLECT_AVOID_PATTERNS_RE.test(lines[index] ?? "")) {
-            kept.push(lines[index] ?? "");
-            index += 1;
-            continue;
-        }
-        stripped = true;
-        index += 1;
-        while (index < lines.length && !SECTION_BOUNDARY_RE.test(lines[index] ?? ""))
-            index += 1;
-    }
-    return { content: kept.join(newline).replace(/(?:\r?\n){3,}/g, `${newline}${newline}`), stripped };
-}
 /**
  * Return a secret-free rejection reason when generated content is unsafe to
  * persist. `redactedContent` is the same body after applying the dispatch

package/dist/core/redaction.js

@@ -19,6 +19,10 @@ const ENV_PASSTHROUGH_REDACTION_POLICY = {
     OPENCODE_CONFIG: "path",
     CLAUDE_CONFIG: "path",
     CODEX_CONFIG: "path",
+    XDG_CONFIG_HOME: "path",
+    XDG_DATA_HOME: "path",
+    XDG_CACHE_HOME: "path",
+    XDG_STATE_HOME: "path",
     AWS_PROFILE: "identifier",
     AWS_REGION: "identifier",
     LLM_MODEL: "identifier",

package/dist/core/spawn-env.js

@@ -45,6 +45,31 @@ export const COMMON_SPAWN_ENV_PASSTHROUGH = [
     "TMPDIR",
     "AKM_EVENT_SOURCE",
 ];
+/**
+ * The XDG base-directory variables. opencode resolves its config, data, cache
+ * and state directories from them, so it must receive them: an akm that runs
+ * under a custom `XDG_CONFIG_HOME` otherwise spawns an opencode that reads
+ * `$HOME/.config/opencode` and misses the provider config its caller named.
+ *
+ * Deliberately NOT part of {@link COMMON_SPAWN_ENV_PASSTHROUGH}, which is every
+ * harness's baseline, the workflow exec unit's default allowlist (a documented
+ * list) and, through profile `envPassthrough`, frozen into workflow plans.
+ * codex, gemini and pi keep their own dotdirs under `$HOME`, and handing the
+ * names to a shell command would redirect the `git` and `gh` config it reads. A
+ * harness that reads them asks for them by name: the opencode profile's list,
+ * and the opencode-sdk server's allowlist (`opencodeSdkServerEnvironmentNames`).
+ *
+ * A name added to a profile's list changes the plans frozen after it (their
+ * bytes, so their `plan_hash`); a stored plan keeps the list it was frozen with
+ * and still resumes, because nothing gates on that hash. The SDK server's
+ * allowlist is not part of a plan, so it takes the names by code.
+ */
+export const XDG_BASE_DIR_ENV_PASSTHROUGH = [
+    "XDG_CONFIG_HOME",
+    "XDG_DATA_HOME",
+    "XDG_CACHE_HOME",
+    "XDG_STATE_HOME",
+];
 /**
  * The names Windows itself requires of ANY child, whatever the caller's
  * allowlist says. Applied at build time rather than added to

package/dist/core/structured.js

@@ -34,7 +34,7 @@ import { parseEmbeddedJsonResponse } from "./parse.js";
 export function withSchemaInstruction(prompt, schema) {
     return `${prompt}\n\nRespond with ONLY a JSON value matching this JSON Schema (no prose, no code fences):\n${JSON.stringify(schema)}`;
 }
-function defaultFeedback(failure) {
+export function defaultFeedback(failure) {
     if (failure.reason === "parse_error") {
         return "Your previous response contained no parseable JSON. Respond with ONLY a JSON value that matches the requested schema — no prose, no code fences.";
     }

package/dist/execution/source.js

@@ -6,19 +6,15 @@ export const EXECUTION_SOURCE_SCHEMA_VERSION = 1;
 /** Current internal adapter identifiers are lowercase kebab-case registry keys. */
 export const EXECUTION_ADAPTER_ID_PATTERN = /^[a-z][a-z0-9]*(?:-[a-z0-9]+)*$/;
 /**
- * The model-work tool policy: unattended model work (improve, the judges,
- * index passes, remember) may read, edit only inside the dispatch's own
- * scratch working directory, and run `akm search` and `akm show`. The stash
- * stays read-only to it. A transport grants what it can confine and refuses
- * the policy at build when it can confine nothing; an LLM has no tools.
+ * The model-work tool policy: unattended model work (improve, the judges, index
+ * passes, remember) may read, edit only inside the dispatch's own scratch
+ * working directory, and run `akm search` and `akm show`; the stash stays
+ * read-only to it. A transport grants what it can confine and refuses the policy
+ * at build when it can confine nothing; an LLM has no tools. A request carries it
+ * as `authorization.policy.id`, set only when its caller asks (`modelWork`): no
+ * `tools` value names it, so an asset's own `tools:` cannot.
  */
-export const MODEL_WORK_TOOLS = Object.freeze(["read", "edit", "akm search", "akm show"]);
-/** Whether a selection names the model-work tool policy. */
-export function isModelWorkTools(tools) {
-    return (Array.isArray(tools) &&
-        tools.length === MODEL_WORK_TOOLS.length &&
-        tools.every((tool, index) => tool === MODEL_WORK_TOOLS[index]));
-}
+export const MODEL_WORK_POLICY_ID = "model-work";
 function requireRecord(value, path) {
     if (value === null || typeof value !== "object" || Array.isArray(value)) {
         throw new TypeError(`${path} must be an object`);