agentfootprint 9.85.0 → 9.86.0

package/dist/esm/core/agent/stages/toolCalls.js

@@ -49,9 +49,11 @@ import { typedEmit } from '../../../recorders/core/typedEmit.js';
 import { extractSequence } from '../../../security/extractSequence.js';
 import { skillTarget } from '../../../security/skillTarget.js';
 import { menuOutstanding } from '../../../lib/injection-engine/routingPolicy.js';
+import { classifySkillTarget, } from '../../../lib/injection-engine/skillGraph.js';
 import { skillActivationConfirmation } from '../../../lib/injection-engine/skillToolDescriptors.js';
 import { selfCallNotice, selfSkillTools } from '../selfCallNotice.js';
 import { parkedMemberIds } from '../../../maps/engagement/types.js';
+import { spoken } from '../../../lib/spokenIds.js';
 import { unconfiguredCredentialProvider } from '../../../identity/types.js';
 import { bindArtifacts, unconfiguredArtifacts, } from '../../../artifacts/capability.js';
 import { resolveToolWants, wantsNeedsStoreRefusal } from '../../../artifacts/wants.js';
@@ -130,60 +132,208 @@ function reportingCredentials(provider, scope, toolName) {
     };
 }
 /**
- * The re-prompt a refused `read_skill` gets back. It is the model's only feedback,
- * so it names what IS allowed rather than only what isn't.
+ * Is this `read_skill` target the cursor's own skill AND is that cursor
+ * MOUNTED — the one owner of "this call would change nothing" (9.86.0).
  *
- * Three shapes, because "not reachable" has three different reasons and only the
- * first two share a fix:
- *   • something is reachable → name it, and ask for one of those;
- *   • a decision `tree()` → nothing is EVER reachable by `read_skill`, because a
- *     tree has no cursor to move (8.5.0). Saying "no skills are reachable from
- *     here" would invite the model to try again from somewhere else; there is no
- *     "elsewhere", so the message explains the tree instead;
- *   • a flat graph that happens to be at a dead end → the original message.
+ * ── WHY POSITION ALONE IS NOT THE ANSWER ────────────────────────────────
+ *
+ * `classifySkillTarget` answers a question about POSITION: is the target where
+ * the cursor stands? For most of the loop that settles it — the cursor's body
+ * is in that call's system prompt and its tools are on that call's wire, so
+ * asking to read it activates nothing, moves nothing and reveals nothing the
+ * request did not already carry.
+ *
+ * A PARKED map breaks that. Parking suppresses a map's contribution and leaves
+ * the cursor exactly where it was, so at a parked cursor the very same id is a
+ * RE-ENGAGEMENT: the gate admits it, the map engages, and the member's body and
+ * tools ride again on the next pass. That call moves something, which makes it
+ * a capability, which makes it the policy's question to answer.
+ *
+ * The dispatch gate has known this since 9.59.0 — it tests re-engagement BEFORE
+ * its self-call arm, which is this same fact written as two ordered arms. The
+ * `skill_read` permission gate, 480 lines upstream, has no arms to order: it
+ * asks one question and either spends the policy check or skips it, and it was
+ * asking `classifySkillTarget` alone. So a role whose checker hid a skill could
+ * have the model un-park and re-activate it. This function is the question that
+ * gate now asks, over the parked set `parkedNow` hands both of them.
+ *
+ * @example
+ * ```ts
+ * atMountedCursor({ cursor: 'audit', target: 'audit' });                          // true  — a stay
+ * atMountedCursor({ cursor: 'audit', target: 'audit', parked: new Set(['audit']) }); // false — a re-engagement
+ * atMountedCursor({ cursor: 'audit', target: 'billing' });                        // false — a move
+ * ```
  */
-function skillRefusal(requestedId, allowed, isTree) {
-    const head = `read_skill("${requestedId}") is not reachable from here. `;
-    if (allowed.length > 0) {
-        return `${head}Reachable skills: ${allowed.join(', ')}. Pick one of these, or finish.`;
-    }
-    if (isTree) {
-        return (`read_skill("${requestedId}") cannot move a decision tree. A tree routes by ` +
-            'predicate on every iteration — it has no cursor to jump, so this skill would ' +
-            'not activate even though the tool accepted the name. Answer with the skill the ' +
-            'tree routed to, or finish.');
+export function atMountedCursor(args) {
+    const stay = classifySkillTarget({
+        ...(args.cursor !== undefined && { cursor: args.cursor }),
+        target: args.target,
+    }) === 'self';
+    return stay && args.parked?.has(args.target) !== true;
+}
+/**
+ * What a call to a name nothing can dispatch gets back (9.86.0).
+ *
+ * `Unknown tool: X` was the whole sentence, on both dispatch doors, while
+ * `mcpServe` — the same library, one directory away — had been answering
+ * "Unknown tool 'rm_rf'. Served tools: echo, delete_account." for releases. The
+ * model that mistyped a name, or named one from a restored transcript, was told
+ * only that it was wrong, never what would have worked, and the cheapest repair
+ * (read the roster) was the one it could not make.
+ *
+ * The roster is the DISPATCH set — `registryByName` plus the provider cache —
+ * because that is the true statement about resolution: these are the names that
+ * answered to something. It is deliberately NOT the offer: the wire is narrowed
+ * by steps, parks and postures, and a held-out tool still dispatches by name, so
+ * an offer-shaped list would be a different (and here, false) claim.
+ *
+ * ── AND IT SAYS "RESOLVED", NOT "COULD BE DISPATCHED" ───────────────────
+ *
+ * The first draft said "Tools that could be dispatched on that call". Two gates
+ * sit between resolution and a tool running — the `tool_call` permission check
+ * and the middleware chain — so a name in this map can resolve and still be
+ * refused, and the sentence promised the model a dispatch the very next gate
+ * would decline. Neither gate is asked here: a `PermissionChecker` is a port
+ * (a hub call per name, with its own events and its own audit rows), and
+ * spending a whole permission sweep to phrase an error is a cost the sentence
+ * does not need. It reports what this map DID — resolve a name — and claims
+ * nothing about what would have happened next.
+ *
+ * The roster is role-filtered before it gets here (`dispatchRoster`), so a tool
+ * belonging to a skill the caller's policy hides is never named; and when the
+ * filter empties a stocked roster the clause is DROPPED, because "no tool name
+ * resolved" would be a denial of what the dispatch map is holding.
+ *
+ * Anchored past tense, like every other persistent result: "on that call" binds
+ * the roster to the finished call the result answers, so a later iteration that
+ * adds or drops a provider tool cannot make this sentence wrong. No exhortation
+ * — the model does not need to be told to try again.
+ *
+ * Exported for the producer registry, and used by BOTH dispatch doors so they
+ * cannot drift.
+ */
+export function unknownToolResult(toolName, dispatchable) {
+    const head = `Unknown tool '${toolName}' on that call.`;
+    if (dispatchable.named.length > 0) {
+        return (`${head} Tool names that resolved to an implementation on that ` +
+            `call: ${dispatchable.named.join(', ')}.`);
     }
-    return `${head}No skills are reachable from here — answer with the current skill, or finish.`;
+    // Held-but-unnameable: the map had entries this caller may not be told about.
+    // Omission is free; the negative below would be false.
+    return dispatchable.held
+        ? head
+        : `${head} No tool name resolved to an implementation on that call.`;
 }
 /**
- * The re-prompt a POSTURE-refused `read_skill` gets back (SG-C `strictness`).
- * The pick was REACHABLE — the graph would have granted it — so the message
- * must teach the posture, not the map: who routes here, and what the model may
- * still do (open skills, staying, finishing).
+ * THE ONE COMPOSER of every `read_skill` refusal (9.86.0) — reachability,
+ * posture and tree, from one set of inputs.
+ *
+ * ── WHY ONE ─────────────────────────────────────────────────────────────
+ *
+ * There were two, and they disagreed. `skillRefusal` named "Reachable skills:
+ * beta" from the graph's raw hop set; `postureRefusal`, forty lines below it in
+ * the same gate, answered a model that took that offer up with "read_skill here
+ * reaches only the open skills: gamma". Same call, same turn, opposite claims —
+ * and the second one cost a refusal from the escalation budget to discover.
+ * Neither had heard of the role filter, so both could name a skill the caller's
+ * policy hides. One composer, handed the FILTERED sets and the posture that
+ * will judge the next pick, cannot contradict itself that way.
+ *
+ * ── THE TENSE DISCIPLINE (the law `selfCallNotice` is written against) ──
+ *
+ * A tool result is composed on iteration N and re-read by the model on every
+ * later call of the turn, including the tool-less wrap-up. So every clause here
+ * is a PAST fact about ONE named call — the call the model made `read_skill`
+ * on — anchored by the phrase "that call", never by deixis ("this turn", "the
+ * call you just made") and never by a forecast ("you can call", "pick one of
+ * these"). What the model may do NEXT is owned by the `read_skill` DESCRIPTION,
+ * which is recomposed for every single request and may therefore speak in the
+ * present. Two owners, one tense each, and they can no longer disagree.
+ *
+ * ── AND WHY IT NEVER NAMES A HOP THE POSTURE WOULD DECLINE ──────────────
+ *
+ * Under `'rails'` the model does not route at all; under `'guard'` it routes
+ * only from an outstanding menu. Naming hops there is an offer the very next
+ * arm of this gate refuses. So those arms name no hop: they say who DID route,
+ * and (when there are any) which open skills were admitted anyway — open skills
+ * never reach the posture arm, so that clause cannot be contradicted.
+ *
+ * ── AND WHY AN EMPTY LIST IS NEVER A DENIAL ─────────────────────────────
+ *
+ * Every set arrives as {@link SpokenIds}, which carries whether the UNFILTERED
+ * set held anything. That is the difference between "nothing was reachable" and
+ * "nothing reachable may be named", and this composer used to say the first
+ * when the second was true — over a graph that was routing the cursor the whole
+ * time. Where the filter empties a set, the clause is omitted. A model told the
+ * map is a dead end stops asking for the door it may not be shown.
+ *
+ * Exported so a producer registry can list it beside the other model-facing
+ * composers rather than re-deriving what it says.
  */
-function postureRefusal(requestedId, posture, turnRoute, currentSkillId, openIds) {
-    const openClause = openIds.length > 0
-        ? ` read_skill here reaches only the open skills: ${openIds.join(', ')}.`
+export function composeReadSkillRefusal(args) {
+    const { requestedId, targetClass, cursorId, hops, openIds, isTree, posture, menuOffered, routedDecisively, } = args;
+    const head = `read_skill("${requestedId}") was not granted on that call: `;
+    // Open skills are the one list every arm may name: they are admitted from
+    // every cursor and no posture governs them (the posture arm below is reached
+    // only by hops), so naming them cannot be falsified by the next call. There is
+    // no negative arm here and there never was — an unnamed open set is simply
+    // omitted, which is the shape the two arms below now share.
+    const openClause = openIds.named.length > 0
+        ? ` Open skills were admitted on that call: ${openIds.named.join(', ')}.`
         : '';
-    if (posture === 'rails') {
-        return (`read_skill("${requestedId}") was declined: this graph runs on rails — turn starts ` +
-            `resolve by declared rule or scorer and transitions by declared routes; the model ` +
-            `does not route itself.${openClause} Continue with the skill you are in` +
-            `${currentSkillId !== undefined ? ` ('${currentSkillId}')` : ''}, or finish.`);
+    if (posture !== undefined) {
+        if (posture === 'rails') {
+            return (`${head}this graph's 'rails' posture reserves routing to the framework — turn ` +
+                `starts resolve by declared rule or scorer and transitions by declared routes — ` +
+                `so a model pick was not admitted on that call.${openClause}`);
+        }
+        if (menuOffered !== undefined && menuOffered.named.length > 0) {
+            return (`${head}this graph's 'guard' posture admits a routing pick only from the menu the ` +
+                `framework offered, and '${requestedId}' was not on it. The menu outstanding when ` +
+                `that call was made: ${menuOffered.named.join(', ')}.${openClause}`);
+        }
+        // A menu WAS outstanding and not one of its ids may be named to this
+        // caller. Neither of the other two sentences is available: naming the menu
+        // would leak it, and "no menu was outstanding" is false — as is the
+        // "declared routes moved the cursor instead" clause that rides with it,
+        // which is a second assertion about a turn that did not route that way. So
+        // this arm says the one thing that is true of that call and stops.
+        if (menuOffered !== undefined && menuOffered.held) {
+            return (`${head}this graph's 'guard' posture admits a routing pick only from the menu the ` +
+                `framework offered, and '${requestedId}' was not admitted on that call.${openClause}`);
+        }
+        return (`${head}this graph's 'guard' posture admits a routing pick only while the framework ` +
+            `has declared ambiguity, and no menu was outstanding when that call was made` +
+            `${routedDecisively === true ? " — the turn's start had already been resolved decisively" : ''}` +
+            `. Declared routes moved the cursor instead.${openClause}`);
     }
-    // guard — either no menu is outstanding, or the pick was off it.
-    if (turnRoute?.offered !== undefined && menuOutstanding(turnRoute, currentSkillId)) {
-        return (`read_skill("${requestedId}") was declined: under this graph's 'guard' posture a ` +
-            `routing pick must come from the offered menu — ${turnRoute.offered.join(', ')} — or ` +
-            `stay (answer without calling read_skill).${openClause}`);
+    if (isTree === true) {
+        return (`${head}this map is a decision tree. A tree routes by predicate on every iteration ` +
+            `and keeps no cursor, so there was nothing for that pick to move and ` +
+            `'${requestedId}' was not activated by it.${openClause}`);
     }
-    const decidedBy = turnRoute?.by === 'intent' || turnRoute?.by === 'entry'
-        ? `the turn's start was resolved decisively this turn`
-        : `no routing ambiguity is open this turn`;
-    return (`read_skill("${requestedId}") was declined: this graph's 'guard' posture admits a ` +
-        `routing pick only while the framework has declared ambiguity (an offered menu), and ` +
-        `${decidedBy}; declared routes handle transitions.${openClause} Continue with the ` +
-        `skill you are in${currentSkillId !== undefined ? ` ('${currentSkillId}')` : ''}, or finish.`);
+    // Reachability. `targetClass` is 'unreachable' by the time the gate calls
+    // this arm — 'self' is answered by the notice, 'hop' and 'open' are admitted
+    // — and it is carried rather than re-derived so the sentence and the verdict
+    // read the same classification.
+    const from = cursorId !== undefined ? `'${cursorId}'` : "the turn's start";
+    // The gate reaches this arm only with `'unreachable'`; the weaker sentence is
+    // for a caller that refused an admissible class for a reason of its own, and
+    // it is deliberately not an invented explanation of one.
+    const reason = targetClass === 'unreachable'
+        ? `'${requestedId}' was not reachable from ${from}.`
+        : `'${requestedId}' was not admitted from ${from}.`;
+    // OMIT, NEVER DENY. Three cases, not two: hops to name, no hops at all, and
+    // hops the graph holds that this caller may not be told about. The third used
+    // to fall into the second and assert "No skill was reachable from 'alpha'"
+    // while the graph routed 'alpha' to a skill the role hides — a denial of what
+    // the run holds, and the one shape omission is always safe against.
+    const hopClause = hops.named.length > 0
+        ? ` Skills reachable from ${from} when that call was made: ${hops.named.join(', ')}.`
+        : hops.held
+            ? ''
+            : ` No skill was reachable from ${from} when that call was made.`;
+    return `${head}${reason}${hopClause}${openClause}`;
 }
 /**
  * The args a `pauseHere` / `askHuman` call was running with, read back on the
@@ -775,6 +925,35 @@ export function buildToolCallsHandler(deps) {
                 }
                 const currentSkillId = scope.currentSkillId;
                 const hops = deps.allowedSkillIds(currentSkillId);
+                // ── A PROPOSAL TO THE CURSOR'S OWN SKILL IS A STAY (9.86.0) ──────
+                // `makeReachableSkills` filters the cursor out of its own successor
+                // set — right for a MOVE, and this judge read it as "not reachable"
+                // and refused. A tool saying "stay where you are" asks for the state
+                // the run is already in: there is no cursor to move, no activation to
+                // append, and nothing to refuse. It is accepted as a no-op — no
+                // `pendingToolTransition`, no cursor move, and no `[tool effect
+                // refused: …]` suffix on a result that proposed nothing wrong.
+                //
+                // The event carries the existing `'accepted'` outcome plus an
+                // ADDITIVE `stay: true`, deliberately not a new enum member: an
+                // exhaustive switch over `outcome` in a consumer must keep compiling.
+                if (classifySkillTarget({
+                    ...(currentSkillId !== undefined && { cursor: currentSkillId }),
+                    target,
+                    hops,
+                }) === 'self') {
+                    typedEmit(scope, 'agentfootprint.tools.effect', {
+                        kind: 'propose-transition',
+                        outcome: 'accepted',
+                        stay: true,
+                        toolName: call.toolName,
+                        toolCallId: call.toolCallId,
+                        iteration: call.iteration,
+                        targetSkillId: target,
+                        reason: effect.reason,
+                    });
+                    continue;
+                }
                 if (!hops.includes(target)) {
                     refuse('propose-transition', `propose-transition → '${target}' was refused: '${target}' is not reachable from ` +
                         `${currentSkillId !== undefined ? `'${currentSkillId}'` : 'the turn start'} per ` +
@@ -1424,6 +1603,62 @@ export function buildToolCallsHandler(deps) {
         const cached = providerToolCache?.current ?? [];
         return cached.find((t) => t.schema.name === toolName);
     };
+    /**
+     * Every member id of every map the kernel currently holds PARKED (9.86.0
+     * fix pass) — asked of `parkedMemberIds`, the kernel's own owner of the fact.
+     *
+     * ONE reader for TWO gates. The `skill_read` permission gate and the graph's
+     * dispatch gate both have to tell a stay from a re-engagement, and they sit
+     * 480 lines apart; the second computed this and the first did not, which is
+     * exactly how a role-hidden parked member became re-activatable. Answers with
+     * `undefined` when no engagement plan is mounted, which is every agent that
+     * never called `.maps()`.
+     */
+    const parkedNow = (scope) => deps.engagementPlan === undefined
+        ? undefined
+        : parkedMemberIds(deps.engagementPlan, scope.mapEngagement);
+    /**
+     * Every name `lookupTool` would resolve, in resolution order, through the
+     * role filter (9.86.0) — the roster `unknownToolResult` names.
+     *
+     * Derived from the SAME two sources the lookup reads, so a name it reports is
+     * a name that resolved. Then filtered by the SAME fact every other sentence in
+     * this stage filters by: `scope.hiddenSkillIds`, the per-iteration answer to
+     * "which skills may this caller be told about". The first draft applied
+     * neither, and named a hidden skill's tool to a role that may not see the
+     * skill — the leak the `read_skill` refusals had just closed, reappearing one
+     * sentence over because the roster read the dispatch map raw.
+     *
+     * A tool is unnameable only when EVERY skill declaring it is hidden
+     * (`deps.toolDeclaringSkills`, the build-time owner of that fact). A tool two
+     * skills share is still named when either is visible — the same
+     * sole-owner rule the step hold-out uses, for the same reason: a shared name
+     * is somebody's escape hatch. Names nobody declares from a skill — static
+     * `.tool()` registrations, `read_skill`, `skip_step`, provider-delivered
+     * tools — have no owning skill and are never filtered.
+     *
+     * DISPATCH IS UNTOUCHED. `lookupTool` still resolves every name in the map:
+     * a narrowing may take a name out of a sentence, never out of the dispatch
+     * map. This filter governs only what the model is TOLD.
+     */
+    const dispatchRoster = (scope) => {
+        const names = [...registryByName.keys()];
+        const seen = new Set(names);
+        for (const t of externalToolProvider ? providerToolCache?.current ?? [] : []) {
+            if (seen.has(t.schema.name))
+                continue;
+            seen.add(t.schema.name);
+            names.push(t.schema.name);
+        }
+        const hidden = new Set(scope.hiddenSkillIds ?? []);
+        if (hidden.size === 0)
+            return { named: names, held: names.length > 0 };
+        const declaredBy = deps.toolDeclaringSkills;
+        return spoken(names, (name) => {
+            const owners = declaredBy?.get(name);
+            return owners === undefined || owners.some((id) => !hidden.has(id));
+        });
+    };
     /**
      * The after-tool moment: the chain's last word, on a call that RAN.
      *
@@ -1481,7 +1716,7 @@ export function buildToolCallsHandler(deps) {
     // resume can't re-pause (resume returns void), so that is surfaced as an error.
     const resolveCredentialAndExecute = async (scope, tool, toolName, args, toolCallId, iteration, env) => {
         if (!tool)
-            return { result: `Unknown tool: ${toolName}`, error: true };
+            return { result: unknownToolResult(toolName, dispatchRoster(scope)), error: true };
         // Declared artifact arguments (9.22.0) — the same resolution the batch
         // loop applies, at this door: a resumed call's refs are judged exactly
         // as an inline call's, BEFORE credentials, and a refusal means the tool
@@ -2114,7 +2349,37 @@ export function buildToolCallsHandler(deps) {
                     checkerGoverns(permissionChecker, 'skill_read')) {
                     const requestedSkillId = callArgs.id;
                     if (typeof requestedSkillId === 'string' && requestedSkillId.length > 0) {
-                        await askCapability('skill_read', skillTarget(requestedSkillId), `Skill '${requestedSkillId}' is not available in this context.`);
+                        // ── A SELF-CALL AT A MOUNTED CURSOR EXERCISES NO CAPABILITY ───
+                        // (9.86.0.) A MOUNTED cursor's own skill is already served: its
+                        // body is in that call's system prompt and its tools are on that
+                        // call's wire. Asking to read it activates nothing, moves nothing,
+                        // and reveals nothing the request did not already carry — so there
+                        // is no capability here for a policy to grant or deny, and denying
+                        // it told the model its current skill was "not available in this
+                        // context" while it was standing in it. That is the exact sentence
+                        // the self-call arm exists to stop the model reading. The gate's
+                        // self-call arm answers instead.
+                        //
+                        // A PARKED cursor is the one case where the same id is not a
+                        // no-op, and the skip is scoped to exclude it (9.86.0 fix pass):
+                        // parking suppressed the map's contribution, so the dispatch gate
+                        // 480 lines below reads that call as a RE-ENGAGEMENT and serves the
+                        // body and tools again on the next pass. That moves something, so
+                        // the policy keeps the question. `atMountedCursor` is where that
+                        // rule lives; the dispatch gate writes the same rule as two ordered
+                        // arms over the same `parkedNow` set. Every OTHER id still goes
+                        // through the policy, unchanged — this narrows nothing a role could
+                        // see.
+                        const stay = atMountedCursor({
+                            ...(typeof scope.currentSkillId === 'string' && {
+                                cursor: scope.currentSkillId,
+                            }),
+                            target: requestedSkillId,
+                            parked: parkedNow(scope),
+                        });
+                        if (!stay) {
+                            await askCapability('skill_read', skillTarget(requestedSkillId), `Skill '${requestedSkillId}' is not available in this context.`);
+                        }
                     }
                 }
                 // ── Check-in gate (evidence-carrying human consent) ──────────────
@@ -2318,7 +2583,7 @@ export function buildToolCallsHandler(deps) {
                     if (!credentialBlocked && !wantsBlocked) {
                         try {
                             if (!tool)
-                                throw new Error(`Unknown tool: ${tc.name}`);
+                                throw new Error(unknownToolResult(tc.name, dispatchRoster(scope)));
                             // Set BEFORE the await: a tool that throws has still run, and a
                             // tool that does not exist has not. This flag is the entire
                             // precondition of the after-tool moment below.
@@ -2490,9 +2755,44 @@ export function buildToolCallsHandler(deps) {
                     const reqId = callArgs.id;
                     if (typeof reqId === 'string' && reqId.length > 0) {
                         const currentSkillId = scope.currentSkillId;
+                        // ── ADMISSION reads the graph; every SENTENCE reads the filter ──
+                        // Two sets, deliberately (9.86.0). `admissible` is the routing law:
+                        // what the graph and the mount actually grant, unfiltered, because a
+                        // narrowing may take a schema off the wire but never a name out of
+                        // the dispatch map — role visibility governs what the model is TOLD,
+                        // not what the gate admits. `hopsNamed` / `openNamed` are what any
+                        // sentence, and the `skill.rejected` payload, may NAME: the same
+                        // hidden set the description filtered itself with, resolved once by
+                        // the tools slot and carried here on scope. Without it the refusals
+                        // read the raw sets and could name a skill this role may never see —
+                        // the leak the description closed, one stage downstream. (A hidden
+                        // skill cannot reach this gate anyway: the `skill_read` permission
+                        // gate above denies it first, on the same checker. The filter is
+                        // what keeps that true by construction rather than by coincidence.)
+                        const hiddenIds = new Set(scope.hiddenSkillIds ?? []);
+                        const mayName = (id) => !hiddenIds.has(id);
                         const hops = deps.allowedSkillIds(currentSkillId);
-                        skillHop = hops.includes(reqId);
-                        const allowed = dedupeIds([...hops, ...(deps.openSkillIds ?? [])]);
+                        const openAll = deps.openSkillIds ?? [];
+                        // `spoken` keeps BOTH halves — what may be named, and whether the
+                        // graph held anything before the filter ran. A composer handed only
+                        // the filtered array cannot tell "nothing was reachable" from
+                        // "nothing reachable may be named", and it used to assert the first.
+                        const hopsSpoken = spoken(hops, mayName);
+                        const openSpoken = spoken(openAll, mayName);
+                        const hopsNamed = hopsSpoken.named;
+                        const openNamed = openSpoken.named;
+                        // The ONE owner of "is this target the cursor?" — the same function
+                        // the description, the tool-effects judge and the permission gate
+                        // switch on, so a stay means the same thing in all four.
+                        const targetClass = classifySkillTarget({
+                            ...(currentSkillId !== undefined && { cursor: currentSkillId }),
+                            target: reqId,
+                            hops,
+                            open: openAll,
+                        });
+                        skillHop = targetClass === 'hop';
+                        const admissible = dedupeIds([...hops, ...openAll]);
+                        const allowed = dedupeIds([...hopsNamed, ...openNamed]);
                         // ── The RE-ENGAGEMENT arm (9.59.0) — a third admission class ──
                         // A pick must be routed by INTENT. The reachable set deliberately
                         // excludes the node the cursor already occupies (`makeReachableSkills`
@@ -2510,15 +2810,13 @@ export function buildToolCallsHandler(deps) {
                         // false, `pendingSkillPick` stays unwritten, and the pick lands on
                         // `acceptedSkillPicks`, which the kernel's renewal feed reads as
                         // explicit evidence and re-engages on the very next pass.
-                        const parked = deps.engagementPlan === undefined
-                            ? undefined
-                            : parkedMemberIds(deps.engagementPlan, scope.mapEngagement);
+                        const parked = parkedNow(scope);
                         const reengaging = parked?.has(reqId) === true;
                         if (reengaging) {
                             // Admitted. Nothing to refuse and no posture to apply: this is
                             // not a hop, and a posture governs routing, not engagement.
                         }
-                        else if (currentSkillId !== undefined && reqId === currentSkillId) {
+                        else if (targetClass === 'self') {
                             // ── The SELF-CALL arm (9.84.0) — BEFORE reachability ─────────
                             // The cursor is in neither half of `allowed`, by construction:
                             // `makeReachableSkills` filters it out of its own successor set
@@ -2593,9 +2891,16 @@ export function buildToolCallsHandler(deps) {
                             });
                             noteSkillRefusal(scope, iteration);
                         }
-                        else if (!allowed.includes(reqId)) {
+                        else if (!admissible.includes(reqId)) {
                             skillRejected = true;
-                            result = skillRefusal(reqId, allowed, deps.skillGraphIsTree === true);
+                            result = composeReadSkillRefusal({
+                                requestedId: reqId,
+                                targetClass,
+                                ...(currentSkillId !== undefined && { cursorId: currentSkillId }),
+                                hops: hopsSpoken,
+                                openIds: openSpoken,
+                                ...(deps.skillGraphIsTree === true && { isTree: true }),
+                            });
                             typedEmit(scope, 'agentfootprint.skill.rejected', {
                                 requestedId: reqId,
                                 ...(currentSkillId !== undefined && { currentSkillId }),
@@ -2617,7 +2922,30 @@ export function buildToolCallsHandler(deps) {
                             if (refusedByPosture) {
                                 skillRejected = true;
                                 skillHop = false; // a refused pick must not move the cursor below
-                                result = postureRefusal(reqId, deps.skillStrictness, turnRoute, currentSkillId, deps.openSkillIds ?? []);
+                                // The posture arm names NO hop — under 'rails' the model does
+                                // not route, and under 'guard' it routes only from a menu — so
+                                // the composer is handed the menu when there IS one and an
+                                // empty hop list either way. Open skills are still named:
+                                // they never reach this arm, so nothing here can decline them.
+                                const menuNamed = turnRoute?.offered !== undefined && menuOutstanding(turnRoute, currentSkillId)
+                                    ? spoken(turnRoute.offered, mayName)
+                                    : undefined;
+                                result = composeReadSkillRefusal({
+                                    requestedId: reqId,
+                                    targetClass,
+                                    ...(currentSkillId !== undefined && { cursorId: currentSkillId }),
+                                    // No hop is named under a posture, and the graph's hop set
+                                    // is not this arm's subject — `held: false` is the honest
+                                    // shape, not a claim that nothing was reachable (this arm
+                                    // never composes the reachability clause at all).
+                                    hops: { named: [], held: false },
+                                    openIds: openSpoken,
+                                    posture: deps.skillStrictness,
+                                    ...(menuNamed !== undefined && { menuOffered: menuNamed }),
+                                    ...((turnRoute?.by === 'intent' || turnRoute?.by === 'entry') && {
+                                        routedDecisively: true,
+                                    }),
+                                });
                                 typedEmit(scope, 'agentfootprint.skill.rejected', {
                                     requestedId: reqId,
                                     ...(currentSkillId !== undefined && { currentSkillId }),