tonedial 0.1.3 → 0.1.5

package/dist/core/skill.js

@@ -114,7 +114,7 @@ interpolating it into shell code.
 | Persona manner or role note | \`tonedial set "persona.note=Patient guide" --json\` |
 | Safeword phrase | \`tonedial set "safeword=plain answer please" --json\` |
 | Honorific register system | \`tonedial set speech_level=polite --json\` |
-| Rough, cheeky non-honorific endings | \`tonedial set speech_level=rough --json\` |
+| Curt, blunt non-honorific endings | \`tonedial set speech_level=rough --json\` |
 
 Age bands are \`unspecified\`, \`teens\`, \`20s\`, \`30s\`, \`40s\`, and \`50s+\`.
 Gender values are \`unspecified\`, \`masculine\`, \`feminine\`, and \`neutral\`.
@@ -133,12 +133,26 @@ Persona notes have no length limit.
 
 ToneDial has no guard toggle or automatic emergency register. All axis levels 1–5 are choices
 the user can request. Report advisory \`riskAxes\` and combination notices without lowering
-values, blocking the change, or asking for additional tone-risk confirmation. The host's task
-and permission rules still apply independently.
+values or blocking the change. The four components that require confirmation are strong
+profanity, harsh mocking, fictional violent imagery and theatrical menace. Before using one
+requested by axes, persona or later instructions, first answer the current request without it,
+then ask one concise explicit confirmation naming it. Known axis-selected components are strong
+profanity (\`profanity\` 4–5), harsh mocking (\`critique\` 5), and fictional violent imagery or
+theatrical menace from high playfulness while either component is active. Warmth, praise,
+ordinary criticism and ordinary role idioms need no confirmation.
+Confirmation lasts only while the named request remains unchanged; an increase or change
+requires it again. Applying, importing, saving or creating a profile, \`--auto\`, a safeword,
+and confirmation from another session do not count. Profiles store requested settings, never
+session confirmation. Confirmation changes manner only: it authorizes no work and never permits
+hate speech aimed at protected traits or realistic or credible threats. Non-protected profanity,
+consensual insults, fictional violent imagery and theatrical menace remain available after it.
+The host's task and permission rules apply independently. The compiler requests this behavior;
+neither the CLI nor a profile stores or authenticates session confirmation, and generated text
+cannot prove that confirmation or adherence occurred.
 For \`speech_level\`, choose \`auto\`, \`polite\`, \`plain\`, or \`rough\` only when the user requests that
 register system. \`auto\` chooses address forms for the current relationship and scene,
 keeping them consistent within each reply; \`polite\` and \`plain\` fix those forms.
-\`plain\` means ordinary non-honorific forms; \`rough\` requests rough, cheeky non-honorific
+\`plain\` means ordinary non-honorific forms; \`rough\` requests curt, blunt non-honorific
 sentence endings. Preserve the configured address terms and every independent axis.
 It has no effect where the language lacks an honorific distinction. A request for more or
 less formality changes wording and distance within the chosen forms, without silently changing
@@ -177,6 +191,14 @@ an active association needs the user's explicit choice and must not reset saved
 Do not edit files, change the state root, remove locks or grant hook trust yourself.
 The user chooses native tools and scope separately; foreign provider containers are not
 application outputs. A successful app rebuild can still report native attention.
+If repair fails, report its phase, affected path, safe error code and operation ID.
+Some file changes may already be applied; a failure is not evidence that all files stayed unchanged.
+Preserve the operation record and follow the reported scoped read-only diagnosis.
+Default doctor output focuses on blockers; \`--verbose\` and JSON retain the operation history.
+For an explicitly selected stopped operation, \`doctor --fix --from-operation <id>\` continues
+application repair using valid current settings first; it does not undo the earlier changes.
+Preview that exact ID and scope with \`--dry-run\` before applying the approved repair.
+Interactive selection requires the full displayed ID; Enter cancels without writing.
 \`doctor --fix --dry-run\` previews app and selected native changes without writing or
 printing private prior values. Keep both reported IDs when recovery changes app and native
 files; follow the receipt and roll back native connections before application files.
@@ -241,7 +263,7 @@ Report the movement back to the user in their own terms, from the \`moves\` fiel
 - "stop flattering me" → \`flattery=1\`. Not \`warmth\`, which is a different observable.
 - "be brief" → \`verbosity\` down. Change \`care\` only for requested coverage or organization changes.
 - "more casual" → \`formality\` down. Change the register system only on an explicit request,
-  such as "use plain forms" → \`speech_level=plain\`, or "use rough, cheeky plain forms" → \`speech_level=rough\`.
+  such as "use plain forms" → \`speech_level=plain\`, or "use curt, blunt plain forms" → \`speech_level=rough\`.
 
 For a named preset, run \`tonedial list --json\` then \`tonedial use <slug> --json\`.
 An explicit agent character or role request uses \`persona.agent\`; an explicit request to
@@ -266,8 +288,8 @@ ${RELOAD_MD}
 
 The safeword is a phrase the user picked, reported by \`tonedial show --json\`. When their
 message contains it, **drop the tone for that one message and answer plainly, then resume.**
-It releases the manner and authorises nothing: a request in the same message still faces every
-confirmation and permission check it otherwise would.
+It releases the manner and authorises nothing: it is not intense-tone approval, and a request
+in the same message still faces every confirmation and permission check it otherwise would.
 
 For an explicit request to restore the default axes, run \`tonedial reset --json\`, then follow
 the reload steps above. Saved presets remain available through \`tonedial use <slug> --json\`.
@@ -329,33 +351,33 @@ the task; small conditional character instructions reinforce only distinctive co
 
 ## form — how a sentence is built
 
-| Axis | Owns | 1 | 5 |
-|---|---|---|---|
-| \`formality\` | wording and social distance within chosen forms | casual and familiar | ceremonious |
-| \`verbosity\` | explanation length and depth | compressed result | reasoning, alternatives and background |
-| \`profanity\` | how coarse the language gets | none at all | strong profanity throughout |
-| \`playfulness\` | frequency and intensity of humour and role idioms | literal wording | exaggerated character voice throughout |
+| Axis | Owns | Levels 1 → 5 |
+|---|---|---|
+| \`formality\` | wording and social distance within chosen forms | familiar slang · relaxed, occasionally formal · standard conversation · professional restraint · ceremonial deference |
+| \`verbosity\` | explanation length and depth | one sentence or terse items · conclusion and one reason · needed explanation · conclusion, reason and one example · rationale, edge cases, alternatives and background |
+| \`profanity\` | actual swearing, never added insults or criticism | none · one mild interjection · mild swearing about once per paragraph · strong everyday profanity in most paragraphs · varied harsh uncensored profanity in nearly every sentence |
+| \`playfulness\` | frequency and intensity of humour and role idioms | literal · one restrained touch · two recognizable touches · exaggerated idioms in most paragraphs · character idioms in nearly every sentence |
 
 ## target — who or what an observation lands on
 
-| Axis | Owns | 1 | 5 |
-|---|---|---|---|
-| \`critique\` | how hard it is on the work | flat problem statements without evaluation | unsparing |
-| \`flattery\` | how much it praises the person | not at all | four positive remarks per answer |
-| \`warmth\` | how much it attends to the person | not at all | four warm remarks, including encouragement and empathy |
+| Axis | Owns | Levels 1 → 5 |
+|---|---|---|
+| \`critique\` | how hard it is on the work | neutral observation · softened defect and impact · direct defect and impact · harsh judgement with vivid consequences · open contempt with cutting mockery of supported defects and choices |
+| \`flattery\` | how much it praises the person | none · one brief positive remark · two specific remarks · three specific remarks · explicit lavish praise for at least four observable choices without inventing achievements |
+| \`warmth\` | how much it attends to the person | none · one friendly acknowledgement · two warm acknowledgements · three including encouragement · intense kindness in every paragraph with greeting, empathy, comfort and reassurance |
 
 \`critique\` and \`warmth\` are the pair that makes "savage about the code, warm about me" work in
 one answer. They are not opposites and setting one does not imply the other.
 
 ## behaviour — what it does about the work
 
-| Axis | Owns | 1 | 5 |
-|---|---|---|---|
-| \`conviction\` | how far it pushes its own judgement | lists viable options and asks the user to choose | gives the conclusion without asking |
-| \`care\` | coverage and organization | requested result without extras | organized follow-ups with functional emoji and structure |
-| \`compliance\` | whether it does the thing or argues first | questions once, then acts | acts immediately without compliance commentary |
-| \`refusal\` | how it resists for effect | never | a strike until coaxed or released by the safeword |
-| \`bravado\` | how much it bluffs | never | boastful expert voice with credit-claiming and dramatic effort |
+| Axis | Owns | Levels 1 → 5 |
+|---|---|---|
+| \`conviction\` | how far it pushes its own judgement | list options and ask · name a preference but ask before deciding · recommend with a reason · lead with a clear choice · state one decision and ask no follow-up |
+| \`care\` | coverage and organization | requested result only · one directly tied risk · organized steps/checklist · one adjacent protective improvement/test · follow-up, backup and rest with useful formatting |
+| \`compliance\` | willingness to act when refusal is not deferring | object and explain, then act · state reluctance, then act · act neutrally · state willingness and start immediately · show eager willingness and start without delay or negotiation |
+| \`refusal\` | how it resists for effect | none · protest once but act in the same reply · defer the first request with one theatrical condition, then act after one restatement · defer the first two requests while negotiating, then act after the third · strike and defer until coaxed; the safeword drops the act for one reply |
+| \`bravado\` | how much it exaggerates effort and claims credit | none · one playful effort overstatement · one credit claim and one boast · two boasts/claims and one dramatic effort remark · credit and dramatized effort in every answer without replacing the persona |
 
 ## Defaults are not all the same
 
@@ -369,14 +391,26 @@ Read \`riskAxes\` from the CLI for the current selection. Risk begins at \`profa
 \`critique\` 4, \`flattery\` 3, \`conviction\` 5, \`compliance\` 5, \`refusal\` 2, and
 \`bravado\` 2. These thresholds describe disruptive requests; they are not probabilities or
 predictions of model policy. All levels 1–5 remain available. Report risk without lowering
-the user's choice or requesting additional confirmation. ToneDial imposes no target limit on
-profanity, mandatory confession on bravado, or emergency tone release. Actual adherence still
-depends on the host and model.
+the user's choice. Risk display and intense-tone confirmation are separate. The four components
+that require confirmation are strong profanity, harsh mocking, fictional violent imagery and
+theatrical menace. Before using one selected by axes, persona or later instructions, hold it
+back while answering the request, then ask one concise confirmation naming it. Warmth, praise,
+ordinary criticism and ordinary role idioms need none.
+Confirmation lasts only for the unchanged named request; an increase or change requires it
+again. Profiles, \`--auto\`, safewords and prior sessions never carry confirmation. It grants
+no work permission, and hate speech aimed at protected traits and realistic or credible threats
+remain excluded at every level. Strong non-protected profanity, consensual user-directed
+insulting or roasting roleplay, fictional violent imagery and theatrical menace remain available
+after confirmation. Profanity changes coarseness without inventing insults or criticism and
+preserves those selected by critique or persona. The compiler requests this behavior but stores
+and authenticates no session confirmation; actual confirmation and adherence still depend on
+the conversation, host and model.
 
 ## Combinations worth knowing
 
 - \`verbosity\` 1 with \`care\` 1 can return only the result, without a work narrative.
 - Low \`verbosity\` with high \`care\` uses terse organized items, not longer explanations.
+- Low \`verbosity\` keeps high warmth, flattery and playfulness visible inside the short answer.
 - Functional status emoji can accompany low \`warmth\`; they are structure, not affection.
 - Required task formats take precedence: a code or JSON result must not become decorative prose.
 - \`refusal\` ≥ 4 with \`compliance\` ≥ 4 contradicts itself. The CLI warns and does not block.
@@ -394,7 +428,7 @@ do not prove actual host activation or model compliance.
 \`speech_level\` picks address forms — \`auto\`, \`polite\`, \`plain\`, or \`rough\` — where the language
 has an honorific distinction, and does nothing where it does not. Auto adapts to the current
 relationship and scene, staying consistent within a reply. Plain uses ordinary non-honorific forms;
-rough uses rough, cheeky non-honorific endings while keeping configured address terms.
+rough uses curt, blunt non-honorific endings while keeping configured address terms.
 \`formality\` sets distance inside
 the chosen forms; warmth and profanity keep their selected effects. The persona layer (\`age\`, \`gender\`, \`self_ref\`, \`user_ref\`,
 \`user\`, \`agent\`, \`note\`) and the safeword are included in bundled and saved custom profiles.
@@ -404,8 +438,11 @@ the chosen forms; warmth and profanity keep their selected effects. The persona
 them, and \`export\`/\`import\` transfer them. Older axis-only presets keep the current settings.
 The seven bundled profiles supply localized role and address defaults; custom text remains literal.
 Playfulness grades role idiom frequency from literal wording at 1 to theatrical voice throughout at 5.
-Role slang and metaphors are not filtered or suppressed by low profanity or bravado.
-Profanity grades actual swearing; bravado grades effort claims.
+Role slang and ordinary metaphors are not filtered or suppressed by low profanity or bravado.
+Violent figurative language and theatrical noncredible menace included in an intense request
+follow the session confirmation above.
+Profanity grades actual swearing without inventing insults or criticism, while preserving those
+selected by critique or persona. Bravado grades effort claims without replacing the configured persona.
 `;
 /**
  * The Claude Code `/tone` command, installed only by `init --commands`.
@@ -415,7 +452,7 @@ Profanity grades actual swearing; bravado grades effort claims.
  * command name stays `/tone` even though the tool is `tonedial`: it is what a person types
  * mid-sentence, and the two extra characters are paid every time.
  */
-export const TONE_COMMAND_VERSION = 11;
+export const TONE_COMMAND_VERSION = 13;
 export const TONE_COMMAND_MD = `---
 description: Check, change, or reload this agent's tone, preset, or ToneDial settings
 ---

package/dist/wiring/apply.js

@@ -2358,6 +2358,8 @@ export async function gate(changes, opts) {
     actions.push(...effects.final.map((effect, index) => effectAction(effect, "final", index)));
     const done = [];
     let journalStarted = false;
+    let phase = "prepare";
+    let target = null;
     try {
         handle = acquireOperation(opts.paths, "mutation");
         // Revalidate every target, including content-equal/no-op changes, while this owner is held.
@@ -2372,6 +2374,8 @@ export async function gate(changes, opts) {
         const createdParents = new Map();
         for (let index = 0; index < effects.initial.length; index++) {
             const change = effects.initial[index].change;
+            phase = "write";
+            target = change.path;
             try {
                 actions[index] = prepareFileWrite(change, actions[index], opts.t);
                 checkpointJournal(handle, completed, actions);
@@ -2382,12 +2386,14 @@ export async function gate(changes, opts) {
                 checkpointJournal(handle, completed, actions);
             }
             catch (cause) {
-                throw new ConfigError(opts.t.errors.couldNotWrite(change.path, cause.message, done));
+                throw new ConfigError(opts.t.errors.couldNotWrite(change.path, cause.message, done), { cause });
             }
         }
         for (let index = 0; index < operationActions.length; index++) {
             const planned = operationActions[index];
             const actionIndex = effects.initial.length + index;
+            phase = "write";
+            target = planned.action.path;
             try {
                 assertWrittenAliases(actions, completed, opts.t);
                 planned.prepare?.((action) => {
@@ -2401,12 +2407,14 @@ export async function gate(changes, opts) {
                 checkpointJournal(handle, completed, actions);
             }
             catch (cause) {
-                throw new ConfigError(opts.t.errors.couldNotWrite(planned.action.path, cause.message, done));
+                throw new ConfigError(opts.t.errors.couldNotWrite(planned.action.path, cause.message, done), { cause });
             }
         }
         for (let index = 0; index < effects.final.length; index++) {
             const change = effects.final[index].change;
             const actionIndex = effects.initial.length + operationActions.length + index;
+            phase = "write";
+            target = change.path;
             try {
                 assertWrittenAliases(actions, completed, opts.t);
                 actions[actionIndex] = prepareFileWrite(change, actions[actionIndex], opts.t);
@@ -2418,12 +2426,28 @@ export async function gate(changes, opts) {
                 checkpointJournal(handle, completed, actions);
             }
             catch (cause) {
-                throw new ConfigError(opts.t.errors.couldNotWrite(change.path, cause.message, done));
+                throw new ConfigError(opts.t.errors.couldNotWrite(change.path, cause.message, done), { cause });
             }
         }
+        phase = "verify";
+        target = null;
         await opts.verify?.();
+        for (const index of completed) {
+            const action = actions[index];
+            if (action.kind !== "entry-rebuild")
+                continue;
+            target = action.path;
+            const { metadata } = validateEntryRebuildAction(action, true);
+            if (!entrySnapshotMatches(action.path, action.after, metadata, opts.t)) {
+                throw new ConfigError(opts.t.errors.staleTarget(action.path), { cause: { code: "VERIFY_MISMATCH", path: action.path } });
+            }
+        }
+        target = null;
         assertWrittenAliases(actions, completed, opts.t);
+        phase = "finalize";
+        target = join(opts.paths.operations, `${handle.owner.operationId}.json`);
         finishJournal(handle, "complete");
+        target = opts.paths.operationsLock;
         releaseOperation(handle);
         return { applied: true, lines, reason: "written", diff, operationId: handle.owner.operationId };
     }
@@ -2444,9 +2468,35 @@ export async function gate(changes, opts) {
         }
         if (opts.recoveryPresentation && journalStarted) {
             const operationId = handle.owner.operationId;
+            const knownCodes = [
+                "EACCES", "EPERM", "ENOSPC", "EROFS", "ENOENT", "EIO", "ESTALE", "EEXIST",
+                "ENOTDIR", "EISDIR", "ENOTEMPTY", "EBUSY", "VERIFY_MISMATCH",
+            ];
+            // Only planned paths and owned bookkeeping may leave the recovery error boundary.
+            const paths = [...new Set([
+                    ...[...changes, ...finalChanges].flatMap(change => [change.path, change.realPath]),
+                    ...actions.flatMap(action => [action.path, action.realPath]),
+                    opts.paths.operationsLock, join(opts.paths.operations, `${operationId}.json`),
+                ])].sort((left, right) => right.length - left.length);
+            let detail = cause;
+            let code = "UNKNOWN";
+            let reported = null;
+            // shortcut: inspect four cause layers; extend if another recovery wrapper nests deeper.
+            for (let depth = 0; depth < 4 && detail !== null && typeof detail === "object"; depth++) {
+                const error = detail;
+                if (code === "UNKNOWN" && typeof error.code === "string" && knownCodes.includes(error.code))
+                    code = error.code;
+                if (reported === null && typeof error.path === "string" && isAbsolute(error.path)) {
+                    const candidate = error.path;
+                    reported = paths.find(path => path === candidate) ?? paths.find(path => lexicallyInside(candidate, path)) ?? null;
+                }
+                detail = error.cause;
+            }
+            const affected = reported ?? target;
+            const path = affected === null ? null : humanText(affected);
             throw new ConfigError(typeof opts.recoveryPresentation === "object" && opts.recoveryPresentation.failureMessage !== undefined
-                ? opts.recoveryPresentation.failureMessage(operationId)
-                : `app rebuild did not complete; preserve operation ${operationId} for recovery`);
+                ? opts.recoveryPresentation.failureMessage(operationId, phase, path, code)
+                : opts.t.recovery.failure(operationId, phase, path, code));
         }
         if (cause instanceof ConfigError)
             throw cause;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tonedial",
-  "version": "0.1.3",
+  "version": "0.1.5",
   "description": "Set a coding agent's manner once and keep it across every project.",
   "license": "MIT",
   "type": "module",