ai-whisper 0.5.6 → 0.5.7

package/README.md

@@ -80,7 +80,7 @@ pnpm install
 pnpm build
 ```
 
-Install the bundled agent skills once (they let the agents verify, kick off, and report on workflows). This also installs `ai-whisper-code-review`, the skill workflow code-review handoffs use to evaluate agent-written code:
+Install the bundled agent skills once (they let the agents verify, kick off, and report on workflows). This also installs `ai-whisper-code-review`, the skill workflow code-review handoffs use to evaluate agent-written code, and `ai-whisper-plan-execution`, the skill plan-execution handoffs use to structure how the implementer executes an approved plan:
 
 ```bash
 whisper skill install

package/dist/bin/broker-daemon.js

@@ -2215,8 +2215,10 @@ Non-blocking risks:
 - <risk that does NOT block this gate, or "None.">
 --- end protocol ---`;
 var CODE_REVIEW_SKILL_GUIDANCE = "Use the ai-whisper-code-review skill to evaluate the delivered code. The workflow review protocol below controls your output format and evaluator semantics; the skill controls how you inspect code and decide which code-quality issues are blocking.\n\n";
+var PLAN_EXECUTION_SKILL_GUIDANCE = "Use the ai-whisper-plan-execution skill to structure HOW you execute this plan: subagent fan-out with model allocation where your harness supports it, disciplined inline execution otherwise. The handback contract above remains authoritative \u2014 settle all delegated work before handing back, and state which execution mode you used.";
 var SDD_SPEC_REVIEW = "Review the spec at {specPath}. This is an autonomous workflow with no human in the loop.\n\n" + WORKFLOW_REVIEW_PROTOCOL;
 var SDD_CODE_REVIEW = "Review the implementer's changes for this phase \u2014 the commits in {commitRange}. The upper bound is a LIVE `HEAD`: resolve it against the current repository at review time and INCLUDE any commits added during this review round (e.g. fixes for your prior findings); do not pin the review to an earlier tip. Verify against the spec's acceptance criteria and run the project's verification/tests. This is an autonomous workflow with no human in the loop.\n\n" + CODE_REVIEW_SKILL_GUIDANCE + WORKFLOW_REVIEW_PROTOCOL;
+var SDD_PLAN_EXECUTION = "Execute the plan at {planPath} now: make all changes it specifies, run the verification command the plan or spec defines and ensure it passes, then commit. This is an autonomous workflow \u2014 no human will respond. Do the work yourself; never ask for confirmation, permission, or clarification. Hand back the commit SHAs and the verification output, plus a 1-2 sentence summary; your reply must be at least two sentences, well over 100 characters \u2014 never hand back only a single word.\n\n" + PLAN_EXECUTION_SKILL_GUIDANCE;
 var SPEC_DRIVEN_DEVELOPMENT = {
   type: "spec-driven-development",
   displayName: "Spec-Driven Development",
@@ -2261,9 +2263,9 @@ var SPEC_DRIVEN_DEVELOPMENT = {
       reviewerRole: null,
       maxRounds: 1,
       initialHandoffStep: "execute",
-      kickoffTemplate: "Execute the plan at {planPath} now: make all changes it specifies, run the verification command the plan or spec defines and ensure it passes, then commit. This is an autonomous workflow \u2014 no human will respond. Do the work yourself; never ask for confirmation, permission, or clarification. Hand back the commit SHAs and the verification output, plus a 1-2 sentence summary; your reply must be at least two sentences, well over 100 characters \u2014 never hand back only a single word.",
+      kickoffTemplate: SDD_PLAN_EXECUTION,
       stepTemplates: {
-        execute: "Execute the plan at {planPath} now: make all changes it specifies, run the verification command the plan or spec defines and ensure it passes, then commit. This is an autonomous workflow \u2014 no human will respond. Do the work yourself; never ask for confirmation, permission, or clarification. Hand back the commit SHAs and the verification output, plus a 1-2 sentence summary; your reply must be at least two sentences, well over 100 characters \u2014 never hand back only a single word."
+        execute: SDD_PLAN_EXECUTION
       },
       evaluatorPromptKey: "execution-gate",
       artifactOut: { kind: "commit-range" }

package/dist/bin/companion-agent.js

@@ -2248,8 +2248,10 @@ Non-blocking risks:
 - <risk that does NOT block this gate, or "None.">
 --- end protocol ---`;
 var CODE_REVIEW_SKILL_GUIDANCE = "Use the ai-whisper-code-review skill to evaluate the delivered code. The workflow review protocol below controls your output format and evaluator semantics; the skill controls how you inspect code and decide which code-quality issues are blocking.\n\n";
+var PLAN_EXECUTION_SKILL_GUIDANCE = "Use the ai-whisper-plan-execution skill to structure HOW you execute this plan: subagent fan-out with model allocation where your harness supports it, disciplined inline execution otherwise. The handback contract above remains authoritative \u2014 settle all delegated work before handing back, and state which execution mode you used.";
 var SDD_SPEC_REVIEW = "Review the spec at {specPath}. This is an autonomous workflow with no human in the loop.\n\n" + WORKFLOW_REVIEW_PROTOCOL;
 var SDD_CODE_REVIEW = "Review the implementer's changes for this phase \u2014 the commits in {commitRange}. The upper bound is a LIVE `HEAD`: resolve it against the current repository at review time and INCLUDE any commits added during this review round (e.g. fixes for your prior findings); do not pin the review to an earlier tip. Verify against the spec's acceptance criteria and run the project's verification/tests. This is an autonomous workflow with no human in the loop.\n\n" + CODE_REVIEW_SKILL_GUIDANCE + WORKFLOW_REVIEW_PROTOCOL;
+var SDD_PLAN_EXECUTION = "Execute the plan at {planPath} now: make all changes it specifies, run the verification command the plan or spec defines and ensure it passes, then commit. This is an autonomous workflow \u2014 no human will respond. Do the work yourself; never ask for confirmation, permission, or clarification. Hand back the commit SHAs and the verification output, plus a 1-2 sentence summary; your reply must be at least two sentences, well over 100 characters \u2014 never hand back only a single word.\n\n" + PLAN_EXECUTION_SKILL_GUIDANCE;
 var SPEC_DRIVEN_DEVELOPMENT = {
   type: "spec-driven-development",
   displayName: "Spec-Driven Development",
@@ -2294,9 +2296,9 @@ var SPEC_DRIVEN_DEVELOPMENT = {
       reviewerRole: null,
       maxRounds: 1,
       initialHandoffStep: "execute",
-      kickoffTemplate: "Execute the plan at {planPath} now: make all changes it specifies, run the verification command the plan or spec defines and ensure it passes, then commit. This is an autonomous workflow \u2014 no human will respond. Do the work yourself; never ask for confirmation, permission, or clarification. Hand back the commit SHAs and the verification output, plus a 1-2 sentence summary; your reply must be at least two sentences, well over 100 characters \u2014 never hand back only a single word.",
+      kickoffTemplate: SDD_PLAN_EXECUTION,
       stepTemplates: {
-        execute: "Execute the plan at {planPath} now: make all changes it specifies, run the verification command the plan or spec defines and ensure it passes, then commit. This is an autonomous workflow \u2014 no human will respond. Do the work yourself; never ask for confirmation, permission, or clarification. Hand back the commit SHAs and the verification output, plus a 1-2 sentence summary; your reply must be at least two sentences, well over 100 characters \u2014 never hand back only a single word."
+        execute: SDD_PLAN_EXECUTION
       },
       evaluatorPromptKey: "execution-gate",
       artifactOut: { kind: "commit-range" }

package/dist/bin/relay-monitor.js

@@ -2218,8 +2218,10 @@ Non-blocking risks:
 - <risk that does NOT block this gate, or "None.">
 --- end protocol ---`;
 var CODE_REVIEW_SKILL_GUIDANCE = "Use the ai-whisper-code-review skill to evaluate the delivered code. The workflow review protocol below controls your output format and evaluator semantics; the skill controls how you inspect code and decide which code-quality issues are blocking.\n\n";
+var PLAN_EXECUTION_SKILL_GUIDANCE = "Use the ai-whisper-plan-execution skill to structure HOW you execute this plan: subagent fan-out with model allocation where your harness supports it, disciplined inline execution otherwise. The handback contract above remains authoritative \u2014 settle all delegated work before handing back, and state which execution mode you used.";
 var SDD_SPEC_REVIEW = "Review the spec at {specPath}. This is an autonomous workflow with no human in the loop.\n\n" + WORKFLOW_REVIEW_PROTOCOL;
 var SDD_CODE_REVIEW = "Review the implementer's changes for this phase \u2014 the commits in {commitRange}. The upper bound is a LIVE `HEAD`: resolve it against the current repository at review time and INCLUDE any commits added during this review round (e.g. fixes for your prior findings); do not pin the review to an earlier tip. Verify against the spec's acceptance criteria and run the project's verification/tests. This is an autonomous workflow with no human in the loop.\n\n" + CODE_REVIEW_SKILL_GUIDANCE + WORKFLOW_REVIEW_PROTOCOL;
+var SDD_PLAN_EXECUTION = "Execute the plan at {planPath} now: make all changes it specifies, run the verification command the plan or spec defines and ensure it passes, then commit. This is an autonomous workflow \u2014 no human will respond. Do the work yourself; never ask for confirmation, permission, or clarification. Hand back the commit SHAs and the verification output, plus a 1-2 sentence summary; your reply must be at least two sentences, well over 100 characters \u2014 never hand back only a single word.\n\n" + PLAN_EXECUTION_SKILL_GUIDANCE;
 var SPEC_DRIVEN_DEVELOPMENT = {
   type: "spec-driven-development",
   displayName: "Spec-Driven Development",
@@ -2264,9 +2266,9 @@ var SPEC_DRIVEN_DEVELOPMENT = {
       reviewerRole: null,
       maxRounds: 1,
       initialHandoffStep: "execute",
-      kickoffTemplate: "Execute the plan at {planPath} now: make all changes it specifies, run the verification command the plan or spec defines and ensure it passes, then commit. This is an autonomous workflow \u2014 no human will respond. Do the work yourself; never ask for confirmation, permission, or clarification. Hand back the commit SHAs and the verification output, plus a 1-2 sentence summary; your reply must be at least two sentences, well over 100 characters \u2014 never hand back only a single word.",
+      kickoffTemplate: SDD_PLAN_EXECUTION,
       stepTemplates: {
-        execute: "Execute the plan at {planPath} now: make all changes it specifies, run the verification command the plan or spec defines and ensure it passes, then commit. This is an autonomous workflow \u2014 no human will respond. Do the work yourself; never ask for confirmation, permission, or clarification. Hand back the commit SHAs and the verification output, plus a 1-2 sentence summary; your reply must be at least two sentences, well over 100 characters \u2014 never hand back only a single word."
+        execute: SDD_PLAN_EXECUTION
       },
       evaluatorPromptKey: "execution-gate",
       artifactOut: { kind: "commit-range" }

package/dist/bin/whisper.js

@@ -2585,7 +2585,7 @@ function derivePlanPath(specPath, dateIso) {
   const date = dateIso.slice(0, 10);
   return `docs/superpowers/plans/${date}-${slug}.md`;
 }
-var RALPH_ITEM_DELIVERED_MARKER, RALPH_GOAL_COMPLETE_MARKER, WORKFLOW_REVIEW_PROTOCOL, WORKFLOW_OPERATOR_CONTROL, WORKFLOW_DIAGNOSIS_PROTOCOL, CODE_REVIEW_SKILL_GUIDANCE, SDD_SPEC_REVIEW, SDD_CODE_REVIEW, SPEC_DRIVEN_DEVELOPMENT, RALPH_KICKOFF, RALPH_FIX, RALPH_ITEM_REVIEW, RALPH_ACCEPTANCE_REVIEW, RALPH_LOOP, BUGFIX_DIAGNOSIS_KICKOFF, BUGFIX_DIAGNOSIS_REVIEW, BUGFIX_DIAGNOSIS_FIX, BUGFIX_FIX_KICKOFF, BUGFIX_FIX_REVIEW, BUGFIX_FIX_FIX, BUGFIX_POSTMORTEM_KICKOFF, BUGFIX_POSTMORTEM_REVIEW, BUGFIX_POSTMORTEM_FIX, COMPLEX_BUG_FIXING, REGISTRY;
+var RALPH_ITEM_DELIVERED_MARKER, RALPH_GOAL_COMPLETE_MARKER, WORKFLOW_REVIEW_PROTOCOL, WORKFLOW_OPERATOR_CONTROL, WORKFLOW_DIAGNOSIS_PROTOCOL, CODE_REVIEW_SKILL_GUIDANCE, PLAN_EXECUTION_SKILL_GUIDANCE, SDD_SPEC_REVIEW, SDD_CODE_REVIEW, SDD_PLAN_EXECUTION, SPEC_DRIVEN_DEVELOPMENT, RALPH_KICKOFF, RALPH_FIX, RALPH_ITEM_REVIEW, RALPH_ACCEPTANCE_REVIEW, RALPH_LOOP, BUGFIX_DIAGNOSIS_KICKOFF, BUGFIX_DIAGNOSIS_REVIEW, BUGFIX_DIAGNOSIS_FIX, BUGFIX_FIX_KICKOFF, BUGFIX_FIX_REVIEW, BUGFIX_FIX_FIX, BUGFIX_POSTMORTEM_KICKOFF, BUGFIX_POSTMORTEM_REVIEW, BUGFIX_POSTMORTEM_FIX, COMPLEX_BUG_FIXING, REGISTRY;
 var init_workflow_registry = __esm({
   "../broker/dist/runtime/workflow-registry.js"() {
     "use strict";
@@ -2666,8 +2666,10 @@ Non-blocking risks:
 - <risk that does NOT block this gate, or "None.">
 --- end protocol ---`;
     CODE_REVIEW_SKILL_GUIDANCE = "Use the ai-whisper-code-review skill to evaluate the delivered code. The workflow review protocol below controls your output format and evaluator semantics; the skill controls how you inspect code and decide which code-quality issues are blocking.\n\n";
+    PLAN_EXECUTION_SKILL_GUIDANCE = "Use the ai-whisper-plan-execution skill to structure HOW you execute this plan: subagent fan-out with model allocation where your harness supports it, disciplined inline execution otherwise. The handback contract above remains authoritative \u2014 settle all delegated work before handing back, and state which execution mode you used.";
     SDD_SPEC_REVIEW = "Review the spec at {specPath}. This is an autonomous workflow with no human in the loop.\n\n" + WORKFLOW_REVIEW_PROTOCOL;
     SDD_CODE_REVIEW = "Review the implementer's changes for this phase \u2014 the commits in {commitRange}. The upper bound is a LIVE `HEAD`: resolve it against the current repository at review time and INCLUDE any commits added during this review round (e.g. fixes for your prior findings); do not pin the review to an earlier tip. Verify against the spec's acceptance criteria and run the project's verification/tests. This is an autonomous workflow with no human in the loop.\n\n" + CODE_REVIEW_SKILL_GUIDANCE + WORKFLOW_REVIEW_PROTOCOL;
+    SDD_PLAN_EXECUTION = "Execute the plan at {planPath} now: make all changes it specifies, run the verification command the plan or spec defines and ensure it passes, then commit. This is an autonomous workflow \u2014 no human will respond. Do the work yourself; never ask for confirmation, permission, or clarification. Hand back the commit SHAs and the verification output, plus a 1-2 sentence summary; your reply must be at least two sentences, well over 100 characters \u2014 never hand back only a single word.\n\n" + PLAN_EXECUTION_SKILL_GUIDANCE;
     SPEC_DRIVEN_DEVELOPMENT = {
       type: "spec-driven-development",
       displayName: "Spec-Driven Development",
@@ -2712,9 +2714,9 @@ Non-blocking risks:
           reviewerRole: null,
           maxRounds: 1,
           initialHandoffStep: "execute",
-          kickoffTemplate: "Execute the plan at {planPath} now: make all changes it specifies, run the verification command the plan or spec defines and ensure it passes, then commit. This is an autonomous workflow \u2014 no human will respond. Do the work yourself; never ask for confirmation, permission, or clarification. Hand back the commit SHAs and the verification output, plus a 1-2 sentence summary; your reply must be at least two sentences, well over 100 characters \u2014 never hand back only a single word.",
+          kickoffTemplate: SDD_PLAN_EXECUTION,
           stepTemplates: {
-            execute: "Execute the plan at {planPath} now: make all changes it specifies, run the verification command the plan or spec defines and ensure it passes, then commit. This is an autonomous workflow \u2014 no human will respond. Do the work yourself; never ask for confirmation, permission, or clarification. Hand back the commit SHAs and the verification output, plus a 1-2 sentence summary; your reply must be at least two sentences, well over 100 characters \u2014 never hand back only a single word."
+            execute: SDD_PLAN_EXECUTION
           },
           evaluatorPromptKey: "execution-gate",
           artifactOut: { kind: "commit-range" }
@@ -11340,8 +11342,8 @@ function compareSemver(a, b) {
 var EZIO_PROVENANCE = {
   ezioCliVersion: "0.2.0-beta.2",
   ezioGitSha: "bb14110",
-  builtAt: "2026-06-11T15:57:51.888Z",
-  whisperVersion: "0.5.6"
+  builtAt: "2026-06-12T01:11:16.623Z",
+  whisperVersion: "0.5.7"
 };
 
 // src/ezio-provenance-types.ts

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ai-whisper",
-  "version": "0.5.6",
+  "version": "0.5.7",
   "description": "Terminal-first relay for paired AI coding agents (Claude + Codex), driven by structured workflows.",
   "license": "Apache-2.0",
   "repository": {
@@ -53,10 +53,10 @@
     "ink-testing-library": "^4.0.0",
     "@ai-whisper/adapter-ai-ezio": "0.0.0",
     "@ai-whisper/adapter-claude": "0.0.0",
-    "@ai-whisper/adapter-codex": "0.0.0",
-    "@ai-whisper/companion-core": "0.0.0",
     "@ai-whisper/broker": "0.0.0",
-    "@ai-whisper/shared": "0.0.0"
+    "@ai-whisper/companion-core": "0.0.0",
+    "@ai-whisper/shared": "0.0.0",
+    "@ai-whisper/adapter-codex": "0.0.0"
   },
   "files": [
     "dist",

package/skills/ai-whisper-plan-execution/SKILL.md

@@ -0,0 +1,65 @@
+---
+name: ai-whisper-plan-execution
+description: Use when executing an approved implementation plan as the implementer inside an ai-whisper autonomous workflow (plan-execution phase). Structures HOW to execute — subagent fan-out with model allocation where the harness supports it, disciplined inline execution otherwise.
+---
+
+# ai-whisper-plan-execution
+
+Structure HOW you execute an approved implementation plan inside an ai-whisper autonomous workflow. The workflow handoff prompt stays authoritative for WHAT to deliver and the handback contract; this skill only governs execution structure.
+
+## Gating
+
+- Implementer role only, inside an ai-whisper autonomous workflow (e.g. the SDD plan-execution phase).
+- Never as the reviewer; never for ad-hoc edits outside a workflow.
+
+## Step 0 — fan-out decision
+
+Read the plan once; extract every task with its full text. Then:
+
+- **Default: fan out** (Path A below, or its built-in fallback).
+- **Inline escape — only if** the plan has ≤2 tasks, OR is purely mechanical single-concern work (one repetitive transformation, e.g. a rename or codemod sweep). Inline execution still follows the plan task-by-task with per-task verification.
+- Whichever branch you take, your handback MUST name the execution mode (see Handback).
+
+## Path A — your harness supports subagent dispatch (e.g. Claude Code)
+
+1. **If `superpowers:subagent-driven-development` is available in your session, you MUST invoke it** and follow its per-task loop: fresh implementer subagent per task → spec-compliance review → code-quality review, with fix-and-re-review loops. Skipping it to "just execute directly" is a contract violation, not a judgment call. Known rationalizations, pre-refuted:
+   - "It pauses for humans" — false: continuous execution between tasks is built into that skill.
+   - "Its endgame doesn't fit the workflow" — the workflow overrides below resolve every misfit.
+2. **Workflow overrides** (apply while inside an ai-whisper workflow):
+   - "Escalate to the human" (BLOCKED because the plan itself is wrong) → hand back that you CANNOT PROCEED and why; the workflow escalation path handles it.
+   - `superpowers:finishing-a-development-branch` → do not use; the workflow handback replaces it. Merging is the operator's decision.
+   - "Never start implementation on main/master without consent" → the collab-mounted workspace/branch is operator-provided; mounting is that consent.
+3. **Fallback — built-in minimal protocol** (ONLY when subagent-driven-development is not available; it mimics the same loop). For each task, sequentially:
+   1. Dispatch a fresh implementer subagent carrying the full task text plus scene-setting context. Never make a subagent read the plan file.
+   2. Dispatch a spec-compliance reviewer subagent, then a code-quality reviewer subagent; loop fixes until both approve.
+   3. Never dispatch implementation subagents in parallel.
+4. **Model allocation** (applies to subagent dispatch under either protocol; your own model is fixed by the mount). Use model aliases (haiku / sonnet / opus), never versioned model names:
+
+| Subagent role | Default | Flip condition |
+| --- | --- | --- |
+| Task implementer | sonnet | ↑ strongest tier if ambiguous / novel / cross-cutting / security-sensitive |
+| Reviewer (spec or quality) | strongest available (opus or better) | never below the task implementer's tier |
+| Mechanical / codemod | haiku | ↑ sonnet if it touches shared types or public API |
+| Test-writer | sonnet | ↑ strongest tier for adversarial/repro tests; never haiku |
+| Verification runner | haiku | ↑ sonnet if it must diagnose failures, not just report them |
+
+Invariants: **reviewer tier ≥ implementer tier, always.** On the **2nd failed review of the same task**, bump the task implementer one tier and pass the failure history into the new dispatch.
+
+## Path B — no subagent dispatch (e.g. Codex, ezio)
+
+Execute the plan inline: task-by-task in plan order, run per-task verification, commit as the plan specifies, and hand back per the workflow prompt with the mode line `Execution mode: inline (no subagent dispatch)`. Path A does not apply to you — nothing in this skill blocks or changes your handback.
+
+## Handback (both paths)
+
+The workflow prompt is authoritative. Before handing back: all subagents settled, verification green, work committed. Include the commit SHAs, the verification output, the 1-2 sentence summary the workflow prompt demands, and one line naming the execution mode:
+
+- `Execution mode: subagent-driven (superpowers)`
+- `Execution mode: subagent-driven (built-in protocol)`
+- `Execution mode: inline (<reason>)`
+
+## Failure handling
+
+- A subagent dies or hangs → retry once; then do that task inline at your own tier and disclose it in the handback.
+- 2nd failed review of the same task → tier escalation per the model table.
+- Plan unreadable or not task-shaped → execute inline with disclosure; never block the workflow on skill-internal structure.
+- Verification red after all tasks → fix before handing back; never hand back red.