killeros 3.0.2 → 3.0.4

package/CHANGELOG.md

@@ -4,6 +4,39 @@ All notable changes to KillerOS are documented here. Releases use New Features f
 
 ## [Unreleased]
 
+## [3.0.4] - 2026-10-07
+
+### Changed
+
+- Made the explicitly enabled `prepublishOnly` lifecycle the single release-validator invocation at npm publication. The release workflow now publishes the source directory with `npm publish . --ignore-scripts=false`, preserving the existing release checks and trusted publishing.
+- Added current release-metadata validation and isolated publication-lifecycle tests with release-pinned npm to the Node-floor quality check. Extended the existing packed-package Pi test to check all required files and package identity without another pack operation. The package smoke job remains until replacement coverage lands and the required-check migration is approved and verified.
+
+### Fixed
+
+- Allowed later verified ordinary work to complete the same saved blocked goal without resume, clear, or extra automatic turns. Completion retains the original objective, file baseline, and accounting, respects native tool selection, and prevents compaction from restarting completed work.
+- Stopped false CRLF change counts in the footer and task receipts for text files with `eol=lf`. Added a differential Git-oracle matrix covering attributes, `core.autocrlf`, text and binary contents, plus mode and symlink comparisons. Fixes #47.
+- Gave each real Pi decision-tool exposure scenario its own bounded timeout so slower Windows TUI scenarios do not cancel the remaining codemode and RPC coverage. All behavioral assertions remain unchanged.
+
+## [3.0.3] - 2026-10-07
+
+### Changed
+
+- Raised the minimum supported Pi version and all four locked Pi development packages to 1.0.4. Compatibility CI now targets the exact 1.0.4 minimum and latest `>=1.0.4 <2`, with unchanged stable check names and package-file immutability guards.
+- Documented Pi 1.0.4's native wildcard tool filters, MCP availability under explicit tool allowlists, and codemode image reads.
+- Updated Azure footer naming and fast-mode exclusion tests for Pi's renamed `azure` provider. Documented the required Azure configuration migration, the unchanged Responses API identifier, native editor/transcript navigation, and saved codemode images.
+- Gave Pi compatibility CI checks stable minimum/latest names so future Pi version changes do not require renaming GitHub's required checks. Documented the one-time protection migration without weakening existing safeguards.
+
+### Added
+
+- Added real Pi regressions for wildcard decision-tool selection and exclusion, codemode-only provider prompt rules, and pending-stream cancellation retries that preserve one goal turn and defer receipts and sounds until settlement.
+- Added real Pi contract coverage for Azure Responses and Foundry Chat Completions requests both before and after reload, preserving sampling settings and deployment names without applying OpenAI priority service.
+- Added custom-editor coverage for Pi 1.0.3's separate line-navigation and fullscreen transcript Home/End shortcuts.
+
+### Fixed
+
+- Fixed a process-tree test startup race by waiting for the complete descendant PID instead of readiness-file creation. Hook termination assertions remain unchanged.
+- Reused bounded cleanup retries throughout real Pi integration tests so transient Windows directory locks do not fail CI after the assertions pass. Added a repository contract against unbounded fixture cleanup.
+
 ## [3.0.2] - 2026-10-05
 
 ### Changed

package/README.md

@@ -19,9 +19,11 @@ A TypeScript extension for the [Pi coding agent](https://github.com/earendil-wor
 ## Requirements
 
 - Node.js 22.19.0+
-- Pi 1.0.2 or later below 2.0.0
+- Pi 1.0.4 or later below 2.0.0
 - An interactive TUI session for the custom header, editor, footer, and `question`
 
+Users on older Pi versions must upgrade.
+
 ## Install
 
 ```bash
@@ -34,7 +36,7 @@ Or from GitHub:
 pi install git:github.com/KyrosHendrix/pi-KillerOS
 ```
 
-Pin a release by appending its tag, for example `@v3.0.2`. Add `-l` to install only for the current project. Restart Pi after installing.
+Pin a release by appending its tag, for example `@v3.0.4`. Add `-l` to install only for the current project. Restart Pi after installing.
 
 ## Commands
 
@@ -69,7 +71,11 @@ Pi 1.0 defaults to fullscreen. Use `--tui-mode regular` for one invocation or se
 | RPC | Goals, proactive compaction; no TUI components, sounds, title indicator |
 | Print/JSON | No interactive questions, `/goal`, or proactive compaction |
 
-`question` and the active `killeros_goal_update` tool use Pi's `model-only` exposure. They stay directly declared to the model when codemode is disabled or enabled in `on` or `only` mode, but scripts and other tools cannot call them through `ctx.executeTool()`. The goal tool remains inactive without an active goal. Questions still require TUI mode; a direct RPC call fails with `The question tool requires interactive TUI mode`. KillerOS does not enable or configure codemode.
+With Pi 1.0.3's default [keybindings](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/keybindings.md#fullscreen), `Home` and `End` move within the current editor line. In fullscreen mode, `Ctrl+Home` and `Ctrl+End` scroll to the transcript's beginning or end without moving the editor cursor. KillerOS inherits these bindings; user overrides still apply.
+
+`question` and the active `killeros_goal_update` tool use Pi's `model-only` exposure. They stay directly declared to the model when codemode is disabled or enabled in `on` or `only` mode, but scripts and other tools cannot call them through `ctx.executeTool()`. The goal tool is active for active goals and, during a later ordinary request, for completion only of a saved blocked goal. It is inactive outside those cases. Questions still require TUI mode; a direct RPC call fails with `The question tool requires interactive TUI mode`. KillerOS does not enable or configure codemode.
+
+Pi 1.0.4's native `--tools` and `--exclude-tools` support `*` patterns. Quote patterns such as `"killeros_*"` so the shell does not expand them. An explicit allowlist must include `question` for interactive questions and permit `killeros_goal_update` for goals. Selection does not make questions usable outside TUI mode. Excluding the goal tool pauses goals that are active before any model request. A later ordinary request with a blocked goal still runs, but cannot record its completion when the tool is excluded. See Pi's [tool selection](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/cli.md#tools).
 
 ## Configuration
 
@@ -103,7 +109,9 @@ A direct quoted path-shaped file target binds silent file proof. For an extensio
 /goal Fix `killeros/footer.ts`, verified by npm test
 ```
 
-Bare quoted prose remains model-reported. KillerOS captures the file baseline at goal start and only completes when the file is created or changed. A normal response never continues a goal by itself: the agent must record `continue`, `complete`, or a blocker decision through `killeros_goal_update`. Repeated continuation reports and unavailable goal tools pause the goal. New goals pause after 20 turns without warning. An explicit `/goal resume` on an exhausted goal grants another 20 turns; compaction recovery never grants turns. Same-turn recovery requires an actual interruption; successful or skipped compaction cannot restart a normally stopped goal without a decision. A request that completes or blocks a goal remains goal-owned through settlement, so automatic compaction cannot restart it as ordinary work. Later ordinary requests still support compaction continuation. Restored goals keep their persisted limit.
+Bare quoted prose remains model-reported. KillerOS captures the file baseline at goal start and only completes when the file is created or changed. File proof verifies that deliverable, not every natural-language acceptance criterion. The assistant must audit the whole objective before reporting completion; KillerOS cannot independently prove arbitrary objectives from prose. A normal response never continues a goal by itself: the agent must record `continue`, `complete`, or a blocker decision through `killeros_goal_update`. Repeated continuation reports and unavailable goal tools pause the goal. New goals pause after 20 turns without warning. An explicit `/goal resume` on an exhausted goal grants another 20 turns; compaction recovery never grants turns. Same-turn recovery requires an actual interruption; successful or skipped compaction cannot restart a normally stopped goal without a decision. A request that completes or blocks a goal remains goal-owned through settlement, so automatic compaction cannot restart it as ordinary work. Later ordinary requests still support compaction continuation. Restored goals keep their persisted limit.
+
+A blocked goal stays stopped. In a later ordinary TUI or RPC request in a saved session, the assistant can complete the same unchanged objective through `killeros_goal_update` after verification and current evidence. Only `complete` is accepted on this path. Completion does not resume automatic work, grant turns, restart the active clock, or require `/goal clear`. It retains the original file baseline and records either `file` proof or `model-reported` completion. Failed proof and ordinary responses without a completion decision leave the goal blocked. Paused goals still require explicit user controls. A completed goal leaves the footer but remains inspectable through `/goal` and session history.
 
 Session replacement, reload, and committed tree navigation discard unfinished task receipts. Late receipt results do not write or notify through the old session context or attach to a different branch. Cancelled navigation preserves the pending receipt.
 
@@ -111,6 +119,8 @@ Pending `/goal` commands discard their mutation if committed navigation changes
 
 Failed Git metadata reads keep the footer's last successful file counts and mark task changes unavailable. They are not treated as an empty repository.
 
+Passive Git scans compare content without running repository filters. CRLF text that normalizes to the indexed LF content stays unchanged, including with `text=auto eol=lf` or bare `eol=lf`. Git's `status` can still report a modification after a size-changing rewrite because of its index stat cache, even when `git diff` is empty.
+
 Completion sounds are off by default; change with `/notification` in TUI mode. The tab-title indicator requires a Nerd Font.
 
 ### Pi model and MCP settings
@@ -129,8 +139,16 @@ Use Pi's native `/thinking` selector for reasoning levels. Pi 1.0.2 can also app
 
 Only configure parameters the endpoint accepts. These settings apply to `openai-completions`, `openai-responses`, and `azure-openai-responses`, not legacy `openai-codex`. Missing levels inherit model defaults; request-level sampling parameters take precedence. `/codex-fast` preserves sampling parameters when it adds the priority service tier. See Pi's [sampling configuration](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/models.md#configure-sampling-by-thinking-level) for the full file format and merge rules.
 
+Pi 1.0.3 renames the Azure **provider** from `azure-openai-responses` to `azure` and adds Azure Foundry Chat Completions, including `azure/deepseek-v4-pro`. Rename the provider key in `auth.json` or run `/login azure` again, rename it in `models.json`, and update `defaultProvider`, `enabledModels` patterns, and `modelThinkingLevels` keys in `settings.json`. The Responses **API** identifier remains `azure-openai-responses`; Foundry Chat Completions uses `openai-completions`. The `AZURE_OPENAI_*` environment variables are unchanged. Old Azure sessions fall back to another model on resume and lose prompt-cache reuse. KillerOS does not migrate credentials or Pi settings. See Pi's [Azure configuration](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md#azure-openai).
+
+In Pi 1.0.4 codemode, `tools.read()` on an image returns an image block that you can pass to native `image()`. See Pi's [codemode tool calls](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/codemode.md#call-tools).
+
+When codemode is enabled, Pi 1.0.3's `image()` also saves images to temporary files and includes their paths in the result, so later turns can copy or move them. Pi creates these output files with user-only permissions on POSIX systems; Windows access follows inherited filesystem ACLs. KillerOS uses Pi's native codemode behavior; see [image generation](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/codemode.md#generate-images).
+
 Use Pi's `/mcp` command and `.pi/mcp.json` to enable, disable, or change the exposure of a user-level MCP server for a trusted project without copying its credentials. See [project MCP overrides](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/mcp.md#configure-servers). KillerOS does not manage MCP configuration.
 
+In Pi 1.0.4, a nonempty `--tools` list with no `mcp__` entry preserves MCP tool availability according to each server's exposure. For example, `--tools codemode` does not disable MCP connections or remove their tools. Use `--no-mcp` to disable MCP connections through Pi's built-in MCP support for one run. Explicit MCP selection and exclusion remain Pi's responsibility; see [MCP tool exposure](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/mcp.md#control-tool-exposure).
+
 ## Development
 
 Strict TypeScript throughout. Tests run on Node's built-in test runner:
@@ -147,9 +165,17 @@ npm ci && npm run check && npm test
 4. Merge using **Create a merge commit**. Squash and rebase merging are disabled because the automated `main`-to-`dev` sync requires shared ancestry. Avoid advancing `dev` until that sync finishes.
 5. Check the `main` CI run and the following Release run. Successful CI on the current `main` commit triggers npm publication with provenance, creates the GitHub release, and fast-forwards `dev` to the released commit. A green PR alone never publishes.
 
-`main` requires pull requests and the CI workflow's required checks, including for administrators. Force pushes and branch deletion are blocked. If CI job names change, update GitHub's required status checks to match.
+`main` requires pull requests and the CI workflow's required checks, including for administrators. Force pushes and branch deletion are blocked. Pi compatibility checks use the stable names `Pi compatibility (minimum)` and `Pi compatibility (latest)`, independent of the tested versions.
+
+When adopting these names, first confirm both checks passed on the release PR's exact head commit. Replace only the two old versioned Pi required checks with their corresponding stable names, preserving their GitHub Actions app bindings and every other protection.
+
+`Quality (Node 22.19.0)` validates current release metadata before the full test suite. It also tests the publication lifecycle with the release workflow's pinned npm CLI in an isolated installation. The packed-package Pi test checks the installed archive's entry point, README, changelog, theme, package name, and current version before activation and reload. These archive assertions run throughout the Node, Windows, and Pi compatibility matrix.
+
+`Package smoke test` remains during the required-check migration. Before removing it, land the replacement coverage and confirm the Node-floor quality check passed on the replacement PR's exact head commit and that the coverage is present on `main`. Inspect live branch protection and rulesets, obtain approval, and remove only the smoke requirement while preserving the required quality check's GitHub Actions binding and every other protection. Remove the job in a follow-up PR and recheck that PR's final head. Do not remove the job if these prerequisites are unmet.
+
+Releases go through CI on `main`; do not push version tags manually. The release workflow publishes the source directory with `npm publish . --ignore-scripts=false`. The package's `prepublishOnly` lifecycle invokes the existing release validator once at that boundary, even if an inherited npm setting disables scripts. A validator failure stops publication and release creation. The earlier verified-commit, provenance, version, tag, and release-note checks remain.
 
-Releases go through CI on `main`; do not push version tags manually. The package smoke job validates release metadata before packing, and the release workflow checks it again before publication. The prepublish check rejects ordinary direct `npm publish`, but `--ignore-scripts` can bypass it. Configure npm's trusted publisher for `release.yml`, set package publishing access to "Require two-factor authentication and disallow tokens", and revoke unused publish tokens. npm maintainers can still publish interactively with 2FA, so workflow-only publishing also depends on maintainer policy.
+The prepublish guard rejects ordinary direct publication, but it does not prove GitHub authorization. Disabling scripts or publishing a tarball bypasses `prepublishOnly`. Configure npm's trusted publisher for `release.yml`, set package publishing access to "Require two-factor authentication and disallow tokens", and revoke unused publish tokens. npm maintainers can still publish interactively with 2FA, so workflow-only publishing also depends on maintainer policy.
 
 ## Security
 

package/killeros/auto-compaction.ts

@@ -23,6 +23,7 @@ export interface AutoCompactionPreference {
 
 export interface AutoCompactionGoalHandlers {
   isActive(ctx: ExtensionContext): boolean;
+  hasAcceptedTerminalDecision(): boolean;
   onRequested(): void;
   onCompleted(ctx: ExtensionContext): void;
   onFailed(ctx: ExtensionContext, error: unknown): void;
@@ -246,7 +247,20 @@ export function registerAutoCompaction(
     goalRequestOwned = dependencies.goal?.isActive(ctx) === true;
   });
 
+  // Capture later blocked completion before goal settlement discards its permission.
+  // A compaction already pending as ordinary work must also stop after that decision.
+  const observeTerminalDecision = (): void => {
+    if (!dependencies.goal?.hasAcceptedTerminalDecision()) return;
+    goalRequestOwned = true;
+    if (request?.phase === "compacting") {
+      request.goal = true;
+      if (request.compactionCompleted) request = undefined;
+    }
+  };
+  pi.on("agent_end", observeTerminalDecision);
+
   pi.on("turn_end", (_event, ctx) => {
+    observeTerminalDecision();
     if (!supportedMode(ctx) || request) return;
 
     let preference: AutoCompactionPreference;
@@ -305,9 +319,10 @@ export function registerAutoCompaction(
     try {
       ctx.compact({
         onComplete: () => {
+          observeTerminalDecision();
           if (request?.phase !== "compacting" || request.token !== token) return;
           rearmOnNextLogicalRequest = preference.percentRemaining === 100;
-          if (goal) {
+          if (goal || request.goal) {
             request = undefined;
             if (activeGoal && dependencies.goal) {
               try {

package/killeros/footer.ts

@@ -280,7 +280,7 @@ function sumSessionCost(ctx: ExtensionContext): number {
 
 const PROVIDER_LABELS: Readonly<Record<string, string>> = {
   "amazon-bedrock": "Amazon Bedrock",
-  "azure-openai-responses": "Azure OpenAI",
+  azure: "Azure",
   "github-copilot": "GitHub Copilot",
   "google-vertex": "Google Vertex",
   "openai-codex": "OpenAI",

package/killeros/goal-interface.ts

@@ -13,7 +13,7 @@ import { safeTerminalText } from "./safe-terminal-text.ts";
 
 const GoalUpdateParams = Type.Object({
   status: StringEnum(["complete", "continue", "blocked"] as const, {
-    description: "Record exactly one active-goal decision: complete, continue, or blocked",
+    description: "Record an active-goal decision, or complete a blocked goal during an eligible later ordinary request",
   }),
   evidence: Type.String({
     minLength: 1,
@@ -148,7 +148,7 @@ export function registerGoalInterface(
   pi.registerTool<typeof GoalUpdateParams, GoalUpdateDetails>({
     name: GOAL_UPDATE_TOOL,
     label: "Goal update",
-    description: "Record exactly one active-goal decision: complete after verification, continue with evidence and one next action, or audit the same blocker before blocking it.",
+    description: "Record exactly one active-goal decision: complete after verification, continue with evidence and one next action, or audit the same blocker before blocking it. During an eligible later ordinary request for a blocked goal, only verified completion of the whole unchanged objective is permitted; it does not resume work or grant turns.",
     parameters: GoalUpdateParams,
     exposure: "model-only",
     executionMode: "sequential",
@@ -157,14 +157,27 @@ export function registerGoalInterface(
       if (!isGoalModeSupported(ctx)) throw new Error("KillerOS goals require TUI or RPC mode");
       if (!isSavedSession(ctx)) throw new Error("KillerOS goals require a saved session");
       const state = runtime.state;
-      if (!state || state.status !== "active") throw new Error("There is no active KillerOS goal to update");
-      if (!runtime.goalTurnInFlight
-        || runtime.goalTurn?.turn !== state.turns
-        || runtime.goalTurn.revision !== state.revision) {
-        throw new Error("A goal decision can only be recorded during an active KillerOS goal turn");
-      }
-      if (state.turnDecision !== undefined) {
-        throw new Error("Only one goal decision may be accepted per logical goal turn");
+      if (!state || state.status !== "active" && state.status !== "blocked") throw new Error("There is no active KillerOS goal to update");
+      const generation = runtime.lifecycleGeneration;
+      const permission = runtime.blockedCompletion;
+      const ownsDecision = (): boolean => runtime.state === state
+        && runtime.lifecycleGeneration === generation
+        && (state.status === "blocked"
+          ? permission?.kind === "eligible"
+            && runtime.blockedCompletion === permission
+            && permission.state === state
+            && permission.generation === generation
+          : runtime.goalTurnInFlight
+            && runtime.goalTurn?.turn === state.turns
+            && runtime.goalTurn.revision === state.revision
+            && state.turnDecision === undefined);
+      if (state.status === "blocked") {
+        if (!ownsDecision()) throw new Error("A blocked goal can only be completed during an eligible later ordinary request");
+        if (params.status !== "complete") throw new Error("A blocked goal accepts only complete; automatic continuation remains stopped");
+        if (state.turns === 0) throw new Error("A zero-turn blocked goal cannot record completion; no accounted goal turn exists");
+      } else {
+        if (state.turnDecision !== undefined) throw new Error("Only one goal decision may be accepted per logical goal turn");
+        if (!ownsDecision()) throw new Error("A goal decision can only be recorded during an active KillerOS goal turn");
       }
       if (params.status !== "complete" && params.status !== "continue" && params.status !== "blocked") {
         throw new Error("Goal update status is invalid");
@@ -175,11 +188,7 @@ export function registerGoalInterface(
       if (params.status === "complete") {
         if (state.verification) await verifyGoalDeliverable(state.verification);
         signal?.throwIfAborted();
-        if (runtime.state !== state
-          || runtime.goalTurn?.turn !== state.turns
-          || runtime.goalTurn.revision !== state.revision
-          || state.turnDecision !== undefined
-          || !runtime.goalTurnInFlight) throw new Error("Goal changed while completion was being verified");
+        if (!ownsDecision()) throw new Error("Goal changed while completion was being verified");
         const verification: "file" | "model-reported" = state.verification ? "file" : "model-reported";
         const decision = { kind: "complete" as const, turn: state.turns, evidence, verification };
         try {
@@ -188,6 +197,7 @@ export function registerGoalInterface(
           pauseGoalAfterFailure(pi, runtime, ctx, `goal completion could not be saved: ${error instanceof Error ? error.message : String(error)}`);
           throw error;
         }
+        if (state.status === "blocked") runtime.blockedCompletion = { kind: "accepted", generation };
         const text = state.verification
           ? `Goal verified complete at ${safeTerminalText(state.verification.path)}: ${evidence}`
           : `Goal marked complete (model-reported): ${evidence}`;
@@ -197,6 +207,7 @@ export function registerGoalInterface(
         };
       }
 
+      if (state.status !== "active") throw new Error("A blocked goal accepts only complete");
       if (params.status === "continue") {
         if (typeof params.nextAction !== "string") throw new Error("A continue decision requires a nextAction");
         const nextAction = normalizeGoalText(params.nextAction, GOAL_EVIDENCE_LIMIT, "Goal nextAction");

package/killeros/goal-runtime.ts

@@ -97,7 +97,11 @@ function setGoalUpdateToolActive(pi: ExtensionAPI, active: boolean): void {
 }
 
 export function syncGoalUpdateTool(pi: ExtensionAPI, runtime: GoalRuntime): void {
-  setGoalUpdateToolActive(pi, runtime.state?.status === "active");
+  const permission = runtime.blockedCompletion;
+  setGoalUpdateToolActive(pi, runtime.state?.status === "active"
+    || permission?.kind === "eligible"
+      && permission.state === runtime.state
+      && permission.generation === runtime.lifecycleGeneration);
 }
 
 function goalUpdateToolIsAvailable(pi: ExtensionAPI): boolean {
@@ -119,6 +123,7 @@ export function persistGoalState(
   const data: GoalEntryData = { version: GOAL_VERSION, event, state: state ?? null };
   pi.appendEntry(GOAL_ENTRY_TYPE, data);
   runtime.state = state;
+  runtime.blockedCompletion = undefined;
   if (event === "clear") {
     runtime.lifecycleGeneration += 1;
     runtime.continuationHeld = undefined;
@@ -372,6 +377,21 @@ function goalInstructions(state: GoalState, heading: string): string {
   ].join("\n");
 }
 
+function blockedGoalSystemPrompt(state: Extract<GoalState, { status: "blocked" }>): string {
+  return [
+    "# Blocked KillerOS goal",
+    "Status: blocked",
+    "Objective:",
+    state.objective,
+    "",
+    "Treat the exact saved objective above as authoritative. Automatic goal continuation remains stopped; this ordinary request grants no goal turns.",
+    "If killeros_goal_update is available, you may call it only with status complete and current evidence after auditing and verifying the whole unchanged objective.",
+    "File proof checks only the persisted deliverable against its original baseline, not every acceptance criterion. General-objective completion is model-reported.",
+    "The disappearance of a blocker alone does not prove completion. If useful work remains or this request is unrelated, finish normally without a goal update. No decision is mandatory.",
+    "Never use the goal tool to continue, audit a blocker, pause, resume, edit, replace, or clear this goal. User controls remain unchanged.",
+  ].join("\n");
+}
+
 function goalSystemPrompt(state: GoalState): string {
   return goalInstructions(state, "Active KillerOS goal");
 }
@@ -406,6 +426,7 @@ export function registerGoalRuntime(
     const generation = runtime.lifecycleGeneration;
     const restored = restoreGoalState(ctx);
     runtime.state = isGoalModeSupported(ctx) ? restored.state : undefined;
+    runtime.blockedCompletion = undefined;
     syncGoalUpdateTool(pi, runtime);
     runtime.continuationScheduled = false;
     runtime.continuationHeld = undefined;
@@ -434,6 +455,7 @@ export function registerGoalRuntime(
 
   pi.on("session_shutdown", (_event, ctx) => {
     runtime.lifecycleGeneration += 1;
+    runtime.blockedCompletion = undefined;
     if (runtime.state?.status === "active") {
       const checkpoint = checkpointActiveGoalState(runtime.state, Date.now());
       try {
@@ -453,7 +475,13 @@ export function registerGoalRuntime(
     runtime.continuationScheduled = false;
     if (!runtime.goalTurnInFlight && isGoalModeSupported(ctx) && isSavedSession(ctx) && pauseGoalAtTurnLimit(pi, runtime, ctx)) return;
     const current = runtime.state;
-    if (!isGoalModeSupported(ctx) || !isSavedSession(ctx) || !current || current.status !== "active") return;
+    if (!isGoalModeSupported(ctx) || !isSavedSession(ctx) || !current) return;
+    if (current.status === "blocked") {
+      runtime.blockedCompletion ??= { kind: "eligible", state: current, generation: runtime.lifecycleGeneration };
+      syncGoalUpdateTool(pi, runtime);
+      return { systemPrompt: `${event.systemPrompt}\n\n${blockedGoalSystemPrompt(current)}` };
+    }
+    if (current.status !== "active") return;
     if (runtime.goalTurnInFlight) {
       if (!goalUpdateToolIsAvailable(pi)) {
         pauseGoalBeforeTurn(pi, runtime, ctx);

package/killeros/goal-settlement.ts

@@ -220,6 +220,8 @@ export function registerGoalSettlement(
   runtime: GoalRuntime,
 ): AutoCompactionGoalHandlers {
   pi.on("agent_settled", (_event, ctx) => {
+    runtime.blockedCompletion = undefined;
+    syncGoalUpdateTool(pi, runtime);
     const wasGoalTurn = runtime.goalTurnInFlight;
     const continuationWasScheduled = runtime.continuationScheduled;
     const agentEndObserved = runtime.agentEndObserved;
@@ -311,6 +313,8 @@ export function registerGoalSettlement(
   });
 
   return {
+    hasAcceptedTerminalDecision: (): boolean => runtime.blockedCompletion?.kind === "accepted"
+      && runtime.blockedCompletion.generation === runtime.lifecycleGeneration,
     isActive: (ctx: ExtensionContext): boolean => isGoalModeSupported(ctx)
       && isSavedSession(ctx)
       && runtime.state?.status === "active",

package/killeros/passive-git-status.ts

@@ -325,7 +325,8 @@ function matchesIndex(content: Buffer, entry: IndexEntry, autoCrlf: string): boo
   if (blobObjectId(content, entry.objectId.length) === entry.objectId) return true;
   if (entry.mode === "120000" || !content.includes(Buffer.from("\r\n"))) return false;
   const attributes = entry.eolAttributes;
-  if (attributes.endsWith("eol=lf") || attributes === "-text") return false;
+  // eol controls checkout bytes; text files still normalize CRLF when compared with the index.
+  if (attributes === "-text") return false;
   const automatic = attributes.startsWith("text=auto") || !attributes;
   if (!attributes && !["true", "input"].includes(autoCrlf)) return false;
   if (automatic) {

package/killeros/runtime.ts

@@ -121,6 +121,10 @@ export interface GoalTurnExecution {
   revision: number;
 }
 
+export type BlockedGoalCompletion =
+  | { kind: "eligible"; state: Extract<GoalState, { status: "blocked" }>; generation: number }
+  | { kind: "accepted"; generation: number };
+
 export interface GoalRuntime {
   state?: GoalState;
   continuationScheduled: boolean;
@@ -128,6 +132,7 @@ export interface GoalRuntime {
   goalTurnInFlight: boolean;
   agentEndObserved: boolean;
   goalTurn?: GoalTurnExecution;
+  blockedCompletion?: BlockedGoalCompletion;
   automaticCompaction?: AutomaticGoalCompaction;
   persistenceRetryNeeded: boolean;
   lastStopReason?: string;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "killeros",
-  "version": "3.0.2",
+  "version": "3.0.4",
   "description": "TUI, goals, and workflow automation for the Pi coding agent",
   "type": "module",
   "keywords": [
@@ -45,16 +45,16 @@
     ]
   },
   "peerDependencies": {
-    "@earendil-works/pi-ai": ">=1.0.2 <2",
-    "@earendil-works/pi-coding-agent": ">=1.0.2 <2",
-    "@earendil-works/pi-tui": ">=1.0.2 <2",
+    "@earendil-works/pi-ai": ">=1.0.4 <2",
+    "@earendil-works/pi-coding-agent": ">=1.0.4 <2",
+    "@earendil-works/pi-tui": ">=1.0.4 <2",
     "typebox": ">=1.1.38 <2"
   },
   "devDependencies": {
-    "@earendil-works/pi-ai": "1.0.2",
-    "@earendil-works/pi-coding-agent": "1.0.2",
-    "@earendil-works/pi-server": "1.0.2",
-    "@earendil-works/pi-tui": "1.0.2",
+    "@earendil-works/pi-ai": "1.0.4",
+    "@earendil-works/pi-coding-agent": "1.0.4",
+    "@earendil-works/pi-server": "1.0.4",
+    "@earendil-works/pi-tui": "1.0.4",
     "@types/node": "24.12.4",
     "eslint": "^10.9.1",
     "typebox": "1.3.20",