@pentoshi/clai 3.7.4 → 3.7.6

package/README.md

@@ -117,6 +117,7 @@ Inside a session:
 | **Network** | `net.scan` (nmap wrapper, SYN with privilege / TCP fallback), `net.context`, `net.pingSweep`, `pentest.recon` (whois + dig + top ports) |
 | **HTTP evidence** | `http.fetch` — status, headers, cookies, TLS, body for **raw protocol / pentest** work (not casual page reading) |
 | **Web reading / OSINT** | `web.search`, `web.fetch` (readable pages), plus shell for specialized CLIs |
+| **Batch recon** | `tool.batch` — up to 20 lookups in one call; optional `on_fail` fail-fast or selective cancel when dependents need a prerequisite |
 | **Discovery** | `tool.check`, `pkg.install`, `wordlist.find` — install only what is missing; locate wordlists per OS (no Kali-only path guesses) |
 | **DNS / ownership** | `dns.lookup`, `whois.lookup` for narrow questions |
 | **Shell** | Full toolbox: `nmap`, `ffuf`, `gobuster`, `feroxbuster`, `sqlmap`, `hydra`, `nikto`, `masscan`, `nuclei`, `tshark`, … via `shell.exec` |
@@ -156,18 +157,19 @@ Default posture is **non-destructive proof**. Escalate impact only when you ask
 - Full plan + notes pager (`Ctrl+P` / `/plan`)  
 - Approve with `/implement`, revise in chat, cancel with `/discard`  
 - Plans **survive context compaction** and **reload with `/history`**  
+- Agent mode: working checklist + evidence-before-done; plan mode: roadmap you approve before execution  
 
 ---
 
 ## Terminal UI (operator console)
 
-Full-screen UI by default: streaming chat, tool cards, plan pane, pickers, history, secure key prompts. Falls back to a classic line REPL if the terminal cannot host the UI.
+Full-screen OpenTUI console by default: streaming chat, nested tool cards (including `tool.batch` sub-sections), file diffs, plan pane, pickers, history, secure key prompts. Falls back to a classic line REPL if the terminal cannot host the UI.
 
 | Action | How |
 |--------|-----|
 | Send | `Enter` |
 | Newline | `Shift+Enter` |
-| Abort turn | `Esc` |
+| Abort turn | `Esc` / `Ctrl+C` (cancels in-flight tools cleanly) |
 | Expand thinking | `Ctrl+T` (clickable on status strip) |
 | Expand tool / compacted output | `Ctrl+O` |
 | Plan pane | `Ctrl+H` |
@@ -176,9 +178,9 @@ Full-screen UI by default: streaming chat, tool cards, plan pane, pickers, histo
 | Commands / files | `/` · `@` |
 | Exit | `Ctrl+C` twice |
 
-Tool cards show **command/input** clearly and keep long scan tails in **OUTPUT** (expand or open pager). Compaction cards preserve engagement memory without dropping the plan.
+Tool cards show **command/input** clearly and keep long scan tails in **OUTPUT** (expand or open pager). File writes show a **diff preview**. Deletes always ask for confirmation (`y`/`n`); press **`v`** to preview the path before confirming. Compaction cards preserve engagement memory without dropping the plan.
 
-**`/history`** restores full sessions — prompts, tools, findings context, and the matching plan when present.
+**`/history`** restores full sessions — prompts, tool results, findings context, and the matching plan when present (including after abort / autosave).
 
 ---
 
@@ -198,14 +200,18 @@ Coding and general sysadmin work use the same agent (scaffold, debug, packages)
 ## Features (summary)
 
 - **Pentest-first agent loop** — recon-before-plan, stack-aware enum, evidence-backed findings  
-- **Durable plans** — `plan.create` / `task.update`, side pane, approve/refine/discard  
-- **11 LLM providers** with streaming (many free tiers + local Ollama)  
-- **Safety gate** + pentest authorization + optional engagement scope  
+- **Durable plans** — `plan.create` / `task.update`, side pane, approve/refine/discard; agent vs plan task workflows  
+- **Parallel multi-tool turns** — independent reads run together; failures do **not** cancel siblings by default  
+- **`tool.batch` fail policy** — opt-in `on_fail=cancel_pending` or selective `cancel_on_fail` / rules when later work depends on earlier success  
+- **Native + text tool calling** — `toolCalling: auto|native|text`  
+- **11+ LLM providers** with streaming (free tiers + local Ollama)  
+- **Safety gate** + pentest authorization + optional engagement scope (deletes always confirm with preview)  
 - **OS-aware** installs and wordlist discovery (macOS / Linux / Windows)  
 - **Context compaction** (auto + `/compact`) that keeps the plan alive  
-- **Session history** with full transcript restore  
+- **Session history** with full transcript + plan restore  
 - **Background jobs** for long scanners and listeners  
 - **Web OSINT** — `web.search` / `web.fetch` alongside raw `http.fetch`  
+- **Stall / cancel robustness** — tool heartbeats, hard deadlines, clean Esc abort with results recorded  
 
 ---
 
@@ -265,15 +271,44 @@ CLI mirrors: `clai authorize-pentest`, `clai scope add`, `clai doctor` (missing
 
 | Tool | Role in engagements |
 |------|---------------------|
-| `shell.exec` / `shell.start` | nmap, ffuf, sqlmap, hydra, custom PoCs, listeners |
-| `net.scan` · `net.context` · `pentest.recon` | Host/port/service discovery |
+| `shell.exec` / `shell.start` · `shell.jobs` / `tail` / `stop` | nmap, ffuf, sqlmap, hydra, custom PoCs, listeners, background jobs |
+| `net.scan` · `net.context` · `net.pingSweep` · `pentest.recon` | Host/port/service discovery |
 | `http.fetch` | Raw HTTP/TLS evidence |
 | `web.search` · `web.fetch` | OSINT / docs (readable), not raw exploit traffic |
 | `dns.lookup` · `whois.lookup` | Narrow DNS / ownership |
+| `tool.batch` | Fan-out up to 20 tools; `concurrency` 1–6; `on_fail` continue (default) / cancel_pending / rules; per-call `cancel_on_fail` |
 | `tool.check` · `pkg.install` · `wordlist.find` | Tooling readiness |
-| `fs.*` | Loot, notes, report files (sandboxed roots) |
-| `plan.create` · `task.update` | Engagement checklist |
+| `fs.read` · `fs.list` · `fs.search` · `fs.write` · `fs.writeMany` · `fs.edit` · `fs.replaceLines` · `fs.append` · `fs.delete` | Files (sandboxed roots; delete always confirms + optional preview) |
+| `plan.create` · `task.update` | Engagement checklist / working tasks |
 | `sysinfo` · `image.ocr` · `pdf.read` | Host context, report/screenshot OCR |
+| `agent.handoff` | Ask mode → offer agent mode when the user wants action, not explanation |
+
+### `tool.batch` fail policy (opt-in)
+
+Default is **continue** — one failed lookup never kills the rest (best for recon).
+
+```json
+// Fail-fast: stop remaining calls after the first failure
+{"name":"tool.batch","args":{
+  "on_fail":"cancel_pending",
+  "calls":[
+    {"name":"net.scan","args":{"target":"lab.example"}},
+    {"name":"http.fetch","args":{"url":"https://lab.example/"}}
+  ]
+}}
+
+// Selective: if scan fails, cancel only fuzz (dns still runs)
+{"name":"tool.batch","args":{
+  "calls":[
+    {"id":"dns","name":"dns.lookup","args":{"target":"lab.example"}},
+    {"id":"scan","name":"net.scan","args":{"target":"lab.example"},
+     "cancel_on_fail":["fuzz"]},
+    {"id":"fuzz","name":"shell.exec","args":{"command":"ffuf …"}}
+  ]
+}}
+```
+
+Top-level multi-tool messages (several separate tool blocks) never cancel siblings; use `tool.batch` when you need a fail policy.
 
 ### Search providers (OSINT)
 
@@ -334,10 +369,10 @@ Node.js ≥ 20.
 Tag-driven CI (`.github/workflows/release.yml`): tests → multi-platform binaries → GitHub Release → npm `@pentoshi/clai` → Homebrew tap.
 
 ```sh
-npm version 2.0.34 --no-git-tag-version
+npm version 3.7.6 --no-git-tag-version
 # bump FALLBACK_VERSION / manifests as needed
-git commit -am "v2.0.34" && git push origin main
-git tag -a v2.0.34 -m "clai v2.0.34" && git push origin v2.0.34
+git commit -am "v3.7.6" && git push origin main
+git tag -a v3.7.6 -m "clai v3.7.6" && git push origin v3.7.6
 ```
 
 Secrets: `NPM_TOKEN`, `TAP_GITHUB_TOKEN`. Optional: `NPM_PROVENANCE=true`.
@@ -352,13 +387,13 @@ clai/
 │  ├─ index.ts              # CLI entry
 │  ├─ modes/                # ask · agent
 │  ├─ agent/                # loop, plans, compaction, tool parsing
-│  ├─ llm/                  # providers + streaming
-│  ├─ tools/                # shell, net, http, web, fs, pentest, …
+│  ├─ llm/                  # providers + streaming + native tools
+│  ├─ tools/                # shell, net, http, web, fs, batch, pentest, …
 │  ├─ safety/               # classifier + patterns
 │  ├─ store/                # config, history, keys, plans, scope, logs
-│  ├─ tui/                  # full-screen terminal UI
+│  ├─ tui-v2/               # full-screen OpenTUI (primary)
 │  ├─ app/                  # session, commands, events
-│  └─ prompts/              # agent methodology (incl. pentest)
+│  └─ prompts/              # agent methodology (incl. pentest; embedded for bun)
 ├─ bin/clai.mjs
 ├─ install/ · manifests/
 └─ package.json

package/dist/agent/confirm-port.js

@@ -77,8 +77,9 @@ export async function ensurePentestAuthorization(call, autoConfirm, session, con
     return true;
 }
 export async function confirmToolExecution(call, autoConfirm, session, confirmPort, options) {
-    // Outside-cwd writes / deletes always prompt — even under allow-all / -y.
-    if (options?.forceConfirm) {
+    // fs.delete + outside-cwd mutates always prompt — even under allow-all / -y.
+    // Deletion is irreversible; never auto-approve.
+    if (options?.forceConfirm || call.name === "fs.delete") {
         return confirmPort.confirmTool(call);
     }
     const config = getConfig();

package/dist/agent/confirm-port.js.map

@@ -1 +1 @@
-{"version":3,"file":"confirm-port.js","sourceRoot":"","sources":["../../src/agent/confirm-port.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,OAAO,EAAE,MAAM,mBAAmB,CAAC;AAC5C,OAAO,KAAK,MAAM,OAAO,CAAC;AAC1B,OAAO,EAAE,SAAS,EAAE,MAAM,oBAAoB,CAAC;AAC/C,OAAO,EAAE,iBAAiB,EAAE,MAAM,yBAAyB,CAAC;AAG5D,OAAO,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAC;AAkBvD;;;;;;;;GAQG;AACH,MAAM,UAAU,uBAAuB;IACrC,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC,KAAK;QAAE,OAAO;IACjC,IAAI,CAAC;QACH,IAAI,CAAE,OAAO,CAAC,KAAiD,CAAC,KAAK,EAAE,CAAC;YACtE,OAAO,CAAC,KAAK,CAAC,UAAU,CAAC,IAAI,CAAC,CAAC;QACjC,CAAC;QACD,OAAO,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC;IACzB,CAAC;IAAC,MAAM,CAAC;QACP,YAAY;IACd,CAAC;AACH,CAAC;AAED,MAAM,CAAC,MAAM,mBAAmB,GAAgB;IAC9C,KAAK,CAAC,WAAW,CAAC,IAAc;QAC9B,OAAO,OAAO,CAAC;YACb,OAAO,EAAE,KAAK,CAAC,MAAM,CAAC,SAAS,IAAI,CAAC,IAAI,KAAK,cAAc,CAAC,IAAI,CAAC,GAAG,CAAC;YACrE,OAAO,EAAE,IAAI;SACd,CAAC,CAAC;IACL,CAAC;IACD,KAAK,CAAC,cAAc;QAClB,OAAO,OAAO,CAAC;YACb,OAAO,EAAE,KAAK,CAAC,GAAG,CAChB,0HAA0H,CAC3H;YACD,OAAO,EAAE,KAAK;SACf,CAAC,CAAC;IACL,CAAC;IACD,KAAK,CAAC,eAAe,CAAC,KAAa;QACjC,OAAO,OAAO,CAAC;YACb,OAAO,EAAE,KAAK,CAAC,MAAM,CAAC,KAAK,KAAK,4BAA4B,CAAC;YAC7D,OAAO,EAAE,IAAI;SACd,CAAC,CAAC;IACL,CAAC;IACD,KAAK,CAAC,kBAAkB,CAAC,IAGxB;QACC,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC;QACzE,OAAO,OAAO,CAAC;YACb,OAAO,EAAE,KAAK,CAAC,MAAM,CACnB,0BAA0B,KAAK,uBAAuB,CACvD;YACD,OAAO,EAAE,IAAI;SACd,CAAC,CAAC;IACL,CAAC;CACF,CAAC;AAEF,MAAM,CAAC,KAAK,UAAU,0BAA0B,CAC9C,IAAc,EACd,WAAoB,EACpB,OAAsB,EACtB,WAAwB;IAExB,IAAI,CAAC,iBAAiB,CAAC,IAAI,CAAC;QAAE,OAAO,IAAI,CAAC;IAC1C,MAAM,MAAM,GAAG,SAAS,EAAE,CAAC;IAC3B,IAAI,MAAM,CAAC,WAAW,KAAK,WAAW;QAAE,OAAO,IAAI,CAAC;IACpD,6DAA6D;IAC7D,IAAI,MAAM,CAAC,iBAAiB;QAAE,OAAO,IAAI,CAAC;IAC1C,+DAA+D;IAC/D,IAAI,OAAO,CAAC,iBAAiB,CAAC,KAAK;QAAE,OAAO,IAAI,CAAC;IAEjD,IAAI,WAAW,EAAE,CAAC;QAChB,sEAAsE;QACtE,oEAAoE;QACpE,OAAO,CAAC,iBAAiB,CAAC,KAAK,GAAG,IAAI,CAAC;QACvC,OAAO,IAAI,CAAC;IACd,CAAC;IAED,MAAM,EAAE,GAAG,MAAM,WAAW,CAAC,cAAc,EAAE,CAAC;IAC9C,IAAI,CAAC,EAAE;QAAE,OAAO,KAAK,CAAC;IACtB,OAAO,CAAC,iBAAiB,CAAC,KAAK,GAAG,IAAI,CAAC;IACvC,OAAO,IAAI,CAAC;AACd,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACxC,IAAc,EACd,WAAoB,EACpB,OAAsB,EACtB,WAAwB,EACxB,OAAgD;IAEhD,0EAA0E;IAC1E,IAAI,OAAO,EAAE,YAAY,EAAE,CAAC;QAC1B,OAAO,WAAW,CAAC,WAAW,CAAC,IAAI,CAAC,CAAC;IACvC,CAAC;IACD,MAAM,MAAM,GAAG,SAAS,EAAE,CAAC;IAC3B,IAAI,MAAM,CAAC,WAAW,KAAK,WAAW;QAAE,OAAO,IAAI,CAAC;IACpD,IAAI,WAAW;QAAE,OAAO,IAAI,CAAC;IAC7B,IAAI,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,OAAO,IAAI,CAAC;IAC9C,uEAAuE;IACvE,wEAAwE;IACxE,qDAAqD;IACrD,IAAI,MAAM,CAAC,gBAAgB,CAAC,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,OAAO,IAAI,CAAC;IAE7D,OAAO,WAAW,CAAC,WAAW,CAAC,IAAI,CAAC,CAAC;AACvC,CAAC"}
+{"version":3,"file":"confirm-port.js","sourceRoot":"","sources":["../../src/agent/confirm-port.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,OAAO,EAAE,MAAM,mBAAmB,CAAC;AAC5C,OAAO,KAAK,MAAM,OAAO,CAAC;AAC1B,OAAO,EAAE,SAAS,EAAE,MAAM,oBAAoB,CAAC;AAC/C,OAAO,EAAE,iBAAiB,EAAE,MAAM,yBAAyB,CAAC;AAG5D,OAAO,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAC;AAkBvD;;;;;;;;GAQG;AACH,MAAM,UAAU,uBAAuB;IACrC,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC,KAAK;QAAE,OAAO;IACjC,IAAI,CAAC;QACH,IAAI,CAAE,OAAO,CAAC,KAAiD,CAAC,KAAK,EAAE,CAAC;YACtE,OAAO,CAAC,KAAK,CAAC,UAAU,CAAC,IAAI,CAAC,CAAC;QACjC,CAAC;QACD,OAAO,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC;IACzB,CAAC;IAAC,MAAM,CAAC;QACP,YAAY;IACd,CAAC;AACH,CAAC;AAED,MAAM,CAAC,MAAM,mBAAmB,GAAgB;IAC9C,KAAK,CAAC,WAAW,CAAC,IAAc;QAC9B,OAAO,OAAO,CAAC;YACb,OAAO,EAAE,KAAK,CAAC,MAAM,CAAC,SAAS,IAAI,CAAC,IAAI,KAAK,cAAc,CAAC,IAAI,CAAC,GAAG,CAAC;YACrE,OAAO,EAAE,IAAI;SACd,CAAC,CAAC;IACL,CAAC;IACD,KAAK,CAAC,cAAc;QAClB,OAAO,OAAO,CAAC;YACb,OAAO,EAAE,KAAK,CAAC,GAAG,CAChB,0HAA0H,CAC3H;YACD,OAAO,EAAE,KAAK;SACf,CAAC,CAAC;IACL,CAAC;IACD,KAAK,CAAC,eAAe,CAAC,KAAa;QACjC,OAAO,OAAO,CAAC;YACb,OAAO,EAAE,KAAK,CAAC,MAAM,CAAC,KAAK,KAAK,4BAA4B,CAAC;YAC7D,OAAO,EAAE,IAAI;SACd,CAAC,CAAC;IACL,CAAC;IACD,KAAK,CAAC,kBAAkB,CAAC,IAGxB;QACC,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC;QACzE,OAAO,OAAO,CAAC;YACb,OAAO,EAAE,KAAK,CAAC,MAAM,CACnB,0BAA0B,KAAK,uBAAuB,CACvD;YACD,OAAO,EAAE,IAAI;SACd,CAAC,CAAC;IACL,CAAC;CACF,CAAC;AAEF,MAAM,CAAC,KAAK,UAAU,0BAA0B,CAC9C,IAAc,EACd,WAAoB,EACpB,OAAsB,EACtB,WAAwB;IAExB,IAAI,CAAC,iBAAiB,CAAC,IAAI,CAAC;QAAE,OAAO,IAAI,CAAC;IAC1C,MAAM,MAAM,GAAG,SAAS,EAAE,CAAC;IAC3B,IAAI,MAAM,CAAC,WAAW,KAAK,WAAW;QAAE,OAAO,IAAI,CAAC;IACpD,6DAA6D;IAC7D,IAAI,MAAM,CAAC,iBAAiB;QAAE,OAAO,IAAI,CAAC;IAC1C,+DAA+D;IAC/D,IAAI,OAAO,CAAC,iBAAiB,CAAC,KAAK;QAAE,OAAO,IAAI,CAAC;IAEjD,IAAI,WAAW,EAAE,CAAC;QAChB,sEAAsE;QACtE,oEAAoE;QACpE,OAAO,CAAC,iBAAiB,CAAC,KAAK,GAAG,IAAI,CAAC;QACvC,OAAO,IAAI,CAAC;IACd,CAAC;IAED,MAAM,EAAE,GAAG,MAAM,WAAW,CAAC,cAAc,EAAE,CAAC;IAC9C,IAAI,CAAC,EAAE;QAAE,OAAO,KAAK,CAAC;IACtB,OAAO,CAAC,iBAAiB,CAAC,KAAK,GAAG,IAAI,CAAC;IACvC,OAAO,IAAI,CAAC;AACd,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACxC,IAAc,EACd,WAAoB,EACpB,OAAsB,EACtB,WAAwB,EACxB,OAAgD;IAEhD,6EAA6E;IAC7E,gDAAgD;IAChD,IAAI,OAAO,EAAE,YAAY,IAAI,IAAI,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;QACvD,OAAO,WAAW,CAAC,WAAW,CAAC,IAAI,CAAC,CAAC;IACvC,CAAC;IACD,MAAM,MAAM,GAAG,SAAS,EAAE,CAAC;IAC3B,IAAI,MAAM,CAAC,WAAW,KAAK,WAAW;QAAE,OAAO,IAAI,CAAC;IACpD,IAAI,WAAW;QAAE,OAAO,IAAI,CAAC;IAC7B,IAAI,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,OAAO,IAAI,CAAC;IAC9C,uEAAuE;IACvE,wEAAwE;IACxE,qDAAqD;IACrD,IAAI,MAAM,CAAC,gBAAgB,CAAC,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,OAAO,IAAI,CAAC;IAE7D,OAAO,WAAW,CAAC,WAAW,CAAC,IAAI,CAAC,CAAC;AACvC,CAAC"}

package/dist/agent/runner.js

@@ -34,7 +34,7 @@ import { stripSentinelTokens, parseToolCall, recognizeBareToolJson, looksLikeTru
 import { createSessionPolicy, isPreApprovalAllowedTool, isPlanModeAllowedShellCommand, isPlanModeAllowedTool, isPlanApprovedByStatus, planHasOpenWork, isAbortError, shouldEnableImageOcr, } from "./session-policy.js";
 import { saveToolOutput, summarizeOutput, formatToolContext, } from "./tool-output-formatting.js";
 import { renderPlanForTerminal, planContextMessage, handlePlanTool, resolvePlanTaskId, } from "./plan-tool.js";
-import { absorbLooseWorkIntoLedger, applyDestinationCwd, canMarkTaskDone, hasLocalRuntimeProof, hasRemoteWorkProof, isBatchSoftFailTool, isDevServerCall, isEvidenceWorkTool, isFeatureImplementationCall, isPackageInstallCommand, isPlanPreflightTool, isPortListeningOutput, isReadOnlyReconTool, isRemoteActiveTestCall, isRemoteObservationTask, isRemoteReconToolCall, isRuntimeObservationTask, isScaffoldCreateCommand, isServerReadyOutput, ledgerFromTaskEvidence, pickPendingTaskForToolCall, recordTaskWorkSuccess, resolveUserDestinationHint, taskEvidenceFromLedger, TOOL_ABORT_GRACE_MS, toolHardBudgetMs, toolStallBudgetMs, userAskedForFeatureApp, } from "./task-evidence.js";
+import { absorbLooseWorkIntoLedger, applyDestinationCwd, canMarkTaskDone, hasLocalRuntimeProof, hasRemoteWorkProof, isDevServerCall, isEvidenceWorkTool, isFeatureImplementationCall, isPackageInstallCommand, isPlanPreflightTool, isPortListeningOutput, isReadOnlyReconTool, isRemoteActiveTestCall, isRemoteObservationTask, isRemoteReconToolCall, isRuntimeObservationTask, isScaffoldCreateCommand, isServerReadyOutput, ledgerFromTaskEvidence, pickPendingTaskForToolCall, recordTaskWorkSuccess, resolveUserDestinationHint, taskEvidenceFromLedger, TOOL_ABORT_GRACE_MS, toolHardBudgetMs, toolStallBudgetMs, userAskedForFeatureApp, } from "./task-evidence.js";
 import { buildSessionStateBlock, inferNextHint, upsertSessionStateMessage, } from "./session-state.js";
 import { detectPackageManager } from "./workspace-orient.js";
 import { budgetRemaining, consumeBudget, createRecoveryBudgets, freestyleClaimsAppReady, looksLikeShallowPentestReport, recoveryForErrorDiagnosis, recoveryForFailedProbe, recoveryForFreshness, recoveryForMissingFeature, recoveryForMissingPlan, recoveryForNarration, recoveryForPrematureComplete, recoveryForRuntimeVerify, recoveryForShallowPentest, } from "./must-continue.js";
@@ -2861,7 +2861,7 @@ export async function runAgentTurn(prompt, options = {}) {
                 // Notice BEFORE tool cards so the transcript reads:
                 // thinking → response → "N tool calls…" → tool cards (not tools then info).
                 if (allCalls.length > 1) {
-                    writeNotice("info", `${allCalls.length} tool calls in this message — running scoped (independent read-only lookups in parallel, everything else in order)`, chalk.dim(`  ℹ ${allCalls.length} tool calls — read-only lookups in parallel, the rest in order\n`));
+                    writeNotice("info", `${allCalls.length} tool calls in this message — read-only in parallel, writes in order (failures do not cancel siblings)`, chalk.dim(`  ℹ ${allCalls.length} tool calls — parallel reads, ordered writes; failures continue\n`));
                 }
                 for (const deferred of activeDeferredToolCalls.slice(0, allCalls.length)) {
                     if (!deferred.call.name || deferred.call.name === "…")
@@ -2897,24 +2897,19 @@ export async function runAgentTurn(prompt, options = {}) {
                         return false;
                     }
                 };
-                /** Tools whose failure must NOT cancel sibling calls in this turn. */
-                const shouldSoftFailTool = (name) => {
-                    if (isBatchSoftFailTool(name))
-                        return true;
-                    if (BATCH_SAFE_TOOLS.has(name))
-                        return true;
-                    return false;
-                };
                 // Recon waves often emit 6–10 lookups; 4 forced a second sequential wave.
                 const PARALLEL_LIMIT = 8;
                 let aborted = false;
-                let blocked = false;
-                let blockedResult = null;
-                let failed = false;
                 let awaitingPlanApproval = false;
                 /** Native tool_call ids that already have a role:tool history entry. */
                 const recordedNativeIds = new Set();
-                const recordResult = (boundCall, res, continueAfterFailure = false) => {
+                /**
+                 * Record a tool result into history. Failures / user declines are
+                 * always returned to the model — we never cancel later siblings or
+                 * force-end the turn as "blocked" because one tool failed. Only an
+                 * explicit user abort stops remaining calls.
+                 */
+                const recordResult = (boundCall, res) => {
                     recordedNativeIds.add(boundCall.id);
                     const toolContent = `Tool ${res.call.name} result (exit=${res.result.exitCode ?? 0}, ok=${res.result.ok}):\n${res.contextOutput}`;
                     if (historyNativeCalls.length) {
@@ -3021,22 +3016,14 @@ export async function runAgentTurn(prompt, options = {}) {
                             session.planApproved.value = true;
                         }
                     }
+                    // User Esc/Ctrl+C only — never cancel siblings because a delete failed
+                    // or a confirm was declined; the model must see every tool result.
                     if (res.lastAnswer === "Aborted.")
                         aborted = true;
-                    else if (res.blockOrCancel) {
-                        blocked = true;
-                        blockedResult = res;
-                    }
-                    else if (!res.ok && !continueAfterFailure)
-                        failed = true;
                 };
                 const groups = groupToolCallsForExecution(allCalls, isParallelSafe, PARALLEL_LIMIT);
                 for (const group of groups) {
-                    if (aborted ||
-                        blocked ||
-                        failed ||
-                        awaitingPlanApproval ||
-                        governorPauseReason)
+                    if (aborted || awaitingPlanApproval || governorPauseReason)
                         break;
                     if (group.length === 1) {
                         const call = group[0];
@@ -3048,8 +3035,7 @@ export async function runAgentTurn(prompt, options = {}) {
                         }
                         const id = callIds[bc.index];
                         const res = await executeSingleTool(call, id, options.signal || new AbortController().signal);
-                        const softFail = shouldSoftFailTool(call.name);
-                        recordResult(bc, res, softFail);
+                        recordResult(bc, res);
                     }
                     else {
                         // Concurrent group — BoundCall via Map; record in document order.
@@ -3067,11 +3053,12 @@ export async function runAgentTurn(prompt, options = {}) {
                         }
                         const results = await Promise.all(groupBound.map((bc, k) => executeSingleTool(bc.call, uiIds[k], options.signal || new AbortController().signal)));
                         for (let k = 0; k < results.length; k += 1) {
-                            recordResult(groupBound[k], results[k], true);
+                            recordResult(groupBound[k], results[k]);
                         }
                     }
                 }
                 // Cards still "running" get a terminal UI result; history always pairs.
+                // Only abort / plan-gate / governor leave calls un-run now.
                 for (let i = 0; i < toRun.length; i += 1) {
                     const bc = toRun[i];
                     if (recordedNativeIds.has(bc.id))
@@ -3082,15 +3069,11 @@ export async function runAgentTurn(prompt, options = {}) {
                     const uiId = callIds[i];
                     const reason = aborted
                         ? "Cancelled — turn aborted before this call ran."
-                        : blocked
-                            ? "Cancelled — earlier tool was blocked or declined."
-                            : awaitingPlanApproval
-                                ? "Deferred — waiting for plan approval."
-                                : governorPauseReason
-                                    ? `Deferred — progress governor paused execution: ${governorPauseReason}`
-                                    : failed
-                                        ? "Cancelled — earlier tool in this batch failed."
-                                        : "Cancelled — not executed.";
+                        : awaitingPlanApproval
+                            ? "Deferred — waiting for plan approval."
+                            : governorPauseReason
+                                ? `Deferred — progress governor paused execution: ${governorPauseReason}`
+                                : "Cancelled — not executed.";
                     const result = {
                         ok: false,
                         output: reason,
@@ -3138,13 +3121,8 @@ export async function runAgentTurn(prompt, options = {}) {
                     writeAbort();
                     return finishTurn(lastAnswer, productiveSteps, "aborted");
                 }
-                if (blocked && blockedResult) {
-                    lastAnswer = blockedResult.lastAnswer || "Blocked or Cancelled.";
-                    outcomeState.outcome.status = "blocked";
-                    await saveOutcomeState(outcomeState);
-                    moveTurn("blocked", lastAnswer);
-                    return finishTurn(lastAnswer, productiveSteps, "blocked");
-                }
+                // Confirm declines / tool failures already have role:tool results —
+                // continue the agent loop so the model can adapt (do not force "blocked").
                 await maybeAutoCompact("post-tool-token-budget");
                 if (options.onMessages) {
                     try {