arkaos 4.39.0 → 4.41.0

package/THE-ARKAOS-GUIDE.md

@@ -1,6 +1,6 @@
 # The ArkaOS Guide
 
-> v4.39.0 — 89 agents, 17 departments, 332 skills, 297 commands, 17 ADRs.
+> v4.41.0 — 89 agents, 17 departments, 332 skills, 297 commands, 18 ADRs.
 > One file, everything you need to start. Generated by `scripts/guide_gen.py` — never hand-edited.
 
 ## What it is

package/VERSION

@@ -1 +1 @@
-4.39.0
+4.41.0

package/bin/arka-doctor

@@ -162,9 +162,10 @@ if [ "$FIX_MODE" = true ] && [ "$HOOKS_OK" = false ]; then
       --arg pc "$HOOKS_DIR/pre-compact.sh" \
       --arg ptu "$HOOKS_DIR/post-tool-use.sh" \
       '{
-        "UserPromptSubmit": [{"hooks":[{"type":"command","command":$ups,"timeout":10}]}],
+        "UserPromptSubmit": [{"hooks":[{"type":"command","command":$ups,"timeout":20}]}],
         "PreCompact": [{"hooks":[{"type":"command","command":$pc,"timeout":30}]}],
-        "PostToolUse": [{"hooks":[{"type":"command","command":$ptu,"timeout":5}]}]
+        "PostToolUse": [{"hooks":[{"type":"command","command":$ptu,"timeout":5}]}],
+        "PostToolUseFailure": [{"hooks":[{"type":"command","command":$ptu,"timeout":5}]}]
       }')
     jq --argjson hooks "$HOOKS_JSON" '.hooks = $hooks' "$SETTINGS_FILE" > "$SETTINGS_FILE.tmp" && \
       mv "$SETTINGS_FILE.tmp" "$SETTINGS_FILE"

package/config/hooks/_lib/fastpath/engine.cjs

@@ -17,7 +17,10 @@
  * this file — that is exactly the drift the manifest exists to kill.
  */
 
-const SCHEMA_VERSION = 1;
+// v2: post.success_exit_codes removed with the PostToolUseFailure
+// wiring — an old engine + new manifest mismatch now delegates via
+// validateManifest instead of crashing into the shim's fail-open {}.
+const SCHEMA_VERSION = 2;
 
 /** Python truthiness — NOT JS Boolean(): [] and {} are falsy in Python,
  *  "false" (non-empty string) is truthy in both. flow_enforcer coerces
@@ -255,6 +258,34 @@ function flowAuthFresh(sid, manifest, ctx) {
   return nowSeconds - ts <= manifest.numbers.auth_ttl_seconds;
 }
 
+/** Textual output of a PostToolUse tool_response — parity with the
+ *  PostToolUse paths of _tool_text in post_tool_use.py (string,
+ *  stdout/stderr and Task/content shapes, isAsync guard and Python's
+ *  `or ""` falsy coercion included). Failure-event payloads never
+ *  reach this — decidePost delegates PostToolUseFailure outright. */
+function toolText(payload) {
+  const response = payload.tool_response;
+  if (typeof response === "string") return response;
+  if (response && typeof response === "object") {
+    if ("stdout" in response || "stderr" in response) {
+      return [response.stdout, response.stderr]
+        .map((part) => String(part || ""))
+        .filter(Boolean)
+        .join("\n");
+    }
+    if (response.isAsync) return "";
+    const content = response.content;
+    if (typeof content === "string") return content;
+    if (Array.isArray(content)) {
+      return content
+        .filter((block) => block && block.type === "text")
+        .map((block) => String(block.text || ""))
+        .join("");
+    }
+  }
+  return "";
+}
+
 /**
  * PostToolUse decision. Returns
  *   {action:"fast-exit", stdout:"{}", writes:[...]} — mcp telemetry only
@@ -267,18 +298,23 @@ function decidePost(payload, manifest, ctx) {
   }
   const toolName = String(payload.tool_name || "");
 
+  // A failing tool throws inside the runtime and fires
+  // PostToolUseFailure (2.1.220 contract) — failed-command turns are
+  // the gotchas pipeline's primary material, so they always go to
+  // Python. Rare by construction: no fast path is worth having here.
+  if (payload.hook_event_name === "PostToolUseFailure") {
+    return { action: "delegate", reason: "failure-event" };
+  }
+
   // Q3/Q4 — plan approval (G2) + subagent tracking live in Python.
   if (manifest.tools.post_delegate_always.includes(toolName)) {
     return { action: "delegate", reason: "post-stateful-tool" };
   }
 
   // Q5 — error path: gotchas, violations, cognition capture, metrics.
-  const exitCode = String(payload.exit_code ?? "") || "0";
-  if (!manifest.post.success_exit_codes.includes(exitCode)) {
-    return { action: "delegate", reason: "error-exit-code" };
-  }
-  const toolOutput = String(payload.tool_output ?? "");
-  if (compileRegex(manifest.post.error_trigger).test(toolOutput)) {
+  // PostToolUse never carries an exit code (failures fire the event
+  // above), so the only error signal on this path is trigger text.
+  if (compileRegex(manifest.post.error_trigger).test(toolText(payload))) {
     return { action: "delegate", reason: "error-trigger" };
   }
 
@@ -333,5 +369,6 @@ module.exports = {
   pythonTruthy,
   readJsonFile,
   safeSessionId,
+  toolText,
   validateManifest,
 };

package/config/hooks/gate-manifest.json

@@ -635,13 +635,9 @@
       "flags": "i",
       "js": "(error:|fatal:|exception:|failed|ENOENT|EACCES|EPERM|panic:)",
       "py": "(error:|fatal:|exception:|failed|ENOENT|EACCES|EPERM|panic:)"
-    },
-    "success_exit_codes": [
-      "0",
-      ""
-    ]
+    }
   },
-  "schema_version": 1,
+  "schema_version": 2,
   "session_id": {
     "flags": "",
     "js": "^[A-Za-z0-9._-]{1,128}$",

package/config/hooks/post-tool-use.ps1

@@ -6,8 +6,16 @@
 # which mistakes repeat across projects.
 #
 # Contract:
-# - Reads a hook payload JSON from stdin (tool_name, tool_output, exit_code,
-#   cwd).
+# - Reads a hook payload JSON from stdin. Registered for two events
+#   (2.1.220 contract): PostToolUse delivers a structured tool_response
+#   (Bash dicts carry stdout/stderr; this port handles the string and
+#   stdout/stderr shapes only — no Task/content path); a failing tool
+#   throws inside the runtime and fires PostToolUseFailure instead,
+#   which carries the output in `error`, prefixed "Exit code N" for
+#   Bash. The payload has no top-level output or exit-code fields.
+#   Also read: tool_name, cwd, session_id, and a legacy
+#   assistant_message the payload never carries (dead read, see
+#   core/hooks/post_tool_use.py `_confirm_flow_authorization`).
 # - Always exits 0 and writes `{}` on stdout — PostToolUse does not inject
 #   context.
 # - Side effect: updates %USERPROFILE%\.arkaos\gotchas.json and
@@ -100,7 +108,7 @@ $stdinText = [Console]::In.ReadToEnd()
 $shouldProcessGotchas = $true
 $payload = $null
 $toolName = ''
-$toolOutput = ''
+$toolText = ''
 $exitCode = '0'
 $cwd = ''
 
@@ -116,10 +124,37 @@ if ([string]::IsNullOrWhiteSpace($stdinText)) {
 
 if ($shouldProcessGotchas -and $null -ne $payload) {
     $toolName   = if ($null -ne $payload.tool_name)   { [string]$payload.tool_name   } else { '' }
-    $toolOutput = if ($null -ne $payload.tool_output) { [string]$payload.tool_output } else { '' }
-    $exitCode   = if ($null -ne $payload.exit_code)   { [string]$payload.exit_code   } else { '0' }
     $cwd        = if ($null -ne $payload.cwd)         { [string]$payload.cwd         } else { '' }
 
+    # Event → (text, exit code). Mirrors the string and stdout/stderr
+    # paths of _tool_text / _tool_exit_code in core/hooks/post_tool_use.py
+    # (no Task/content path in this port). On PostToolUseFailure the
+    # output arrives in `error`; the runtime's "Exit code N" envelope is
+    # its wrapper, not command output, so it is stripped from the text
+    # and surfaced as the exit code. A user interrupt is not a failure.
+    $eventName = if ($null -ne $payload.hook_event_name) { [string]$payload.hook_event_name } else { '' }
+    if ($eventName -eq 'PostToolUseFailure') {
+        $errText = if ($null -ne $payload.error) { [string]$payload.error } else { '' }
+        $envMatch = [regex]::Match($errText, '^Exit code (\d+)\n?')
+        if ($envMatch.Success) {
+            $exitCode = $envMatch.Groups[1].Value
+            $toolText = $errText.Substring($envMatch.Length)
+        } else {
+            $exitCode = '1'
+            $toolText = $errText
+        }
+        if ($payload.is_interrupt) { $exitCode = '0' }
+    } else {
+        $resp = $payload.tool_response
+        if ($resp -is [string]) {
+            $toolText = $resp
+        } elseif ($null -ne $resp) {
+            $stdoutPart = if ($null -ne $resp.stdout) { [string]$resp.stdout } else { '' }
+            $stderrPart = if ($null -ne $resp.stderr) { [string]$resp.stderr } else { '' }
+            $toolText = (@($stdoutPart, $stderrPart) | Where-Object { $_ }) -join "`n"
+        }
+    }
+
     # --- Flow marker cache write (v2 ALLOW accelerator) ------------------
     # Mirror of the bash hook: detect [arka:routing] or [arka:trivial] in
     # $payload.assistant_message and persist via core.workflow.marker_cache.
@@ -211,7 +246,7 @@ except Exception:
     # ─── Only process when there is actually an error ─────────────────
     $errorPattern = '(?i)(error:|fatal:|exception:|failed|ENOENT|EACCES|EPERM|panic:)'
     if ($exitCode -eq '0' -or [string]::IsNullOrEmpty($exitCode)) {
-        if ($toolOutput -notmatch $errorPattern) {
+        if ($toolText -notmatch $errorPattern) {
             $shouldProcessGotchas = $false
         }
     }
@@ -230,7 +265,7 @@ $null = New-Item -ItemType Directory -Force -Path $arkaosRuntimeDir -ErrorAction
 if ($shouldProcessGotchas) { do {
 
 # ─── Extract first meaningful error line ─────────────────────────────
-$lines = $toolOutput -split "`r?`n"
+$lines = $toolText -split "`r?`n"
 $errorLineRegex = '(?i)(error|fatal|exception|failed|ENOENT|EACCES|EPERM|panic|cannot|not found|permission denied)'
 $errorLine = $null
 foreach ($ln in $lines) {

package/config/settings-template.json

@@ -52,6 +52,17 @@
         ]
       }
     ],
+    "PostToolUseFailure": [
+      {
+        "hooks": [
+          {
+            "type": "command",
+            "command": "{{HOOKS_DIR}}/post-tool-use.sh",
+            "timeout": 5
+          }
+        ]
+      }
+    ],
     "CwdChanged": [
       {
         "hooks": [

package/core/governance/evidence_checks.py

@@ -36,6 +36,7 @@ import time
 from dataclasses import asdict, dataclass, field, replace
 from pathlib import Path
 
+from core.governance.qg_digest import evidence_digest
 from core.shared.test_evidence import coverage_percent_from_xml
 
 TIMEOUT_SECONDS = 300
@@ -119,11 +120,16 @@ class EvidenceReport:
     results: list[CheckResult] = field(default_factory=list)
 
     def to_dict(self) -> dict:
-        return {
+        payload = {
             "project_dir": self.project_dir,
             "overall": self.overall,
             "results": [asdict(r) for r in self.results],
         }
+        # PR-B2: the digest excludes its own key, so recomputing over
+        # the embedded dict reproduces it — reviewers can quote it and
+        # the aggregate verifies it once PR-B3 wires the check.
+        payload["report_digest"] = evidence_digest(payload)
+        return payload
 
 
 # ─── Subprocess plumbing ────────────────────────────────────────────────

package/core/governance/judge.py

@@ -24,6 +24,7 @@ it.
 
 from __future__ import annotations
 
+import re
 from typing import Literal
 
 from pydantic import BaseModel, Field, model_validator
@@ -69,12 +70,22 @@ class JudgeVerdict(BaseModel):
         max_length=120,
         description="Opening excerpt of the judged artifact (auditability)",
     )
+    # PR-B2: a real content hash alongside the prose excerpt above —
+    # judged_digest stays an excerpt (not repurposed); judged_sha256 is
+    # the verifiable one. Optional so existing verdicts stay valid.
+    judged_sha256: str = Field(
+        default="",
+        description=(
+            "sha256 of the judged artifact's full text; empty when the "
+            "caller did not hash it"
+        ),
+    )
     reviewer: str = Field(description="Judge id, e.g. plan-judge-g2")
     model_used: str = Field(description="Model tier the judgment ran on")
     notes: str = ""
 
     @model_validator(mode="after")
-    def revise_requires_actionable_findings(self) -> "JudgeVerdict":
+    def revise_requires_actionable_findings(self) -> JudgeVerdict:
         """REVISE must cite at least one blocker/major finding —
         a judge cannot loop work back on narrative alone."""
         if self.verdict == "REVISE":
@@ -91,5 +102,16 @@ class JudgeVerdict(BaseModel):
                 )
         return self
 
+    @model_validator(mode="after")
+    def judged_sha256_is_hex_or_empty(self) -> JudgeVerdict:
+        if self.judged_sha256 and not re.fullmatch(
+            r"[0-9a-f]{64}", self.judged_sha256
+        ):
+            raise ValueError(
+                "judged_sha256 must be 64 lowercase hex chars or empty, "
+                f"got {self.judged_sha256!r}"
+            )
+        return self
+
 
 JUDGE_VERDICT_JSON_SCHEMA: dict = JudgeVerdict.model_json_schema()

package/core/governance/qg_digest.py

@@ -0,0 +1,64 @@
+"""Integrity digests for the Quality Gate chain (PR-B2).
+
+Two digest primitives supply the material for proving — once PR-B3
+wires the verification that consumes them — that reviewers and
+verdict refer to the same evidence report:
+
+- ``evidence_digest`` — hashes an EvidenceReport dict.
+- ``verdict_digest``  — hashes a QGVerdict dict.
+
+Both return a 64-char lowercase sha256 hex string over canonical JSON
+(sorted keys, compact separators): key order and formatting never
+change a digest; a content change does — except under the
+``report_digest`` key, which ``evidence_digest`` excludes by design so
+the digest is stable whether computed before or after embedding.
+
+A working-tree digest deliberately does NOT live here. Three designs
+of a ``tree_digest`` were driven through six wrong-digest routes
+over ten QG rounds, each reproduced at the gate before acceptance;
+``docs/adr/2026-07-29-tree-digest-corpus.md`` records the corpus
+durably and is the acceptance spec for the dedicated PR.
+``QGVerdict.tree_digest`` is enforced empty until that primitive
+exists and survives its own gate.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+
+
+def _sha256(data: bytes) -> str:
+    return hashlib.sha256(data).hexdigest()
+
+
+def _canonical(payload: dict) -> bytes:
+    """Canonical JSON: sorted keys, compact separators, UTF-8-encoded
+    (surrogatepass, so surrogate-bearing strings serialize rather than
+    raise).
+
+    Key order and formatting never change the digest; content does.
+    ``allow_nan=False``: NaN/Infinity are invalid JSON and would make
+    the canonical form implementation-defined — they raise instead.
+    Other non-serializable values raise too, on purpose — a digest
+    over a lossy repr would silently stop meaning anything.
+    """
+    return json.dumps(
+        payload, sort_keys=True, separators=(",", ":"),
+        ensure_ascii=False, allow_nan=False,
+    ).encode("utf-8", "surrogatepass")
+
+
+def evidence_digest(report: dict) -> str:
+    """Digest of an EvidenceReport dict, canonically normalized.
+
+    The report's own ``report_digest`` key (if present) is excluded so
+    the digest is stable whether computed before or after embedding.
+    """
+    payload = {k: v for k, v in report.items() if k != "report_digest"}
+    return _sha256(_canonical(payload))
+
+
+def verdict_digest(verdict: dict) -> str:
+    """Digest of a QGVerdict dict, canonically normalized."""
+    return _sha256(_canonical(verdict))

package/core/governance/qg_verdict.py

@@ -15,10 +15,13 @@ schema when dispatching reviewers via the Agent tool.
 
 from __future__ import annotations
 
+import re
 from typing import Literal
 
 from pydantic import BaseModel, Field, model_validator
 
+_SHA256_HEX_RE = re.compile(r"[0-9a-f]{64}")
+
 
 class QGBlocker(BaseModel):
     """One concrete issue that blocks approval."""
@@ -66,9 +69,55 @@ class QGVerdict(BaseModel):
             "evidence is insufficient"
         ),
     )
+    # PR-B2 integrity digests — optional so the pre-B2 verdict corpus
+    # stays valid. evidence_digest and reviewer_output_sha256 must be
+    # real sha256 hex when present; tree_digest is RESERVED and
+    # rejects ANY value (see its description and the validator).
+    evidence_digest: str = Field(
+        default="",
+        description=(
+            "sha256 of the EvidenceReport the reviewer interpreted "
+            "(report_digest from the --json output); empty when the "
+            "report carried none"
+        ),
+    )
+    tree_digest: str = Field(
+        default="",
+        description=(
+            "RESERVED — must be empty. The producing primitive will "
+            "ship in its own PR (acceptance spec: "
+            "docs/adr/2026-07-29-tree-digest-corpus.md); until it "
+            "exists, a non-empty value is rejected so no unverifiable "
+            "digest can enter the corpus."
+        ),
+    )
+    reviewer_output_sha256: str = Field(
+        default="",
+        description=(
+            "sha256 of the reviewer's own raw output, as captured by the "
+            "ledger (raw_sha256); empty when not yet ledgered"
+        ),
+    )
+
+    @model_validator(mode="after")
+    def digests_are_sha256_hex_or_empty(self) -> QGVerdict:
+        for name in ("evidence_digest", "reviewer_output_sha256"):
+            value = getattr(self, name)
+            if value and not _SHA256_HEX_RE.fullmatch(value):
+                raise ValueError(
+                    f"{name} must be 64 lowercase hex chars or empty, "
+                    f"got {value!r}"
+                )
+        if self.tree_digest:
+            raise ValueError(
+                "tree_digest is RESERVED and must be empty until the "
+                "tree-digest primitive ships in its own PR — an "
+                "unverifiable digest must not enter the corpus"
+            )
+        return self
 
     @model_validator(mode="after")
-    def enforce_evidence_floor(self) -> "QGVerdict":
+    def enforce_evidence_floor(self) -> QGVerdict:
         """APPROVED cannot coexist with failing evidence (evidence floor)."""
         if self.verdict == "APPROVED" and self.evidence_report.overall == "fail":
             raise ValueError(

package/core/governance/reviewer_ledger.py

@@ -338,8 +338,17 @@ def _build_record(
         "raw_sha256": digest,
         "verdict": verdict,
         "parse_error": parse_error,
-        "evidence_digest": (verdict or {}).get("evidence_digest"),
-        "tree_digest": (verdict or {}).get("tree_digest"),
+        # Digest columns are harvested ONLY from a VALIDATED verdict:
+        # _validated is fail-soft (it keeps the raw dict so no review
+        # text is ever lost), but a value the validator rejected must
+        # not enter the corpus through these columns (QG r12 — a
+        # rejected tree_digest was reproduced landing on disk here).
+        "evidence_digest": (
+            None if parse_error else (verdict or {}).get("evidence_digest")
+        ),
+        "tree_digest": (
+            None if parse_error else (verdict or {}).get("tree_digest")
+        ),
         "source": source,
         "sanitized": sanitized,
     }

package/core/hooks/gate_manifest.py

@@ -34,7 +34,12 @@ import re
 from dataclasses import asdict
 from pathlib import Path
 
-_SCHEMA_VERSION = 1
+# v2: `post.success_exit_codes` removed (PostToolUse never carries an
+# exit code; failures fire PostToolUseFailure, which the shim delegates
+# outright). The bump makes an old engine reading a new manifest
+# delegate everything instead of crashing into the shim's fail-open
+# `{}` — a partial deploy must degrade to Python, never drop turns.
+_SCHEMA_VERSION = 2
 
 # Python-only regex constructs that would compile differently (or not at
 # all) under JavaScript's RegExp. The generator refuses to export any
@@ -177,7 +182,7 @@ def _tool_sets() -> dict:
     from core.hooks import pre_tool_use
     from core.workflow import flow_enforcer, research_gate
 
-    if pre_tool_use._FLOW_GATED_TOOLS != flow_enforcer.EFFECT_TOOLS_ALWAYS | {"Bash"}:
+    if flow_enforcer.EFFECT_TOOLS_ALWAYS | {"Bash"} != pre_tool_use._FLOW_GATED_TOOLS:
         raise ValueError(
             "pre_tool_use._FLOW_GATED_TOOLS no longer equals "
             "flow_enforcer.EFFECT_TOOLS_ALWAYS + Bash — re-derive the "
@@ -258,9 +263,9 @@ def _flags_and_budget() -> dict:
 
 
 def _state_and_paths() -> dict:
+    from core.runtime import mcp_telemetry
     from core.shared.temp_paths import arkaos_temp_dir
     from core.workflow import flow_authorization, flow_enforcer, research_gate
-    from core.runtime import mcp_telemetry
 
     auth_dir = arkaos_temp_dir("arkaos-flow-auth")
     if str(auth_dir.parent) != "/tmp":  # pragma: no cover — POSIX only
@@ -290,11 +295,11 @@ def _state_and_paths() -> dict:
 
 def build_manifest() -> dict:
     """Assemble the manifest from the real gate constants (import-time)."""
-    from core.hooks import post_tool_use, pre_tool_use  # noqa: F401
+    from core.hooks import post_tool_use, pre_tool_use
     from core.shared import safe_session_id as ssid
 
     state = _state_and_paths()
-    if pre_tool_use._ASSISTANT_WINDOW != state["numbers"]["assistant_window"]:
+    if state["numbers"]["assistant_window"] != pre_tool_use._ASSISTANT_WINDOW:
         raise ValueError("ASSISTANT_WINDOW diverged between hook and enforcer")
     return {
         "schema_version": _SCHEMA_VERSION,
@@ -312,9 +317,12 @@ def build_manifest() -> dict:
         ],
         "tools": _tool_sets(),
         "bash": _bash_classifier(),
+        # No exit-code corpus: PostToolUse never carries one (a failing
+        # tool throws and fires PostToolUseFailure, which the shim
+        # delegates outright), so trigger text is the only error signal
+        # the fast path needs.
         "post": {
             "error_trigger": _js_pattern(post_tool_use._ERROR_TRIGGER_RE),
-            "success_exit_codes": ["0", ""],
         },
         "session_id": {
             **_js_pattern(ssid.SAFE_SESSION_ID_RE),

package/core/hooks/post_tool_use.py

@@ -144,8 +144,8 @@ def _confirm_flow_authorization(session_id: str, transcript_path: str) -> None:
 # ─── Section 2+3: CQO verdicts, pattern stubs, activation ───────────────
 
 
-def _record_cqo_rejected(tool_output: str, prompt: str, session_id: str) -> None:
-    if not _REJECTED_RE.search(tool_output):
+def _record_cqo_rejected(tool_text: str, prompt: str, session_id: str) -> None:
+    if not _REJECTED_RE.search(tool_text):
         return
     match = _REVIEWING_RE.search(prompt)
     if not match:
@@ -153,7 +153,7 @@ def _record_cqo_rejected(tool_output: str, prompt: str, session_id: str) -> None
     try:
         from core.governance.cqo_experience_recorder import record_from_verdict
         record_from_verdict(
-            verdict_text=tool_output,
+            verdict_text=tool_text,
             agent_id=match.group(1),
             session_id=session_id,
             context="auto-recorded via PostToolUse hook (cqo dispatch REJECTED)",
@@ -162,8 +162,8 @@ def _record_cqo_rejected(tool_output: str, prompt: str, session_id: str) -> None
         pass
 
 
-def _record_pattern_stub(tool_output: str, prompt: str) -> None:
-    if not _APPROVED_RE.search(tool_output):
+def _record_pattern_stub(tool_text: str, prompt: str) -> None:
+    if not _APPROVED_RE.search(tool_text):
         return
     match = _PATTERN_SUGGEST_RE.search(prompt)
     if not match:
@@ -227,6 +227,59 @@ def _task_result_text(stdin_json: dict) -> str:
     return ""
 
 
+# A failing tool THROWS inside the runtime and fires PostToolUseFailure
+# — PostToolUse never sees it (proven live against 2.1.220: a logger
+# registered on both events fired only PostToolUseFailure for failing
+# tools; the binary's dispatcher reaches the failure event from its
+# catch block). On that event the output arrives in `error`, prefixed
+# with this envelope for Bash ("Exit code N\n<output>") and as a plain
+# message for other tools. The envelope is the runtime's wrapper, not
+# tool output.
+_EXIT_ENVELOPE_RE = re.compile(r"^Exit code (\d+)\n?")
+
+
+def _tool_text(stdin_json: dict) -> str:
+    """Textual output of the tool call under the two-event contract.
+
+    PostToolUseFailure carries it in ``error`` (envelope stripped here,
+    surfaced by ``_tool_exit_code``). PostToolUse only ever delivers a
+    structured ``tool_response`` — Bash dicts carry stdout/stderr, Task
+    dicts carry content; a plain string passes through defensively.
+    """
+    if get_str(stdin_json, "hook_event_name") == "PostToolUseFailure":
+        error = get_str(stdin_json, "error")
+        return _EXIT_ENVELOPE_RE.sub("", error, count=1)
+    response = stdin_json.get("tool_response")
+    if isinstance(response, str):
+        return response
+    if isinstance(response, dict) and (
+        "stdout" in response or "stderr" in response
+    ):
+        parts = (
+            str(response.get(key) or "") for key in ("stdout", "stderr")
+        )
+        return "\n".join(part for part in parts if part)
+    return _task_result_text(stdin_json)
+
+
+def _tool_exit_code(stdin_json: dict) -> str:
+    """Exit code derived from the event, not from a payload field.
+
+    Neither event carries an exit_code field. PostToolUseFailure is a
+    failure by definition: a leading ``Exit code N`` line in ``error``
+    yields N, any other message maps to "1" — except a user interrupt
+    (``is_interrupt``), which is not a command failure. PostToolUse is
+    always "0"; error text on successful turns is caught downstream by
+    the error-trigger scan.
+    """
+    if get_str(stdin_json, "hook_event_name") != "PostToolUseFailure":
+        return "0"
+    if stdin_json.get("is_interrupt"):
+        return "0"
+    match = _EXIT_ENVELOPE_RE.match(get_str(stdin_json, "error"))
+    return match.group(1) if match else "1"
+
+
 def _record_reviewer_output(
     subagent_type: str, stdin_json: dict, session_id: str
 ) -> None:
@@ -261,8 +314,8 @@ def _record_activation(subagent_type: str, session_id: str) -> None:
 # ─── Section 5: gotchas memory ───────────────────────────────────────────
 
 
-def _extract_error_line(tool_output: str) -> str:
-    lines = tool_output.splitlines()
+def _extract_error_line(tool_text: str) -> str:
+    lines = tool_text.splitlines()
     for line in lines:
         if _ERROR_LINE_RE.search(line):
             return line
@@ -425,16 +478,16 @@ def _detect_rule_violations(input_data: dict) -> tuple[str, list[tuple]]:
     if state is None:
         return "", []
     tool_name = get_str(input_data, "tool_name")
-    tool_output = get_str(input_data, "tool_output")
+    tool_text = _tool_text(input_data)
     violation_msg = ""
     persist: list[tuple] = []
 
     if tool_name == "Bash":
         on_master = any(
             re.match(r"^\[(master|main)", line)
-            for line in tool_output.splitlines()
+            for line in tool_text.splitlines()
         )
-        cmd_text = get_str(input_data, "command")
+        cmd_text = get_str(input_data, "tool_input", "command")
         if on_master and re.search(r"git commit", cmd_text):
             persist.append((
                 "branch-isolation",
@@ -446,7 +499,7 @@ def _detect_rule_violations(input_data: dict) -> tuple[str, list[tuple]]:
             )
 
     if tool_name in ("Write", "Edit"):
-        file_path = get_str(input_data, "file_path")
+        file_path = get_str(input_data, "tool_input", "file_path")
         if _CODE_FILE_RE.search(file_path):
             if _phase_status(state, "spec") != "completed":
                 persist.append((
@@ -507,11 +560,13 @@ def _enforcer_messages(input_data, enforce_tool, add_violation, msg: str) -> str
         except Exception:
             extra["git_branch"] = ""
     try:
+        # user_input is a prompt-side enforcer parameter with no source
+        # in either post-tool payload (2.1.220 contract).
         result = enforce_tool(
             tool_name=tool_name,
-            command=get_str(input_data, "command"),
-            file_path=get_str(input_data, "file_path"),
-            user_input=get_str(input_data, "user_input"),
+            command=get_str(input_data, "tool_input", "command"),
+            file_path=get_str(input_data, "tool_input", "file_path"),
+            user_input="",
             **extra,
         )
     except Exception:
@@ -543,7 +598,7 @@ def _forge_violation(input_data: dict, yaml_mod) -> str:
     forge_file = Path.home() / ".arkaos" / "plans" / f"{forge_id}.yaml"
     if not forge_file.is_file():
         return ""
-    edited = get_str(input_data, "file_path")
+    edited = get_str(input_data, "tool_input", "file_path")
     if not edited:
         return ""
     try:
@@ -601,8 +656,8 @@ def _workflow_sections_with_fallback(
 # ─── Section 9: cognition capture (detached, fire-and-forget) ───────────
 
 
-def _enqueue_cognition_capture(session_id: str, tool_output: str) -> None:
-    if not session_id or not tool_output:
+def _enqueue_cognition_capture(session_id: str, tool_text: str) -> None:
+    if not session_id or not tool_text:
         return
     repo = repo_path()
     if not repo or not Path(repo).is_dir():
@@ -619,7 +674,7 @@ def _enqueue_cognition_capture(session_id: str, tool_output: str) -> None:
         )
         if proc.stdin is not None:
             try:
-                proc.stdin.write(tool_output.encode("utf-8", "replace"))
+                proc.stdin.write(tool_text.encode("utf-8", "replace"))
                 proc.stdin.close()
             except (BrokenPipeError, OSError):
                 pass
@@ -672,8 +727,8 @@ def main(stdin_json: dict | None = None) -> int:
     ensure_root_on_path(root)
 
     tool_name = get_str(stdin_json, "tool_name")
-    tool_output = get_str(stdin_json, "tool_output")
-    exit_code = get_str(stdin_json, "exit_code") or "0"
+    tool_text = _tool_text(stdin_json)
+    exit_code = _tool_exit_code(stdin_json)
     cwd = get_str(stdin_json, "cwd")
     session_id = get_str(stdin_json, "session_id")
     transcript_path = get_str(stdin_json, "transcript_path")
@@ -717,17 +772,17 @@ def main(stdin_json: dict | None = None) -> int:
         _record_reviewer_output(subagent_type, stdin_json, session_id)
         if subagent_type in _CQO_IDS:
             prompt = get_str(stdin_json, "tool_input", "prompt")
-            _record_cqo_rejected(tool_output, prompt, session_id)
-            _record_pattern_stub(tool_output, prompt)
+            _record_cqo_rejected(tool_text, prompt, session_id)
+            _record_pattern_stub(tool_text, prompt)
         _record_activation(subagent_type, session_id)
 
     # Only process further if there was an error signal (same early exit
     # as the bash version — violations/metrics only run on error turns).
-    if exit_code in ("0", "") and not _ERROR_TRIGGER_RE.search(tool_output):
+    if exit_code in ("0", "") and not _ERROR_TRIGGER_RE.search(tool_text):
         print("{}")
         return 0
 
-    error_line = _extract_error_line(tool_output)
+    error_line = _extract_error_line(tool_text)
     if not error_line:
         print("{}")
         return 0
@@ -746,11 +801,17 @@ def main(stdin_json: dict | None = None) -> int:
         stdin_json, root, persist, violation_msg
     )
 
-    _enqueue_cognition_capture(session_id, tool_output)
+    _enqueue_cognition_capture(session_id, tool_text)
     _log_metrics(int((time.monotonic() - start) * 1000))
 
     if violation_msg:
-        emit_additional_context("PostToolUse", violation_msg)
+        # The runtime drops context whose hookEventName differs from the
+        # event it invoked (live probe, QG round 3) — echo the payload's
+        # own event so violations on failure turns are delivered.
+        emit_additional_context(
+            get_str(stdin_json, "hook_event_name") or "PostToolUse",
+            violation_msg,
+        )
     else:
         print("{}")
     return 0

package/harness/codex/AGENTS.md

@@ -1,6 +1,6 @@
 # ArkaOS — The Operating System for AI Agent Teams
 
-> v4.39.0 — 89 agents, 17 departments, 332 skills. Generated by `scripts/harness_gen.py`; do not edit.
+> v4.41.0 — 89 agents, 17 departments, 332 skills. Generated by `scripts/harness_gen.py`; do not edit.
 
 You are operating within ArkaOS. Every request routes through the
 appropriate department squad — never respond as a generic assistant.

package/harness/copilot/copilot-instructions.md

@@ -1,6 +1,6 @@
 # ArkaOS — The Operating System for AI Agent Teams
 
-> v4.39.0 — 89 agents, 17 departments, 332 skills. Generated by `scripts/harness_gen.py`; do not edit.
+> v4.41.0 — 89 agents, 17 departments, 332 skills. Generated by `scripts/harness_gen.py`; do not edit.
 
 You are operating within ArkaOS. Every request routes through the
 appropriate department squad — never respond as a generic assistant.

package/harness/cursor/rules/arkaos.mdc

@@ -1,11 +1,11 @@
 ---
-description: ArkaOS v4.39.0 agent-team contract
+description: ArkaOS v4.41.0 agent-team contract
 alwaysApply: true
 ---
 
 # ArkaOS — The Operating System for AI Agent Teams
 
-> v4.39.0 — 89 agents, 17 departments, 332 skills. Generated by `scripts/harness_gen.py`; do not edit.
+> v4.41.0 — 89 agents, 17 departments, 332 skills. Generated by `scripts/harness_gen.py`; do not edit.
 
 You are operating within ArkaOS. Every request routes through the
 appropriate department squad — never respond as a generic assistant.

package/harness/gemini/GEMINI.md

@@ -1,6 +1,6 @@
 # ArkaOS — The Operating System for AI Agent Teams
 
-> v4.39.0 — 89 agents, 17 departments, 332 skills. Generated by `scripts/harness_gen.py`; do not edit.
+> v4.41.0 — 89 agents, 17 departments, 332 skills. Generated by `scripts/harness_gen.py`; do not edit.
 
 You are operating within ArkaOS. Every request routes through the
 appropriate department squad — never respond as a generic assistant.

package/harness/opencode/AGENTS.md

@@ -1,6 +1,6 @@
 # ArkaOS — The Operating System for AI Agent Teams
 
-> v4.39.0 — 89 agents, 17 departments, 332 skills. Generated by `scripts/harness_gen.py`; do not edit.
+> v4.41.0 — 89 agents, 17 departments, 332 skills. Generated by `scripts/harness_gen.py`; do not edit.
 
 You are operating within ArkaOS. Every request routes through the
 appropriate department squad — never respond as a generic assistant.

package/harness/opencode/agents/arka-architect-gabriel.md

@@ -3,7 +3,7 @@ description: "Software Architect — ArkaOS /dev department"
 mode: subagent
 ---
 
-You are Gabriel, Software Architect of the ArkaOS /dev department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Gabriel, Software Architect of the ArkaOS /dev department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: system design, system visualization via dev/diagram (architecture + dataflow diagrams delivered as browser artifacts), domain modeling (event storming, bounded contexts), design patterns (GoF, PoEAA), business / domain analysis, API design.
 

package/harness/opencode/agents/arka-brand-director-valentina.md

@@ -3,7 +3,7 @@ description: "Creative Director — ArkaOS /brand department"
 mode: subagent
 ---
 
-You are Valentina, Creative Director of the ArkaOS /brand department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Valentina, Creative Director of the ArkaOS /brand department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: brand identity creation, reference-video visual analysis via dev/watch (complete frames + transcript — motion and art direction judged on evidence, never on screenshots), visual design direction, UX/UI strategy, design systems, brand voice & tone.
 

package/harness/opencode/agents/arka-cfo-helena.md

@@ -3,7 +3,7 @@ description: "Chief Financial Officer — ArkaOS /fin department"
 mode: subagent
 ---
 
-You are Helena, Chief Financial Officer of the ArkaOS /fin department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Helena, Chief Financial Officer of the ArkaOS /fin department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: financial planning & analysis, valuation & investment, unit economics & SaaS metrics, risk management & ERM, fundraising & cap tables, cash flow management.
 

package/harness/opencode/agents/arka-chief-of-staff-afonso.md

@@ -3,7 +3,7 @@ description: "Chief of Staff & Governance Lead — ArkaOS /org department"
 mode: subagent
 ---
 
-You are Afonso, Chief of Staff & Governance Lead of the ArkaOS /org department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Afonso, Chief of Staff & Governance Lead of the ArkaOS /org department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: meeting cadence (daily/weekly/quarterly/annual), OKR & CFR orchestration cross-department, decision records & RACI, premortem / blameless postmortem rituals, governance, board & founder-CEO succession, strategic alignment & single-threaded leadership.
 

package/harness/opencode/agents/arka-community-strategist-beatriz.md

@@ -3,7 +3,7 @@ description: "Community Strategist — ArkaOS /community department"
 mode: subagent
 ---
 
-You are Beatriz, Community Strategist of the ArkaOS /community department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Beatriz, Community Strategist of the ArkaOS /community department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: community strategy & design, platform selection (Discord, Telegram, Skool, Circle), member onboarding & retention, monetization (membership, courses, coaching), gamification & engagement, niche communities (betting, AI, vertical).
 

package/harness/opencode/agents/arka-content-strategist-rafael.md

@@ -3,7 +3,7 @@ description: "Content Strategist — ArkaOS /content department"
 mode: subagent
 ---
 
-You are Rafael, Content Strategist of the ArkaOS /content department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Rafael, Content Strategist of the ArkaOS /content department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: viral content design, reference-video analysis via dev/watch (frames + timestamped transcript before judging any video), hook writing & packaging, script structure, content operating systems, platform-specific optimization.
 

package/harness/opencode/agents/arka-conversion-strategist-ines.md

@@ -3,7 +3,7 @@ description: "Conversion Strategist — ArkaOS /landing department"
 mode: subagent
 ---
 
-You are Ines, Conversion Strategist of the ArkaOS /landing department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Ines, Conversion Strategist of the ArkaOS /landing department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: sales funnels, landing page optimization, offer creation, copywriting (direct response), launch sequences, affiliate marketing.
 

package/harness/opencode/agents/arka-coo-sofia.md

@@ -3,7 +3,7 @@ description: "Chief Operations Officer — ArkaOS /org department"
 mode: subagent
 ---
 
-You are Sofia, Chief Operations Officer of the ArkaOS /org department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Sofia, Chief Operations Officer of the ArkaOS /org department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: organizational design, process optimization, cross-department coordination, culture & team health, scaling operations, workflow automation.
 

package/harness/opencode/agents/arka-copy-director-eduardo.md

@@ -3,7 +3,7 @@ description: "Copy & Language Director — ArkaOS /quality department"
 mode: subagent
 ---
 
-You are Eduardo, Copy & Language Director of the ArkaOS /quality department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Eduardo, Copy & Language Director of the ArkaOS /quality department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: spelling and grammar (EN, PT-PT, PT-BR, ES, FR), tone and voice consistency, AI pattern detection and removal, accentuation and orthography, copywriting quality, factual accuracy in text.
 

package/harness/opencode/agents/arka-cqo-marta.md

@@ -3,7 +3,7 @@ description: "Chief Quality Officer — ArkaOS /quality department"
 mode: subagent
 ---
 
-You are Marta, Chief Quality Officer of the ArkaOS /quality department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Marta, Chief Quality Officer of the ArkaOS /quality department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: quality assurance orchestration, cross-department quality standards, text quality (spelling, grammar, tone), technical quality (code, UX, data), compliance and audit.
 

package/harness/opencode/agents/arka-cto-marco.md

@@ -3,7 +3,7 @@ description: "Chief Technology Officer — ArkaOS /dev department"
 mode: subagent
 ---
 
-You are Marco, Chief Technology Officer of the ArkaOS /dev department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Marco, Chief Technology Officer of the ArkaOS /dev department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: software architecture, system design, tech strategy, cloud infrastructure, ai/ml systems.
 

package/harness/opencode/agents/arka-design-ops-lead-iris.md

@@ -3,7 +3,7 @@ description: "Design Ops Lead — ArkaOS /brand department"
 mode: subagent
 ---
 
-You are Iris, Design Ops Lead of the ArkaOS /brand department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Iris, Design Ops Lead of the ArkaOS /brand department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: design tokens (JSON + CSS variables), component libraries (shadcn/ui, Radix, Headless UI), design system governance, figma → code pipelines, accessibility compliance (WCAG 2.2 AA/AAA), cross-platform tokenisation (Style Dictionary, Tailwind).
 

package/harness/opencode/agents/arka-ecom-director-ricardo.md

@@ -3,7 +3,7 @@ description: "E-Commerce Director — ArkaOS /ecom department"
 mode: subagent
 ---
 
-You are Ricardo, E-Commerce Director of the ArkaOS /ecom department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Ricardo, E-Commerce Director of the ArkaOS /ecom department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: e-commerce strategy, conversion optimization, marketplace operations, pricing strategy, fulfillment & logistics, email & retention.
 

package/harness/opencode/agents/arka-knowledge-director-clara.md

@@ -3,7 +3,7 @@ description: "Knowledge Director — ArkaOS /kb department"
 mode: subagent
 ---
 
-You are Clara, Knowledge Director of the ArkaOS /kb department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Clara, Knowledge Director of the ArkaOS /kb department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: knowledge management, research methodology, persona building, content curation, taxonomy & ontology, Obsidian vault management.
 

package/harness/opencode/agents/arka-leadership-director-rodrigo.md

@@ -3,7 +3,7 @@ description: "Leadership & People Director — ArkaOS /lead department"
 mode: subagent
 ---
 
-You are Rodrigo, Leadership & People Director of the ArkaOS /lead department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Rodrigo, Leadership & People Director of the ArkaOS /lead department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: team assessment & health, leadership development, hiring & onboarding, performance management, feedback & 1-on-1s, culture building.
 

package/harness/opencode/agents/arka-marketing-director-luna.md

@@ -3,7 +3,7 @@ description: "Marketing Director — ArkaOS /mkt department"
 mode: subagent
 ---
 
-You are Luna, Marketing Director of the ArkaOS /mkt department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Luna, Marketing Director of the ArkaOS /mkt department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: growth strategy, video-ad teardown via dev/watch (hook, pacing and spoken-copy evidence from frames + transcript), content marketing, SEO, paid acquisition, social media.
 

package/harness/opencode/agents/arka-ops-lead-daniel.md

@@ -3,7 +3,7 @@ description: "Operations Lead — ArkaOS /ops department"
 mode: subagent
 ---
 
-You are Daniel, Operations Lead of the ArkaOS /ops department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Daniel, Operations Lead of the ArkaOS /ops department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: workflow automation (Zapier, Make, n8n), SOP/process visualization via dev/diagram (workflow + lifecycle diagrams for automations and runbooks), process mapping & optimization, SOP creation & management, bottleneck analysis, integration design.
 

package/harness/opencode/agents/arka-pm-director-carolina.md

@@ -3,7 +3,7 @@ description: "Product Manager — ArkaOS /pm department"
 mode: subagent
 ---
 
-You are Carolina, Product Manager of the ArkaOS /pm department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Carolina, Product Manager of the ArkaOS /pm department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: continuous product discovery (daily habit), deliverable visualization via dev/diagram (workflow diagrams so stakeholders see scope before build), weekly customer interviewing, dual-track agile (discovery + delivery), product risk assessment (value/usability/feasibility/viability), framing problems for empowered teams (not features).
 

package/harness/opencode/agents/arka-revops-lead-vicente.md

@@ -3,7 +3,7 @@ description: "RevOps Lead — ArkaOS /saas department"
 mode: subagent
 ---
 
-You are Vicente, RevOps Lead of the ArkaOS /saas department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Vicente, RevOps Lead of the ArkaOS /saas department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: revenue operations (cross mkt + sales + CS), unified funnel & CRM hygiene, SLA MQL→SQL between marketing and sales, revenue metrics (LTV/CAC, NRR, payback), lead scoring & routing, commission & forecast modeling.
 

package/harness/opencode/agents/arka-saas-strategist-tiago.md

@@ -3,7 +3,7 @@ description: "SaaS Strategist — ArkaOS /saas department"
 mode: subagent
 ---
 
-You are Tiago, SaaS Strategist of the ArkaOS /saas department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Tiago, SaaS Strategist of the ArkaOS /saas department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: SaaS metrics & benchmarking, product-led growth, pricing strategy, customer success, micro-SaaS validation, go-to-market for SaaS.
 

package/harness/opencode/agents/arka-sales-director-miguel.md

@@ -3,7 +3,7 @@ description: "Sales Director — ArkaOS /sales department"
 mode: subagent
 ---
 
-You are Miguel, Sales Director of the ArkaOS /sales department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Miguel, Sales Director of the ArkaOS /sales department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: consultative selling, pipeline management, proposal writing, negotiation, discovery calls, deal qualification.
 

package/harness/opencode/agents/arka-strategy-director-tomas.md

@@ -3,7 +3,7 @@ description: "Chief Strategist — ArkaOS /strat department"
 mode: subagent
 ---
 
-You are Tomas, Chief Strategist of the ArkaOS /strat department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Tomas, Chief Strategist of the ArkaOS /strat department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: competitive strategy, business-flow visualization via dev/diagram (architecture + dataflow diagrams of business models and value chains), market analysis, business model design, positioning, innovation strategy.
 

package/harness/opencode/agents/arka-tech-director-francisca.md

@@ -3,7 +3,7 @@ description: "Technical & UX Quality Director — ArkaOS /quality department"
 mode: subagent
 ---
 
-You are Francisca, Technical & UX Quality Director of the ArkaOS /quality department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Francisca, Technical & UX Quality Director of the ArkaOS /quality department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: code quality (SOLID, Clean Code, DRY), test coverage and quality, UX/UI review (heuristics, accessibility), security review (OWASP), performance review (CWV, API latency), data integrity and API contracts.
 

package/harness/opencode/agents/arka-tech-lead-paulo.md

@@ -3,7 +3,7 @@ description: "Tech Lead — ArkaOS /dev department"
 mode: subagent
 ---
 
-You are Paulo, Tech Lead of the ArkaOS /dev department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Paulo, Tech Lead of the ArkaOS /dev department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: workflow orchestration, visual spec/plan companions via dev/diagram (typed IR -> interactive HTML the user opens before build), code quality enforcement, sprint/cycle management, technical decision-making, developer experience.
 

package/harness/opencode/agents/arka-video-producer-simao.md

@@ -3,7 +3,7 @@ description: "Video Producer & Production Lead — ArkaOS /content department"
 mode: subagent
 ---
 
-You are Simão, Video Producer & Production Lead of the ArkaOS /content department (v4.39.0; generated by scripts/harness_gen.py — do not edit).
+You are Simão, Video Producer & Production Lead of the ArkaOS /content department (v4.41.0; generated by scripts/harness_gen.py — do not edit).
 
 Expertise: video production pipelines (script → storyboard → assets → edit → render), cut review via dev/watch (frame + transcript evidence on own renders before the Quality Gate), Hyperframes video-as-code editing (HTML/CSS/JS + GSAP → MP4), Higgsfield generation orchestration (image, video, audio, motion control, upscale, reframe), shot lists and EDLs (scene/shot/VO/on-screen-text columns), transcription-synced cuts and word-level captions.
 

package/harness/zed/.rules

@@ -1,6 +1,6 @@
 # ArkaOS — The Operating System for AI Agent Teams
 
-> v4.39.0 — 89 agents, 17 departments, 332 skills. Generated by `scripts/harness_gen.py`; do not edit.
+> v4.41.0 — 89 agents, 17 departments, 332 skills. Generated by `scripts/harness_gen.py`; do not edit.
 
 You are operating within ArkaOS. Every request routes through the
 appropriate department squad — never respond as a generic assistant.

package/installer/adapters/claude-code.js

@@ -166,6 +166,14 @@ export default {
       { hooks: [hookEntry(hooksDir, "post-tool-use", 5)] },
     ];
 
+    // PostToolUseFailure — a failing tool throws inside the runtime and
+    // fires THIS event, never PostToolUse (2.1.220 contract), so
+    // failed-command turns only reach the hook when it is registered
+    // here too. Same entrypoint; Python branches on hook_event_name.
+    settings.hooks.PostToolUseFailure = [
+      { hooks: [hookEntry(hooksDir, "post-tool-use", 5)] },
+    ];
+
     // PreToolUse — Flow enforcement gate (gated by hooks.hardEnforcement
     // feature flag in ~/.arkaos/config.json; no-op when flag is false).
     settings.hooks.PreToolUse = [

package/knowledge/skills-manifest.json

@@ -1,7 +1,7 @@
 {
   "_meta": {
     "generator": "scripts/marketplace_gen.py",
-    "version": "4.39.0",
+    "version": "4.41.0",
     "marketplace": "arkaos"
   },
   "structural": {

package/mcps/arka-tools/pyproject.toml

@@ -4,7 +4,7 @@ version = "1.0.0"
 description = "ArkaOS programmatic MCP tools — KB search, workflow state, Quality Gate, recipes, session memory, telemetry"
 requires-python = ">=3.11"
 dependencies = [
-    "mcp[cli]>=1.2.0",
+    "mcp[cli]>=1.2.0,<2",  # <2: mcp 2.0.0 removed mcp.server.fastmcp (1.x import path)
     "pydantic>=2.0",
     "pyyaml>=6.0",
 ]

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "arkaos",
-  "version": "4.39.0",
+  "version": "4.41.0",
   "description": "The Operating System for AI Agent Teams",
   "type": "module",
   "bin": {

package/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "arkaos-core"
-version = "4.39.0"
+version = "4.41.0"
 description = "Core engine for ArkaOS — The Operating System for AI Agent Teams"
 readme = "README.md"
 license = {text = "MIT"}
@@ -38,7 +38,10 @@ mcp = [
     # import the FastMCP decorators so the arka-tools security tests
     # (OWASP LLM01 neutralization, write-gating) run on every Python
     # version, not just where mcp happens to be installed.
-    "mcp[cli]>=1.2.0",
+    # <2: mcp 2.0.0 (2026-07-28) removed mcp.server.fastmcp — the 1.x
+    # import path server.py uses. Migrating to the 2.x API is its own
+    # change; the ceiling keeps CI deterministic until then.
+    "mcp[cli]>=1.2.0,<2",
 ]
 dashboard = [
     "fastapi>=0.115.0",