Skip to content

The agent that grows with you

fix: a refused skill_manage write reads neutral, not as a red failure

Teknium74cc8a105fe8
diff --git a/agent/display.py b/agent/display.py
index 590be47d763..7ee2b75b035 100644
--- a/agent/display.py
+++ b/agent/display.py
@@ -1015,6 +1015,13 @@ def _degraded_suffix(data: dict) -> str:
return f" [{_tail_trunc(text, _DEGRADED_SUFFIX_MAX_LEN)}]"
+# Tools whose refused calls are the agent's own retry loop, not something the user must act on: a
+# skill_manage write refused by the size cap, a stale old_string or a guard is re-planned by the
+# agent. Renderers show these rows neutrally; the failure still reaches the model and the loop
+# guardrail unchanged (display-only).
+NEUTRAL_FAILURE_TOOLS = frozenset({"skill_manage"})
+
+
def _detect_tool_failure(tool_name: str, result: Any) -> tuple[bool, str]:
"""Return ``(is_failure, suffix)`` for a tool result, e.g. ``(True, " [exit 1]")``."""
if result is None or file_mutation_result_landed(tool_name, result):
diff --git a/apps/desktop/src/components/assistant-ui/tool/fallback-model.test.ts b/apps/desktop/src/components/assistant-ui/tool/fallback-model.test.ts
index bda833c25ff..5ccd512f74a 100644
--- a/apps/desktop/src/components/assistant-ui/tool/fallback-model.test.ts
+++ b/apps/desktop/src/components/assistant-ui/tool/fallback-model.test.ts
@@ -107,6 +107,7 @@ describe('buildToolView error confidence', () => {
[{ result: { meta: { error: 'a previous attempt' }, data: { count: 1 } } }, 'success'],
[{ toolName: 'read_file', result: { error: 'Permission denied reading /repo/private.ts' } }, 'error'],
[{ toolName: 'patch', result: { error: 'File not found: /repo/session-view.ts' } }, 'error'],
+ [{ toolName: 'skill_manage', result: { success: false, error: 'SKILL.md content is over the limit' } }, 'notice'],
[{ result: { success: false, result: { output: { error: { message: 'Connection refused' } } } } }, 'error']
]
diff --git a/apps/desktop/src/components/assistant-ui/tool/fallback-model/format.ts b/apps/desktop/src/components/assistant-ui/tool/fallback-model/format.ts
index c75b6800126..d4d152ec123 100644
--- a/apps/desktop/src/components/assistant-ui/tool/fallback-model/format.ts
+++ b/apps/desktop/src/components/assistant-ui/tool/fallback-model/format.ts
@@ -105,12 +105,18 @@ export function parseMaybeObject(value: unknown): Record<string, unknown> {
}
}
+/** Tools whose refused calls are the agent's own retry loop (a skill write
+ * refused by the size cap or a stale old_string), not something the user must
+ * act on. Mirrors agent/display.py::NEUTRAL_FAILURE_TOOLS. */
+export const NEUTRAL_FAILURE_TOOLS: ReadonlySet<string> = new Set(['skill_manage'])
+
/** A call that reported failure, from the part's flag or its result body.
* Explicit success beats stale envelope errors, as in individual rows. */
-export function toolCallFailed(part: { isError?: boolean; result?: unknown }): boolean {
+export function toolCallFailed(part: { isError?: boolean; result?: unknown; toolName?: string }): boolean {
const result = parseMaybeObject(part.result)
return (
+ !NEUTRAL_FAILURE_TOOLS.has(part.toolName ?? '') &&
result.success !== true &&
result.ok !== true &&
Boolean(part.isError || result.success === false || result.ok === false || extractToolErrorMessage(part.result))
diff --git a/apps/desktop/src/components/assistant-ui/tool/fallback-model/index.ts b/apps/desktop/src/components/assistant-ui/tool/fallback-model/index.ts
index 91f86eefb7b..68f57c82f23 100644
--- a/apps/desktop/src/components/assistant-ui/tool/fallback-model/index.ts
+++ b/apps/desktop/src/components/assistant-ui/tool/fallback-model/index.ts
@@ -17,6 +17,7 @@ import {
contextValue,
formatDurationSeconds,
isRecord,
+ NEUTRAL_FAILURE_TOOLS,
numberValue,
parseMaybeObject,
unwrapToolPayload
@@ -735,6 +736,10 @@ function toolStatus(part: ToolPart, resultRecord: Record<string, unknown>): Tool
return 'notice'
}
+ if (NEUTRAL_FAILURE_TOOLS.has(part.toolName)) {
+ return 'notice'
+ }
+
// A rejected memory write is a budget negotiation, not a failure: the store
// refuses an over-limit batch and the agent retries smaller. Soft warning β€”
// never destructive-red beside routine bookkeeping.
diff --git a/hermes_cli/cli_modern_layout.py b/hermes_cli/cli_modern_layout.py
index 096a911617d..e61cc570426 100644
--- a/hermes_cli/cli_modern_layout.py
+++ b/hermes_cli/cli_modern_layout.py
@@ -166,20 +166,22 @@ def completed_tool_line(cli, function_name: str, args: dict | None, duration: fl
def tool_row(function_name: str, args: dict | None, duration: float, result=None, is_error: bool = False) -> str:
"""``πŸ“– Read calc.py 0.2s`` / ``βœ— $ pytest -q 0.4s [exit 1]`` β€” one row per completed call, led by
- the tool's emoji (the classic ``β”Š`` row's), or a red ``βœ—`` when that entry failed."""
- from agent.display import get_tool_row_parts
+ the tool's emoji (the classic ``β”Š`` row's), or a red ``βœ—`` when that entry failed. A tool in
+ ``NEUTRAL_FAILURE_TOOLS`` keeps its emoji and shows the reason dim."""
+ from agent.display import NEUTRAL_FAILURE_TOOLS, get_tool_row_parts
rows = get_tool_row_parts(function_name, args or {}, result, is_error=bool(is_error))
+ neutral = function_name in NEUTRAL_FAILURE_TOOLS
out = []
for index, (emoji, verb, detail, suffix) in enumerate(rows):
verb = (verb or function_name).strip()
- mark = f"{_RED}βœ—{_RST}" if suffix is not None else (emoji or "●")
+ mark = f"{_RED}βœ—{_RST}" if suffix is not None and not neutral else (emoji or "●")
line = f"{mark} {_BOLD}{verb[:1].upper()}{verb[1:]}{_RST}"
if detail:
line += f" {detail}"
if index == len(rows) - 1 and duration >= 0.1:
line += f" {_DIM}{duration:.1f}s{_RST}"
if suffix:
- line += f"{_RED}{suffix}{_RST}"
+ line += f"{_DIM if neutral else _RED}{suffix}{_RST}"
out.append(line)
return "\n".join(out)
diff --git a/tests/hermes_cli/test_cli_modern_layout.py b/tests/hermes_cli/test_cli_modern_layout.py
index 907315b84b0..abcd2fd2e99 100644
--- a/tests/hermes_cli/test_cli_modern_layout.py
+++ b/tests/hermes_cli/test_cli_modern_layout.py
@@ -209,6 +209,19 @@ def test_tool_rows_keep_the_classic_tool_emoji():
assert failed.startswith("βœ— ") and "[exit 1]" in failed, failed
+def test_refused_skill_write_reads_neutral_not_failed():
+ """A skill_manage write the agent will re-plan (size cap, stale old_string) keeps the tool's own
+ emoji and a dim reason: no red βœ— for the agent's routine retry loop."""
+ from agent.display import get_tool_emoji
+ from hermes_cli.cli_modern_layout import _RED, tool_row
+
+ refused = ('{"success": false, "error": "operations[0] (patch on \'hermes-agent-dev\') failed: '
+ 'SKILL.md content is 100,193 characters (limit: 100,000)."}')
+ row = tool_row("skill_manage", {"operations": [{"action": "patch", "name": "hermes-agent-dev"}]}, 1.1, refused)
+ assert _plain(row).startswith(get_tool_emoji("skill_manage") + " "), row
+ assert "[operations[0]" in _plain(row) and _RED not in row, row
+
+
def test_every_block_is_framed_by_bars_and_blank_lines(monkeypatch):
"""User message, reasoning and answer each read as: blank line, bar, content, bar, blank line.
The blank line after a closing bar is paid by whatever prints next (a tool row, the footer)
diff --git a/tests/tui_gateway/test_show_reasoning_display_gate.py b/tests/tui_gateway/test_show_reasoning_display_gate.py
index 37720ec5bc9..9cc861c27a9 100644
--- a/tests/tui_gateway/test_show_reasoning_display_gate.py
+++ b/tests/tui_gateway/test_show_reasoning_display_gate.py
@@ -126,6 +126,17 @@ def test_tool_progress_off_keeps_only_essential_tool_rows(monkeypatch):
assert failed[0][2]["result"]["error"] == "disk full"
+def test_tool_progress_off_does_not_surface_a_refused_skill_write(monkeypatch):
+ """Answer-only mode forwards failures the user must see; a skill write the agent re-plans is not one."""
+ events = _capture(monkeypatch)
+ _session(monkeypatch, "tools-off-skill", show_reasoning=True, tool_progress="off")
+
+ refused = json.dumps({"success": False, "error": "SKILL.md content is 100,193 characters (limit: 100,000)."})
+ server._on_tool_complete("tools-off-skill", "tool-skill", "skill_manage", {"operations": []}, refused)
+
+ assert _lifecycle(events, "tool-skill") == []
+
+
def test_subagent_tool_mirror_follows_tool_progress_not_reasoning(monkeypatch):
for sid, show_reasoning, tool_progress, expected in (
("child-reasoning-off", False, "all", ["tool.start", "tool.complete"]),
diff --git a/tui_gateway/tool_progress.py b/tui_gateway/tool_progress.py
index db024b80a4e..8d82c87f5a7 100644
--- a/tui_gateway/tool_progress.py
+++ b/tui_gateway/tool_progress.py
@@ -362,9 +362,10 @@ def _on_tool_complete(sid: str, tool_call_id: str, name: str, args: dict, result
payload.update(todo_state)
if session is not None:
_cache_todo_state(session, todo_state)
+ from agent.display import NEUTRAL_FAILURE_TOOLS
if (_tool_progress_enabled(sid) or payload.get("inline_diff") or _tool_lifecycle_required_for_ui(name)
or is_todo_tool_name(name) or _connector_tool_lifecycle(name, args)
- or _tool_result_needs_user(result)):
+ or (name not in NEUTRAL_FAILURE_TOOLS and _tool_result_needs_user(result))):
_emit_tool_lifecycle("tool.complete", sid, name, args, payload)
# Task state is application data, not tool-progress chrome: a dedicated full-snapshot event lets
# every client reconcile without parsing tool args.
diff --git a/website/docs/user-guide/cli.md b/website/docs/user-guide/cli.md
index cf4abba893f..2adca4b4c50 100644
--- a/website/docs/user-guide/cli.md
+++ b/website/docs/user-guide/cli.md
@@ -207,7 +207,7 @@ Added multiply; pytest isn't installed here, so I ran the tests directly: 2 pass
gpt-5-mini Β· ctx [β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘] 4% Β· ~/project (main) enter interrupt Β· tab queue Β· ctrl+c stop
```
-- A 40-column bar opens each user message and answer (accent) and each reasoning block (dim); reasoning streams live as in classic and ends with a `∴ Thought for 3s Β· <title>` line. `❯` your message, a finished tool call led by its emoji (`πŸ“– Read`, `πŸ’» $`; `βœ—` in red when it failed, with its exit code), `β”‚` an edit's diff under its row, `☀` the answer, and `─` the turn's summary line.
+- A 40-column bar opens each user message and answer (accent) and each reasoning block (dim); reasoning streams live as in classic and ends with a `∴ Thought for 3s Β· <title>` line. `❯` your message, a finished tool call led by its emoji (`πŸ“– Read`, `πŸ’» $`; `βœ—` in red when it failed, with its exit code; a refused `skill_manage` write keeps its emoji and shows the reason dim, since the agent retries it itself), `β”‚` an edit's diff under its row, `☀` the answer, and `─` the turn's summary line.
- The composer's top rule carries the live activity (spinner, what's running, elapsed time, streamed tokens) on the left and the session title on the right; no spinner row is printed into the scrollback.
- The status line sits under the composer: the same fields as the classic status bar (model, context, cache hit, latency, tokens/s, background work, goal, duration, YOLO…) plus the working directory (with git branch) on the left; the keys that do something right now on the right.
- The welcome banner is the same in both layouts.