Update changelog for v2.1.305

This commit is contained in:
Mike
2026-10-20 16:43:20 -06:00
parent b85e92d67a
commit d0b3a40a33
+29
View File
@@ -4,6 +4,35 @@ Note: Only use **NEW:** for entirely new prompt files, NOT for new additions/sec
### Claude Code System Prompts Changelog
# [2.1.305](https://github.com/Piebald-AI/claude-code-system-prompts/commit/b85e92d)
_+3,943 tokens_
- **NEW:** Agent Prompt: Security monitor email direct send example, Security monitor email group reply example, and Security monitor email group reply exception — Group reply, where enabled, lets "reply all" asks reach people the sender put on the thread; examples confine asks to their stated recipients and content.
- **NEW:** Agent Prompt: Security monitor email reply approval scope and Security monitor email sender-only approval scope — With email replies enabled, a sender's bare "yes" approves only what the one mail they are answering asked; otherwise a bare approval approves nothing.
- **NEW:** Agent Prompt: Security monitor email reply outbound content rule, Security monitor email reply outbound recipient rule, and Security monitor email sender-only outbound clearance rule — Mail to anyone besides the sender passes only if the sender's own words cover its recipients and content, read with the answered mail when replies are enabled.
- **NEW:** Agent Prompt: Security monitor email sender turns — Gives the auto-mode classifier feature-gated rules for emails a session's user sends: only the sender's own new words establish intent; quoted, forwarded, and attached content is data.
- **NEW:** Data: attributionProject setting — Documents a new setting naming the claude.ai Project credited in organization usage reports, sent only with claude.ai login and direct API access, outside cloud sessions.
- **NEW:** Skill: /doctor /usage flagged check — Adds Check 10, reporting /usage's flagged rows from the past week as untrusted data; rows with fixes become proposed edits, the rest advice for the user.
- **NEW:** Skill: Mod plugin packaging (any plan), Mod plugin packaging (organization plans), and Mod plugin packaging (personal plans) — Give plan-specific after-saving steps: press Share on the file's card (Team and Enterprise can publish to the organization where allowed) or send the file; Pro and Max only send.
- **NEW:** Skill: Mod plugin packaging consent — In flag-gated remote Cowork sessions, offers once to package a working mod as a plugin file handed over in chat, after a key scan; GitHub only on request.
- **NEW:** System Prompt: Claude in Chrome task-specific tool loading — Splits out, unchanged, the advice to add task-specific tools to the same ToolSearch call; Chrome browser MCP tools and Claude in Chrome deferred tool loading now share it.
- **NEW:** System Prompt: Windows sandbox localhost access (local binding off) — Splits out the blocked-localhost opening shared by the retry and user-alternatives prompts; it no longer says that no setting allows localhost.
- **NEW:** System Prompt: Windows sandbox localhost access (local binding on) and Windows sandbox server reachability (local binding on) — With `sandbox.network.allowLocalBinding` on, these replace the whole blocked-localhost guidance: commands reach localhost on any port but one the sandbox reserves; off-machine traffic still uses the proxy.
- **NEW:** System Prompt: Windows sandbox localhost workarounds (local binding off) — Splits out the shared list of what still works, now with lower-case proxy variables and `sandbox.network.allowLocalBinding` as the user's call, needing a set-up rerun and restart.
- **NEW:** System Reminder: Managed Agents ant command refused (altered --transform command) and Managed Agents ant command refused (staging, vrtx or bdrk ID) — Newly cataloged refusals the `ant` command guard, active once /claude-api loads, already returned; the staging, vrtx or bdrk one no longer lists `local_` IDs.
- **NEW:** System Reminder: Managed Agents ant command refused (local ID form or placement) and Managed Agents ant command refused (local servers unsupported) — Local Managed Agents server IDs, previously always refused, now pass the guard only in one exact form and placement; /claude-api managed-agents-secure still refuses them.
- Agent Prompt: Claude Test author — Tests for a suspected bug assert the right outcome and fail while it exists; only code contradicting itself or its visible promise warrants suspecting a bug.
- Agent Prompt: Claude Test explorer — Reports suspected bugs, which need strong evidence, beside the action they affect instead of omitting it, and maps files on a "Changed lately:" line to journeys.
- Data: allowManagedPermissionRulesOnly allowed-tools frontmatter scope — In Claude Desktop, host-delivered `--plugin-dir` plugins keep their `allowed-tools` frontmatter only while the organization's own managed settings leave this setting unset.
- Data: allowedMcpServers setting — Specifies which fields count: a `managed-mcp.json` entry with `${VAR}` in its command, args, URL, or header values must match the allowlist.
- Skill: /doctor slash command — When /usage has flagged issues, adds Check 10 to the report, the consolidated cleanup question, and the checks that may propose edits to checked-in files.
- System Prompt: Claude in Chrome browser automation and Claude in Chrome browser automation (non-deferred tools) — Now hold, respectively, the shared heading and introduction and the shared guidelines body; Claude Code reassembles them, deferred-loading steps included when needed, exactly as before.
- System Prompt: Windows sandbox localhost retry guidance and Windows sandbox localhost user alternatives — Now start from the shared blocked-localhost opening and workarounds, keeping their own retry or user-decision closings, which apply only while local binding is off.
- System Prompt: Windows sandbox profile tool retry guidance and Windows sandbox profile tool user guidance — Now say sandbox set-up normally makes the profile readable, except credential stores Claude must never read even when possible, and writable only where `write.allowOnly` allows.
- System Reminder: Proactivity level high — Claude now asks only before irreversible steps, like deleting customer data, not over costly ambiguity, and treats pushing or landing an approved PR as reversible.
- Tool Description: Plugin organization submission and replacement — Adds two results to relay: publishing allowed only from claude.ai (no retry) and an earlier submission awaiting review (retry only on request, once withdrawn).
# [2.1.304](https://github.com/Piebald-AI/claude-code-system-prompts/commit/7f5fe8d)
_+3,747 tokens_