Skip to content

docs: correct scheduled-routines.md's trigger tool-access claim - #668

Merged
rosscado merged 1 commit into
mainfrom
docs/654-trigger-tool-access-reality
Oct 12, 2026
Merged

rosscado merged 1 commit into
mainfrom
docs/654-trigger-tool-access-reality

Conversation

@rosscado

Copy link
Copy Markdown
Contributor

Why

doc/scheduled-routines.md claimed both scheduled triggers (weekly maintenance, monthly audit) run with tools "limited to Bash, Read, Glob, Grep, Edit, Write, TodoWrite — no browser or MCP tools," and called this "the mechanical half of 'provably excludes Layer 4'." Issue #654 flagged that this session (and prior routine runs) observably had GitHub MCP access the doc said shouldn't exist, but couldn't pull the trigger's actual config to pin down the real list.

This run has mcp__Claude_Code_Remote__get_trigger / list_triggers available, so it queried both trigger IDs directly.

What

Ground truth from the live trigger definitions: neither trig_01VgDQW3FWV5Ln4HUutsHBRt (weekly) nor trig_01RoxWz7WbEqtHxe4y62A99M (monthly) carries an allowed_tools restriction at all. Both have:

  • the full session tool surface (Bash, Agent, Artifact, Workflow, ToolSearch, ScheduleWakeup, SendUserFile, Skill, ListAgents, the github MCP server, …)
  • nine additional MCP connectors (Google_Calendar, Slack, Gmail, Google_Drive, Figma, Vercel, Linear, Claude_Code_Remote, Format), each with an empty permitted_tools (i.e. unrestricted)

So the doc's specific tool list and its "no MCP tools" claim were both wrong, and the "mechanical half of 'provably excludes Layer 4'" framing doesn't hold — Bash alone is sufficient to invoke layer4cdp/dev-rig/the sweeps. Updated doc/scheduled-routines.md to state this accurately: the hermetic/no-Layer-4 guarantee rests entirely on the routine prompt's HARD RULES (the same trust basis as every other "never" in AGENTS.md), not on a tool allowlist.

No code changes; docs only.

Verification

  • Read the trigger JSON directly via mcp__Claude_Code_Remote__get_trigger for both IDs and list_triggers for cross-check — no secrets in the output, only connector names/UUIDs and the stored prompt text (already public in this doc).
  • git diff reviewed; change is confined to doc/scheduled-routines.md.

Closes #654.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TgrUL5Je2LLNyxN61gE9qt


Generated by Claude Code

The doc claimed both cloud triggers run with tools limited to Bash,
Read, Glob, Grep, Edit, Write, TodoWrite and no MCP tools, calling that
the mechanical half of "provably excludes Layer 4". Querying the live
trigger definitions (mcp__Claude_Code_Remote__get_trigger/list_triggers)
shows neither carries an allowed_tools restriction: both have the full
session tool surface plus nine MCP connectors with empty
permitted_tools (unrestricted). There is no tool-level exclusion of
Layer 4 — Bash alone can invoke it. The hermetic guarantee is enforced
by the routine prompt's HARD RULES, not by tool access control.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TgrUL5Je2LLNyxN61gE9qt
@rosscado

Copy link
Copy Markdown
Contributor Author

Review verdict: approve (docs-only)

Reviewed the diff (git diff main...HEAD), which touches exactly one file, doc/scheduled-routines.md (15 insertions, 4 deletions), with no code, config, or CI changes.

Factual accuracy: The new text matches the live trigger reality described in the issue: the old claim of a Bash, Read, Glob, Grep, Edit, Write, TodoWrite-only allowlist with no MCP tools is removed, and replaced with an accurate statement that neither trigger has a tool allowlist — both carry the full session tool surface plus nine MCP connectors (Google_Calendar, Slack, Gmail, Google_Drive, Figma, Vercel, Linear, Claude_Code_Remote, Format) with unrestricted permitted_tools, alongside the github MCP server. The count of "nine more MCP connectors" checks out. The doc now correctly attributes the hermetic/no-Layer-4 guarantee to the routine prompt's HARD RULES rather than to tool-access control, and that framing is consistent with the rest of the file (the weekly/monthly prompt text quoted later in the same doc already states the HARD RULES this way, so there's no internal contradiction). The one remaining mention of allowed_tools (in the "To (re)create one" section, describing the RemoteTrigger creation API's accepted fields) is a separate, still-accurate statement about the API shape, not a claim about the current triggers' configuration — it doesn't conflict with the correction.

Secrets/sensitive data: None introduced. The content added is connector names, a HARD RULES restatement, and a dated correction note — no tokens, keys, or credentials.

Scope: Docs-only, single file, no unrelated changes bundled in.

Prose/Markdown: Reads cleanly; bold/code-span markup is balanced, the dated correction note is clearly flagged, and links/formatting are consistent with the rest of the file.

No changes requested.


Generated by Claude Code

@rosscado
rosscado marked this pull request as ready for review October 12, 2026 06:16
@rosscado
rosscado merged commit e7498a3 into main Oct 12, 2026
5 checks passed
@rosscado
rosscado deleted the docs/654-trigger-tool-access-reality branch October 12, 2026 06:16
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Oct 12, 2026 •

Copy link
Copy Markdown

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-10-12T06:19:42.731504Z ea247bc Draft marked ready
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ea247bcf24

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Comment thread doc/scheduled-routines.md
`Artifact`, `Workflow`, `ToolSearch`, `ScheduleWakeup`, `SendUserFile`, `Skill`,
`ListAgents`, the `github` MCP server, plus nine more MCP connectors (`Google_Calendar`,
`Slack`, `Gmail`, `Google_Drive`, `Figma`, `Vercel`, `Linear`, `Claude_Code_Remote`,
`Format`) each with an empty `permitted_tools` (i.e. unrestricted). There is **no

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Verify connector access before calling it unrestricted

An empty permitted_tools field does not by itself establish that every connector tool is callable; current remote-trigger behavior has also treated [] as “nothing permitted,” and attached connectors can remain filtered or approval-gated (upstream reproduction). The trigger definitions therefore support saying these connectors are attached, but not that the routines carry their unrestricted runtime tool surface without verifying a fired session. This distinction matters when this section is used as the routine's access/threat inventory.

AGENTS.md reference: AGENTS.md:L154-L154

Useful? React with 👍 / 👎.

Comment thread doc/scheduled-routines.md
Comment on lines +63 to +65
"never" in `AGENTS.md`, not a tool-level allowlist. Treat "hermetic" as a prompted
behavior, not a provable one, until/unless the trigger's `mcp_connections` are
actually pared down. Manage the triggers at https://claude.ai/code/routines.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Keep hermeticity conditional on blocking the Bash paths

The until/unless qualification implies that paring down mcp_connections would make hermeticity mechanically provable, but the preceding sentence correctly notes that Bash alone can launch layer4cdp, dev-rig, or the sweeps. Even with every MCP connector removed, the unattended routine retains the exact real-host path forbidden by the repository policy, so mechanical exclusion must also block those Bash executions or their required capabilities rather than treating connector removal as sufficient.

AGENTS.md reference: AGENTS.md:L49-L49

Useful? React with 👍 / 👎.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

doc-drift: doc/scheduled-routines.md claims cloud routines have no MCP tools, but this session has GitHub MCP access

2 participants