[safeoutputs] Disambiguate field names and document constraints for create_code_scanning_alert, push_to_pull_request_branch, noop, create_disc [Content truncated due to length]

[github-actions]

Analysis of agentic workflow runs from the last 24 hours (75 runs, about 40 with safe-output activity) surfaced 6 safe-output schema rejections at the MCP server (captured in rpc-messages.jsonl as rpc_response.error). Four are tool-description gaps: the field name or constraint the agent intuitively reaches for is not disambiguated by the description. None duplicate the open issues (the create_discussion parameter-wrapping issue from 2026-06-05 or the submit_pull_request_review targeting-field issue from 2026-06-02).

The fix style is the one already used for issue_number to item_number, category_id to category, and the name-on-create_check_run cases: an explicit "this field is named X, NOT Y" note in the field description.

Summary

Tool	Agent sent	Schema expects	Type	Source
create_code_scanning_alert	level	severity	wrong field	Smoke Claude (test)
push_to_pull_request_branch	commit_message	message	wrong field	Smoke Claude (test)
noop	structured metric fields	message (string)	missing required field	PR Sous Chef (prod)
create_discussion	body over 10 KB	body 10 KB or less	undocumented constraint	2 prod report workflows

All six agents received an instructive error and, except the noop case, recovered. So this is about cutting avoidable round-trips and the one noop run that ended on the rejection, not about hard failures.

Honest scoping note

Two of the four (level, commit_message) come from Smoke Claude, a test harness that deliberately exercises every safe-output tool with dummy data, so they are single-run, test-context evidence. They are included because the field-name choice was the agent own decision (the prompt does not dictate it), which is exactly the disambiguation signal we act on. The noop and create_discussion gaps are from production workflows and are the higher-priority items.

Error Pattern Details and Evidence

Pattern 1 - create_code_scanning_alert: level used instead of severity. Run 27067065704 (https://github.com/github/gh-aw/actions/runs/27067065704, Smoke Claude, claude). Sent file=README.md, level=warning, line=1, message=dummy. Server returned: Invalid arguments: missing or empty severity. The word level is a near-universal synonym for severity (Python logging, syslog, most linters); the severity field description lists the enum but does not warn against level.

Pattern 2 - push_to_pull_request_branch: commit_message used instead of message. Run 27067065704 (same run). Sent commit_message=test add smoke test file, pr_number=37338. Server returned: Invalid arguments: missing or empty message. The field is named message but its description reads "Commit message describing the changes", so the agent reaches for commit_message (also the git and GitHub API convention).

Pattern 3 - noop: structured summary fields instead of the required message string. Run 27067988983 (https://github.com/github/gh-aw/actions/runs/27067988983, PR Sous Chef, copilot/gpt-5-mini). Sent branch_update_attempts, formatter_pushes, nudged_other, nudged_review_comments, processed, skipped, and more, with no message. Server returned: Invalid arguments: missing or empty message. The same workflow in run 27073954344 (https://github.com/github/gh-aw/actions/runs/27073954344) did it correctly by serializing the summary into the message string. So the tool works fine when the agent serializes; a description nudge should close the gap.

Pattern 4 - create_discussion: body exceeds the undocumented 10 KB limit. Two distinct production report workflows in the same window: run 27070604637 (https://github.com/github/gh-aw/actions/runs/27070604637, Copilot Agent PR Analysis, claude) body 13712 bytes; run 27067333745 (https://github.com/github/gh-aw/actions/runs/27067333745, Daily Security Observability Report, copilot/claude-sonnet-4.6) body 10620 bytes. Server returned: Input string parameter(s) exceed the 10 KB limit for tool create_discussion: body N bytes. The create_discussion body field documents no size limit, whereas add_comment documents its 65536-character limit inline. Report-generating workflows naturally produce bodies over 10 KB.

Recommended Improvements

All edits in pkg/workflow/js/safe_outputs_tools.json, then propagate the identical change to the runtime copy actions/setup/js/safe_outputs_tools.json (per the 2026-05-30 deployment-gap lesson, both copies must stay in sync).

1. create_code_scanning_alert.severity - append: This field is named severity, NOT level.
2. push_to_pull_request_branch.message - append: This field is named message, NOT commit_message.
3. noop.message - append: This is the only content field noop accepts. Serialize any metrics or structured run summary into this single string; additional named fields (for example processed, skipped) are not accepted and cause a missing-message error.
4. create_discussion.body - append: Maximum 10 KB (10240 bytes). If your content is larger, trim it, link out to a gist or artifact, or attach it via upload_asset rather than inlining. Mirror the inline-limit style already used by add_comment.

Root Cause

Each gap is the same class: the description omits the one disambiguation or constraint the agent then violates. The level and commit_message synonyms and the absent size limit are all predictable from common API conventions, and the established remediation (explicit named-X-not-Y notes) has worked for the prior issue_number, category_id, and name cases.

Implementation Checklist

• Update the 4 field descriptions in pkg/workflow/js/safe_outputs_tools.json
• Propagate identical changes to actions/setup/js/safe_outputs_tools.json
• Run make build and make recompile
• Run make test (including the tools-JSON sync test)
• Monitor rpc-messages.jsonl rejections for these 4 tools over the next 2-3 days

Notes and Out of Scope

• submit_pull_request_review empty-body rejection (Matt Pocock Skills Reviewer, run 27068395228) - existing ERR_VALIDATION message is already clear; tracked, no change.
• Empty-argument discovery probes (PR Code Quality Reviewer, run 27066992596) - known probe pattern; message already optimal; not actionable.
• mcpscripts query-script failures (run 27073816278) - command failures, not safe-output schema errors; out of scope.

References: run 27067988983, run 27070604637, run 27067065704.

Generated by ⚡ Daily Safe Output Tool Optimizer · 606 AIC · ⌖ 32.7 AIC · ⊞ 10.5K ambient context · ◷

• expires on Jun 8, 2026, 9:34 PM UTC

[pelikhan]

Extend the safe-outputs tool definition to include a set of accepted synonyms for each parameter. If the agent calls a parameter using an alternate name that clearly maps to a supported parameter, accept the request.

If a parameter does not match any supported name or synonym, the SafeOutput MCP server should reject the request more strictly. The error message should clearly indicate the unknown parameter and suggest the closest valid parameters supported by the tool.