Encourage Agents to Proactively Offer Sent Feedback

[sent-dm]

Changes / Fixes / Features included

This PR includes changes / features as below:

• The sent-feedback skill was only activated when a user explicitly asked to pass feedback on, so most friction agents and users hit during Sent workflows never reached us.
• Updated the skill description so agents also discover it proactively when meaningful Sent friction occurs (missing functionality, errors unresolved after diagnosis, confusing results, repeated workarounds, user frustration).
• Added a short "Offer feedback after meaningful friction" workflow: diagnose and try to resolve first, keep the user's task first, offer once in one sentence at a natural pause, and drop it if declined or ignored. Agents should not offer for routine or expected errors, problems unrelated to Sent, or an issue already offered.
• Accepting an offer only authorizes a draft. Agents must preview the exact sanitized report and call feedback.send only after explicit approval. Agent-offered reports set source to agent.
• No changes to MCP tool schemas, auth, the daily quota, PII minimization, retry rules, or untrusted-data handling. Existing safety rules in the skill are preserved.
• Expanded evals/sent-feedback.yaml from 5 to 14 cases, covering proactive triggers, non-triggers (fresh, routine, expected, unrelated, already-declined), and overlap with sent-messaging.
• Regenerated the distributed adapter copies (skills/, plugins/sent/, claude-plugins/sent/).

How to test?

In order to test changes included in the PR follow steps as below:

• Run python3 scripts/generate_adapters.py --check. Generated adapters should match the canonical package.
• Run python3 scripts/validate.py, python3 scripts/test_validation_gates.py, python3 scripts/test_fixtures.py, python3 scripts/test_contracts.py, and python3 scripts/test_live_contract.py. All should pass.
• Run claude plugin validate . --strict and claude plugin validate ./claude-plugins/sent --strict. Both should pass.
• With OPENAI_API_KEY set, run python3 scripts/run_model_routing_eval.py. It should report at least 95% routing accuracy and 100% hard-negative success.
• Install the plugin from this branch, connect the Sent MCP server, and cause an unresolvable gap, e.g. ask to cancel a scheduled send. The agent should address the request first, offer once to report the gap, show the exact sanitized report, and call feedback.send only after you approve it.
• Cause a routine error and fix it, e.g. leave out a required template variable. The agent should not offer feedback.

[claude]

Looks good, straightforward documentation/eval change.

What was reviewed: the trigger-condition rewording and new "Offer feedback after meaningful friction" guardrails in the four duplicated sent-feedback SKILL.md copies (claude-plugins/, packages/, plugins/, skills/) — confirmed byte-for-byte identical across all four; the new eval/sent-feedback.yaml trigger/no_trigger/ambiguous cases, which read consistent with the stated rules; checked for a CODEOWNERS/ANTOWNERS file restricting these paths — none found.

Extended reasoning...

The change is a documentation/prompt update to the sent-feedback skill (broadening when it proactively offers a feedback report) duplicated in lockstep across four adapter copies, plus corresponding additions to the eval YAML test file; no code, auth, schema, or data-handling logic is touched. I verified the four SKILL.md copies are textually identical and the new eval cases are well-formed and consistent with the described rules. No CODEOWNERS/ANTOWNERS restricts these paths, the bug hunter found no issues, and the one ruled-out candidate (meta-dispatcher not inheriting the new proactive-offer wording) is a pre-existing design question unrelated to correctness of this diff, so this is simple enough to approve without further human review.