Fix intent classification from semantic main request

[giscebot]

Summary

• update the intent classifier prompt to separate the human comment's main request from subordinate rationale before choosing work_intent
• include main_request and subordinate_reason in classifier metadata for easier audits
• keep the prompt renderer safe for literal JSON examples and add regression coverage

Verification

• .venv/bin/pytest tests/test_intent_classifier.py
• .venv/bin/pytest tests/test_intent_classifier.py tests/test_queue.py tests/test_policy.py tests/test_packaged_resources.py
• real gpt-5.4-mini prompt check on @giscebot fes una pull request per cada notificació... returned reply_comment/work_allowed with confidence 0.96

[pilipilisbot]

Approved. I reviewed the classifier prompt/resource changes, the new metadata propagation in IntentClassification.to_metadata(), and the queue integration path that stores the classifier metadata and applies the LLM result only when valid/confident. The decomposition fields are optional in metadata, so existing callers remain compatible, while the packaged prompt now explicitly asks the classifier to separate the main request from subordinate rationale before choosing work_intent.

Checks run locally against the PR worktree using the repository venv with PYTHONPATH=src:

../github-agent-bridge/.venv/bin/python -m pytest tests/test_intent_classifier.py tests/test_queue.py tests/test_policy.py tests/test_packaged_resources.py

Result: 47 passed. GitHub Actions are also green for dashboard and Python 3.11/3.12 pytest.

[pilipilisbot]

Post-merge cleanup complete for PR #143: removed the clean dedicated worktree /home/openclaw/.openclaw/workspace/github/github-agent-bridge-pr143. No code changes were made; the canonical checkout was left untouched.

[pilipilisbot]

Reviewed the merged PR state again: current head c2da9f2 had already been formally approved, the PR is merged, and the focused classifier/queue/policy/resource suite passes locally. No further changes requested.