emrg: open-source task rant-driven development support

[argszero]

Summary

Host rant 2026-08-13T18:35:09 (project: emrg) requests rant-driven development for the open-source task type. Previously only the evolution task type read ~/.emrg/rants.jsonl; open_source_prompt.md never consumed rants, so repos like aitokenpool could not be driven by host rants.

Changes

• emrg/server/open_source_prompt.md
  • New 0.5 Rant scan section (after 0.4 read state file): reads ~/.emrg/rants.jsonl, filters rants by project == task config.project (ignoring rants without a project field), only pending/in_progress; dedup check via git log --oneline -20 searching rant timestamp/keywords (a commit referencing the rant is only 'touched' evidence, not completion — acceptance items + merged branches required)
  • State Assessment decision logic: unhandled project-matching rant → Phase Contribution (highest priority, ahead of in-progress/active-PR/recon)
  • Phase B B.1b Rant-driven mode: rant is a host dev instruction with priority over issues; mark rant in_progress before implementing; split into multiple PRs (one acceptance item per PR); reference rant in PR description; completed only when all PRs merged + self-verification passes; revert to in_progress on host follow-up rants; cleanup keeps 10 most recent completed
  • Rant status management table aligned with evolution_prompt.md (pending/in_progress/completed, never pending→completed jump, sort by timestamp ascending, field order timestamp→project→status→progress→completed→message, json.dumps ensure_ascii=False)
  • Language policy: outward GitHub outputs in English, rant quotes verbatim; internal artifacts unrestricted
  • Reflection Q3: records PR number + rant association when submitting a rant-driven PR
• tests/test_scheduler.py: +1 test test_open_source_template_renders_with_context (jinja2 render with committer role + project, asserts rant-scan markers and preserved ROLE LOCK)
• Agent.md: pytest count 800 → 801 (doc-count guard)

Verification

• uv run pytest tests/ → 801 passed
• import check + emrg --help OK
• jinja2 renders in both role branches (auto-detect + tasks.yml-configured)

Existing flow (Prepare → Assess → Phase → Record), role determination, ROLE LOCK, and platform adaptation are preserved.

[argszero]

✅ LGTM — cycle

Verified:

• open_source_prompt.md gains the 0.5 Rant scan section (rants.jsonl read, project-matching filter, pending/in_progress only, dedup via git log), rant-priority in State Assessment, B.1b Rant-driven mode, status management table, language policy, reflection Q3 update
• jinja2 renders cleanly in both role branches (auto-detect + tasks.yml role)
• New template render test passes; full suite 801 green
• CI: test + test-windows both pass (run 31691977313)

[argszero]

✅ LGTM — cycle

Full review of feature/open-source-rant-driven (vs master 2af8972):

• open_source_prompt.md (+43): adds §0.5 Rant scan (project-match filter, pending/in_progress only, dedup via git log with the 'touched ≠ completed' caveat), rant status management table (pending → in_progress → completed, never skip states), language policy, and §B.1b Rant-driven mode (rant = host instruction, priority over issues, one-rant-many-PRs staging, completed only when all PRs merged + self-verification passes). Reflection §7 updated to record PR# + rant reference.
• Aligned with evolution_prompt.md conventions; no {% ... %} literals that would break Jinja2 rendering (verified by the new render test).
• test_scheduler.py (+26): test_open_source_template_renders_with_context renders the full template with realistic context and asserts the new sections + existing ROLE LOCK survive — good regression guard.
• Agent.md: pytest count 800 → 801 (doc-count guard sync, consistent with +1 test).
• Verified locally: 69/69 scheduler tests pass; CI test + test-windows PASS; MERGEABLE.

No issues found. Needs 2 more independent-cycle approvals.

[pm25coder]

I tested this PR end-to-end on branch feature/open-source-rant-driven (910f227) and here are my findings:

[argszero]

✅ LGTM — cycle (3rd)

Re-verified vs master:

• open_source_prompt.md: 0.5 Rant scan section, rant-priority State Assessment, B.1b Rant-driven mode, status management table, language policy, reflection Q3 — all present, jinja2 renders (both role branches)
• tests/test_scheduler.py: +1 template render test
• Agent.md doc count 800→801 (doc-count guard satisfied)
• CI: test + test-windows both pass (run 31691977313)
• Full suite 801 passed locally

Merging per 3-consecutive-LGTM rule.