amirkiarafiei/open-scholar-peer/5-osp-qa

Tool capability: This tool documents a subagent framework, and subagent isolation is preferred here. Try it first: delegate each question to osp-answer-generator-agent as a subagent with a fresh, minimal context bundle.

If delegation is unavailable in your session — the persona is not reachable as a subagent, the call errors, or the capability is simply absent — fall back to self-reflection with strict turn markers (=== Query Agent === ... === END === === Answer Generator === ...) in the main context window, and carry on. Do not stop the phase over it.

Either way, record which one you used in the ## Method section of each 05_qa_<slug>.md (Mode: subagent or Mode: self-reflection), so the artifact says how the answers were actually produced.

/5-osp-qa — Multi-Aspect Q&A Engine

For each criterion in session.json.qa_criteria[], generate N probing Q&A pairs (N = qa_pairs_per_criterion, default 2). The Query Agent runs in the main thread; the Answer Generator runs as a subagent (or self-reflects on the few tools that lack them).

Activation

Invoke the osp-query-agent skill (main thread). The Query Agent will spawn osp-answer-generator-agent per question.

Inputs — none of these is a gate

• phases.summary.status, phases.literature.status, phases.historian.status and

phases.baseline_scout.status. Each one that is missing removes a source of evidence, not the phase: no summary means you question the paper directly; no literature or narrative means novelty claims cannot be checked against prior work; no baseline scout means no missing-baseline questions. Run with what exists, name each gap in the artifact's Provenance, and let the answers say "could not be verified" rather than guessing.

• qa_criteria[] in session.json. If it is empty, onboarding never ran — use the generic criteria from

.agents/defaults/generic_review_guidelines.md and record that you did.

Record any skip before you start. The orchestrator is not in the loop when the user runs this command directly, so it falls to you: for every earlier phase still pending, set phases.<name>.status = "skipped" and phases.<name>.skip_reason to their reason, or "user ran /5-osp-qa first". A phase left pending reads as "not reached yet", and the final review cannot tell the difference.

Step 0 — Resource check and pair count (run BEFORE any Q&A work)

1. Read session.json. Count qa_criteria[] items (call it C).
2. Read qa_pairs_per_criterion from session.json (default 2 if absent).
3. Print the resource estimate:

``` ⚠️ Q&A Engine — resource estimate Criteria: C Pairs/criterion: N (currently set in session.json) Total subagent calls: C × N = <total> Estimated time: ~<total × 45s> at typical API latency

Pair count guide: 2 — quick scan, catches the most obvious issues (default) 5 — thorough coverage, good for most reviews 10 — exhaustive, suitable for high-stakes decisions ```

1. Ask the user: "How many Q&A pairs per criterion? Press Enter for [N] or type a number (2–10):"
2. If the user enters a number, update session.json.qa_pairs_per_criterion to that value and use it.

If the user presses Enter, use the existing value.

Mode selection

The banner at the top of this command says which mode your tool is in.

• Subagent mode (default): the Query Agent delegates each question to the Answer Generator as a subagent with a fresh, minimal context bundle.
• Prefer-subagent mode: try delegation first; if it is unavailable in your session, fall back to self-reflection and keep going rather than stopping the phase.
• Self-reflection mode: the Query Agent uses strict turn markers (=== Query Agent === ... === END === === Answer Generator === ...) within the main context window.

Whichever you end up using, record it as Mode: in each file's ## Method section.

Opening block (print before step 1)

Render the opening block exactly as .agents/defaults/phase_block_template.md defines it — that file holds the rail, the rules and the widths, and it is the only place they are written down. This is phase 6 of 7 (qa); read the rail's state from session.json. Values:

DOING probe each criterion, and verify the answers READS .brain/raw/ summary, narrative, baselines WRITES .brain/raw/05_qa_<slug>.md (one per criterion) COST <C>x<N> subagent calls, ~45 s each

Steps

1. Read all input artifacts listed in the frontmatter.
2. Activate the osp-query-agent skill.
3. For each criterion in qa_criteria[]:

• Open or initialize .brain/raw/05_qa_<criterion_slug>.md from .agents/defaults/qa_pair_template.md.
• Generate exactly qa_pairs_per_criterion Q&A pairs:
• For each, the Query Agent formulates a probing question grounded in the structured summary, narrative, and missing baselines.
• The Query Agent delegates to the Answer Generator (subagent or self-reflection per mode).
• The Answer Generator returns (answer, citations, discrepancy_flag).
• The pair is appended to the file.
• Update phases.qa.criteria_progress[<slug>] = "completed".

1. After all criteria are done:

• phases.qa.status = "completed"
• phases.qa.notes = "<C> criteria × <N> pairs each; <M> discrepancies flagged"
• resume_from = "review"

Closing block (print when the phase ends)

Render the closing block from .agents/defaults/phase_block_template.md. Build the rail from `session.json` — ● only where status == "completed", ○ for pending and skipped. A phase the user skipped must not show as done. Drop BLOCKED and NOTE when there is nothing to put on them. Values:

DONE <N> pairs, <C> criteria, <M> discrepancies .brain/raw/05_qa_<slug>.md NEXT /6-osp-review write the review

Re-run behavior

Re-running overwrites 05_qa_<slug>.md for every criterion it covers, and these files cost the most of anything here — one subagent call per pair. Say what will be lost and wait for an answer before overwriting. To redo a single criterion, pass its slug as an argument; the rest are left alone.

Pitfalls

• The Answer Generator must NOT see prior questions in subagent mode. Each invocation is stateless.
• The Query Agent must NOT answer its own questions.
• Do not generate fewer pairs than qa_pairs_per_criterion. The count must match exactly.