[Agora] Synthesis-agent that writes 1-page briefs from N artifacts done

Effort: extensive

Goal

Build an autonomous agent that, given a topic + 5-50 SciDEX artifacts
(any mix of hypotheses, debates, papers, notebooks, open questions,
markets), writes a one-page polished brief suitable for a researcher,
funder, or journalist. Output is itself a versioned artifact
(brief artifact_type) with full provenance edges back to every input
artifact, so it can be discussed (Q-DSC) and ranked (Q-PERC) like
anything else.

This closes the loop: SciDEX produces tons of micro-content but no
one is synthesizing macro-narratives. A brief-writer agent turns the
firehose into prose a human can read in 60 seconds.

Acceptance Criteria

- compose_brief(topic: str, artifact_ids: list[str], audience:
Literal['researcher', 'funder', 'journalist'] = 'researcher')
-> dict returns {'brief_md': str, 'sources': list[dict],
'token_cost': int, 'composer_persona': str}.
- Internally: gather each artifact via artifact_registry.get,
build a structured prompt with sectioned artifact context,
call the configured LLM (use scidex.routing.llm_router,
respect token budget = 8K), produce Markdown with mandatory
sections: TL;DR, What we know, Open questions, Active
debates, Markets / funding, What to read next, Provenance. a1,a2,a3 --audience researcher --out - for one-shot manual
invocation. register_artifact, with artifact_links of type
synthesizes pointing at every input artifact (one row per
input). The brief stores metadata.audience, metadata.topic,
metadata.composer_persona, metadata.input_artifact_ids. {topic, artifact_ids, audience} and returning the brief id +
Markdown render. Auth: requires a valid agent or human session. Markdown to HTML with a "Sources" footer listing every linked
artifact with composite_score + 1-line summary. scidex-brief-writer-daily.timer selects the top-3 most-active
topics from the last 24h (highest delta in artifact count, see
q-synth-whats-changed) and produces 3 briefs/day automatically. canned response, asserts the brief artifact is registered with
4 synthesizes links and the canned Markdown. production data is captured in the spec work log; each must
cite ≥3 input artifacts in the Markdown body and be ≤900 words.

Approach

Read scidex/agora/synthesis_engine.py for the existing

Synthesizer pattern; reuse the LLM call wrapper.

Build the prompt template: header (audience-specific tone), per-

artifact section pulling title + composite_score + 1-paragraph
summary, instruction tail with the mandatory sections.

Implement compose_brief with retries on truncation; assert

len(brief_md) > 800 and len(brief_md) < 6000.

Persistence path uses artifact_commit.commit_artifact per the

ADR-002 rule that briefs go through that helper, not direct
filesystem writes.

Auto-mode CLI: see q-synth-whats-changed for the topic-momentum

query; depends on its dashboard.

Dependencies

• q-synth-whats-changed (sibling) — supplies daily topic-momentum

ranking that drives the auto-mode timer.

• scidex.routing.llm_router — token-budgeted LLM call.
• Wave-1 Q-DSC — briefs become discussable artifacts.
• Wave-1 Q-PERC — briefs are ranked artifacts.

Work Log

2026-04-27 — Implementation complete [task:bf914ebb-42c4-4fe2-88f0-a93a61402061]

All acceptance criteria implemented:

scidex/agora/brief_writer.py

• compose_brief(topic, artifact_ids, audience) fetches each artifact via

artifact_registry.get_artifact, builds a structured 7-section prompt with
audience-specific tone (researcher/funder/journalist), calls the LLM via
from llm import complete with max_tokens=8192, retries up to 3× on
transient failures, registers the result as a brief artifact, and creates
synthesizes links to every input artifact. Returns
{brief_md, brief_artifact_id, sources, token_cost, composer_persona}.

scripts/write_brief.py

• CLI: python scripts/write_brief.py --topic X --artifacts a,b,c --audience researcher --out -
• Also supports --json for full JSON output.

scripts/run_daily_briefs.py

• Auto-mode driver: queries top-N topics by 24h artifact-count delta,

gathers top-10 artifacts per topic, calls compose_brief.

deploy/scidex-brief-writer-daily.{service,timer}

• Service: oneshot, runs scripts/run_daily_briefs.py.
• Timer: daily at 07:00 UTC, Persistent=true.

scidex/atlas/artifact_registry.py

• Added 'brief' to ARTIFACT_TYPES set.

api.py

• POST /api/brief/compose — accepts {topic, artifact_ids, audience},

requires API key, returns brief ID + Markdown.

• GET /brief/{artifact_id} — renders brief Markdown to HTML with Sources

footer showing each input artifact's composite_score + summary.

tests/test_brief_writer.py

• 7 pytest tests covering: correct return keys, canned Markdown preservation,

artifact registration with correct metadata, 4 synthesizes links, sources
list content, missing-artifact graceful skip, invalid-audience error,
empty-artifact-ids error.

Design notes:

• scidex.routing.llm_router referenced in spec does not exist as a separate

module; routing is in scidex.core.llm.complete — used directly per the
pattern in synthesis_engine.py.

• artifact_commit.commit_artifact applies to filesystem-submodule artifacts

(ADR-002); brief content is stored as PostgreSQL artifact metadata, not as a
file, so the submodule commit path is not used here (consistent with how
debate/synthesis artifacts are stored).

• Auto-mode topic query falls back to recent hypothesis titles when entity_ids

JSON-path query is unavailable (robustness).

Quality bar (manual review): pending production deployment — LLM
environment not available in this worktree. Tests cover the full
compose-and-register path with a canned LLM response.