Media Curator → Research Complete: Acquisition Handoff

Status: Interface specification Related issues: #817 (enforce PDF acquisition), #818 (this handoff)

Design Principle

Media curation is upstream. Research is downstream. Acquisition of files from external sources is media curation's responsibility — research consumes what media curation provides.

This handoff formalizes the boundary that removed ~40 embedded `curl` calls from research induction agents and eliminated duplicate functionality between `/research-acquire` and `/acquire`. For time-based media, the same boundary now extends through `transcribe-media` and `induct-media`: media-curator acquires and transcribes; research-complete turns the result into citable REF artifacts.

Responsibilities

Media Curation (upstream)

Discovery: Finding source candidates (`/find-sources`)
Acquisition: Downloading files from any source (`/acquire`)
Format normalization: Converting to canonical formats (PDF, HTML, MP3, MP4)
Storage: Placing files in canonical locations
Metadata extraction: Pulling title, authors, dates, checksums from the file itself
Integrity: SHA-256 hashing, archival packaging

Media curation answers: "Given a URL or query, produce a file on disk."

Research Complete (downstream)

Analysis: Reading the file and writing 15-section analysis docs (`/induct-research`)
Citation extraction: Parsing references from analyzed papers
Indexing: Building topic/year/author/citation graphs (`/corpus-index-build`)
Linking: Bidirectional citations (`/citation-backfill`), clusters (`/research-gap-detect`)
Quality assessment: GRADE scoring (`/research-quality`), depth audits (`/research-quality-audit`)
Reporting: Corpus snapshots (`/corpus-snapshot`), gap analyses

Research answers: "Given a file produced by media curation, extract knowledge and integrate it into the corpus."

For videos, lectures, podcasts, talks, and interviews, research-complete also answers: "Given acquired media plus a transcript sidecar, produce a REF, sidecars, timestamp citations, and searchable corpus metadata."

Canonical Storage Locations

Media curation writes to these paths; research reads from them:

.aiwg/research/sources/
├── pdfs/
│   ├── full/                    # Full PDF files (primary artifact)
│   │   └── REF-016-autogen.pdf
│   └── preview/                 # Abstract-only PDFs (fallback)
├── web/                         # Saved web pages
│   └── REF-298-blog-post.html
├── text/                        # Extracted plain text (primary input for analysis)
│   └── REF-016.txt
└── metadata/                    # Extracted metadata sidecars
    └── REF-016.yaml
media/
├── video/                       # Copied or LFS-managed video when permitted
├── audio/                       # Copied or LFS-managed audio when permitted
└── transcripts/                 # Timestamped transcript sidecars

Handoff Protocol

1. Research requests acquisition

When `/induct-research` encounters a paper URL, it delegates to media curation:

# Research invokes media-curator's acquire skill:
/acquire --url <paper-url> --format pdf --output .aiwg/research/sources/pdfs/full/ --extract-text

# On success, media curator returns:
{
  "status": "acquired",
  "ref_id": "REF-016",
  "pdf_path": ".aiwg/research/sources/pdfs/full/REF-016-autogen.pdf",
  "text_path": ".aiwg/research/sources/text/REF-016.txt",
  "metadata_path": ".aiwg/research/sources/metadata/REF-016.yaml",
  "checksum": "a1b2c3d4...",
  "size_bytes": 2457600
}

# On failure:
{
  "status": "acquisition-failed",
  "reason": "paywall|dead-link|parse-error|timeout",
  "suggestion": "Try /find-sources for alternative URLs"
}

2. Research consumes the acquired file

Once acquisition succeeds, research runs its analysis pipeline:

# Research reads the extracted text and writes analysis:
/induct-research --ref-id REF-016 --source-text .aiwg/research/sources/text/REF-016.txt

3. Research requests targeted hunting

When research has gap analysis results, it asks media curation to hunt for specific papers:

# Research generates bridge candidate list from /research-gap-detect
# Media curation finds and acquires them:
/find-sources --query "LLM agent GUI interaction orchestration" --limit 10
/acquire --plan .curator/sources/plan-XXX.yaml

4. Time-based media induction

When the source is a recording rather than a paper, media-curator produces an acquisition record and transcript sidecar before research-complete induction:

# Acquire through media-curator patterns (YouTube, Internet Archive, direct URL).
/acquire --url "https://example.invalid/watch?v=lecture" --format video \
  --output .aiwg/media/acquisitions/

# Produce a research-grade transcript sidecar.
/transcribe-media .aiwg/media/acquisitions/REF-123-lecture.mp4 \
  --source-url "https://example.invalid/watch?v=lecture" \
  --output .aiwg/research/media/transcripts/REF-123-lecture.transcript.json

# Hand off to research-complete.
/induct-media .aiwg/media/acquisitions/REF-123-lecture.metadata.json \
  --transcript .aiwg/research/media/transcripts/REF-123-lecture.transcript.json \
  --source-url "https://example.invalid/watch?v=lecture" \
  --storage hash-only --ref-id REF-123

Expected research outputs:

`.aiwg/research/findings/REF-123.md` from `reference-media.md`
`.aiwg/research/citations/REF-123-citations.md` with spoken references
`.aiwg/research/radar/REF-123-radar.md`
index/search metadata for title, source type, speakers, channel/venue, source

URL, transcript path, and storage policy

Timestamp citations use the corpus REF and exact transcript quote:

@.aiwg/research/findings/REF-123.md @ 00:12:34 - "exact transcript quote"

`citation-guard` accepts this form only after the REF and transcript sidecar exist and the timestamped quote matches the transcript segment. `research-query` can then ground answers against the REF metadata and transcript text when the project index includes transcript content.

Cross-Pollination

The handoff is not one-directional:

Research signals media curator	Media curator signals research
Gap analysis → acquisition priorities	New acquisitions → induction queue
Stub docs with missing PDFs → re-acquire requests	Format availability (PDF vs preview)
Bridge candidate searches	Quality indicators from acquisition (paywall, scan quality)

Deprecation Path

`/research-acquire` → delegated to `/acquire`

The existing `/research-acquire` skill is retained as a thin wrapper that delegates to `/acquire`:

`/research-acquire` stays as the entry point for research workflows
Internally, it calls `/acquire` for the actual download
Research-specific post-processing (REF assignment, finding doc generation) stays in `/research-acquire`

This preserves backwards compatibility while eliminating the duplicate implementation.

Embedded `curl` calls → removed

All ~40 embedded `curl`/`wget` calls in research induction agents should be replaced with calls to `/acquire`. This is tracked separately.

Interface Contract

Contract element	Value
Input protocol	URL, DOI, arXiv ID, or search query
Output artifacts	File in canonical location + metadata sidecar + extracted text
Failure modes	`acquisition-failed` status with reason + suggestion
Idempotency	Re-running `/acquire` on same URL returns cached result unless `--force`
Concurrency	Media curator handles rate limiting; research does not need to
Checksums	SHA-256, stored in metadata sidecar and corpus fixity manifest

Integration Points

Skill	How it uses the handoff
`/induct-research`	Requests acquisition in Phase 2 (#817)
`/research-gap-detect`	Delegates bridge candidate acquisition to `/find-sources` + `/acquire`
`/corpus-export`	Reads acquired files from canonical locations for packaging (#819)
`/research-quality-audit`	Flags stubs with missing PDFs; triggers re-acquisition

References

`@agentic/code/frameworks/media-curator/skills/acquire/SKILL.md` — Acquisition command
`@agentic/code/frameworks/media-curator/skills/find-sources/SKILL.md` — Source discovery
`@agentic/code/frameworks/research-complete/skills/research-acquire/SKILL.md` — Research-specific wrapper
`@agentic/code/frameworks/research-complete/skills/induct-research/SKILL.md` — Consumes acquired files
Issue #817 — Enforce PDF acquisition before analysis
Issue #818 — This handoff specification