Tooling
What ARC Is
ARC means Agent Research Copilot. In this workspace it is a cache-first research toolkit for theoretical-physics papers and research-domain construction. ARC is available on GitHub at github.com/tririver/arc.
ARC connects agent work to INSPIRE metadata/citations, arXiv identifiers, ar5iv/full-text sections, local paper caches, source downloads, equation/context search, LLM paper summaries, and report/export helpers.
ARC was used as both MCP tools and local workflow discipline: gather paper data first, cache it under the project dir, verify metadata, then synthesize reports from local artifacts instead of relying on memory.
How ARC Was Used Here
- Seed paper:
arXiv:0911.3380. - Initial ARC run fetched INSPIRE citer records and filtered
year >= 2024. - ARC metadata and summary artifacts separated arXiv records from non-arXiv records.
- For the focused full-shape report, ARC metadata verified nine paper titles/authors/arXiv IDs, while local PDFs/source archives supplied text, tables, and figures.
- ARC self-reflection files recorded outcomes and improvement notes.
Exact prompts
User Prompts Relevant to the Two HTMLs
use arc to download papers that cited 0911.3380 since 2024. And create a full summary of these papers. Use subagents and parallize if possible
Summarize the above results into research directions, for each direction provide the summary, and list related papers. Make sure the title, authors and arxiv numbers are accurate for each paper listed. Present the summary as a html report.
For the research direction "Build full-shape collider searches that combine CMB, LSS, and estimator systematics", read every paper carefully, and provide a html summary on observational constraints on various cosmological collider shapes.
For the research direction "Build full-shape collider searches that combine CMB, LSS, and estimator systematics", read every paper carefully, and provide a html summary on observational constraints on various cosmological collider shapes. In the summary, provide explicit numbers for the constraint on each shape, and use figures for observational constraints. Use original figures from these papers and cite them. Use consistent theme with the previous html summary.
Cost estimate
Estimated Time and Token Usage
Times are from user-prompt history, ARC context timestamps, and output file modification times. Token counts are from local Codex token_count events where available. Counts include cached input tokens; cached tokens are not the same as uncached billable tokens.
| Phase | Time window | Elapsed | Main-session token estimate | Known subagent token estimate | Notes |
|---|---|---|---|---|---|
| ARC citer summary evidence base | 00:52:08 to 01:48:06 UTC | ~56 min | ~8.01M total tokens; ~7.77M cached input; ~26.6k output. | Some early subagent/ARC calls exist, but ARC-LLM usage fields in summary JSON are null. | Produced report.md, summaries_all.jsonl, metadata files. Exact ARC-LLM independent-call cost unavailable. |
| Research-directions HTML synthesis | 02:03:11 to 02:05:56 UTC | ~2 min 45 sec | ~458k total tokens; ~422k cached input; ~9.8k output. | Earlier four synthesis/check subagent logs sum to ~1.79M total tokens, but they were part of upstream direction preparation, not solely final HTML emission. | Produced research_directions.html and manifest. |
| Full-shape HTML, strict final prompt | 02:23:08 to 02:41:47 UTC | ~18 min 39 sec | ~5.50M total tokens; ~4.66M cached input; ~39.4k output. | Three focused subagents sum to ~3.29M total tokens. | Produced PDFs/source/text cache, 39 rendered/copy figures, final HTML, self-reflection. |
| Whole two-HTML work interval | 00:52:08 to 02:41:47 UTC | ~1 hr 49 min 39 sec | ~18.58M total tokens; ~17.33M cached input; ~89.8k output. | Known report-related subagent logs sum to ~5.08M total tokens. | Does not include exact ARC-LLM independent Codex calls because local JSON usage fields are null. |