How many tokens are required to complete a reverse project with this REA?

[klamiefileous]

The invocation of tools and the repetitive output will waste many unnecessary tokens. How can we achieve the most efficient output?

[jayhemnani9910]

There's no fixed number. It depends on the target size, how broad the question is, and the model. What you can control:

Tool definitions. In #2019 the default tool catalog measured about 520 KB. Either use a client that loads MCP tools lazily, or set REA_MCP_INPUT_SCHEMA_PROFILE=compact in the server's env (about 40% smaller in that measurement). See mcp-contracts.md.

Large results. analyze_javascript_application and inspect_binary_layout take detail: "summary". The full Evidence stays in the session, and inspect_analysis_view pages into it (offset / limit) without rerunning the analysis. On the CLI, --filter-output keeps only the fields you need.

Repeated calls. Install the bundled skill (npx rea-agents@latest setup). It tells the agent to work summary-first and reuse existing Evidence.

The biggest saving is a narrow question ("how does X compute Y") instead of "reverse this whole app".

[pablodiazjorge]

Adding some measured numbers, since "it depends" is true but hard to plan with. I've been running preregistered benchmarks of rea in Claude Code (rea-lab): rea 6.4.0, Claude Code 2.1.296, Sonnet 5.5 at medium effort, stripped x64 builds of small open-source C programs (48 to 329 KB), and one question per run ("where is the function that does X?"). A whole reverse project is many such questions, so read these as per-question costs.

How many tokens

• Study 3: a run with rea took a median of about 300,000 input tokens (cache reads and writes included) and 0.19 USD, against about 113,000 tokens and 0.08 USD with only the LLVM command-line tools. Accuracy was similar: 68 of 72 runs right with rea, 71 of 72 without.
• Study 5, 30 new tasks of the same kind: a median of 0.21 to 0.24 USD per run with rea, 15 to 16 turns and 8 to 10 rea calls.

Where they go

• About 70 % of the cost was cache writes: each turn writes the newest messages, tool results included, into the prompt cache, and the context keeps growing. The number of turns matters more than any single large result.
• The tool catalog: even with Claude Code's tool search loading schemas on demand, rea's 153 tools added about 3,200 input tokens to every request, a median of 12 % of a run's input. I didn't measure the compact schema profile.
• Large results: every list_procedures call on these binaries returned 51 to 224 KB and the client saved it to a file; a few procedure_pseudo_code calls returned 58 to 471 KB.

What did and didn't help

• With the bundled skill's text in the system prompt, cost didn't change measurably (ratio 0.98 against no skill, p = 0.84), although list_procedures calls halved. The skill was injected because, installed normally, the agent never loaded it on its own in study 1 (0 of 10 runs).
• A revised skill that pushed a targeted route (search strings and procedures, decompile only candidates, don't list everything) changed the process as written but made runs 14 % more expensive (p = 0.012): fewer large results, more steps, each one carrying a bigger context.

So in this data the lever is fewer turns, which mostly comes from a narrow question, as suggested above.

Caveats: costs are Claude Code's estimates at list prices, not money spent. Study 3 very likely ran with Claude Code's microcompaction off, so its absolute counts may be high; study 5 used the defaults. It's one model, one client and small binaries; other clients cache and load tools differently.