reqshape

What it is#

A req is one model call — the same unit the opencode sidebar counts: a single assistant row in opencode's store. An agentic ask is many reqs, because the model is called again after every tool result.

reqshape reads that history, drops the requests that ask for nothing, averages the rest, and hands you the answer as a token vector: one of my requests is 8.2K input, 309 output, 139 reasoning, 248K cache read. Then it prices that vector on every model OpenCode Go and CommandCode sell, and divides each plan's allowance and window caps by the result.

mpc answers the same question with a fixed assumption (800 in / 50K cache / 200 out). This answers it with your actual traffic — and usually disagrees, because context is re-read on every call and yours is deep.

Install#

Install the published CLI globally:

bun add -g @jeffreyjyz/reqshape

Or run it from a checkout — Bun runs the TypeScript entry directly, or link the binary:

bun install
bun run src/index.ts --help
bun link

Usage#

reqshape

reqshape  7.3K reqs · every req weighted equally
measured from the opencode v2 store · priced by mpc
 
MEASURED  1,084 asks · 38 sessions · 12 projects · 2026-07-10 → 2026-09-29
DROPPED   synthetic prompt 260 · trivial prompt 138 · system prompt 130 · empty prompt 14 · compaction prompt 9 · shell prompt 7
 
PER REQ   input 8.2K · output 309 · reasoning 139 · cache read 248.0K · cache write 8.1
          p10/p90  input 67/9.2K · output 46/701 · cache read 1.9K/648.7K
          one vote per conversation instead: cache read 64.0K · input 10.4K · output 216
 
PER SIDE  each side priced on its own traffic
          OpenCode Go   input     7.2K  output     296  cache read    134.8K  3,684 reqs
          CommandCode   input     3.7K  output     337  cache read    397.0K  3,360 reqs
 
POSITION  how far into its session the req sat — every call re-reads the context
          pos 1        cache read      579  input    8.5K  output     75  28 reqs
          pos 2-5      cache read     9.9K  input    4.9K  output    180  107 reqs
          pos 6-20     cache read    21.5K  input    8.2K  output    279  356 reqs
          pos 21-100   cache read    62.2K  input   10.9K  output    388  1,043 reqs
          pos 101+     cache read   301.5K  input    7.8K  output    301  5,737 reqs
 
MODELS    deepseek-v4.1-flash 2.8K · minimax-m3 1.7K · glm-5.2 1.4K · best-coding 447 · hy3-free 391 · big-pickle 140 …2 more
 
PROJECTED what the plan's allowance buys at the measured shape
 
MODEL                       PLAN                $/req      req/mo      req/5h      req/wk
──────────────────────────  ───────────  ────────────  ──────────  ──────────  ──────────
Jev                         CC GOAT       $0.00034461       58.0K       11.6K       29.0K
Muse Spark 1.3 Contributor  OC Go         $0.00140623       42.7K        8.5K       21.3K
MiMo V2.6 Flash             OC Go         $0.00197864       30.3K        6.1K       15.2K
DeepSeek V4.1 Flash         OC Go          $0.0022531       26.6K        5.3K       13.3K
...
Grok 4.6                    OC Go             $0.1431         105          21          52
 
CC account  1,898 reqs this period · GOAT · ends 2026-10-27

Why "cache read" is the whole story#

Every request re-sends the conversation, and providers bill the re-sent part as cached tokens. So cache read is not a property of how you prompt; it is a function of how far into the conversation you are:

The first request of a conversation is essentially free to re-send; the hundredth re-sends 300K tokens. On this history 78% of all requests sit at 101+, which is why the per-request average (248K) is four times the per-conversation average (64K) — one enormous session otherwise sets the profile.

--weight turn (the default) treats every req as one data point, because that is the unit you are billed in. --weight session gives each conversation one vote and prices that instead. Both are printed either way, so you can see the gap.

A model with no published cache-read rate is billed for that context at its input rate, and flagged with *. This is the single biggest swing in the table: a model with no caching support is not slightly more expensive for this traffic, it is unusable.

Noise filtering#

"hi", "thanks", "ok", "continue" are real requests but say nothing about the work you do. reqshape drops an ask — the prompt and every req it drove — when:

• its prompt normalises to a keyword (hi, hey there, thanks, great, lgtm, continue please, …) — edit the list with --keywords
• its prompt is nothing but emoji or punctuation (--min-chars adds a length floor)
• it produced almost no output (--min-output N)

Structural noise is always dropped: auto-generated synthetic prompts, the compaction summary turn, shell commands, and system rows are not requests you made. --keep-trivial turns off the prompt filters and keeps everything else the same, and --explain prints every reason with its count — nothing is silently discarded.

Options#

A model nobody sells#

reqshape --in 0.15 --out 0.6 --cache-read 0.003 --budget 60 --five-hour 12 --weekly 30

--budget is the monthly allowance; --five-hour and --weekly are the plan's caps, and without them those two columns show — rather than inventing a window. Omit --cache-read and the context is billed at the input rate, flagged *.

Where the numbers come from#

• Your traffic — opencode's own store, read-only (~/.local/share/opencode/opencode.db). v2's session_message is read when present, the pre-v2 message table otherwise, and session metadata is joined across session and session_v2.
• Model rates, allowances and window caps — mpc --json. The 5-hour and weekly caps are read back from mpc's own figures, so the provider's window rule is not restated here where it could drift.
• The account line — cmduse -1 --json, purely as context for how much of your usage this profile covers. If it cannot answer, the line is simply absent.

Per side, and back into mpc#

--format json carries sides.oc and sides.cc: the same measured profile split by whose traffic it is (requests to opencode* vs the CC_PREFIXES ids), with each side's own request count. mpc --shape measured reads exactly that payload and prices each plan on its side's shape, so its estimated req/mo is "how many of my requests fit" instead of a fixed 800-in / 50K-cache / 200-out assumption.

Runs take about 20 seconds: both mpc and cmduse go to the network.

Links#

• Source (GitHub)
• npm: @jeffreyjyz/reqshape
• MIT license