skills-hub
Tool · Claude Code · runs on your machine

Runway

Can I afford this prompt? When you press Enter, Runway tells you, before Claude sees the prompt, how long you can keep working once it has run and whether it will finish before your 5-hour limit. Then you decide whether to send it. It also shows how much context and quota is left, and how long that lasts at your current pace.

npx -y github:soumyaRauth/skills-hub runway install-hooks --statusline

The problem

a big refactor, started at 90% of the window  →  cut off halfway, with nothing committed

Claude Code tells you when you have hit a limit, not how close you are to one or what the next request will spend. A one-line question and a multi-file refactor look the same in the prompt box. In the history Runway was first tested on, one turn cost 95k tokens and another 1.8M.

Decide before it runs

you press Enter  →  Runway estimates the prompt and holds it back
   →  you see how long you can keep working once it has run: if it turns out quick, typical or heavy
   →  and the verdict: fits, may not fit or won’t fit
   →  nothing has been sent, so nothing is spent
   →  you decide: send it again (↑, Enter) to run it, or drop it

⛽ Runway · waiting for you to confirm this prompt

⏳ TIME LEFT after this prompt: ~1h05m
   if it turns out typical, at your pace over the last hour · ~1h07m left right now

✅ Should fit, even if it turns out heavy.

   If it turns out  Steps   Tokens   Time left after
   quick                1   360.7k   ~1h06m
 ▸ typical              3     1.1M   ~1h05m
   heavy               19     6.9M   ~55m

   5h window   46% used · resets in 3h53m
   Context     36% used
   Based on    your last 20 prompts · 13 of 15 landed in this range

👉 To run it: press ↑ then Enter within 2 min. To drop it, do nothing; it was not sent.

Why three numbers, not one. What a prompt costs depends almost entirely on how many steps Claude takes, and that is decided while it works. On real history, neither the prompt’s wording nor the prompts before it predicted the number of steps; once the steps are known, the cost is close to exact. So Runway shows the time left after a quick, a typical and a heavy outcome, with step counts from your own prompts, and how often past prompts landed inside that range. Time left is the room left after the prompt, divided by your pace over the last hour (or your average this window, after a break), and it stops at the reset. When a prompt could run out partway, the message says so (here with the quota set to 92% used, to show it):

⛽ Runway · waiting for you to confirm this prompt

⏳ TIME LEFT after this prompt: ~3m
   if it turns out typical, at your pace over the last hour · ~5m left right now

⚠️  May not finish before your 5-hour limit if it turns out heavy.

   If it turns out  Steps   Tokens   Time left after
   quick                1   360.7k   ~4m
 ▸ typical              3     1.1M   ~3m
   heavy               19     6.9M   limit hit partway

   5h window   92% used · resets in 2h00m
   Context     36% used
   Based on    your last 20 prompts · 13 of 15 landed in this range

👉 Run it anyway: press ↑ then Enter within 2 min. To drop it, do nothing; it was not sent.

That is the default, "gate": "always". In ~/.claude/runway/config.json, set "risky" to be asked only about prompts that may not fit, with the rest running at once, or "off" to never be asked. Only the quota can stop a turn partway through. A full context does not, because Claude Code compacts it and carries on, so context never pauses a prompt.

To estimate a prompt without sending it, from any terminal:

npx -y github:soumyaRauth/skills-hub runway check "refactor the auth module"

Install

npx -y github:soumyaRauth/skills-hub runway install-hooks --statusline
npx -y github:soumyaRauth/skills-hub runway calibrate

It needs Python 3 and nothing else: no packages, no API key, no network. The first command copies Runway to ~/.claude/runway/ and registers it in ~/.claude/settings.json, keeping everything already there and backing the file up first. --statusline sets the status line only if you have none. The second measures your own history, so predictions fit how you work.

To remove it: npx -y github:soumyaRauth/skills-hub runway uninstall-hooks, then delete ~/.claude/runway.

What you see

Once a prompt runs, one line, shown to you and not to the model (all the numbers in this section are illustrative):

⛽ Runway · running it, as you confirmed · ⏳ TIME LEFT after this prompt: ~1h05m (quick: ~1h06m · heavy: ~55m) · 5h used 46%, resets in 3h53m · context 36%

In the status line, all the time:

⛽ context 10% · 5h used 33% · resets in 4h00m · time left ~1h05m at this pace

And in the terminal, whenever you ask:

$ npx -y github:soumyaRauth/skills-hub runway status
Context    151,777 / 1,000,000 tokens (15.2%), ~42 turns to full at ~20.1k/turn
Burn       266.2k tokens/min over the last 60 min
5h window  33.2% used (Claude Code's rate limits), resets in 4h00m
Time left  ~21m (at the current burn rate); ~2 typical turns

How it counts

one prompt  →  several API calls (the tool loop)
   →  each call re-reads the whole context
   →  cost ≈ steps × (context + prompt) + extra per step

How many steps a prompt takes matters far more than how long it is, and nothing in the prompt reliably says how many it will take. So Runway measures your own history: the step count at the 10th percentile, the median and the 90th percentile become the quick, typical and heavy cases. N steps re-read the context N times, so the quick case is a floor. The measurements refresh in the background every six hours.

It checks itself

$ npx -y github:soumyaRauth/skills-hub runway backtest
  ...
  quick-to-heavy range   13 of 15 turns landed inside it (87%)

backtest replays every transcript you have, turn by turn. Each prompt is predicted using only the prompts that came before it, and the report says how many landed inside the quick-to-heavy range: on the history Runway was built on, 13 of 15. accuracy does the same for the live prompts the hook saw.

What it does not know

Plan weightingPlans weigh models differently, and Runway counts raw tokens. It converts with the rate your last quota reading implies. When another device also used the window, that rate runs high, so the estimate errs toward less time left, not more.
Auto-compactClaude Code compacts the context before it is full, at a point Runway cannot see. So turns to full is an upper bound.
Unseen callsBackground calls that never reach the transcript, such as generating a session title, are not counted.
SurprisesA prompt heavier than nine in ten of your past prompts lands above the range. That is the case no estimate can foresee. accuracy shows you how often it happens.
Early daysUntil you have five prompts of history, the range uses starting defaults (1, 4 and 8 steps), and the message says so.

Runway is a tool, not a skill. It does arithmetic on your logs, which code does exactly and for free, so the model is never asked to estimate it.