Don't break the streak.

  THERMAL  — Don't break the streak.

  Token Warriors
   #    Tool           Strk    Best    Days     Tokens      Cost
   ─────────────────────────────────────────────────────────────────
   1. OpenCode          10d     10d     26d   8.2B tok    $50.54
   2. Codex              6d      7d     43d   938.7M tok  —
   3. ZCode              3d      3d      5d   676.2M tok  —
   4. Devin              2d     33d     45d   19.2B tok   —
   5. MiMoCode           1d      7d     16d   1.2B tok    $56.14
   6. DeepSeek (DSH)     1d      1d      5d   18.0M tok   ~$0.07
   7. Grok               1d      1d      1d   692.7K tok  $0.44
   8. codewhale          1d      1d      3d   110.1K tok  $0.02
   9. Claude             1d      1d      3d   0 tok       —

  Activity Hunters
   #    Tool           Strk    Best    Days     Activity
   ───────────────────────────────────────────────────────
   1. Agy               36d     36d     43d   60.1K step
   2. command-code       2d      9d     44d   10.4K msg
   3. Droid              1d      1d      2d   14 msg
   4. Muse               0d      0d      0d   0 prompt

  >> Agy is on fire with a 36-day streak!

  Keep the heat going. Don't break the streak.

thermal --sort tokens   # biggest token consumers first
thermal --sort cost     # biggest recorded spend first
thermal --weeks 26      # narrower heatmap, 4 to 104
thermal --no-color      # strip ANSI for logs and CI

thermal

A day where an agent ran but recorded no tokens still counts, so a push that produced no usage telemetry never breaks your streak.

thermal dashboard

  q        quit                          r    range: 30d, 90d, 1y, all
  1-5      jump to a view               t    metric: tokens or cost
  tab      next view                     s    sort: tokens, cost, name
  j / k    move down and up              ?    help
  f        tool filter (Projects)        enter  open a project detail
  l        log scale (Stats)             v      mix by tool or model

The Overview view opens with totals, then ranks every tool that recorded usage in the window, with a bar, a twelve-week sparkline, and the current streak. The Projects view ranks repositories and opens a detail on enter, with the tool split, the model split, a weekly trend, and a heatmap scoped to the selected range. Mix stacks the window over time by tool or by model and reports who led and how often that changed. Models ranks every model with an estimated cost. Stats shows the daily distribution with a log scale for long tails, the weekday profile, the top days, the outliers, and a month-end projection drawn as a range rather than a single number. Every total equals the matching report command for the same window, and a test asserts that rather than eyeballing it.

The dashboard needs a terminal. Piped output is told to use the static commands instead, and --json never opens a terminal UI.

thermal weekly                          # every tool, newest week first
thermal opencode daily --last 7         # one tool, last 7 days
thermal monthly --last 1 --breakdown    # this month, plus a row per model
thermal weekly --since 2026-08-01 --until 2026-08-31
thermal weekly --order asc --start-of-week monday
thermal weekly --json                   # unchanged shape, scripts first

Tables list token-bearing periods. A day with no classified tokens, no recorded cost, and no model attribution is skipped, so the Tokens column and the Total row stay in one unit. Daily, weekly, and monthly totals for the same window are identical. Activity-only periods still count toward streaks and the leaderboard.

thermal weekly

JSON output

Every report command takes --json and writes to stdout, so a warning can never corrupt the pipe. One period looks like this:

{
  "period": "2026-09-18",
  "models": {
    "deepseek-flash": {
      "input": 101725,
      "output": 6712,
      "reasoning": 20800,
      "cacheRead": 2064256
    }
  },
  "inputTokens": 101725,
  "outputTokens": 6712,
  "reasoningTokens": 20800,
  "cacheTokens": 2064256,
  "totalTokens": 2193493,
  "turns": 1,
  "activeDays": 1,
  "storedCost": 0.037958718,
  "cost": 0.037958718
}

thermal weekly --json | jq '.totals.cost'
thermal projects --json | jq -r '.data[:5][] | "\(.project)\t\(.cost)"'

thermal projects                     # rank every project by tokens
thermal projects --sort cost         # or cost, days, recent
thermal projects --last 30 --top 10  # window and row limit
thermal projects --breakdown         # tool split and top models per project
thermal projects --json              # full paths, every row

thermal projects

Merge happens by git root, not by path string. Symlinked homes and duplicate checkouts collapse to a single project. A directory with no repository marker falls back to its recorded path rather than an invented name.

thermal models                # global ranking
thermal models --sort cost    # rank by estimated cost
thermal opencode models       # one tool only
thermal models --top 10

  Thermal · models

  #    Model                           Tools                    Tokens       Cost  Days  Last
  ─────────────────────────────────────────────────────────────────────────────────────────────────
   1.  deepseek-v4.1-flash             OpenCode,ZCode             3.9B     $23.29     8  2026-09-17
   2.  deepseek-v4-flash-0731          MiMoCode,OpenCode +1       1.1B     $26.62     9  2026-09-12
   3.  gpt-5.6-sol                     Codex                    386.8M    $256.25    10  2026-09-11

Cost here is always an estimate. A recorded cost belongs to a session or a day, never to one model, so thermal prices models from list prices instead of pretending the split is known.

Two totals, two questions

The leaderboard and the reports answer different questions, and their totals do not match. That is by design.

The estimate covers only what the sources name. A day is priced when its source records a model and that model has a list price. A day whose source records neither cost nor a model is counted, not priced, and the footer says so: 19.2B tokens have no model recorded, so no price is applied to them. Treat the recorded part as fact, the estimated part as inference from list prices, and read that line to see what is missing rather than assuming the total is complete.

A day is priced when its source names a model and that model has a price. Tools often record the variant they asked for, like claude-sonnet-5-high, while the catalog prices claude-sonnet-5, so a short list of tier words is stripped before the lookup. Only one word is stripped, never a fuzzy match, because folding two models onto one price would be worse than a blank.

A model the catalog genuinely lacks stays unpriced and is named in the No pricing for line, such as Devin's swe-1-7, which is proprietary and will never appear in a public catalog. Give it a price in ~/.config/thermal/pricing.json and the total picks it up on the next run. A model whose tokens arrive with no type breakdown stays unpriced too, rather than being half priced. Nothing is guessed and nothing is treated as free.

Tools that record no tokens at all, such as Agy, Droid and command-code, report steps or messages. Those counts stay out of token totals, because a step is not a token, and they appear in the Activity Hunters leaderboard and in streaks.

thermal weekly --offline      # cached prices only, never fetch
thermal weekly --no-estimate  # recorded cost only, works on the leaderboard too

{
  "codex-auto-review": { "input": 1.25, "output": 10 }
}

thermal stats                               # percentiles, cache efficiency, weekday profile
thermal mix --by tool --grain week          # tool share over time and switching velocity
thermal trend --last 30                     # daily volume fit and 14-day month-end projection
thermal weekly --chart                      # append clean text bars to reports, projects & models

  Thermal · stats (all tools, 2026-06-25 to 2026-09-22)
  Window: 90 days · 67 active (74.4%) · 5.82B total tokens · ~$214.30

  Distribution (tokens / active day)
    Median:   64.2M tok       Outliers: 2 days > 180.4M tok
    p90:     122.8M tok       Max:      248.1M tok (2026-08-14)

  Token composition & cache efficiency
    Cache read:       5.59B tok  ████████████████████████████████████████  (96.1%)
    Cache write:     112.4M tok  █                                         ( 1.9%)
    Uncached input:   64.1M tok  ░                                         ( 1.1%)
    Output:           42.0M tok  ░                                         ( 0.7%)
    Reasoning:        11.5M tok  ░                                         ( 0.2%)
    Cache hit rate: 97.0% of prompt tokens read from cache

  Weekday profile
    Mon:   84.2M tok/day  ████████████
    Tue:   92.5M tok/day  █████████████
    Wed:  104.1M tok/day  ███████████████
    Thu:   88.3M tok/day  ████████████
    Fri:   76.0M tok/day  ███████████
    Sat:   31.4M tok/day  ████
    Sun:   28.1M tok/day  ████

Copy-pasteable Unicode bar charts with --chart. Append --chart to daily, weekly, monthly, projects, or models to render clean proportional bars right in stdout. Bar rows are width-aware, capped at twelve rows with a stated remainder, and never touch --json.

thermal replay                              # default comparison vs standard plans & APIs
thermal replay --against claude-pro         # test a specific subscription
thermal replay --against deepseek-v3        # simulate API pay-as-you-go
thermal replay --compare "claude-pro,chatgpt-plus,deepseek-v3"
thermal replay --last 30 --json             # machine-readable simulation report

  Thermal · replay (2026-08-23 to 2026-09-22, 30 days)
  Workload: 21 active days, 3.84B total tokens (96.2% cached)

  Tier / Plan             Type     Monthly Cost   Est. Spend  Capacity  Throttle%  Verdict
  ──────────────────────────────────────────────────────────────────────────────────────────
  Claude Pro              Sub      $20.00 / mo       $20.00   FAIL          85.7%  Exceeds burst caps
  Claude Max (5x)         Sub      $100.00 / mo     $100.00   DEGRADED      14.3%  Peak days throttled
  Claude Max (20x)        Sub      $200.00 / mo     $200.00   PASS           0.0%  Handles daily peaks
  ChatGPT Plus            Sub      $20.00 / mo       $20.00   FAIL          81.0%  Exceeds burst caps
  Cursor Pro              Sub      $20.00 / mo       $20.00   FAIL          66.7%  Exceeds fast quota
  DeepSeek V3             API      Pay-as-you-go     $18.42   PASS           0.0%  Unmetered burst
  Claude 3.5 Sonnet       API      Pay-as-you-go     $94.30   PASS           0.0%  Unmetered burst

  Recommendation:
  Subscription tier 'Claude Pro' would throttle on 85.7% of active days.
  DeepSeek V3 API would handle 100% of workload for ~$18.42/mo (saving $1.58/mo vs $20 flat sub).

Disjoint token simulation with cache read discounts. Thermal separates uncached prompt tokens from cached reads and reasoning. Because 95%+ of agent tokens are prompt cache reads, pay-as-you-go API simulation accurately applies tiered caching discounts (e.g. $0.14/M vs $0.014/M on DeepSeek) rather than overcharging by 10x–20x.

Pick thermal when the question is about habit. Streaks, active days, and which tools you actually open are its subject. Pick ccusage or tokscale when the question is about accounting down to the session, and pick tku when you want to watch spend while it happens. Nothing stops you running two of them against the same files.

1. Read only, always

   Open SQLite with mode=ro and set PRAGMA mmap_size right after. Never write to a source file. Cache into ~/.cache/thermal/ instead.

2. Keep token types disjoint

   Input, output, reasoning, cache read and cache write must add up to the recorded total. Several tools nest reasoning inside output or cache inside input; subtract the nested part in the loader, never in the renderer or the pricer.

3. Normalize the model id

   Run every model name through the canonical mapper before it becomes a map key, because tools disagree on case and one model must be one row everywhere.

4. Attribute the project

   When the source records where a session ran, emit project rows too, so subdirectories fold into their git root. A source with no project data returns an empty slice rather than an invented name.

5. Bound the work

   Multi-file scanners use a worker pool, capped around sixteen, with thread-safe aggregation. Large databases get a delta cache keyed on file modification time or the highest row id.

6. Safe JSONL ceilings and diagnostics

   JSONL scanners must use newJSONLScanner() to set a 32 MiB line ceiling so multi-megabyte tool results and rollout chunks are not truncated. Check scanner.Err() and record non-fatal errors in Summary.Warnings with formatScanWarning() so diagnostics ride to --verbose stderr without polluting stdout or JSON output.

7. Test the edges

   Missing databases, corrupt files, zero-token sessions, and unusual timestamp formats all need a case in internal/loaders/*_test.go. Then register the tool in internal/loaders/registry.go and add its row to the table above.

The full standard lives in AGENTS.md. It covers the loader rules, the build and release flow, and the terminal rendering conventions. Read it before your first pull request.