dsh-opencode-go-eta

A quota forecast widget for OpenCode Go inside DeepSeek Harness — it tells you when the monthly allowance runs out, and whether that happens before the renewal date.

English · 简体中文

Dark theme

Composer dock

_{The shipped context ring (54%) and the Go capsule (18%) share the same 14 px / 2 px specification; the capsule stays a text-only trigger, expanding nothing until clicked.}

Why

OpenCode Go is a flat monthly subscription, but its console only shows used percentages and a reset time. It does not tell you whether your current pace will exhaust the allowance before the renewal date — the one question that decides whether you need to top up, renew early, or switch to the pay-as-you-go API for the rest of the period.

This plugin keeps a small local history of the official quota readings and answers that question in the composer dock. It never reads your session logs: the Go allowance is account-wide, so only the official percentage sees usage from every client.

Features

• Projected exhaustion time — a reset-aware Theil–Sen fit over the recorded quota samples gives an ETA per window, with a confidence tier. No guessing from a single reading.
• Renewal-date comparison — says plainly whether the monthly allowance will run out before the renewal date, or leave headroom until then.
• Long-run cost advice — when the quota will run out early, it compares two steady-state policies: renewing OpenCode Go (a prepayment of the next period, unit cost planPrice ÷ allowance per $1 of usage, about $0.125) versus paying the overflow on the official API ($1 per $1). Renewal is recommended unless the new allowance would largely expire unused.
• Merged usage & forecast chart — history, projection, “now”, ETA and renewal markers on one time axis, with a 0–100 % axis.
• Native composer dock capsule — the 14 px / 2 px ring matches the shipped context meter, so both rings read as peers; the card itself is a liquid-glass panel that dims its highlight layers in dark mode.
• Per-model allowance — Go's monthly limit is per model (DeepSeek V4.1 Flash: $80 on Go, $160 on Go Plus; Kimi K3: $15). The card shows the model and its allowance, while percentages stay authoritative from the console.
• No polling storms — samples at four local anchors a day, on task ends, on start-up and on demand.

Install

Requires DeepSeek Harness 0.2.x (Desktop or Web profile) and Node.js ≥ 22.

# from GitHub
dsh plugin --profile desktop add github:HugoOu/dsh-opencode-go-eta

# or from a local checkout
dsh plugin --profile desktop add link:C:\path\to\dsh-opencode-go-eta

Restart DSH completely after installing. The capsule appears below the composer; click it for the card.

The plugin row is a bundle patch: dsh plugin add appends dsh-opencode-go-eta to dsh.profile.bundles automatically, and the bundled cordis.patch.yml inserts the host row.

Configuration

Add to your profile's cordis.patch.yml to override the defaults:

- id: opencode-go-eta
  name: dsh-opencode-go-eta
  config:
    apiKeyRef: OPENCODE_GO_API_KEY   # credential reference, not the key itself
    anchors: [3, 9, 15, 21]          # local hours; [] disables timed sampling
    sampleIntervalMs: 1200000        # minimum gap between implicit samples
    staleAfterMs: 2700000            # a read this old triggers a fresh sample
    retentionDays: 45
    planName: Go                     # shown in the card
    planPriceUsd: 10                 # price of one period, used by the advice
    modelLabel: DeepSeek V4.1 Flash
    modelAllowancesUsd:
      deepseek-v4.1-flash: 80        # Go table; use 160 for Go Plus
    pricingModel: deepseek-v4.1-flash
    pricingUrl: https://models.dev/api.json
    pricingTtlMs: 43200000

The API key is resolved through the Harness credentials service (OPENCODE_GO_API_KEY); it never reaches the browser.

How it works

Stage	Behaviour
Sampling	GET /zen/go/v1/usage (percentages + resets) and GET /console/api/go/status (meters, period start/end) — at 03/09/15/21 local, on turn/end, at start-up, and on demand.
History	Samples are kept for 45 days in $DSH_HOME/opencode-go-eta/samples.json (atomic writes).
Projection	Samples from before the last reset are cut; the Theil–Sen slope of the current cycle gives the ETA. Fewer than 3 points or a span under 6 h is reported as insufficient samples.
Verdict	ETA is compared with the window reset and the subscription's endsAt: before expiry, after expiry, safe within the cycle.
Advice	Two steady-state policies: renew prices every $1 of usage at planPriceUsd ÷ allowance (≈ $0.125), API overflow costs $1 per $1 on top of the plan. Costs are annualised: at $96/month of usage the allowance lasts about 25 days, so renewal means ~14 periods/year (≈$144) while the API path costs ≈$312/year. Renewal wins whenever usage exceeds one allowance; the API only wins when the new allowance would expire unused. Official model prices come from models.dev (12 h cache) for display only. At $96/month of usage the allowance lasts about 25 days, so renewal means ~14 periods/year (≈$144) while the API path costs ≈$312/year.

The console meter is a plan-credit cap (observed $60), while the official Go table lists a per-model monthly allowance (DeepSeek V4.1 Flash: $80). Percentages always follow the console; money follows the model allowance so the two never contradict each other silently.

Development

node test/projection.test.mjs   # 10 pure-math tests: slope, reset cutting, verdicts
node test/host.test.mjs         # 2 host tests: stubbed fetch, route payload
node test/client.test.mjs       # 4 client tests: slot registration, render, fetch

No network or real key is needed. preview.html reproduces the shipped styles with mock data (#shot hides the page chrome; #shot-dark renders the dark theme).

Limitations

• /console/api/go/status is an internal console endpoint; if it changes, the card falls back to percentages only.
• Projections are linear extrapolations from your own sampling history, not vendor forecasts.
• Cost advice assumes the quota-meter value and the official list price share the same scale; the token estimate is an approximation.
• The plugin stores no prompts or session data — only quota readings and the pricing cache.

License

MIT © 2026 HugoOu