v1.4.0

What's new

• Our scripts now share two small helpers. One reads JSON that carries a BOM and writes it atomically (temp file, then rename over the target). The other re-runs a script with a proxy when your system has one. So all four network scripts respect a configured proxy, and a Node runtime that can't use one says so on stderr instead of looking like a network outage.
• Agent discovery is more complete. scripts/discover-agents.mjs now reports mode, hidden, traits (worked out from the agent's description, name and mode, with known roles taking priority) and provenance.fetchedAt. Built-in names it skips show up under skippedBuiltins instead of disappearing, and the workflow now asks you about any entry that might be one of your own agents.
• New flags: --include-builtin-names, --builtin and --extra-builtin for agent discovery, and --rebuild for the snapshot refresh.

Changed

• Refreshes now write the snapshot safely. If the existing snapshot can't be parsed, the script stops with a clear message and leaves your file alone — including your remembered mode and cached scores. It used to quietly start from an empty cache. Use refresh-snapshot.mjs --rebuild to discard it on purpose.
• The catalog fetcher and the snapshot summary use the shared helpers too, so what one script writes is what the next one expects.
• The README version badge no longer drifts from the skill. The badge is generated from metadata.version in SKILL.md, and the doc check fails if the badge or either changelog disagrees — a 1.2.0 badge on a 1.3.0 skill can't return unnoticed.
• Observer.md and observer.md behave the same now: known-role trait matching ignores case.
• Run the three refresh commands one after another, never in parallel — they write the same snapshot.
• The proxy warning now fires for every runtime below the version we verified (22.23.1), not only for major versions below 22.

Fixed

• Reading a JSON file now treats only a missing file as "no file". A permission or sharing error is an error, so a snapshot that exists but can't be read is never replaced by a fresh one. And when an atomic replace fails — a file held open on Windows — the message names the path and says the previous file is intact, instead of a bare system error.

Full diff: v1.3.0...v1.4.0

v1.3.0

Two versions of work are rolled up here (v1.2.0 and v1.3.0).

Better agent discovery and chain safety

• New scripts/discover-agents.mjs: a read-only, zero-dependency CLI that lists your
  custom agents (slim and native sources) and flags risky fallback chains — duplicate
  model ids, chains outside 2–4 entries, non-Go fallback ids not confirmed in the
  local registry, and vision lanes whose model is not marked vision-verified. It
  prints compact JSON instead of making you read config files by hand.
• New scripts/snapshot-summary.mjs: a compact digest of the cached model snapshot
  (~97% shorter than the raw file), with --ids, --trait, and --json.

Clearer recommendations

• The recommendation table now includes a role overview column: a one-line,
  plain-language description of what each agent does, so a suggestion can be judged
  without knowing the role by heart.
• A dominance check now closes model allocation: if another model is no worse on
  price, throughput and the trait-relevant capability (and strictly better on at
  least one), the dominated model is replaced — or must be justified in the "why"
  column (for example, a cross-provider fallback).

Correct, dated vision data

• Image-input support is recorded with sources and dates: MiMo-V2.6 Flash/Pro and
  DeepSeek V4.1 Flash accept images. DeepSeek's Go id, lab version name and API
  request name are now spelled out (the request model name is deepseek-flash).
• JS-rendered provider pages fall back to searching the official domain instead of
  being treated as "no evidence".

Smoother runs on Windows

• Documented the Windows refresh path: set $env:NODE_USE_ENV_PROXY=1 before running
  any scripts/refresh-*.mjs, and judge success from stdout JSON — a stderr warning
  and a PowerShell exit code of 1 are not failures.

Documentation

• SKILL.md is now a focused decision surface; the details live in references/,
  which gained section lists and clearer failure handling.
• If you install by copying files, re-copy SKILL.md, references/, and scripts/.

v1.1.0

opencode-go-model-picker v1.1.0

Reworks the skill's output into a scannable three-part body plus a data appendix. The read-only apply gate, the recommendation modes, and the fallback-chain policy are unchanged.

What's Changed

Changed

• Recommendation table — one row per custom agent with exactly: agent name, capability focus (trait), current model, recommended model, and the cost/throughput/capability reason. ⚠ limited-time / ⚠ geo-restricted / ⚠ trains on your data markers live in the recommended-model cell, with their full explanation in the highlights.
• Paste-ready config — a complete copy-pasteable JSONC block per detected source: a preset block for oh-my-opencode-slim.json, an agent block for opencode.jsonc (or the frontmatter model: line for Markdown agents), both when both are detected — each labeled with its paste target, with fallback chains preserved where the source supports them.
• Highlights — one paragraph covering the recommendation mode and its origin, whether the balanced price ceiling held, the full text of every ⚠ flag, plan changes vs the last snapshot, and any pick that needs a human decision.
• Data appendix — sources with fetch dates (seed values explicitly labeled with the seed's date), the snapshot diff, the relevant-models price table, the first-run decision-factors table, the manual-verification list, and the verify commands.
• The apply gate is unchanged: the skill still stops and asks before touching any config.

Synced across README.md / README-ZH.md, CONTRIBUTING (EN/ZH), docs/index.html, and test-prompts.json.

Full changelog: see CHANGELOG.md.

Full Changelog: v1.0.0...v1.1.0

v1.0.0

opencode-go-model-picker v1.0.0

First stable release of opencode-go-model-picker — an OpenCode Agent Skill that picks or rebalances the models used by your custom OpenCode agents (native agents, oh-my-opencode-slim presets, and agent-injecting plugins) against the latest OpenCode Go plan. Read-only by default: it fetches the plan, produces recommendations with fallback chains, and never edits your config without a preview and an explicit confirmation.

Highlights

• Plan-aware allocation — fetches the live Go plan every run (models, per-model monthly limits, limited-time promos, estimated request counts) and allocates a model per agent by role trait, under budget / balanced / quality recommendation modes with a two-question first-run diagnostic.
• Fallback chains — builds ordered 2–4 entry failover chains (best fit → cheaper Go model → different provider) for chain-capable sources such as oh-my-opencode-slim, and explicitly refuses to invent arrays for sources whose model is a single value.
• Never invents numbers — every price, limit, and capability is reported with its source and fetch date. Committed price and LMArena score seeds act as dated offline fallbacks; anything unverifiable is marked for manual verification.
• Read-only by default — output ends at an explicit 🛑 gate; nothing is written to your configuration without your confirmation.
• Multi-source discovery — native OpenCode agents (opencode.json/opencode.jsonc, Markdown agent files), oh-my-opencode-slim presets, and other plugins that inject agents, through uniform read-only source adapters. OpenCode's built-in agents are always skipped.

What's Changed

Added

• Explicit Failure Modes (if → then → still failing) table in the skill: serial re-run on snapshot write races, dated seed/cache fallbacks when endpoints or the network are unreachable, local-registry status verification for non-Go fallback ids, and a zero-custom-agent stop branch.
• Visible 🛑/🔴 checkpoint markers at the output gate, the first-run mode diagnostic, and the apply gate.
• Committed price seed (references/model-prices.json) with scripts/refresh-prices.mjs, and a committed LMArena score seed (references/model-scores.json) with scripts/refresh-scores.mjs (no API key, no browser).
• Three recommendation modes (budget, balanced, quality) plus an auditable decision-factors table in the report.
• Estimated request counts (per 5h / week / month) as a first-class throughput signal — equal dollar limits do not mean equal request counts.
• Mandatory caveat flags for limited-time multipliers, geo-restricted models, and privacy-for-discount "Contributor" tiers that train on your prompts (opt-in, never recommended silently).
• Generalized agent discovery beyond oh-my-opencode-slim via references/agent-sources.md (uniform inventory record, role-trait mapping, known-role overrides).

Changed

• Refresh strategy deduplicated to a single source of truth in the skill; Failure Modes row 5 disambiguated — an explicit deprecated/removed status blocks an id, while an absent status field is the registry's default-active convention.
• README (English and Chinese) now documents the explicit failure-path handling.

Fixed

• scripts/refresh-snapshot.mjs no longer drops top-level snapshot keys such as preferences on write, so the remembered recommendation mode actually persists across runs.
• Snapshot schemaVersion now writes 2, matching references/model-snapshot.md.
• scripts/refresh-prices.mjs output now separates seed coverage (seedUsed), priced count (priced 40/42), and null-upstream ids instead of a single misleading matched number.

Installation

Clone the repository into your OpenCode skills directory so the folder name matches the skill's name:

Linux / macOS

git clone https://github.com/sogeisetsu/opencode-go-model-picker.git \
  ~/.config/opencode/skills/opencode-go-model-picker

Windows (PowerShell)

git clone https://github.com/sogeisetsu/opencode-go-model-picker.git `
  "$env:USERPROFILE\.config\opencode\skills\opencode-go-model-picker"

Minimal install: the skill only uses SKILL.md, references/, and scripts/ at runtime — copying just those three is enough.

Verify the helper script runs:

node scripts/fetch-go-models.mjs   # prints { fetchedAt, source, count, ids }

Full Changelog

https://github.com/sogeisetsu/opencode-go-model-picker/blob/main/CHANGELOG.md

v0.1.0

OpenCode Go Model Picker v0.1.0

First public release: an OpenCode Agent Skill that picks the most cost-effective
OpenCode Go model for each oh-my-opencode-slim agent, based on the current Go
plan — read-only, source-cited, and fallback-ready.

Highlights

• Fresh on every run — fetches the current Go plan, so it never recommends a
  retired model or a stale price.
• Read-only by default — you see the exact change before anything is written.
• Per-agent recommendations with ordered fallback chains, so a capped model
  does not end the session.
• Source-cited — every number carries a source and a fetch date, and
  unverifiable values are flagged instead of guessed.
• Helper scripts — fetch-go-models.mjs (live catalog), check-docs.mjs
  (relative links and English/Chinese doc-pair validation), and
  generate-assets.mjs (SVG icon and banners).
• Bilingual documentation (English / 中文) and a new GitHub Pages landing
  page: https://sogeisetsu.github.io/opencode-go-model-picker/

Quick start

Install the skill, then invoke it in OpenCode with its slash command — no input
needed:

/opencode-go-model-picker

See the README
for installation.

Full Changelog: https://github.com/sogeisetsu/opencode-go-model-picker/commits/v0.1.0