November 1, 20262026年11月1日

Install

Minimal package: opencode-go-model-picker-3.0.0-min.zip contains only what the skill runs — SKILL.md, references/ (one doc plus two dated seeds), scripts/ (one entry point + modules) and LICENSE. Unzip it into ~/.config/opencode/skills/ so the folder is named opencode-go-model-picker. Verify with node scripts/go.mjs --help (four verbs: run / report / check / apply).

Full notes: CHANGELOG.md · 中文：CHANGELOG-ZH.md

[3.0.0] - 2026-11-01

The size of the simplification

3.0.0 is not a feature release: it is the same job described with **an order of
magnitude less text**. Measured against the 2.0.0 tag:

A typical 2.0.0 run loaded SKILL.md plus two or three references (≈11–15k
tokens); the same run now loads 8.3k characters, about 2.2k tokens — roughly
80–85% less context before any work starts. Token figures are estimates at
3.8 characters/token (no tokenizer ships with the repo); character and line counts
are exact: git show v2.0.0:SKILL.md | wc -c and the same for each reference.

Code did not shrink — it moved: ~4,700 lines of scripts became ~5,300 lines of
modules plus tests, because the rules that used to be prose (refresh order, basis
selection, priced columns, chain shape, apply diff) are now enforced by code.

Breaking

• One entry point. scripts/go.mjs replaces nine scripts — discover-agents,

probe-console, calibrate-usage, quota-math, refresh-snapshot,
refresh-prices, refresh-scores, snapshot-summary and fetch-go-models are
gone. Four verbs: run (agents → Console → usage → refresh → plan diff → basis,
one serial pass, one run.json), report (table, money columns, config blocks),
check (the gate before delivery), apply (exact diff; writes only with
--write, only the named file, only that agent's model value).

• One reference. references/plan-facts.md replaces quota-math.md,

model-snapshot.md, data-sources.md, agent-sources.md, console-mcp.md and
output-format.md. SKILL.md drops from ~19k to ~4k characters and is
model-facing only; teaching moved to the READMEs.

• Code over prose. The refresh order (three writers, one snapshot file), the

Basis: line and its denominator, the priced $/session column, chain-shape
rules, seed staleness, mode resolution and the apply diff are computed, not
described. The nine self-checks became go.mjs check.

• apply refuses a chain containing a dead id and refuses to write a result that

does not re-parse.

Added

• node --test tests/*.test.mjs (built-in node:test, no dependencies):

tests/golden/ holds outputs captured from the 2.0.0 scripts and covers pricing,
discovery and the five Console states; CLI tests cover offline degradation, the
report gate, and dry-run versus --write.

• run --offline always produces a dated, seed-labelled run.json; data.quality

says live / seed / none instead of leaving the reader to infer it.

Changed

• scripts/check-docs.mjs and scripts/generate-assets.mjs moved to tools/:

they maintain the repository, not the skill, and are not part of the minimal
install.

• Moved test-prompts.json to tests/prompts.json and documented it in the

repository layout: it is an evaluation fixture (a request paired with the
behaviour the skill must exhibit), not a runtime asset — nothing loads it.

v1.2.0

skill_management gives every installed OpenCode skill a switch: list what is installed, find out
why one will not load, hide it from the model, and put it back. Disabling writes one global
skill permission rule — no skill file is ever moved, edited or deleted.

No breaking changes in 1.2.0. The four actions, their inputs, the JSON envelope, the written
rule shape, the same-line marker and the single-file layout are unchanged since 1.1.0.

Fixed since v1.1.0

• list no longer reports "N skills differ across other agents" on every call.

summary.diverged counted hidden internal agents (title, summary, …), which carry
{*,*,deny}; on a real machine that made every skill look divergent (33/33) and added a noise
warning to each list. Only visible agents are compared now; inspect.agents[] still lists
every agent, so the detail is not lost.

• Frontmatter booleans no longer contradict each other. disable-model-invocation: "false" is a

string, and Boolean("false") === true, so inspect reported it as true while list reported
autoinvoke: false for the same file. true/yes/on/1 and false/no/off/0 (quoted or not) are
normalized now, and both fields go through the same rule.

• A failed sidecar write no longer turns a successful config change into an error. Once the

config was written and verified, a sidecar write failure used to surface as internal_error: the
file really changed, but the result had no changed / diff. It now returns ok: true,
changed: true, keeps diff, sets sidecar: null, and explains itself in warnings. The config
change is not rolled back.

• enable is no longer blocked by an unrelated file's syntax error. It reads sidecar-recorded

config files alongside the target; before, any of them failing to parse plus "no candidate rule"
produced config_unparsable. The hard refusal now applies only to the target config file;
other files are reported in warnings by name.

Also since v1.1.0

• The installer checks before writing. If opencode.json(c) already lists this package in

plugins, or a second copy sits in plugins/, it refuses and exits with code 2, leaving every
file untouched — two sources make the host report Duplicate plugin ID and fail one of them.

• New installer flags: --check / --dry-run, --uninstall (deletes only this plugin's file),

--npm (asserts the npm channel, never writes), --force, --json. <config> resolves the way
the host resolves it ($OPENCODE_CONFIG_DIR first).

• The plugin exports PLUGIN_VERSION, asserted against package.json by the harness.
• The acceptance harness grew from 260 to 326 assertions, and now validates every action's

output against the declared JSON Schema (required fields, property types, error.code /
error.message) and asserts that content is the same data as output.

• README is down to 51 lines; field tables, installer flags, package contents, usage and the test

workflow moved to the documentation site.

• The documentation site gained Design decisions (why permission rules instead of policies, why

there is no allow, why global-only, why text edits plus a same-line marker, why the sidecar is
only an audit trail — each with its cost) and Development & verification (harness sections,
CI, the performance and release gates, npm package contents, why two copies cannot coexist).
The Permissions page now expands Effects into a deny vs ask comparison with three real
skill examples. English and Chinese stay mirrored.

What it does

Permissions are evaluated for the calling agent only: last match wins (findLast, with *
and ? globs), and no match falls back to ask, not allow.

The rule disable writes is exactly this line (the trailing comment is the provenance marker
enable uses):

{ "action": "skill", "resource": "windows-shell", "effect": "deny" } // skill_management

Writes are atomic, preserve indentation, key order, line endings (CRLF/LF) and BOM, keep one
backup, and are verified by reloading the host and polling the caller's rule set; a write the host
did not pick up is rolled back.

Usage

skill_management({ action: "list" })
skill_management({ action: "inspect", id: "windows-shell" })
skill_management({ action: "disable", id: "windows-shell" })
skill_management({ action: "disable", id: "windows-shell", dryRun: true })
skill_management({ action: "disable", id: "windows-shell", effect: "ask" })
skill_management({ action: "enable", id: "windows-shell" })

Full field tables and error codes are on the four action pages of the docs site:
<https://sogeisetsu.github.io/opencode-skill-management/>

Install

Verify it loaded in a new process:

opencode plugin list --builtin
# skill_management 1.2.0 opencode-skill-management@1.2.0

Compatibility and verification…