October 30, 20262026年10月30日

Minimal install

This release ships a package with exactly what the skill needs at runtime, plus the license:

• opencode-go-model-picker-2.0.0-min.zip → SKILL.md, references/, scripts/, LICENSE

Unzip it into ~/.config/opencode/skills/ (Linux/macOS) or %USERPROFILE%\.config\opencode\skills\ (Windows) so the folder is named opencode-go-model-picker. No dependencies, no build step.

Chinese documentation: README-ZH.md.

---

[2.0.0] - 2026-10-30

Added

• references/quota-math.md and scripts/quota-math.mjs: the one priced-accounting

implementation. Formula v2 prices each assistant message on its own timestamp,
picks the price by the request's prompt length when a model has context-length
tiers, applies the peak-hour multiplier only to models that declare one, and
bills cache writes on the models that charge for them. Reasoning tokens are
billed at the output rate, checked message-by-message against the provider's own
per-message cost — omitting them under-prices a message by 4–34%. Derived quantities
(share of cap, the three independent pool windows, sessions per window, official
and observed capacity), the pool reverse-solve order, the eight known gaps, and
the degradation rules for every missing field.

• references/console-mcp.md and scripts/probe-console.mjs: the five Console MCP

states (authorized / gated / unauthenticated / absent / unknown),
decided cheapest-first (in-run tool result → opencode mcp list → config → one
initialize request), with the fixed per-state guidance copy, the three-level
fallback, and the enabling steps for the OpenCode CLI and OpenChamber. The probe
never prints a token and never classifies a network failure as "not set up".

• scripts/calibrate-usage.mjs: sums the local OpenCode SQLite database per agent

and model, compares it with the Console's own cost to measure the bias (warning
when it drifts more than 10%), reverse-solves the plan pool in the documented
order, projects month-end occupancy from the billing cycle, and — with --write
— stores plan, models.<id>.billing, usage.agentSplit and calibration in
the snapshot.

• references/model-prices.json gained cacheWritePer1M, ctxTiers, capUsd,

peak, promoUntil and a verified status per model, plus current
planVerifiedAt dates. It is now the single home of price, cap, context tiers
and peak hours.

• Snapshot schema v3: plan.go, models.<id>.billing, availability,

capabilities (per-capability source + verifiedAt), usage.agentSplit and
calibration.

• scripts/discover-agents.mjs now reports a run-level deadChainIds summary

(missing-from-catalog, retired, registry-missing, registry-deprecated),
warns on a shared chain head (contention) because the per-model cap is shared,
and warns per agent when a Go id's catalog status is retired (catalog).

• references/data-sources.md gained a Console-bill (priority 0) section, a

two-quota-layer section, official-vs-observed request counts, and a dated
capability/compliance table (image input, training, retention, availability).

• quota-math.mjs reports seed freshness — verifiedAt (the newest

planVerifiedAt / priceVerifiedAt in the table), ageDays and stale — and
marks a seed older than seven days as expired, so an offline run labels its
numbers instead of presenting them as current. calibrate-usage.mjs reports the
same under prices.

Changed

• The report now opens with a mandatory Basis: line (`Console bill

(authoritative) / Estimate / Uncalibrated`) and one denominator for the whole
table, and the table columns are the priced ones: headroom, accounted
$/session (median · P90), share of the model cap, 10 runs as a share of the
three pool windows, actual monthly occupancy, and why.

• A new Usage evidence block is mandatory: last-30-days spend and requests, a

last-7-days pace with a month-end projection, the count of candidates without an
ELO, the dead chain ids for this run, and the main agents' share of spend.

• SKILL.md's workflow was reordered to discover → probe Console → read the bill

and calibrate → refresh the snapshot → detect plan changes → allocate → build
chains → output → stop, and now ends with nine explicit self-checks.

• scripts/refresh-prices.mjs reads the cache-write price and the context-length

tiers from models.dev, and carries the plan-side constants (capUsd /
monthlyLimitUsd, peak, promoUntil, estReq*, status) through a rewrite,
so regenerating the seed no longer deletes them.

• calibrate-usage.mjs --write raises schemaVersion to 3 when it writes v3 keys;

an existing v2 file is read as-is (every v3 key is optional).

• references/agent-sources.md states that a native agent.<name>.model cannot

express a chain, and points at the chain-capable source when failover is wanted.

• Stale plan facts from the 2026-09/10 verifications were corrected against the

2026-10-30 plan page (the DeepSeek V4.1 Flash 4x promo is gone; its cap is $80;
peak/off-peak pricing replaced it).

Breaking

• Output format: the previous table columns (role overview, capability focus,

current model) are replaced by the priced columns, and the Basis: line is
mandatory. A v1 report shape is no longer valid.

• Snapshot schema v2 → v3. Existing files stay readable, but the new fields are

absent (hence null) until calibrate-usage.mjs --write runs.

• references/model-prices.json entry shape grew from 16 to 21 keys; anything

reading that seed directly must tolerate the new fields.

• SKILL.md's workflow numbering changed, so any prompt that referred to

"step 2 = refresh" now points at the Console probe instead.

v1.1.0

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.

Breaking change in 1.1.0

All four actions now return one JSON object instead of Markdown. The tool also declares an
output JSON Schema, so the host validates every result; this is a real structured tool result,
not Markdown wrapped in a JSON string.

• Envelope: schemaVersion (currently 1), action, ok, warnings.
• Failures use ok: false plus error: { code, message }; success is no longer inferred from

✅ / ❌ in the text.

• list returns callerAgent, summary, skills[], config, sidecar, consistency.

description is still the first line, no longer truncated to 60 characters.

• inspect returns skill, idDerivation, agents[], caller, conclusion. Missing and

ambiguous targets use skill_not_found / name_ambiguous.

• disable / enable return changed, reason, before, diff, verification, backup,

sidecar; enable also returns location, rule, candidates, relatedRules.

• enable now returns ok: false with config_unparsable when the target config cannot be

parsed and no candidate rule exists, instead of claiming "nothing to change".

Nothing else broke: actions, input fields, the written rule shape, the same-line marker format
and the single-file layout are unchanged.

Also since v1.0.0

• New bilingual documentation site (Astro + Starlight; English at the root, Chinese under

/zh-cn): <https://sogeisetsu.github.io/opencode-skill-management/>

• README now documents the minimal npm install (opencode plugin add) and the package contents.
• Theme/language menus, card alignment and mobile spacing fixes on the docs site.

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" })

A list result looks like this (abridged):

{
 "schemaVersion": 1,
 "action": "list",
 "ok": true,
 "warnings": [],
 "callerAgent": { "id": "build", "verified": true },
 "summary": { "total": 29, "disabled": 1, "ask": 0, "diverged": 2, "derivedIds": 0 },
 "skills": [
 { "id": "windows-shell", "state": "available", "source": "config", "path": "..." },
 { "id": "pdf-tools", "state": "disabled", "source": "catalog", "path": "..." }
 ],
 "config": { "file": ".../opencode.jsonc", "hadPermissions": true, "notes": [] },
 "sidecar": { "file": ".../skill_management.state.json", "ok": true, "error": null },
 "consistency": { "staleSidecarIds": [], "unrecordedRules": [], "missingFromRegistry": [] }
}

Full field tables and error codes are on the four action pages of the docs site.

Install

Verify it loaded in a new process:

opencode plugin list --builtin
# skill_management 1.1.0 opencode-skill-management@1.1.0

Compatibility and verification

• Targets OpenCode V2; 1.1.0 was accepted on host 2.0.32 on Windows and Linux.
• node tools/verify.mjs: 260 assertions pass, entirely against a temporary config directory.
• New-process runs against the real host returned parseable JSON for both list and the

"skill not found" error path, and both passed the host's output validation.

• One file, pure Node built-ins, no runtime dependencies, no shell calls.

Full diff: <https://github.com/sogeisetsu/opencode-skill-management/compare/v1.0.0...v1.1.0>