* feat(#49): provider-neutral model policy presets Adds model_policy config surface with known-provider presets (openai/anthropic/google/qwen) and generic provider escape hatch. model_policy.runtime_tiers resolves before legacy model_profile_overrides. reasoning_effort is stripped for unsupported runtimes. Closes #49 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(#49): replace unregistered /gsd-settings-advanced token in docs docs-parity-live-registry enforces every /token in docs/*.md maps to a live command. /gsd-settings-advanced is a workflow filename, not a registered command — use /gsd:settings instead. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(#49): update INVENTORY.md count and manifest for config-types.cjs inventory-counts and inventory-manifest-sync tests require the headline count and INVENTORY-MANIFEST.json to reflect every file in bin/lib/. config-types.cjs (new module added by feat(#49)) was missing from both. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
230 lines
8.5 KiB
JavaScript
230 lines
8.5 KiB
JavaScript
'use strict';
|
|
|
|
const path = require('node:path');
|
|
|
|
// Resolve model-catalog.json via a prioritised candidate list so the module
|
|
// works in every layout:
|
|
//
|
|
// 1. Co-located install path — get-shit-done/bin/shared/model-catalog.json
|
|
// Written by bin/install.js (#3288 fix). This is the canonical post-install
|
|
// location across all runtimes (Claude Code, Codex, OpenCode, etc.).
|
|
//
|
|
// 2. Source-repo dev path — sdk/shared/model-catalog.json
|
|
// Three levels up from bin/lib/: works when running directly from the
|
|
// open-gsd/gsd-core clone (the original path introduced by #3230).
|
|
//
|
|
// 3. GSD_MODEL_CATALOG env override — allows test harnesses and custom
|
|
// deployments to point at an arbitrary catalog file.
|
|
//
|
|
// Throws with a diagnostic message that lists all candidates when none resolve,
|
|
// so MODULE_NOT_FOUND surfaces as a clear actionable error (PRED.k301).
|
|
const _catalogCandidates = [
|
|
path.resolve(__dirname, '..', 'shared', 'model-catalog.json'),
|
|
path.resolve(__dirname, '..', '..', '..', 'sdk', 'shared', 'model-catalog.json'),
|
|
process.env.GSD_MODEL_CATALOG ? path.resolve(process.env.GSD_MODEL_CATALOG) : null,
|
|
].filter(Boolean);
|
|
|
|
let catalog = null;
|
|
let _catalogLastErr = null;
|
|
for (const _p of _catalogCandidates) {
|
|
try {
|
|
catalog = require(_p);
|
|
break;
|
|
} catch (e) {
|
|
// Only treat missing-file errors as recoverable — rethrow parse errors,
|
|
// permission errors, and any other real failures so they surface clearly
|
|
// instead of being silently swallowed (CR finding, PR #3293).
|
|
const isMissingCandidate =
|
|
(e && e.code === 'MODULE_NOT_FOUND' && String(e.message || '').includes(_p)) ||
|
|
(e && e.code === 'ENOENT');
|
|
if (!isMissingCandidate) throw e;
|
|
_catalogLastErr = e;
|
|
}
|
|
}
|
|
if (!catalog) {
|
|
throw new Error(
|
|
`model-catalog.json not found. Tried:\n${_catalogCandidates.map((p) => ` ${p}`).join('\n')}\nLast error: ${_catalogLastErr?.message}`
|
|
);
|
|
}
|
|
|
|
const VALID_PROFILES = [...catalog.profiles];
|
|
const VALID_PHASE_TYPES = new Set(catalog.phaseTypes);
|
|
const VALID_AGENT_TIERS = new Set(Object.keys(catalog.adaptiveTierMap));
|
|
|
|
const MODEL_PROFILES = Object.fromEntries(
|
|
Object.entries(catalog.agents).map(([agent, meta]) => [agent, {
|
|
quality: meta.golden,
|
|
balanced: meta.balanced,
|
|
budget: meta.budget,
|
|
adaptive: catalog.adaptiveTierMap[meta.routingTier],
|
|
}])
|
|
);
|
|
|
|
const AGENT_TO_PHASE_TYPE = Object.fromEntries(
|
|
Object.entries(catalog.agents).map(([agent, meta]) => [agent, meta.phaseType])
|
|
);
|
|
|
|
const AGENT_DEFAULT_TIERS = Object.fromEntries(
|
|
Object.entries(catalog.agents).map(([agent, meta]) => [agent, meta.routingTier])
|
|
);
|
|
|
|
const MODEL_ALIAS_MAP = Object.fromEntries(
|
|
Object.entries(catalog.runtimeTierDefaults.claude).map(([tier, entry]) => [tier, entry?.model])
|
|
);
|
|
|
|
const RUNTIME_PROFILE_MAP = Object.fromEntries(
|
|
Object.entries(catalog.runtimeTierDefaults)
|
|
.map(([runtime, tiers]) => [
|
|
runtime,
|
|
Object.fromEntries(
|
|
Object.entries(tiers).filter(([, entry]) => entry).map(([tier, entry]) => [tier, entry])
|
|
),
|
|
])
|
|
.filter(([, tiers]) => Object.keys(tiers).length > 0)
|
|
);
|
|
|
|
const KNOWN_RUNTIMES = new Set(Object.keys(catalog.runtimeTierDefaults));
|
|
const RUNTIMES_WITH_REASONING_EFFORT = new Set(
|
|
Object.entries(catalog.runtimeTierDefaults)
|
|
.filter(([, tiers]) => Object.values(tiers).some((entry) => entry && entry.reasoning_effort))
|
|
.map(([runtime]) => runtime)
|
|
);
|
|
|
|
const PROVIDER_PRESETS = catalog.providerPresets || {};
|
|
|
|
// KNOWN_PROVIDERS excludes 'generic' — it is a sentinel (all null entries) that
|
|
// forces users to supply model IDs via model_profile_overrides. It is not a
|
|
// real catalog-backed provider (#49).
|
|
const KNOWN_PROVIDERS = new Set(
|
|
Object.entries(PROVIDER_PRESETS)
|
|
.filter(([, tiers]) =>
|
|
Object.values(tiers).some((budgets) =>
|
|
budgets && Object.values(budgets).some((entry) => entry && entry.model)
|
|
)
|
|
)
|
|
.map(([name]) => name)
|
|
);
|
|
|
|
function nextTier(currentTier) {
|
|
const order = ['light', 'standard', 'heavy'];
|
|
const idx = order.indexOf(String(currentTier));
|
|
if (idx === -1) return null;
|
|
return order[Math.min(idx + 1, order.length - 1)];
|
|
}
|
|
|
|
function formatAgentToModelMapAsTable(agentToModelMap) {
|
|
const agentWidth = Math.max('Agent'.length, ...Object.keys(agentToModelMap).map((a) => a.length));
|
|
const modelWidth = Math.max('Model'.length, ...Object.values(agentToModelMap).map((m) => m.length));
|
|
const sep = '─'.repeat(agentWidth + 2) + '┼' + '─'.repeat(modelWidth + 2);
|
|
const header = ` ${'Agent'.padEnd(agentWidth)} │ ${'Model'.padEnd(modelWidth)}`;
|
|
let out = `${header}\n${sep}\n`;
|
|
for (const [agent, model] of Object.entries(agentToModelMap)) {
|
|
out += ` ${agent.padEnd(agentWidth)} │ ${model.padEnd(modelWidth)}\n`;
|
|
}
|
|
return out;
|
|
}
|
|
|
|
function getAgentToModelMapForProfile(normalizedProfile) {
|
|
const profile = VALID_PROFILES.includes(normalizedProfile) ? normalizedProfile : 'balanced';
|
|
const out = {};
|
|
for (const [agent, profiles] of Object.entries(MODEL_PROFILES)) {
|
|
out[agent] = profile === 'inherit' ? 'inherit' : (profiles[profile] ?? profiles.balanced);
|
|
}
|
|
return out;
|
|
}
|
|
|
|
// ─── Effort rendering ────────────────────────────────────────────────────────
|
|
//
|
|
// Universal effort ladder: minimal < low < medium < high < xhigh < max
|
|
//
|
|
// Each runtime supports a subset. The unique tails must be clamped when emitting
|
|
// to a runtime that does not support them:
|
|
// - 'max' is Anthropic-only: Codex does not support it -> clamp to 'xhigh'
|
|
// - 'minimal' is Codex-only: Claude does not support it -> clamp to 'low'
|
|
//
|
|
// Rendering maps the universal effort string to the runtime's native parameter.
|
|
|
|
const EFFORT_RENDERING = {
|
|
// Claude Code subagent effort: output_config.effort frontmatter key /
|
|
// CLAUDE_CODE_EFFORT_LEVEL env. Supports: low, medium, high, xhigh, max.
|
|
// Does NOT support 'minimal' (Codex-only) -> clamp to 'low'.
|
|
claude: {
|
|
param: 'output_config.effort',
|
|
channel: 'frontmatter',
|
|
supported: new Set(['low', 'medium', 'high', 'xhigh', 'max']),
|
|
clamp(level) {
|
|
if (level === 'minimal') return 'low';
|
|
return level;
|
|
},
|
|
},
|
|
// Codex Responses API reasoning.effort. Supports: minimal, low, medium, high, xhigh.
|
|
// Does NOT support 'max' (Anthropic-only) -> clamp to 'xhigh'.
|
|
codex: {
|
|
param: 'model_reasoning_effort',
|
|
channel: 'api',
|
|
supported: new Set(['minimal', 'low', 'medium', 'high', 'xhigh']),
|
|
clamp(level) {
|
|
if (level === 'max') return 'xhigh';
|
|
return level;
|
|
},
|
|
},
|
|
};
|
|
|
|
/**
|
|
* Render a universal effort string for a specific runtime.
|
|
*
|
|
* Returns { value (clamped), param, channel } where:
|
|
* - value: the clamped effort string safe to pass to the runtime
|
|
* - param: the native parameter name (e.g. 'output_config.effort')
|
|
* - channel: how the value is propagated ('frontmatter', 'api', null)
|
|
*
|
|
* Unknown runtimes return { value: universalEffort, param: null, channel: null }
|
|
* so callers can always read .value safely.
|
|
*/
|
|
function renderEffortForRuntime(runtime, universalEffort) {
|
|
const spec = EFFORT_RENDERING[runtime];
|
|
if (!spec) {
|
|
return { value: universalEffort, param: null, channel: null };
|
|
}
|
|
return {
|
|
value: spec.clamp(universalEffort),
|
|
param: spec.param,
|
|
channel: spec.channel,
|
|
};
|
|
}
|
|
|
|
// ─── Fast mode propagation ───────────────────────────────────────────────────
|
|
//
|
|
// RUNTIMES_WITH_FAST_MODE is the set of runtimes where fast_mode=true can be
|
|
// propagated to a SPAWNED SUBAGENT via a native mechanism.
|
|
//
|
|
// Claude Code has NO per-subagent fast-mode mechanism — /fast is a session-level
|
|
// toggle only. Emitting a `fast_mode: true` frontmatter key on a Claude subagent
|
|
// would be a SILENT NO-OP, which is why 'claude' is deliberately excluded here.
|
|
//
|
|
// Only API-direct runtimes ('api') accept a speed:"fast" field in the request.
|
|
// Codex and other runtimes do not expose per-call fast_mode either.
|
|
const RUNTIMES_WITH_FAST_MODE = new Set(['api']);
|
|
|
|
module.exports = {
|
|
catalog,
|
|
MODEL_PROFILES,
|
|
VALID_PROFILES,
|
|
AGENT_TO_PHASE_TYPE,
|
|
VALID_PHASE_TYPES,
|
|
AGENT_DEFAULT_TIERS,
|
|
VALID_AGENT_TIERS,
|
|
MODEL_ALIAS_MAP,
|
|
RUNTIME_PROFILE_MAP,
|
|
KNOWN_RUNTIMES,
|
|
RUNTIMES_WITH_REASONING_EFFORT,
|
|
PROVIDER_PRESETS,
|
|
KNOWN_PROVIDERS,
|
|
nextTier,
|
|
formatAgentToModelMapAsTable,
|
|
getAgentToModelMapForProfile,
|
|
EFFORT_RENDERING,
|
|
renderEffortForRuntime,
|
|
RUNTIMES_WITH_FAST_MODE,
|
|
};
|