gstack/scripts/resolvers/gbrain.ts

271 lines
12 KiB
TypeScript

/**
* GBrain resolver — brain-first lookup and save-to-brain for thinking skills.
*
* GBrain is a "mod" for gstack. When installed, coding skills become brain-aware:
* they search the brain for context before starting and save results after finishing.
*
* These resolvers are suppressed on hosts that don't support brain features
* (via suppressedResolvers in each host config). For those hosts,
* {{GBRAIN_CONTEXT_LOAD}}, {{GBRAIN_SAVE_RESULTS}}, {{BRAIN_PREFLIGHT}},
* {{BRAIN_CACHE_REFRESH}}, and {{BRAIN_WRITE_BACK}} all resolve to empty string.
*
* Compatible with GBrain >= v0.10.0 (search CLI, doctor --fast --json, entity enrichment).
*
* Brain-aware planning (T4 / v1.48 plan): adds three new resolvers powered by
* the bin/gstack-brain-cache CLI and scripts/brain-cache-spec.ts. The new
* resolvers fire only for the 5 planning skills registered in
* SKILL_DIGEST_SUBSETS (office-hours, plan-ceo-review, plan-eng-review,
* plan-design-review, plan-devex-review).
*/
import type { TemplateContext } from './types';
import {
SKILL_DIGEST_SUBSETS,
SKILL_CALIBRATION_WEIGHTS,
BRAIN_CACHE_ENTITIES,
getSkillSubset,
getInvalidationTargets,
} from '../brain-cache-spec';
// Per-skill slug + title + tag metadata for SAVE_RESULTS. The full save
// template (heredoc body, entity-stub instructions, throttle handling,
// backlinks) lives in docs/gbrain-write-surfaces.md §Save Template and is
// read on-demand by the agent. Compressing the inline prose keeps the
// token footprint at ~150 tokens per skill (down from ~500), so users with
// gbrain installed pay a small overhead and users without it (whose hosts
// have GBRAIN_SAVE_RESULTS suppressed at gen-time) pay nothing.
interface SkillSaveMeta {
slugPrefix: string;
title: string;
tag: string;
}
const skillSaveMap: Record<string, SkillSaveMeta> = {
'office-hours': { slugPrefix: 'office-hours', title: 'Office Hours', tag: 'design-doc' },
'investigate': { slugPrefix: 'investigations', title: 'Investigation', tag: 'investigation' },
'plan-ceo-review': { slugPrefix: 'ceo-plans', title: 'CEO Plan', tag: 'ceo-plan' },
'plan-eng-review': { slugPrefix: 'eng-reviews', title: 'Eng Review', tag: 'eng-review' },
'plan-design-review': { slugPrefix: 'design-reviews', title: 'Design Review', tag: 'design-review' },
'plan-devex-review': { slugPrefix: 'devex-reviews', title: 'Devex Review', tag: 'devex-review' },
'retro': { slugPrefix: 'retros', title: 'Retro', tag: 'retro' },
'ship': { slugPrefix: 'releases', title: 'Release', tag: 'release' },
'cso': { slugPrefix: 'security-audits', title: 'Security Audit', tag: 'security-audit' },
'design-consultation': { slugPrefix: 'design-systems', title: 'Design System', tag: 'design-system' },
};
export function generateGBrainContextLoad(ctx: TemplateContext): string {
let base = `## Brain Context Load
**Skip this entire section if \`gbrain\` is not on PATH.**
Extract 2-4 keywords from the user's request. Search the brain:
\`gbrain search "<keywords>"\`. Read the top 3 results with
\`gbrain get_page "<slug>"\`. Use that context to inform your analysis.
If \`gbrain search\` returns no results or any non-zero exit, proceed
without brain context. Full search/read protocol + examples:
see \`docs/gbrain-write-surfaces.md\` §Context Load.`;
if (ctx.skillName === 'investigate') {
base += `\n\nFor structured-data extraction requests ("track this", "extract from emails", "build a tracker"), route to GBrain's data-research skill instead: \`gbrain call data-research\`.`;
}
return base;
}
export function generateGBrainSaveResults(ctx: TemplateContext): string {
// gbrain v0.18+ uses `gbrain put <slug>` (NOT the deprecated `put_page`
// MCP op). Compressed in v1.50.0.0: the inline heredoc + entity-stub +
// throttle + backlink prose moved to docs/gbrain-write-surfaces.md
// §Save Template, which the agent reads on demand when it actually
// saves. The compact pointer keeps non-gbrain users' token overhead
// near zero when their host's static suppression is overridden by
// detection.
const meta = skillSaveMap[ctx.skillName];
if (!meta) {
return `## Save Results to Brain
**Skip this entire section if \`gbrain\` is not on PATH.**
If the skill output is worth preserving, save it via
\`gbrain put "<slug>" --content "<frontmatter + markdown>"\`. Full template
(heredoc body, frontmatter shape, entity-stub instructions, throttle
handling): see \`docs/gbrain-write-surfaces.md\` §Save Template.`;
}
return `## Save Results to Brain
**Skip this entire section if \`gbrain\` is not on PATH.**
After completing this skill, save the output:
\`\`\`bash
gbrain put "${meta.slugPrefix}/<feature-slug>" --content "$(cat <<'EOF'
---
title: "${meta.title}: <feature name>"
tags: [${meta.tag}, <feature-slug>]
---
<skill output in markdown>
EOF
)"
\`\`\`
Then extract person/org entities and create stub pages for each one.
Throttle errors (exit 1 with "throttle"/"rate limit"/"busy") and any
other non-zero exit are transient — don't retry inline. Full entity-stub
template, throttle handling, and backlink protocol:
see \`docs/gbrain-write-surfaces.md\` §Save Template.`;
}
// ────────────────────────────────────────────────────────────────────
// Brain-aware planning resolvers (T4 / v1.48 plan)
// ────────────────────────────────────────────────────────────────────
/**
* Returns true when this skill is registered for brain preflight. Skills not
* in SKILL_DIGEST_SUBSETS get an empty BRAIN_PREFLIGHT block (no behavior).
*/
function isPreflightSkill(skillName: string): boolean {
return Object.prototype.hasOwnProperty.call(SKILL_DIGEST_SUBSETS, skillName);
}
/**
* Renders the per-skill BRAIN_PREFLIGHT block. The rendered output is a single
* bash script that:
* 1. Reads each digest file from gstack-brain-cache get (one call per digest)
* 2. Falls back to "(brain context unavailable)" on missing
* 3. Concatenates outputs into a single ## Brain Context block injected
* into the skill's prompt context
* 4. Tells the agent: "use this context to skip already-known questions"
*
* The cache CLI handles cold-refresh + lock dedup + stale-but-usable
* fallback internally. From the resolver's perspective the call is one
* shell command per digest.
*/
export function generateBrainPreflight(ctx: TemplateContext): string {
if (!isPreflightSkill(ctx.skillName)) return '';
const subset = getSkillSubset(ctx.skillName);
const binDir = ctx.paths.binDir;
// Build the bash that loads each digest. Per-skill subset is small (2-5 entries).
const loadLines = subset.map((entityName) => {
const entity = BRAIN_CACHE_ENTITIES[entityName];
if (!entity) return '';
const projectFlag = entity.scope === 'per-project' ? '--project "$SLUG"' : '';
return ` printf '\\n### %s\\n\\n' "${entityName}"\n ${binDir}/gstack-brain-cache get ${entityName} ${projectFlag} 2>/dev/null || printf '_(no ${entityName} digest available yet)_\\n'`;
}).join('\n');
return `## Brain Context (preflight)
Before asking any clarifying questions, load the brain's structured context
for this project. The cache layer handles staleness, refresh, and stale-but-
usable fallback automatically. Skip questions whose answers are already
present in the loaded context; ground recommendations in what the brain
already knows about the user, the product, the goals, and recent decisions.
\`\`\`bash
eval "$(${binDir}/gstack-slug 2>/dev/null)" 2>/dev/null || true
{
printf '## Brain Context\\n\\n'
${loadLines}
} > /tmp/.gstack-brain-context-$$.md 2>/dev/null
[ -s /tmp/.gstack-brain-context-$$.md ] && cat /tmp/.gstack-brain-context-$$.md
rm -f /tmp/.gstack-brain-context-$$.md 2>/dev/null || true
\`\`\`
**How to use this context:**
- If \`product\` digest names the value prop, target user, or stage — don't re-ask.
- If \`goals\` digest lists active goals — frame recommendations against them.
- If \`recent-decisions\` digest names a prior scope/architecture choice — flag if this plan contradicts.
- If \`user-profile\` digest carries calibration pattern statements ("tends to over-engineer security") — surface them when relevant.
- If a digest is \`(no X digest available yet)\`, treat that section as cold; ask the user.
**Privacy:** Salience digest is filtered by allowlist (D9 default: \`projects/\`,
\`gstack/\`, \`concepts/\` only). Personal/family/therapy content never leaks here.
`;
}
/**
* Renders the at-skill-end background refresh hook. Fires after the skill's
* own work completes (telemetry has already logged); kicks any digest whose
* age exceeds half its TTL but hasn't yet expired, so the NEXT invocation
* gets a fresh cache without paying the cold-miss tax.
*
* Subordinate to {{TELEMETRY}} — runs after. Doesn't block the user.
*/
export function generateBrainCacheRefresh(ctx: TemplateContext): string {
if (!isPreflightSkill(ctx.skillName)) return '';
const binDir = ctx.paths.binDir;
return `## Brain Cache Background Refresh
After the skill's work completes (and telemetry has logged), kick a
background refresh of any cache digest that's getting close to its TTL.
This is non-blocking — the user doesn't wait. Next invocation benefits
from the warm cache.
\`\`\`bash
eval "$(${binDir}/gstack-slug 2>/dev/null)" 2>/dev/null || true
(${binDir}/gstack-brain-cache refresh --project "$SLUG" 2>/dev/null &) || true
\`\`\`
`;
}
/**
* Renders the calibration write-back block. ONLY emits when the skill makes
* typed decisions worth a kind=bet take AND the brain trust policy is
* personal. Phase 2 / E5 cross-skill calibration.
*
* Gated behind BRAIN_CALIBRATION_WRITEBACK feature flag in the resolver
* output — the flag stays false until upstream gbrain ships takes_add MCP
* op (T8). When the flag flips, the existing skill templates pick up the
* write-back behavior without any template changes.
*/
export function generateBrainWriteBack(ctx: TemplateContext): string {
if (!isPreflightSkill(ctx.skillName)) return '';
const weight = SKILL_CALIBRATION_WEIGHTS[ctx.skillName];
if (weight == null) return '';
// List the cache digests this skill's writes should invalidate. Multiple
// skills write to multiple entities; the invalidation map captures this.
const invalidatesEntities = getInvalidationTargets(`/${ctx.skillName}`);
const invalidateBash = invalidatesEntities
.map((e) => ` ${ctx.paths.binDir}/gstack-brain-cache invalidate ${e} --project "$SLUG" 2>/dev/null || true`)
.join('\n');
return `## Brain Calibration Write-Back (Phase 2 / gated)
When the skill makes a typed prediction worth tracking (scope decision,
TTHW target, architectural bet, wedge commitment), it MAY write a
\`kind=bet\` take to the brain so a calibration profile builds over time.
**Gated on two things:**
1. Brain trust policy for the active endpoint is \`personal\` (check via
\`${ctx.paths.binDir}/gstack-config get brain_trust_policy@<endpoint-hash>\`).
Shared brains skip write-back to avoid polluting team calibration.
2. Feature flag \`BRAIN_CALIBRATION_WRITEBACK\` is set (today: false; flips
to true when upstream gbrain v0.42+ ships \`takes_add\` MCP op).
When both gates pass, the write-back path uses \`mcp__gbrain__takes_add\`
to record a take with weight ${weight} (per SKILL_CALIBRATION_WEIGHTS).
If the MCP op is unavailable, fall back to \`mcp__gbrain__put_page\` with
a gstack:takes fence block (documented but uglier path).
Mandatory take frontmatter shape:
\`\`\`yaml
kind: bet
holder: <user identity from whoami>
claim: <one-line prediction the skill is making>
weight: ${weight}
since_date: <today's date>
expected_resolution: <date in 1-3 months depending on skill>
source_skill: ${ctx.skillName}
\`\`\`
After write, invalidate the affected digests so the next preflight reflects
the new state:
\`\`\`bash
eval "$(${ctx.paths.binDir}/gstack-slug 2>/dev/null)" 2>/dev/null || true
${invalidateBash || ' # (no per-skill invalidation targets configured)'}
\`\`\`
`;
}