pi-t3-pstack 0.0.0-stage → 0.1.0

package/pstack/skills/typescript-best-practices/references/patterns.md

@@ -0,0 +1,328 @@
+# TypeScript patterns
+
+Code examples for each rule in `SKILL.md`. The underlying principles are language-agnostic. See the **type-system-discipline** and **boundary-discipline** principle skills.
+
+## Branded types
+
+Brand primitives so they can't be mixed up. Validate once at the boundary. Downstream code trusts the type.
+
+```ts
+type AgentId = string & { readonly __brand: "AgentId" };
+
+function isAgentId(input: string): input is AgentId {
+  return isUUID(input);
+}
+
+function parseAgentId(input: string): AgentId {
+  if (!isAgentId(input)) throw new Error(`Invalid agent id: ${input}`);
+  return input;
+}
+
+function focusAgent(id: AgentId): void {
+  /* input is trusted */
+}
+```
+
+Match the `readonly __brand: 'X'` shape. Don't invent a new convention.
+
+## Discriminated unions
+
+Model variants with a literal discriminant. Every variant shares the field name and each variant's value is unique, so impossible combos can't be represented.
+
+```ts
+// Don't. Boolean + optionals lets contradictory states exist.
+type DiffState = { loading: boolean; diff?: GitDiff; error?: string };
+
+// Do. Only valid states exist.
+type DiffState =
+  | { kind: "loading" }
+  | { kind: "ready"; diff: GitDiff }
+  | { kind: "error"; error: string };
+```
+
+Pick one discriminant name (`kind`, `type`, `tag`) and stick to it.
+
+## Constructive modeling
+
+Build the type from parts that are all legal instead of restricting a loose type with runtime checks.
+
+Non-empty, via a variadic tuple:
+
+```ts
+type NonEmpty<T> = [T, ...T[]];
+
+// Don't: T[] plus a length check every caller must repeat
+function pickWinner(entries: string[]): string {
+  if (entries.length === 0) throw new Error("no entries");
+  return entries[Math.floor(Math.random() * entries.length)];
+}
+
+// Do: an empty value of the type can't exist
+function pickWinner(entries: NonEmpty<string>): string {
+  return entries[Math.floor(Math.random() * entries.length)];
+}
+```
+
+Where a plain `T[]` arrives, narrow once with a guard. The fact then travels in the type:
+
+```ts
+const isNonEmpty = <T>(arr: T[]): arr is NonEmpty<T> => arr.length > 0;
+```
+
+Even length, as pairs:
+
+```ts
+type Pairs<T> = [T, T][];
+```
+
+A time range, as start plus duration:
+
+```ts
+// Don't: a comment holds the invariant
+type TimeRange = { start: Date; end: Date }; // start <= end
+
+// Do: a negative range can't be written; derive end when needed
+type TimeRange = { start: Date; durationMs: number };
+```
+
+Keep `durationMs` a plain number. Brand it (per Branded types) only if a raw number could be passed where a duration is expected, not by reflex. Pick the representation that makes the bad state unconstructable, then expose the reading you need on top (`pairs.flat()`, a `rangeEnd()` helper).
+
+## Simplest total type
+
+Don't strengthen everything. Keep `T[]` when every operation on it is total:
+
+```ts
+const sum = (xs: number[]) => xs.reduce((a, b) => a + b, 0); // [] is 0, fine
+```
+
+Strengthen when the loose type forces a lie at a use site. The tells are `!`, `arr[0] as T`, and a "should never happen" throw:
+
+```ts
+// Don't: partiality smuggled past the compiler
+function newestSession(sessions: Session[]): Session {
+  return sessions.at(0)!;
+}
+
+// Do: strengthen the input; the assertion disappears
+function newestSession(sessions: NonEmpty<Session>): Session {
+  return sessions[0];
+}
+```
+
+Weakening the result to `Session | undefined` is the other total signature.
+
+## `unknown` over `any`
+
+External data is always `unknown`. Narrow before use.
+
+```ts
+// Don't
+function handle(input: any) {
+  return input.foo.bar;
+}
+
+// Do
+function handle(input: unknown) {
+  if (typeof input === "object" && input !== null && "foo" in input) {
+    // narrowed; compiler verifies access
+  }
+}
+```
+
+External sources include RPC payloads, `JSON.parse`, `postMessage`, IPC, file contents, environment variables, database results.
+
+## Schemas before hand-rolled guards
+
+Before writing a property-by-property type guard for external data, look for the repository's runtime schema library and existing schemas. Let one schema own validation and derive the TypeScript type from it. Do not maintain a schema, a duplicate interface, and a guard that can drift apart.
+
+```ts
+import { z } from "zod";
+
+const UserSchema = z.object({
+  id: z.string().uuid(),
+  role: z.enum(["admin", "member"]),
+});
+
+type User = z.infer<typeof UserSchema>;
+
+function parseUser(input: unknown): User {
+  return UserSchema.parse(input);
+}
+```
+
+Use `safeParse` when failure is an expected branch. Use the equivalent inference helper when the repository uses another schema library. Do not add a new schema dependency for one guard. This rule prefers the schema system the codebase already trusts.
+
+## No `as` casts
+
+Every `as` is a potential runtime crash. Cast only after the type system has verified the claim.
+
+```ts
+import { z } from "zod";
+
+// Don't
+const user = data as User;
+
+// Don't
+function isUser(data: unknown): data is User {
+  return typeof data === "object" && data !== null && "id" in data;
+}
+
+// Do
+const userSchema = z.object({ id: z.string(), name: z.string() });
+type User = z.infer<typeof userSchema>;
+
+function parseUser(data: unknown): User {
+  return userSchema.parse(data);
+}
+```
+
+When the type comes first, annotate the validator with the type it proves. The compiler then rejects a validator that proves less than the type. Remove `name` from the object below and the assignment fails to compile.
+
+```ts
+type User = { id: string; name: string };
+
+const userSchema: z.ZodType<User> = z.object({ id: z.string(), name: z.string() });
+```
+
+When refactoring an `as` out of existing code, identify why TypeScript can't infer:
+
+- Missing discriminant: add one, switch to a discriminated union.
+- Overly wide source type (e.g. `Record<string, unknown>`): narrow it.
+- Untyped boundary: parse with the schema that owns the shape. Add a schema only where none exists.
+- Genuinely inexpressible: use a branded type or `satisfies`.
+
+## Narrowing hierarchy
+
+From best to last-resort:
+
+1. **Discriminated union switch / if.** Compiler narrows automatically.
+2. **`in` operator.** `"key" in obj` narrows to variants containing that key.
+3. **`typeof` / `instanceof`.** For primitives and class instances.
+4. **User-defined type guard.** When the above aren't enough.
+5. **`as` cast.** Only after validation.
+
+```ts
+function area(s: Shape): number {
+  if ("radius" in s) return Math.PI * s.radius ** 2; // narrowed to circle
+  return s.width * s.height; // narrowed to rect
+}
+```
+
+## Type guards
+
+A guard must actually verify the claim. A lying guard is worse than `as`.
+
+```ts
+function isCircle(s: Shape): s is Shape & { kind: "circle" } {
+  return s.kind === "circle";
+}
+```
+
+Prefer discriminant narrowing when possible.
+
+## Exhaustiveness
+
+In default arms, assign the discriminant to a `never`-typed local.
+
+```ts
+// Value-returning switch
+function area(s: Shape): number {
+  switch (s.kind) {
+    case "circle":
+      return Math.PI * s.radius ** 2;
+    case "rect":
+      return s.width * s.height;
+    default: {
+      const _exhaustive: never = s;
+      return _exhaustive;
+    }
+  }
+}
+
+// Void switch
+function handle(s: Shape): void {
+  switch (s.kind) {
+    case "circle":
+      drawCircle(s);
+      break;
+    case "rect":
+      drawRect(s);
+      break;
+    default: {
+      const _exhaustive: never = s;
+      void _exhaustive;
+    }
+  }
+}
+```
+
+Return-style in value-returning switches, void-style in statement switches.
+
+## `satisfies` over `as`
+
+`satisfies` validates without widening literal types.
+
+```ts
+// Don't. Widens, loses literal types.
+const config = { theme: "dark", cols: 3 } as Config;
+
+// Do. Validates AND preserves literal types.
+const config = { theme: "dark", cols: 3 } satisfies Config;
+// config.theme is "dark" (literal), not string
+```
+
+## Boundary validation
+
+Validate once where data crosses in. Trust types inside. See the **boundary-discipline** principle skill.
+
+- **Wire formats** (proto, JSON-RPC): parse with `ignoreUnknownFields` so forward-compatible changes don't break old clients.
+- **Persisted JSON:** versioned blob with a try/catch around the parse.
+- **Don't re-validate** deep in call chains.
+
+## Schema-derived types
+
+When a `.proto`, OpenAPI spec, GraphQL schema, or database migration already defines a shape, derive from the generated types instead of duplicating them.
+
+```ts
+// Don't. Duplicate shape, drifts when the schema changes.
+type CheckSummary = {
+  totalCount: number;
+  checks: { name: string; status: string }[];
+};
+function renderChecks(s: CheckSummary) {
+  /* ... */
+}
+
+// Do. Derive from the generated schema type.
+import type { ChecksMessage } from "<generated module>";
+function renderChecks(s: Pick<ChecksMessage, "totalCount" | "checks">) {
+  /* ... */
+}
+```
+
+Reach for `Pick`, `Omit`, `Parameters`, `ReturnType`, `Awaited`, `typeof` before writing a new interface.
+
+## Object args
+
+```ts
+// Don't. Swap two args, still compiles.
+openFile(uri, {
+  startLineNumber: 10,
+  startColumn: 1,
+  endLineNumber: 10,
+  endColumn: 1,
+});
+
+// Do. Order-independent, self-documenting.
+openFile({
+  uri,
+  selection: {
+    startLineNumber: 10,
+    startColumn: 1,
+    endLineNumber: 10,
+    endColumn: 1,
+  },
+});
+```
+
+Skip on hot paths: per-frame render, tokenizers, parsers, anything in a tight loop where the allocation cost matters.

package/pstack/skills/unslop/SKILL.md

@@ -0,0 +1,62 @@
+---
+name: unslop
+description: Cut AI tells from any writing.
+disable-model-invocation: true
+---
+
+# Unslop
+
+Edit text to remove AI patterns. Scan for the patterns below, then rewrite. Preserve meaning and match the intended tone.
+
+## Patterns to detect and fix
+
+Rule numbers are stable ids that other skills cite. A removed rule leaves a gap.
+
+### Content
+
+3. **Superficial -ing phrases.** "highlighting...", "ensuring...", "reflecting...", "showcasing...", "fostering...". Delete or expand with real sources.
+5. **Vague attributions.** "Experts believe", "Industry reports suggest", "Some critics argue". Name the source or delete.
+
+### Language
+
+7. **AI vocabulary.** Additionally, crucial, delve, enduring, enhance, fostering, garner, interplay, intricate, landscape (abstract), pivotal, showcase, tapestry (abstract), testament, underscore, vibrant. Replace with plain words.
+8. **Fancy ways to say "is".** "serves as", "stands as", "boasts", "features". Say "is" or "has".
+9. **"Not just X, but Y."** State the point directly instead.
+10. **Rule of three.** Don't force ideas into groups of three. Use the natural number.
+11. **Synonym cycling.** Protagonist, main character, central figure, hero all in one paragraph. Pick one, repeat it.
+12. **False ranges.** "from X to Y" where X and Y aren't on a meaningful scale. List topics directly.
+
+### Style
+
+13. **Em dash overuse.** Avoid em dashes entirely. Use periods or commas only (no parentheses, no en dashes, no hyphen-as-dash substitutes). If a thought needs separation, end the sentence or use a comma.
+14. **Colon overuse.** Colons are fine before a list or example, not as mid-sentence connectors. "If you're coming from traditional automation: instead of registering event handlers, you describe conditions" adds nothing with the colon. Rewrite to let the point stand on its own without comparison framing. "Describing when the scheduler should fire works best as plain English." Same meaning, no crutch punctuation.
+15. **Boldface overuse.** Don't bold every proper noun or acronym.
+16. **Inline-header lists.** The tell is a bold label and colon that restates the line: "**Performance:** Performance improved...". Convert those to prose. A bold lead-in that ends in a period, names the item, and is followed by genuinely new detail ("**Schema in TypeScript.** Tables live in one file.") is fine, not a tell.
+17. **Title case headings.** Use sentence case.
+18. **Decorative emojis.** Remove from headings and bullets.
+19. **Curly quotes.** Use straight quotes.
+
+### Communication artifacts
+
+20. **Chatbot phrases.** "I hope this helps!", "Let me know if...", "Of course!", "Certainly!", "Found the smoking gun!" Remove.
+22. **Sycophantic tone.** "Great question! You're absolutely right!" Respond directly.
+
+### Filler
+
+23. **Filler phrases.** "In order to" becomes "To". "Due to the fact that" becomes "Because". Delete "It is important to note that".
+24. **Excessive hedging.** "could potentially possibly be argued that it might" becomes "may".
+25. **Generic conclusions.** "The future looks bright." State specific plans or facts.
+
+### Jargon
+
+26. **Abstract metaphor nouns.** Substrate (base), wedge, wedge in (add), vector (way, method), locus, vantage, nexus, primitive as a noun, harness as a metaphor, surface as in "API surface", bedrock, scaffolding as a metaphor, modality, paradigm, gold-plating (more than the job needs), ratchet as a metaphor (the mechanism's real name, or a limit that only tightens), evacuate for moving code (move out), endgame (the last phase), north star, flywheel. These read as technical but usually have a plainer concrete word. Pick the concrete word.
+
+### Plain speech
+
+27. **Say what it does, not how it feels.** "the database stays close at hand", "SQL you can read", "types that follow your schema" name a feeling. The fix names the mechanism or a number: "`.toSQL()` returns the exact string sent to the database", "a column rename fails the build". Ask what the sentence tells the reader to do or know, then write that. If you can't restate it as a concrete instruction, fact, or number, cut it. One more check: if the sentence could appear unchanged in another project's docs, it says nothing about this one. Cut it.
+28. **Shorten or split dense sentences.** If the reader has to backtrack to parse a sentence, break it in two or drop clauses. One idea per sentence.
+29. **Active voice.** Prefer it. Catch "is/are/was/were + past participle" and name the actor: "queries are validated" becomes "the compiler validates queries", "the file is parsed by the loader" becomes "the loader parses the file". Passive is fine only when the actor is unknown or genuinely doesn't matter.
+30. **Cut adverbs, or use a stronger verb.** "runs quickly" becomes "is fast" or the number, and "significantly improves" becomes the measured delta. An adverb propping up a weak verb means the verb is wrong.
+31. **Prefer the plain word.** "utilize" and "leverage" become "use", "facilitate" becomes "help", "numerous" becomes "many", "in the event that" becomes "if". The fancier synonym is rarely clearer.
+32. **Mannered prose.** Metaphor or flourish where a literal phrase exists: aphorisms ("wire it or delete it"), rhetorical fragments for effect, personified code ("the plan holds it"), figurative verbs ("rides along", "stands on"), stock framing phrases. "A dial worth turning" becomes "a parameter worth varying". Say what you mean. Rule 26 covers the metaphor nouns.
+33. **Over-compression.** Dropped articles, verbless fragments, symbol-speak, and abbreviations that make the reader decode instead of read. "Parser rejects bad date → exit 2, no write" becomes "The parser rejects a bad date, exits with code 2, and writes nothing." Write whole sentences with their articles and verbs, and spell out arrows and abbreviations.

package/pstack/skills/what-did-i-get-done/SKILL.md

@@ -0,0 +1,31 @@
+---
+name: what-did-i-get-done
+description: Summarize authored commits over a user-specified time period into a concise update
+---
+
+# What did I get done
+
+## Trigger
+
+Need a short, high-signal summary of work completed in a specific time range (for example: yesterday, last 3 days, or last week).
+
+## Workflow
+
+1. Resolve the requested time window into concrete dates.
+2. Read commits authored by the current git user email within that range.
+3. Exclude merge commits and uncommitted changes.
+4. Synthesize the most important shipped changes into a concise status update.
+5. Include the actual date range used in the final summary.
+
+## Guardrails
+
+- Be extremely concise and information-dense.
+- Prioritize substantial behavior or architecture changes.
+- Omit cosmetic-only changes (formatting, imports, minor renames).
+- Do not infer intent or motivation. Describe changes functionally.
+
+## Output
+
+- One short summary suitable for a status update
+- Real date range
+- Optional 2-5 bullets for major changes only

package/pstack/skills/why/SKILL.md

@@ -0,0 +1,105 @@
+---
+name: why
+description: "Use for 'why does X work this way', 'why we picked Y', design rationale, regressions, postmortems, or data-backed thresholds. Discovers available MCPs and queries each evidence category (source control, issue tracker, long-form docs, real-time chat, infrastructure observability, error tracking, product analytics warehouse) in parallel, then returns a cited read on decisions and tradeoffs. Use how for runtime behavior."
+disable-model-invocation: true
+---
+
+# Why
+
+Investigate the motivation and intent behind code. The companion `how` skill answers what the code does and how it works. `why` answers what forces led to its shape.
+
+Each spawn below names a role line in the model sheet and a default. Get each `target` from the `pstack_model` tool with that role name. Omit `target` when it returns `inherit-parent`. Pick each subagent's reasoning level and fast or standard speed for the job, per the Subagents section of the **poteto-mode** skill.
+
+Every spawn is a `delegate_task` child with full tool access, MCP included, which the MCP-backed investigators and the synthesizer's citation spot-checks need. Investigators still shouldn't write anything.
+
+## Operating Posture
+
+Operate as a **careful, cautious, and precise investigator**. Be honest about what you know vs what you're inferring. Read `references/epistemics.md` for the full confidence framework and phrasing guide. The synthesizer must follow it.
+
+## Step 1. Understand the Target and the Question
+
+The **target** is usually a chunk of code, a pattern, a feature, or a named design decision. The **question** is usually a design rationale, a tradeoff, a motivating edge case, an external constraint, dead code, or a broad history sweep.
+
+If the target is vague ("why do we do it this way?" with no clear referent), make your best guess from conversation context (open files, recent edits, cursor location, what was just discussed). State your interpretation briefly so the user can redirect if you're off, then proceed.
+
+## Step 2. Establish the Code Anchor
+
+Before spawning investigators, build the seed context inline: file paths and line ranges, key symbols (functions, classes, constants), the last few commits touching the target, PR numbers from merge commits (pattern `(#1234)` in the subject line), and any linked ticket IDs. Run `git blame -L <start>,<end> <file>` for last-touch commits, `git log --follow -p -- <file>` for full history with patches through renames, `git log --oneline -20 -- <file>` for recent commits with PR numbers visible, and `git log -1 --format=%B <commit>` to extract PR numbers. Then pull the PR bodies and discussion for any substantive commit: `gh pr view <number> --json title,body,author,createdAt,mergedAt,labels,closingIssuesReferences,comments,reviews`. This anchor goes to every investigator and to the synthesizer.
+
+## Step 3. Spawn Parallel Investigators
+
+**Default to the full parallel investigation.** The goal is a complete coverage map, not a minimal one. Document the null, don't skip the search.
+
+### Discovery
+
+Before spawning investigators, list the MCP servers whose tools this session exposes.
+
+Map each available MCP to one evidence category:
+
+1. Source control history
+2. Issue / ticket tracker
+3. Long-form documents
+4. Real-time team chat
+5. Infrastructure observability
+6. Error / exception tracking
+7. Product analytics warehouse
+
+Source control is always available through git and `gh`. For the other six, classify using the MCP name, server instructions, tool names, and resource descriptors. If an MCP could fit more than one category, choose the one matching its primary evidence. Record ambiguous cases in the coverage map.
+
+Spawn one investigator per category that has a matching MCP, all in a single message so they run concurrently, each owning exactly one tool or MCP. Don't ask one agent to cover multiple MCPs. `target`: the `why investigators` line, default in [Models](#models).
+
+Each investigator's prompt is `references/investigator-prompt.md` filled in, plus:
+- the category playbook `references/sources/<source>.md` for the selected MCP, adapted from the examples in `references/source-playbook.md`
+- `references/sources/incident-postmortem.md` **if the target code looks defensive** (null checks, retry logic, timeout handling, rate limiting, feature flags, egress guards, OOM handlers)
+- the code anchor and the user's original question
+
+### Investigator roster, one per available evidence category
+
+Each category below is uniquely positioned to surface one shape of "why", shown after "Best at". Use it to know what to expect back, to name a gap when a category returns empty, and, only in the rare provably-irrelevant case, to justify a skip.
+
+1. **Source control** (git history, `gh` for PRs, code comments, tests). Always spawn, the only guaranteed source. Best at *implementation-time rationale captured during review*.
+2. **Issue / ticket tracker** (e.g. Linear, Jira, GitHub Issues, Plane, Shortcut MCP). Best at *the product or business forcing function*. Strongest when the why is external to engineering.
+3. **Long-form documents** (e.g. Notion, Confluence, Google Docs, Coda MCP). Best at *long-form design rationale*, written out before it becomes code.
+4. **Real-time team chat** (e.g. Slack, Discord, Microsoft Teams, Mattermost MCP). Best at *real-time deliberation that never reached a doc*. Especially important when the source control, ticket, and doc paper trail is thin.
+5. **Infrastructure observability** (e.g. Datadog, New Relic, Honeycomb, Grafana, Splunk MCP). Best at *infrastructure and runtime reality that motivated the code*. Strongest when the target reacts to an infra signal (timeouts, retries, rate limits, circuit breakers).
+6. **Error / exception tracking** (e.g. Sentry, Rollbar, Bugsnag, Airbrake MCP). Best at *the specific exceptions and error trajectories that motivated defensive or corrective code*. Strongest for catch blocks, null guards, type checks, retries, and other defenses.
+7. **Product analytics warehouse** (e.g. Databricks, Snowflake, BigQuery, ClickHouse, dbt, Redshift MCP). Best at *product and data reality that shaped the code*. Strongest for flag-gated code, experiment-driven ships, data migrations, and "where did this number come from" questions.
+
+### When to skip an investigator
+
+Only with an **explicit, written justification** that goes in the final Sources Consulted. Two valid reasons:
+
+- **No MCP is available for that category** in this environment. Flag it as a gap, not a choice: "Real-time team chat skipped. No matching MCP available, so the conversational record was not searchable."
+- **The source is provably irrelevant**, not just "probably irrelevant." A high bar: "Error / exception tracking skipped. Target is a build-time script with no runtime code path."
+
+If your scope assessment suggests a single-commit trivial target where the PR description already contains the complete answer, you may answer inline **only after** confirming all seven available category searches would be redundant. Say so explicitly. This should be rare.
+
+## Step 4. Synthesize
+
+Spawn one synthesizer. `target`: the `why synthesizer` line, default in [Models](#models). It gets:
+1. The investigator findings, including any null results and any categories skipped with justification
+2. The code anchor
+3. The user's original question
+4. The epistemics framework from `references/epistemics.md`
+5. The synthesizer prompt template from `references/synthesizer-prompt.md`
+
+## Step 5. Present
+
+Present the synthesizer's output to the user. You may lightly edit for clarity or add context from the conversation, but **do not rewrite the confidence language**.
+
+## Output Format
+
+The structure is the one in `references/synthesizer-prompt.md`: The Question, The Code in Question, What We Found, What We Can Reasonably Infer, Competing Hypotheses, What We Don't Know, Sources Consulted, Confidence Summary. Adapt as needed, but keep the confidence separation intact, and keep Sources Consulted as one line per investigator, including the ones that returned nothing or were skipped, with the reason.
+
+After Sources Consulted, if the user's `why` question is a precursor to changing this code, convert the lineage findings into a Preserve / Change / Avoid / Risk constraint set for planning the change.
+
+## Common Failure Modes to Avoid
+
+- **Recency bias**. Assuming the most recent commit is authoritative. The current shape is often the accretion of many earlier decisions. Trace back.
+
+## Models
+
+Get every target from the `pstack_model` tool. Pass the role name plus the job type and size from the routing table in poteto-mode's Subagents section. This skill's spawns are `research` jobs for investigators and `hard` jobs for the synthesizer. Write every brief from `<plugin root>/skills/poteto-mode/brief.md`: the extension refuses a `delegate_task` call whose brief is missing its sections or limits. Without a job, pass the complexity: `groundwork` for cheap mechanical work, `bounded` for involved but bounded work, `general` for work that needs substantial judgment. The tool reads the role's line in the model sheet (`~/.config/t3-pstack/models.md`, written by `/setup-pstack`). A line naming a model is used as written, moved to a provider T3 can run right now when needed. A role with no line, or `inherit-parent` or `auto`, is routed by this extension's `models.json`: the parent's model family, sized and reasoned by complexity, and for panels one seat per family. The extension blocks a `delegate_task` call that uses the reserved Fable model or fast mode unless the user named it, reasoning above what `pstack_model` returned, or reasoning above high unless the user's latest message is "allow xhigh". Pass each returned target as `delegate_task`'s `target`. `inherit-parent` means omit `target`. `blocked` means the panel cannot reach two model families on this T3, so record its reason.
+
+- why investigators: one target
+- why synthesizer: one target

package/pstack/skills/why/references/epistemics.md

@@ -0,0 +1,71 @@
+# Epistemics
+
+How to reason about confidence when evidence is historical, fragmentary, and sometimes contradictory, and how to tell the user without flattening it into false certainty. Code shows what it does, not *why it exists*. The why lives in commits, PRs, tickets, docs, and conversations, all incomplete, biased, and sometimes missing. Pretending otherwise produces confident-sounding guesses that mislead the user.
+
+## Confidence Tiers
+
+Every claim in the output must sit in one tier. The tier sets its section and its phrasing.
+
+### 1. Direct
+
+An explicit, textual citation answers the question: something an author actually *wrote* that says why, not "the code does X so the author must have wanted X." For example, a PR description "this fixes the bug where users with >1000 items couldn't paginate", a ticket "we're adding this because customer Acme requested it in their security review", a code comment "// clamp to 100 because the upstream API rejects larger values", a design doc "we chose option A over option B because we need persistence across restarts", or a chat message from the author "switching to this approach since the old one was flaky in tests".
+
+Phrasing: confident, present tense, cited. "This exists because X."
+
+### 2. Supported
+
+Multiple pieces of indirect evidence converge on one conclusion. No single source states it explicitly, but the pattern across sources makes it likely. For example, the PR title says "improve performance", the ticket is labeled "perf", and the surrounding commits all touch the same hot path. Or several tests added with the change all exercise very large inputs. Or the author's other PRs from the same week all name the same incident.
+
+Phrasing: confident but clearly derived, citing each item. "The evidence points strongly to X: [the specific pieces]."
+
+### 3. Inferred
+
+A reasonable reading of the context that nothing explicitly supports. It is *your interpretation*, not a fact from the record. For example, the PR doesn't say why, but the error was happening in production (per the incident channel timing) and the fix was rushed (merged the same day), so it was likely a hotfix. Or the function name suggests retry logic, and its retry count of 3 matches the team's 3-retry convention elsewhere in the codebase.
+
+Phrasing: hedged, with the chain explicit. "Given A and B, C seems likely because D."
+
+### 4. Speculative
+
+A plausible hypothesis on thin evidence, where other explanations fit equally well. Worth presenting, since the user may know which is right, but mark it as a guess. For example, "This might be a workaround for a browser bug that's since been fixed, but we found no contemporary evidence of that", or "It's possible this threshold was chosen to match an SLA commitment, but no SLA doc references it."
+
+Phrasing: "One possibility is X, but we have no direct evidence." Usually lives in Competing Hypotheses, alongside other possibilities.
+
+### 5. Unknown
+
+You looked and couldn't find out. A valid, important outcome. Document it.
+
+Phrasing: "We searched X, Y, and Z and found no evidence of why." Be specific about *what* you searched. "We couldn't find out" is less useful than "We searched the ticket tracker with keywords A and B, scanned the 6 PRs that touched this file since 2023, and grep'd the repo for string literals matching the threshold. None surfaced a rationale."
+
+## Phrasing Guide
+
+**Confidence words** imply Direct or Supported evidence and need a citation immediately adjacent. Never use them for inferences: "because", "the reason is", "was designed to" (claims author intent), "fixes" / "addresses" / "solves" (claims the change achieved its goal), "the team decided" (claims a group decision happened).
+
+**Hedges** signal interpretation, not report. Use them liberally in What We Can Reasonably Infer: "appears to", "seems to", "likely", "suggests", "is consistent with", "one reading is", "plausibly", "may have been", "the evidence points toward".
+
+**Words to avoid:**
+
+- "obviously". If it were obvious, the user wouldn't be asking
+- "clearly". Almost always precedes a claim that isn't clear
+- "of course". Same
+- "just" (as in "it's just X for performance"). Dismissive and usually hides uncertainty
+- "I think" / "I believe". You're synthesizing evidence, not giving a personal opinion. Use "the evidence suggests" instead.
+
+## Avoid Rationalization
+
+Code that "makes sense" today may have been written for reasons that no longer apply, or that were wrong at the time. Don't retrofit a clean rationale onto messy history. Don't assume the author did the "right" thing and work backward to justify it. Don't assume a consistent pattern was intentional when it might be copy-paste. Don't turn an absence of evidence into evidence of absence ("no one mentioned security concerns, so it must not have been a concern").
+
+## The Sycophancy Trap
+
+Users often embed a hypothesis: "Why do we do it this way, I assume it's for performance?" Don't simply confirm it. The guess is a prompt for investigation, not a conclusion to validate. Treat it as one candidate among others and check the evidence independently. If the evidence supports it, say so with citations. If not, say so, and present what the evidence *does* support.
+
+## When Evidence Contradicts
+
+Surface both sources in the output, with citations. Don't pick the one that fits a tidier narrative. For example, the ticket says "we need this for customer X's compliance requirement" and the PR says "cleaning up tech debt in this area". Both may be true (the ticket motivated the work, the PR gave the author's framing), or one may be wrong. Let the user make the call.
+
+## When Evidence Is Missing
+
+An honest "we don't know" is one of the most valuable outputs. The user learns the answer isn't in the obvious places, that they'll need to ask a human (the original author, the product owner, the team lead), or that they can decide the question isn't worth pursuing further. Failing to mark a gap, and instead filling it with a confident guess, actively harms the user, because they'll act on the guess. Name each gap concretely: the question, the sources searched, what you searched for in each, and what you found (nothing, or only tangential material).
+
+## Before Finalizing
+
+Before the synthesizer delivers the output, it checks every claim in "What We Found" and "What We Can Reasonably Infer". Each claim has a citation or moves to Inferred or Hypotheses. Its phrasing matches its tier. It doesn't treat code as evidence for its own intent (if it does, remove or reclassify it). The output has a "What We Don't Know" section. If no gaps are mentioned, that's suspicious. Either the evidence was unusually complete or something is being swept under the rug.