@diegopetrucci/pi-extensions 0.1.78 → 0.1.79

package/extensions/context-inspector/index.ts

@@ -143,7 +143,6 @@ type MinimalMessage = {
   customType?: string;
   display?: boolean;
   summary?: string;
-  fromId?: string;
   tokensBefore?: number;
   command?: string;
   output?: string;
@@ -165,7 +164,6 @@ type MinimalEntry = {
   display?: boolean;
   details?: unknown;
   summary?: string;
-  fromId?: string;
   tokensBefore?: number;
   firstKeptEntryId?: string;
 };
@@ -177,7 +175,6 @@ type MinimalProjectionEntry = {
 
 type MinimalSessionProjection = {
   entries: MinimalProjectionEntry[];
-  messages?: MinimalMessage[];
 };
 
 type AnalysisResult = {
@@ -192,7 +189,6 @@ type Segment = {
   source: string;
   role: string;
   tokens: number;
-  chars: number;
   turn: number;
   sequence: number;
   entryId?: string;
@@ -211,7 +207,6 @@ type Segment = {
 type CategoryStat = {
   id: CategoryId;
   label: string;
-  shortLabel: string;
   color: string;
   description: string;
   tokens: number;
@@ -230,13 +225,9 @@ type AggregateStat = {
 
 type DatasetStats = {
   tokens: number;
-  rawTokens: number;
-  visibleRawTokens: number;
   excludedTokens: number;
   providerDeltaTokens: number;
   estimatorOverageTokens: number;
-  reconciliationScale: number;
-  segmentCount: number;
   messageCount: number;
   categories: CategoryStat[];
   topSegments: Segment[];
@@ -247,8 +238,6 @@ type DatasetStats = {
 
 type Dataset = {
   id: DatasetId;
-  label: string;
-  description: string;
   segments: Segment[];
   stats: DatasetStats;
 };
@@ -259,7 +248,6 @@ type RoutedModel = {
 };
 
 type ReportData = {
-  generatedAt: string;
   cwd: string;
   session: {
     id?: string;
@@ -280,7 +268,6 @@ type ReportData = {
     percent: number | null;
   };
   options: {
-    redacted: boolean;
     defaultDataset: DatasetId;
   };
   categoryMeta: typeof CATEGORY_META;
@@ -507,7 +494,6 @@ function addTextSegment(
     source: base.source,
     role: base.role,
     tokens,
-    chars,
     turn: base.turn ?? state.turn,
     entryId: base.entryId,
     timestamp: base.timestamp,
@@ -543,7 +529,6 @@ function addImageSegment(
     source: base.source,
     role: base.role,
     tokens: IMAGE_TOKEN_ESTIMATE,
-    chars: IMAGE_TOKEN_ESTIMATE * 4,
     turn: base.turn ?? state.turn,
     entryId: base.entryId,
     timestamp: base.timestamp,
@@ -820,7 +805,6 @@ function contextEntryToMessage(entry: MinimalEntry): MinimalMessage | undefined
     return {
       role: "branchSummary",
       summary: entry.summary,
-      fromId: entry.fromId,
       timestamp: entry.timestamp ? new Date(entry.timestamp).getTime() : undefined,
     };
   }
@@ -916,10 +900,7 @@ function resolveSessionProjection(
     (!Array.isArray(projection.messages) || !projection.messages.every(isProjectionMessage))
   )
     return undefined;
-  return {
-    entries,
-    messages: entries.flatMap((entry) => entry.messages),
-  };
+  return { entries };
 }
 
 function resolveCurrentContextEntries(
@@ -1045,8 +1026,6 @@ function aggregateBy(
 
 function finalizeDataset(
   id: DatasetId,
-  label: string,
-  description: string,
   segments: Segment[],
   messageCount: number,
   authoritativeTokens: number | null,
@@ -1091,7 +1070,6 @@ function finalizeDataset(
       return {
         id: categoryId,
         label: meta.label,
-        shortLabel: meta.shortLabel,
         color: meta.color,
         description: meta.description,
         tokens: Math.round(rawTokens),
@@ -1105,18 +1083,12 @@ function finalizeDataset(
 
   return {
     id,
-    label,
-    description,
     segments,
     stats: {
       tokens: Math.round(displayTotal),
-      rawTokens: Math.round(segments.reduce((sum, segment) => sum + segment.tokens, 0)),
-      visibleRawTokens: Math.round(visibleRawTokens),
       excludedTokens: Math.round(excludedTokens),
       providerDeltaTokens,
       estimatorOverageTokens,
-      reconciliationScale,
-      segmentCount: segments.length,
       messageCount,
       categories,
       topSegments: [...visibleSegments].sort((a, b) => b.tokens - a.tokens).slice(0, 24),
@@ -1192,20 +1164,11 @@ function buildReportData(
   const fullSegments = [...overhead, ...fullAnalysis.segments];
   const currentDataset = finalizeDataset(
     "current",
-    "Current model context",
-    "What pi is expected to send to the model on the next turn: latest compaction summary plus unsummarized messages, system prompt, and active tool schemas.",
     currentSegments,
     currentAnalysis.messageCount,
     contextTokens,
   );
-  const fullDataset = finalizeDataset(
-    "full",
-    "Full active branch history",
-    "Every context-bearing entry on the active branch, including pre-compaction history that may no longer be sent verbatim.",
-    fullSegments,
-    fullAnalysis.messageCount,
-    null,
-  );
+  const fullDataset = finalizeDataset("full", fullSegments, fullAnalysis.messageCount, null);
 
   const notes: string[] = [
     "Per-component token counts are local estimates based mostly on characters/4. Providers expose aggregate usage, not exact per-message attribution.",
@@ -1250,7 +1213,6 @@ function buildReportData(
   }
 
   return {
-    generatedAt: new Date().toISOString(),
     cwd: options.redact ? "[redacted cwd]" : ctx.cwd,
     session: {
       id: options.redact ? "[redacted session]" : safeCall(() => ctx.sessionManager.getSessionId()),
@@ -1273,7 +1235,6 @@ function buildReportData(
       percent: contextPercent,
     },
     options: {
-      redacted: options.redact,
       defaultDataset: options.defaultDataset,
     },
     categoryMeta: CATEGORY_META,
@@ -1368,16 +1329,12 @@ function buildHtml(data: ReportData): string {
 	color-scheme: dark;
 	--bg: #070a12;
 	--panel: rgba(15, 23, 42, 0.72);
-	--panel-strong: rgba(15, 23, 42, 0.94);
 	--line: rgba(148, 163, 184, 0.18);
 	--text: #e5edf7;
 	--muted: #94a3b8;
 	--dim: #64748b;
 	--accent: #8b5cf6;
 	--accent-2: #06b6d4;
-	--good: #22c55e;
-	--warn: #f59e0b;
-	--bad: #ef4444;
 	--shadow: 0 24px 80px rgba(0, 0, 0, 0.35);
 }
 * { box-sizing: border-box; }
@@ -1588,7 +1545,6 @@ pre { margin: 0; white-space: pre-wrap; word-break: break-word; color: #cbd5e1;
 		for (const [key, value] of Object.entries(attrs)) {
 			if (key === 'class') node.className = value;
 			else if (key === 'text') node.textContent = value;
-			else if (key === 'html') node.innerHTML = value;
 			else if (key.startsWith('on') && typeof value === 'function') node.addEventListener(key.slice(2), value);
 			else if (value !== undefined && value !== null) node.setAttribute(key, value);
 		}

package/extensions/context-inspector/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@diegopetrucci/pi-context-inspector",
-  "version": "0.1.16",
+  "version": "0.1.17",
   "description": "A pi extension that opens a local HTML dashboard explaining where the current session context is going.",
   "keywords": [
     "pi-package",

package/extensions/contrarian/index.ts

@@ -1944,8 +1944,7 @@ function createContrarianExtension(pi: ExtensionAPI, deps: ContrarianExtensionDe
           onUpdate?.(partial);
         };
 
-        let lastErrorResult: { ok: false; error: string; details: ContrarianDetails } | undefined;
-        for (let index = 0; index < attempts.length; index++) {
+        for (let index = 0; ; index++) {
           const previous = index > 0 ? attempts[index - 1] : undefined;
           const attempt = previous
             ? withFallbackReason(attempts[index], previous)
@@ -1969,7 +1968,6 @@ function createContrarianExtension(pi: ExtensionAPI, deps: ContrarianExtensionDe
             });
           }
 
-          lastErrorResult = result;
           const canFallBack = index < attempts.length - 1 && isModelAvailabilityError(result.error);
           if (!canFallBack) {
             return finalizeResult({
@@ -1978,30 +1976,6 @@ function createContrarianExtension(pi: ExtensionAPI, deps: ContrarianExtensionDe
             });
           }
         }
-
-        return finalizeResult({
-          content: [
-            {
-              type: "text",
-              text: lastErrorResult?.error ?? "Contrarian could not select an available model.",
-            },
-          ],
-          details: lastErrorResult?.details ?? {
-            modelRef: "",
-            provider: ctx.model?.provider ?? "unknown",
-            modelId: "",
-            modelName: undefined,
-            thinkingLevel: configuredThinkingLevel ?? DEFAULT_THINKING_LEVEL,
-            autoSelected: true,
-            selectionReason: "No authenticated models were available to run the contrarian.",
-            includeBash: params.includeBash ?? false,
-            usage: createEmptyUsage(),
-            stderr: "",
-            exitCode: 1,
-            durationMs: 0,
-            cwd: params.cwd ?? ctx.cwd,
-          },
-        });
       } finally {
         activeRuns.delete(toolCallId);
         updateContrarianUi(ctx, activeRuns);

package/extensions/contrarian/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@diegopetrucci/pi-contrarian",
-  "version": "0.1.16",
+  "version": "0.1.17",
   "description": "A read-only contrarian subagent extension for pi that stress-tests plans, designs, assumptions, and conclusions.",
   "keywords": [
     "pi-package",

package/extensions/dynamic-context-pruning/README.md

@@ -278,8 +278,10 @@ Notes:
 - `gate.breakEvenThreshold`'s default of `22` is calibrated from the
   representative-corpus benchmark at `cachedPriceRatio` r=0.1 (aggressive
   prompt caching, the common case). It is **provider/ratio-
-  dependent**: the optimal threshold rises as caching gets weaker (T=29 at
-  r=0.25, T=54 at r=0.5, T=58 at r=0.9 on the same corpus).
+  dependent**. In the **historical, retired turn-end run**, the optimal
+  threshold rose as caching got weaker (T=29 at r=0.25, T=54 at r=0.5, T=58
+  at r=0.9); these are not current turn-start results. See
+  [`docs/v2-design.md` §4](./docs/v2-design.md#4-benchmark-decision-gate).
 
 ## Prompt-cache trade-off
 
@@ -326,6 +328,8 @@ Options:
 - `--limit N`: only process the first N session files found.
 - `--ratio R`: cached-price ratio(s) to model; repeatable and/or
   comma-separated (e.g. `--ratio 0.1,0.2`). Default: `0.1`.
+- `--sweep-max N`: explicit ceiling for the auto-expanding threshold sweep
+  (overrides the corpus-derived ceiling). Default: auto-expand from 30 as needed.
 - `--json`: emit a full machine-readable JSON dump instead of aligned text.
 - `--simulate-compression`: additionally SIMULATE v2-style range compression
   (docs/v2-design.md §1/§4 go/no-go evidence) alongside — never mixed into —
@@ -390,9 +394,11 @@ reports two kinds of numbers:
   prune point, and the *actual* net benefit that prune would have produced —
   `realizedNetBenefit = actualRemainingCalls * recurringSaving - penalty`.
 
-It also sweeps the break-even threshold `T` from 1 to 30 and reports the `T`
-that maximizes total realized net benefit across all gate-eligible
-candidates, both overall and split by idle/mid-loop agent state.
+It also sweeps the break-even threshold `T` from 1 to 30, then keeps doubling
+the range while the best `T` sits at the top edge, up to a corpus-derived
+ceiling (hard cap 500). `--sweep-max N` sets that ceiling explicitly. It
+reports the `T` that maximizes total realized net benefit across all
+gate-eligible candidates, both overall and split by idle/mid-loop agent state.
 
 ## Comparison
 
@@ -428,9 +434,11 @@ net benefit, at sufficient candidate volume, with enough remaining-calls
 runway per session to plausibly amortize a compression summary's own token
 cost — see `docs/v2-design.md` §4 for the exact go/no-go bar.
 
-The representative-corpus run of the benchmark harness (over
+The **historical, retired turn-end run** of the benchmark harness (over
 `~/.the-last-harness/agent/sessions`: 1,390+ session files, 556 gate-eligible
-candidates — see the corpus-choice note above) produced:
+candidates — see the corpus-choice note above) produced the following
+figures, preserved as historical context rather than current turn-start
+results (see [`docs/v2-design.md` §4](./docs/v2-design.md#4-benchmark-decision-gate)):
 
 - **Candidates and tokens removed by strategy**: `dedupe` 111 candidates /
   93,330 tokens removed; `error-purge` 187 / 28,191; `superseded-file-ops`
@@ -440,6 +448,7 @@ candidates — see the corpus-choice note above) produced:
   this corpus is dominated by long, tool-heavy agent sessions rather than
   short orchestrator sessions.
 
+The current turn-start re-derivation has 460 gate-eligible candidates.
 Shipped gate defaults remain `idle=22` / `mid_loop=1` (see
 [Configuration](#configuration)).
 

package/extensions/dynamic-context-pruning/index.ts

@@ -140,12 +140,6 @@ export function normalizeSessionProjection(value: unknown): MinimalSessionProjec
   return { entries, messages: entries.flatMap((entry) => entry.messages) };
 }
 
-export function sessionProjectionToMessages(
-  projection: MinimalSessionProjection,
-): MinimalMessage[] {
-  return projection.entries.flatMap((entry) => entry.messages);
-}
-
 const EXTENSION_ID = "dynamic-context-pruning";
 const CONFIG_FILE_NAME = "dynamic-context-pruning.json";
 const DECISION_ENTRY_TYPE = "dynamic-context-pruning:decision";
@@ -164,37 +158,10 @@ const DEFAULT_RECENT_TURNS = 4;
 /** Cache cost model default: fraction of full (non-cached) price still paid for cached tokens. */
 const DEFAULT_CACHED_PRICE_RATIO = 0.1;
 
-/**
- * Default net-benefit gate threshold (max amortization calls). Recalibrated
- * from the representative-corpus benchmark (~/.the-last-harness/
- * agent/sessions, 1,390+ session files, 556 candidates): this is the
- * hindsight-optimal break-even T at cachedPriceRatio r=0.1 (aggressive
- * prompt caching, the common Anthropic case). IMPORTANT caveats:
- *  - Provider/ratio-dependent: this optimum shifts higher as caching gets
- *    weaker (T=29 at r=0.25, T=54 at r=0.5, T=58 at r=0.9 on the same
- *    corpus) — see docs/v2-design.md section 4 and the extension README.
- *  - At r=0.1 the realized deterministic savings this threshold buys are
- *    economically marginal (~20.6k token-units total across the whole
- *    corpus, i.e. pennies); savings only become material at r>=0.25.
- *  - Re-derive via `scripts/benchmark.mjs <sessions-dir> --ratio 0.1` if
- *    the representative corpus changes meaningfully.
- */
+/** Default idle net-benefit gate threshold; see README.md and docs/v2-design.md §4. */
 const DEFAULT_BREAK_EVEN_THRESHOLD = 22;
 
-/**
- * Default mid_loop break-even threshold. The package's own
- * representative-corpus benchmark (turn-START state definition — the same one
- * the runtime `context` handler uses) favored rejecting mid_loop candidates
- * at cachedPriceRatio r=0.1 (mid_loop-optimal T=1, total realized
- * net benefit ~0.0). This supports a conservative default for the measured
- * corpus, not a universal claim that mid_loop pruning cannot pay off.
- * The reported 2026-09-08 byte-level capture (a mid_loop prune at T=22
- * invalidating 81% of a warm request prefix) illustrates cache invalidation,
- * not an inability to amortize it: enough subsequent calls can recoup a
- * one-time bust. Idle keeps T=22; turn-start classification does not imply
- * the provider cache is cold. Users who want the old behavior can override
- * via gate.breakEvenThresholdByState.mid_loop.
- */
+/** Default mid_loop threshold; see README.md for the runtime-aligned benchmark rationale. */
 const DEFAULT_MID_LOOP_BREAK_EVEN_THRESHOLD = 1;
 
 /** Tool names never pruned by default: orchestration/subagent-style tools and todo-like state. */
@@ -243,18 +210,7 @@ export interface PruneProtections {
 }
 
 export interface PruneThresholds {
-  /**
-   * Minimum raw text characters a fresh AUTOMATIC prune proposal must remove
-   * (original text chars minus placeholder chars) to be considered at all
-   * Enforced by `runDynamicContextPruningPipeline` BEFORE the
-   * net-benefit gate (see `PruneGateConfig`/`gate`, which uses break-even token
-   * math and is a separate, later check): text-only candidates below this floor
-   * are dropped up front and never reach gate/batch consideration. Image-bearing
-   * tool-result candidates bypass this raw-text floor (images receive no
-   * synthetic character count) and proceed to the token-based gate. Manual
-   * prunes and already-persisted/replayed decisions always bypass this floor,
-   * same as the gate. A value of 0 disables the floor (no filtering).
-   */
+  /** Minimum automatic-prune text saving; see README.md Configuration for bypass rules. */
   minCharsSaved: number;
 }
 
@@ -493,29 +449,8 @@ export function estimateTokensForContent(
 export type TokenEstimator = (text: string) => number;
 
 /**
- * Classify the agent's current state from the message payload alone:
- * this is the runtime-observable definition used both by the live `context`
- * event handler and (via re-export) by the offline benchmark, so the two can
- * never drift apart.
- *
- * Walk backwards from the end of `messages`, skipping any message that is
- * neither `user`, `assistant`, nor `toolResult` (e.g. system-role bookkeeping
- * entries). The first such message found determines the state:
- *  - `user` (a plain user turn message) => "idle": this call is the FIRST LLM
- *    call of a turn, i.e. the decision is being made from a turn-start state.
- *  - `assistant` or `toolResult` (mid-chain) => "mid_loop": the agent is
- *    iterating (already made at least one call this turn, or is reacting to a
- *    tool result).
- *  - No such message found (empty history, or system-only) => "idle" (safe
- *    default: nothing suggests an in-progress loop).
- *
- * The benchmark (scripts/benchmark.mjs) reclassifies its candidate labels to
- * this SAME definition, reusing this function directly, so the
- * defaults below rest on evidence measured with the exact definition the
- * runtime observes. (An earlier turn-end-based benchmark definition -- "idle"
- * meant "this IS the turn's last assistant message", only knowable in
- * hindsight -- has been retired: the runtime cannot know in advance whether a
- * call is a turn's last one, so that definition was never runtime-observable.)
+ * Shared runtime/benchmark state classifier; see README.md Configuration and
+ * docs/v2-design.md §4 for the turn-start definition and retired turn-end analysis.
  */
 export function classifyAgentStateFromMessages(messages: MinimalMessage[]): AgentState {
   for (let i = messages.length - 1; i >= 0; i--) {
@@ -677,7 +612,7 @@ export function isWithinRecencyWindow(index: number, boundaryIndex: number): boo
  * Used by the error-input purge strategy's "older than N turns" rule; a
  * message at the very end of the conversation has 0 turns elapsed.
  */
-export function computeTurnsElapsedSince(messages: MinimalMessage[], index: number): number {
+function computeTurnsElapsedSince(messages: MinimalMessage[], index: number): number {
   let count = 0;
   for (let i = index + 1; i < messages.length; i++) {
     if (messages[i].role === "user") count++;
@@ -974,7 +909,7 @@ export function resolvePruneTombstoneState(entries: MinimalSessionEntry[]): Prun
   return { activeDecisions, activeIdempotencyKeys, restoredKeys, lastDecisionByKey };
 }
 
-export function buildPlaceholderText(reason: string, charsRemoved: number): string {
+function buildPlaceholderText(reason: string, charsRemoved: number): string {
   return `[pruned by ${EXTENSION_ID}: ${reason} (${charsRemoved} chars removed)]`;
 }
 
@@ -1581,7 +1516,7 @@ export function collectCompletedToolCallOccurrences(
 }
 
 /** Placeholder text for a duplicate tool result whose content was replaced. */
-export function buildDedupePlaceholder(toolName: string, newestToolCallId: string): string {
+function buildDedupePlaceholder(toolName: string, newestToolCallId: string): string {
   return `[pruned by ${EXTENSION_ID}: duplicate ${toolName} call result; see newest occurrence (call ${newestToolCallId})]`;
 }
 
@@ -1642,46 +1577,6 @@ export const errorPurgeStrategy: PruneStrategy = {
   },
 };
 
-// When the same file path is read and/or written multiple times in a
-// session, older read/write/edit tool *outputs* for that path can become
-// stale relative to a later operation. This strategy proposes replacing
-// those older outputs with a placeholder, per the following CONSERVATIVE,
-// binding rules (do not loosen without an architect decision):
-//
-//   1. A full-file read (no offset/limit args) is superseded by a LATER
-//      full-file read of the same normalized path.
-//   2. A partial read (offset/limit args present) is NOT superseded by a
-//      later partial read UNLESS the later read's range fully covers the
-//      earlier one's range. If range coverage can't be reliably determined
-//      from the args (e.g. earlier read has no limit — reads to EOF — but
-//      the later read is bounded), skip: do not propose supersession.
-//      Full-file reads are represented as the range [1, EOF); comparing
-//      that representation via the same coverage check naturally satisfies
-//      rule 1 as a special case of rule 2 (see `readRangeCovers`).
-//   3. ANY read of a path (full or partial) is superseded once a LATER
-//      *successful* (non-error) write/edit to that same path occurs — the
-//      old read output is now known-stale regardless of read-vs-read range
-//      comparisons. The placeholder explicitly notes the file has since
-//      changed. An ERRORED later write/edit does NOT count (the file may be
-//      unchanged; conservative).
-//   4. Older write/edit outputs for a path are superseded by a LATER
-//      *successful* write/edit to that same path (again, a later errored
-//      write/edit never supersedes). Write/edit tool outputs are usually
-//      small, so proposals are only ever emitted when there is a non-empty
-//      result to prune; the shared net-benefit gate additionally decides
-//      whether the estimated savings are worth the cache-bust cost.
-//
-// The NEWEST operation for a given path (by message order) is never
-// superseded/pruned, regardless of kind — there is nothing later to point
-// the placeholder at.
-//
-// Path extraction is defensive: pi's built-in `read`/`write`/`edit` tools
-// all accept a `path` argument (see packages/coding-agent/src/core/tools/
-// {read,write,edit}.ts); some tool-call renderers also fall back to a
-// legacy `file_path` name, and `filePath` is accepted as a further
-// defensive fallback. `bash` is explicitly OUT of scope: this strategy does
-// not attempt to parse shell commands for file paths.
-
 const SUPERSEDED_FILE_OPS_STRATEGY_ID = "superseded-file-ops";
 
 type FileOpKind = "read" | "write" | "edit";
@@ -2234,7 +2129,7 @@ export function sessionEntriesToMessages(
   entries: MinimalSessionEntry[],
   projection?: MinimalSessionProjection,
 ): MinimalMessage[] {
-  if (projection) return sessionProjectionToMessages(projection);
+  if (projection) return projection.entries.flatMap((entry) => entry.messages);
   const messages: MinimalMessage[] = [];
   for (const entry of entries) {
     if (entry.type === "message" && entry.message) messages.push(entry.message);

package/extensions/dynamic-context-pruning/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-dynamic-context-pruning",
-  "version": "0.1.13",
+  "version": "0.1.14",
   "type": "module",
   "description": "A pi extension that dynamically prunes stale context to keep sessions lean.",
   "keywords": [

package/extensions/fast/README.md

@@ -2,7 +2,7 @@
 
 A pi extension that provides `/fast` and `/ultrafast` toggles for supported direct OpenAI API, OpenAI Codex, and Anthropic Claude models. It selects the provider-specific Fast-mode request shape from the currently selected model.
 
-For eligible direct OpenAI API models it injects:
+For eligible direct OpenAI API-key models it injects:
 
 ```json
 {
@@ -10,6 +10,14 @@ For eligible direct OpenAI API models it injects:
 }
 ```
 
+For eligible direct OpenAI ChatGPT OAuth models it injects:
+
+```json
+{
+  "service_tier": "priority"
+}
+```
+
 For eligible OpenAI Codex OAuth models it injects:
 
 ```json
@@ -39,16 +47,15 @@ Direct OpenAI API Fast mode requires all of the following:
 - Provider `openai`.
 - API `openai-responses` or `openai-completions`.
 - Model `gpt-6.1-sol`, `gpt-6-astra`, `gpt-6-sol`, or `gpt-6-luna`.
-- API-key auth or Pi's `openai` ChatGPT OAuth (`Sign in with ChatGPT`).
+- API-key auth sets `service_tier: "fast"`; Pi's `openai` ChatGPT OAuth sets `priority`.
 - No existing `service_tier` field in the request payload.
 
-The direct OpenAI route stays on OpenAI's direct API contract for both API-key
-auth and Pi's `openai` ChatGPT OAuth. Acceptance of `service_tier: "fast"` by
-Pi's `openai` ChatGPT OAuth sign-in endpoint has not been live-verified.
-This direct route is never forwarded to legacy `openai-codex` or its
-subscription-usage endpoint. The documented request shape is support evidence,
-not live certification; this change was validated only with mocked eligibility
-and request payloads, not live or paid provider calls.
+The direct OpenAI route stays on OpenAI's direct API contract but selects the tier by auth
+type. OAuth uses `priority` because its endpoint rejected `fast`; acceptance or application
+of `priority` on Pi's OAuth endpoint has not been live-verified. This direct route is never
+forwarded to legacy `openai-codex` or its subscription-usage endpoint. The request shape is
+support evidence, not certification; tests use mocked eligibility and payloads only, with
+no live or paid provider calls.
 
 ### Ultrafast support (initial, fail-closed)
 

package/extensions/fast/index.ts

@@ -155,7 +155,7 @@ function describeMode(state: SessionState): string {
   return state.config.enabled ? "on (config default)" : "off (config default)";
 }
 
-function getEligibility(ctx: ExtensionContext): Eligibility {
+function getEligibility(ctx: ExtensionContext, includeServiceTier = true): Eligibility {
   const model = ctx.model;
   if (!model) {
     return { eligible: false, modelKey: "no-model", reason: "no model is selected" };
@@ -237,7 +237,11 @@ function getEligibility(ctx: ExtensionContext): Eligibility {
       eligible: true,
       modelKey: key,
       provider: "openai",
-      serviceTier: OPENAI_API_FAST_SERVICE_TIER,
+      serviceTier: includeServiceTier
+        ? ctx.modelRegistry.isUsingOAuth(model)
+          ? OPENAI_FAST_SERVICE_TIER
+          : OPENAI_API_FAST_SERVICE_TIER
+        : undefined,
     };
   }
 
@@ -328,7 +332,7 @@ function updateStatus(ctx: ExtensionContext, state: SessionState): void {
     return;
   }
 
-  const eligibility = getEligibility(ctx);
+  const eligibility = getEligibility(ctx, false);
   ctx.ui.setStatus(EXTENSION_ID, isFastEnabled(state) && eligibility.eligible ? "fast" : undefined);
 }
 
@@ -346,7 +350,7 @@ function getStatusMessage(ctx: ExtensionContext, state: SessionState): string {
   }
 
   const enabled = isFastEnabled(state);
-  const eligibility = getEligibility(ctx);
+  const eligibility = getEligibility(ctx, enabled);
   const active = enabled && eligibility.eligible;
 
   if (active) {
@@ -405,7 +409,7 @@ async function injectAnthropicFastHeader(
   hasActiveTools: boolean,
 ): Promise<void> {
   if (state.ultrafast || !isFastEnabled(state)) return;
-  const eligibility = getEligibility(ctx);
+  const eligibility = getEligibility(ctx, false);
   if (!eligibility.eligible || eligibility.provider !== "anthropic") return;
 
   const matchingKeys = Object.keys(headers).filter(

package/extensions/fast/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@diegopetrucci/pi-fast",
-  "version": "0.1.9",
+  "version": "0.1.10",
   "description": "A pi extension that enables Fast mode for supported direct OpenAI API, OpenAI Codex, and Anthropic Claude models based on the currently selected model.",
   "keywords": [
     "pi-package",