@pugi/cli 0.1.0-beta.2 → 0.1.0-beta.20

package/dist/core/repl/session.js

@@ -34,11 +34,15 @@ import { parseSlashCommand } from './slash-commands.js';
 import { webFetchTool } from '../../tools/web-fetch.js';
 import { loadSettings } from '../settings.js';
 import { getJobRegistry } from '../jobs/registry.js';
+import { applyCompactMask } from '../compact/buffer-rewriter.js';
+import { evaluateAutoCompact } from '../compact/auto-trigger.js';
+import { estimateTokensInMany } from '../compact/token-counter.js';
 import { extractAskTags, extractPlanReviewTags, signatureForAsk, } from './ask.js';
 import { existsSync, readdirSync, statSync } from 'node:fs';
 import { resolve as resolvePath } from 'node:path';
 import { CancellationToken } from './cancellation.js';
 import { DispatchFSM } from './dispatch-fsm.js';
+import { computeCostUsd, formatCostUsd, formatTokens } from './model-pricing.js';
 const MAX_TRANSCRIPT_ROWS = 500;
 const MAX_TOOL_CALLS = 200;
 const MAX_RECONNECT_ATTEMPTS = 10;
@@ -315,6 +319,19 @@ export class ReplSession {
             toolCalls: [],
             transcript: [],
             tokensDownstreamTotal: 0,
+            // α7 cost-meter sprint — cost accumulators land at zero on boot.
+            // `sessionStartedAtEpochMs` is set at construction time (vs the
+            // server-side `agent.session.opened` event) so the elapsed slot
+            // on the status row starts ticking the moment the REPL mounts.
+            sessionTokensIn: 0,
+            sessionTokensOut: 0,
+            sessionCostUsd: 0,
+            sessionStartedAtEpochMs: this.now(),
+            recentTurns: [],
+            turnTokensIn: 0,
+            turnTokensOut: 0,
+            turnCostUsd: 0,
+            lastTurnDelta: null,
             briefStartedAtEpochMs: undefined,
             pendingAsk: null,
             pendingAskSource: null,
@@ -322,6 +339,7 @@ export class ReplSession {
             pendingPlanReviewSource: null,
             dispatchState: 'idle',
             dispatchToolLabel: null,
+            lastCompletedOutcome: null,
         };
         // α6.9: mirror every FSM transition into the public state so the
         // status-bar surface can rerender on the next frame. Local listener
@@ -359,6 +377,7 @@ export class ReplSession {
                 apiUrl: this.options.apiUrl,
                 apiKey: this.options.apiKey,
                 workspace: this.options.workspace,
+                cyberZoo: this.options.cyberZoo,
             });
             this.patch({ sessionId, connection: 'connecting' });
             this.openStream();
@@ -577,6 +596,18 @@ export class ReplSession {
                 await this.dispatchStop(verdict.persona);
                 return verdict;
             }
+            case 'delegate': {
+                // α7.5 Phase 1: surface the dispatch intent inline. The actual
+                // wire shape (POST /api/pugi/sessions/:id/delegate) requires the
+                // SDK transport extension that ships alongside this PR; the
+                // REPL session module wires the call when the matching transport
+                // method lands (paired CLI follow-up). Today we surface the
+                // delegation intent in the transcript so the operator sees the
+                // verdict echo for muscle-memory before the round-trip lights up.
+                this.appendSystemLine(`delegate ${verdict.persona}: ${verdict.brief.length > 80 ? `${verdict.brief.slice(0, 77)}...` : verdict.brief}`);
+                this.appendSystemLine('Run `pugi delegate <slug> "<brief>"` from a fresh shell while the REPL transport wiring lands.');
+                return verdict;
+            }
             case 'dispatch': {
                 await this.dispatchBrief(verdict.brief);
                 return verdict;
@@ -602,11 +633,15 @@ export class ReplSession {
                 return verdict;
             }
             case 'cost': {
-                this.dispatchCost();
+                await this.dispatchCost();
+                return verdict;
+            }
+            case 'quota': {
+                await this.dispatchQuota();
                 return verdict;
             }
             case 'status': {
-                this.dispatchStatus();
+                await this.dispatchStatus();
                 return verdict;
             }
             case 'consensus': {
@@ -651,12 +686,205 @@ export class ReplSession {
                 await this.dispatchPrivacy();
                 return verdict;
             }
+            case 'init': {
+                // β1 Sl11 → β1a r1 (real inline scaffold, 2026-05-26): invoke
+                // `scaffoldPugiWorkspace` directly so the operator gets the
+                // same .pugi/ setup they would from `pugi init` on a fresh
+                // shell. Already-initialised workspaces (every artifact already
+                // present) get the "Already initialised" copy; partial / fresh
+                // workspaces get the full Created+Skipped breakdown. Default
+                // skills install is best-effort — any error from the bundled
+                // pack is surfaced as a system line and does not break the
+                // REPL session. The dynamic import keeps the slash dispatcher
+                // free of a runtime/cli.ts cycle on every keystroke.
+                try {
+                    const { scaffoldPugiWorkspace } = await import('../../runtime/cli.js');
+                    const lines = [];
+                    const result = await scaffoldPugiWorkspace({
+                        cwd: process.cwd(),
+                        // Slash callers default to the full default-skills pack so
+                        // the in-REPL experience matches `pugi init`. Operators who
+                        // want a minimal scaffold still have the shell command.
+                        noDefaults: false,
+                        log: (line) => {
+                            const trimmed = line.replace(/\n+$/u, '');
+                            if (trimmed.length > 0)
+                                lines.push(trimmed);
+                        },
+                    });
+                    if (result.alreadyInitialized) {
+                        this.appendSystemLine(`.pugi/ already initialised at ${result.root}. ${result.skipped.length} artefact(s) verified.`);
+                    }
+                    else {
+                        this.appendSystemLine(`Pugi initialised at ${result.root}. Created ${result.created.length} artefact(s), skipped ${result.skipped.length}.`);
+                    }
+                    if (result.defaultSkills.length > 0) {
+                        const installed = result.defaultSkills.filter((s) => s.status === 'installed').length;
+                        const skippedSkills = result.defaultSkills.filter((s) => s.status === 'skipped-existing').length;
+                        this.appendSystemLine(`Default skills: ${installed} installed, ${skippedSkills} already present.`);
+                    }
+                    for (const line of lines)
+                        this.appendSystemLine(line);
+                }
+                catch (error) {
+                    const message = error instanceof Error ? error.message : String(error);
+                    this.appendSystemLine(`/init failed: ${message}`);
+                }
+                return verdict;
+            }
+            case 'mcp': {
+                // β4 Sl7 (2026-05-26): /mcp [sub] [args...] forwards to the
+                // runtime command. We deliberately route through the same
+                // entry-point used by `pugi mcp` from a fresh shell so the
+                // surface stays single-sourced. `serve` is refused inline —
+                // booting an MCP server inside an active REPL would compete
+                // with the REPL itself for stdio, which is exactly the wrong
+                // thing to do.
+                if (verdict.args[0] === 'serve') {
+                    this.appendSystemLine('/mcp serve is not safe inside the REPL (it competes for stdio). ' +
+                        'Run `pugi mcp serve` from a fresh shell instead.');
+                    return verdict;
+                }
+                try {
+                    const { runMcpCommand } = await import('../../runtime/commands/mcp.js');
+                    const lines = [];
+                    await runMcpCommand(verdict.args, {
+                        workspaceRoot: process.cwd(),
+                        writeOutput: (_payload, text) => {
+                            const trimmed = text.replace(/\n+$/u, '');
+                            if (trimmed.length > 0)
+                                lines.push(trimmed);
+                        },
+                    });
+                    for (const line of lines)
+                        this.appendSystemLine(line);
+                    if (lines.length === 0) {
+                        this.appendSystemLine('/mcp: no output.');
+                    }
+                }
+                catch (error) {
+                    const message = error instanceof Error ? error.message : String(error);
+                    this.appendSystemLine(`/mcp failed: ${message}`);
+                }
+                return verdict;
+            }
+            case 'doctor': {
+                // L17 (2026-05-27): run the doctor probe sweep inline. We
+                // dynamic-import the runtime/commands/doctor module so the
+                // slash dispatcher does not pull the diagnostics graph
+                // (execFileSync + fs probes) into every keystroke. The
+                // module's output is captured into local lines so we can
+                // render it as system entries in the conversation pane;
+                // an Ink-rendered table inside the REPL frame is a follow-up.
+                try {
+                    const { runDoctorCommand, defaultHome } = await import('../../runtime/commands/doctor.js');
+                    const lines = [];
+                    await runDoctorCommand({
+                        cwd: process.cwd(),
+                        home: defaultHome(),
+                        env: process.env,
+                        json: false,
+                        writeOutput: (_payload, text) => {
+                            const trimmed = text.replace(/\n+$/u, '');
+                            if (trimmed.length > 0)
+                                lines.push(trimmed);
+                        },
+                    });
+                    for (const line of lines)
+                        this.appendSystemLine(line);
+                    if (lines.length === 0) {
+                        this.appendSystemLine('/doctor: no output.');
+                    }
+                }
+                catch (error) {
+                    const message = error instanceof Error ? error.message : String(error);
+                    this.appendSystemLine(`/doctor failed: ${message}`);
+                }
+                return verdict;
+            }
+            case 'permissions': {
+                // Leak L6: handle the `/permissions [mode] [--persist]` flow.
+                // The session module forwards to the runtime helper so the
+                // workspace + global-config writes share one code path with
+                // the CLI's top-level `--mode` resolution. The dynamic import
+                // keeps the dispatcher free of a session.ts -> runtime/cli.ts
+                // cycle.
+                try {
+                    const { runPermissionsCommand } = await import('../../runtime/commands/permissions.js');
+                    const lines = [];
+                    await runPermissionsCommand(verdict, {
+                        workspaceRoot: process.cwd(),
+                        writeOutput: (line) => {
+                            const trimmed = line.replace(/\n+$/u, '');
+                            if (trimmed.length > 0)
+                                lines.push(trimmed);
+                        },
+                    });
+                    for (const line of lines)
+                        this.appendSystemLine(line);
+                }
+                catch (error) {
+                    const message = error instanceof Error ? error.message : String(error);
+                    this.appendSystemLine(`/permissions failed: ${message}`);
+                }
+                return verdict;
+            }
+            case 'compact': {
+                // Leak L8 (2026-05-27): /compact summarises older turns and
+                // appends a boundary marker. We forward to the same runner the
+                // top-level `pugi compact` command uses so the surface stays
+                // single-sourced. The session module owns the in-memory
+                // transcript echo (system line + banner row) so the operator
+                // sees the marker land without a fresh REPL bootstrap.
+                await this.dispatchCompact('manual');
+                return verdict;
+            }
             case 'stub': {
                 this.appendSystemLine(verdict.message);
                 return verdict;
             }
         }
     }
+    /**
+     * Leak L8 (2026-05-27): drive the `/compact` flow from inside the
+     * REPL. Reuses the standalone runner so the wire shape + reason
+     * codes stay single-sourced. The result is echoed into the
+     * transcript as a system line; on success the operator sees the
+     * banner sentinel on next render.
+     *
+     * `trigger='manual'` for explicit `/compact` invocations;
+     * `trigger='auto'` for the threshold gate. The runner records the
+     * trigger in the marker payload so the banner can distinguish them.
+     */
+    async dispatchCompact(trigger) {
+        if (!this.store || !this.localSessionId) {
+            this.appendSystemLine('Local session store is disabled — /compact is unavailable.');
+            return;
+        }
+        try {
+            const { runCompactCommand } = await import('../../runtime/commands/compact.js');
+            const result = await runCompactCommand([], {
+                workspaceRoot: process.cwd(),
+                sessionId: this.localSessionId,
+                store: this.store,
+                trigger,
+                writeOutput: (_payload, text) => {
+                    if (text.length > 0)
+                        this.appendSystemLine(text);
+                },
+            });
+            if (result.status === 'compacted') {
+                // Echo a visible separator into the transcript so the operator
+                // immediately sees where the compaction landed. The Ink banner
+                // renders the row when the session reloads / resumes.
+                this.appendSystemLine(`─── context compacted (${result.turnsBefore} turns → 1 summary, ${trigger}) ───`);
+            }
+        }
+        catch (error) {
+            const message = error instanceof Error ? error.message : String(error);
+            this.appendSystemLine(`/compact failed: ${message}`);
+        }
+    }
     /**
      * In-REPL `/privacy` - alpha 6.13. Prints the full 3-mode contract
      * doc + the current mode banner inline. The current mode is fetched
@@ -938,22 +1166,195 @@ export class ReplSession {
             this.appendSystemLine(`/diff failed: ${this.errorMessage(error)}`);
         }
     }
-    dispatchCost() {
-        const { tokensDownstreamTotal, agents } = this.state;
+    async dispatchCost() {
+        // α7 cost-meter sprint — full breakdown matching the TUI status row
+        // footer. The session totals line mirrors the footer format
+        // (`↑ <in> ↓ <out> · $X.XX · <elapsed>`) so the operator scans the
+        // same numbers in two places. Per-turn list shows the last 5 turns
+        // oldest → newest; an empty list renders one system line so the
+        // operator knows the surface is wired (`No completed turns yet.`).
+        //
+        // L19 (2026-05-27) — after the in-memory recap, also render the
+        // persisted per-model table from `.pugi/cost.json`. That surface
+        // survives a REPL restart and answers the "what did I spend on
+        // claude-opus vs qwen this week?" question the in-memory recap can
+        // not. Errors loading the file collapse to a single warning line so
+        // the in-memory recap (the older, well-tested surface) is never
+        // gated behind a fresh dependency.
+        const { sessionTokensIn, sessionTokensOut, sessionCostUsd, sessionStartedAtEpochMs, recentTurns, agents, } = this.state;
         const active = agents.filter((a) => a.status === 'queued' || a.status === 'thinking').length;
-        const lineTokens = `Tokens this session: ${tokensDownstreamTotal.toLocaleString()} (in+out).`;
-        const lineAgents = `Active dispatches: ${active} of cap.`;
-        this.appendSystemLine(lineTokens);
-        this.appendSystemLine(lineAgents);
-        this.appendSystemLine('Full per-persona budget breakdown lands in α6.5.');
-    }
-    dispatchStatus() {
-        const sessionId = this.state.sessionId ?? '(unbound)';
-        const reach = this.state.connection;
-        this.appendSystemLine(`Backend: ${this.options.apiUrl} (${reach}).`);
-        this.appendSystemLine(`Session: ${sessionId}.`);
-        this.appendSystemLine(`Workspace: ${this.state.workspaceLabel}.`);
-        this.appendSystemLine(`CLI: pugi ${this.state.cliVersion}.`);
+        const elapsedMs = Math.max(0, this.now() - sessionStartedAtEpochMs);
+        const elapsedLabel = formatElapsedShort(elapsedMs);
+        this.appendSystemLine(`Session: ↑ ${formatTokens(sessionTokensIn)} ↓ ${formatTokens(sessionTokensOut)} · ${formatCostUsd(sessionCostUsd)} · ${elapsedLabel}`);
+        this.appendSystemLine(`Active dispatches: ${active} of cap.`);
+        if (recentTurns.length === 0) {
+            this.appendSystemLine('No completed turns yet — brief the workforce to charge the meter.');
+        }
+        else {
+            this.appendSystemLine(`Recent turns (last ${recentTurns.length}):`);
+            for (let i = 0; i < recentTurns.length; i += 1) {
+                const turn = recentTurns[i];
+                const idx = (i + 1).toString().padStart(2, ' ');
+                this.appendSystemLine(`  ${idx}. ↑ ${formatTokens(turn.tokensIn)} ↓ ${formatTokens(turn.tokensOut)} · ${formatCostUsd(turn.costUsd)}`);
+            }
+        }
+        // L19: append the persisted per-model table from .pugi/cost.json.
+        try {
+            const [{ createCostTracker }, { renderCostForSlash }] = await Promise.all([
+                import('../cost/tracker.js'),
+                import('../../runtime/commands/cost.js'),
+            ]);
+            const workspaceRoot = this.options.workspace?.workspaceCwd ?? process.cwd();
+            const sessionId = this.state.sessionId ?? 'no-session';
+            const tracker = createCostTracker({
+                workspaceRoot,
+                sessionIdProvider: () => sessionId,
+                now: () => this.now(),
+            });
+            const current = tracker.current();
+            if (current && Object.keys(current.models).length > 0) {
+                this.appendSystemLine('');
+                const { lines } = renderCostForSlash({
+                    tracker,
+                    allSessions: false,
+                    windowDays: 30,
+                    now: () => this.now(),
+                });
+                for (const line of lines)
+                    this.appendSystemLine(line);
+            }
+        }
+        catch {
+            // best-effort — the persisted view is additive; failure never
+            // breaks the in-memory recap above
+        }
+    }
+    /**
+     * α7 cost-meter sprint — `/quota` slash handler. Fetches the live
+     * `/api/pugi/usage` snapshot and renders three lines: plan tier,
+     * monthly window, and per-counter `used/cap (pct%)`. Failure modes
+     * (offline, unauth, older admin-api) collapse to a single one-line
+     * `Could not fetch quota…` system message so the surface never throws
+     * from a keystroke handler.
+     *
+     * The fetch is best-effort with a 4s timeout — mirrors the `whoami`
+     * pattern in `runtime/cli.ts` so the operator gets the same UX on the
+     * REPL slash and the CLI command.
+     */
+    async dispatchQuota() {
+        const controller = new AbortController();
+        const timer = setTimeout(() => controller.abort(), 4000);
+        try {
+            const url = `${this.options.apiUrl.replace(/\/+$/, '')}/api/pugi/usage`;
+            const res = await fetch(url, {
+                method: 'GET',
+                headers: {
+                    authorization: `Bearer ${this.options.apiKey}`,
+                    accept: 'application/json',
+                },
+                signal: controller.signal,
+            });
+            if (!res.ok) {
+                this.appendSystemLine(`Could not fetch quota: HTTP ${res.status}.`);
+                return;
+            }
+            const body = (await res.json());
+            const tier = typeof body.tier === 'string' ? body.tier : '(unknown)';
+            const tierLabel = QUOTA_TIER_LABELS[tier] ?? tier;
+            const month = typeof body.billingMonth === 'string' ? body.billingMonth : '(unknown month)';
+            const resetAt = typeof body.resetAt === 'string' ? body.resetAt : null;
+            const resetLine = resetAt ? ` · resets ${formatResetWindow(resetAt, this.now())}` : '';
+            this.appendSystemLine(`Plan: ${tierLabel} · ${month}${resetLine}`);
+            const used = body.used ?? {};
+            const caps = body.quotas ?? {};
+            const counters = [
+                ['sync', used.sync, caps.sync],
+                ['review', used.review, caps.review],
+                ['engine', used.engine, caps.engine],
+            ];
+            for (const [name, value, cap] of counters) {
+                const v = typeof value === 'number' ? value : 0;
+                if (cap === null || cap === undefined) {
+                    this.appendSystemLine(`  ${name.padEnd(7, ' ')} ${v.toLocaleString()} / unlimited`);
+                }
+                else {
+                    const pct = cap > 0 ? Math.round((v / cap) * 100) : 0;
+                    this.appendSystemLine(`  ${name.padEnd(7, ' ')} ${v.toLocaleString()} / ${cap.toLocaleString()} (${pct}%)`);
+                }
+            }
+        }
+        catch (error) {
+            const msg = error instanceof Error ? error.message : String(error);
+            this.appendSystemLine(`Could not fetch quota: ${msg}.`);
+        }
+        finally {
+            clearTimeout(timer);
+        }
+    }
+    /**
+     * In-REPL `/status` — Leak L34 (2026-05-27). Surfaces the full
+     * session snapshot (id + age, cwd, permission mode, CLI version,
+     * tokens, dispatches, last cmd, compact boundaries, auth identity,
+     * connection) by delegating к the same `runStatusCommand` the
+     * top-level `pugi status` shell uses. Live REPL state (session
+     * id, token totals, last operator command) flows in through the
+     * context so the slash variant shows MORE than the shell path.
+     *
+     * The renderer routes к the system pane via `appendSystemLine`
+     * so the snapshot lands as a single contiguous block в the
+     * conversation transcript. Migrating к the Ink `<StatusTable>`
+     * mounted directly в the REPL frame is a follow-up sprint —
+     * keeping the line-buffered path here avoids cycling the
+     * conversation pane's render model mid-α7.
+     */
+    async dispatchStatus() {
+        try {
+            const { runStatusCommand, defaultStatusHome } = await import('../../runtime/commands/status.js');
+            // Find the most-recent operator transcript row + its timestamp
+            // so the snapshot's `Last cmd` field has real content в REPL
+            // mode. Walking от newest end is O(transcript) worst case but
+            // bounded by MAX_TRANSCRIPT_ROWS so this stays cheap.
+            let lastCommand = null;
+            let lastCommandAtEpochMs = null;
+            for (let i = this.state.transcript.length - 1; i >= 0; i -= 1) {
+                const row = this.state.transcript[i];
+                if (row.source === 'operator') {
+                    lastCommand = row.text;
+                    lastCommandAtEpochMs = row.timestampEpochMs;
+                    break;
+                }
+            }
+            const liveTokens = this.state.sessionTokensIn + this.state.sessionTokensOut;
+            const lines = [];
+            await runStatusCommand({
+                cwd: process.cwd(),
+                home: defaultStatusHome(),
+                env: process.env,
+                json: false,
+                liveSessionId: this.state.sessionId ?? null,
+                sessionStartedAtEpochMs: this.state.sessionStartedAtEpochMs,
+                liveTokensUsed: liveTokens >= 0 ? liveTokens : 0,
+                lastCommand,
+                lastCommandAtEpochMs,
+                writeOutput: (_payload, text) => {
+                    for (const line of text.split('\n')) {
+                        const trimmed = line.replace(/\s+$/u, '');
+                        if (trimmed.length > 0)
+                            lines.push(trimmed);
+                    }
+                },
+            });
+            if (lines.length === 0) {
+                this.appendSystemLine('/status: no output.');
+                return;
+            }
+            for (const line of lines)
+                this.appendSystemLine(line);
+        }
+        catch (error) {
+            const message = error instanceof Error ? error.message : String(error);
+            this.appendSystemLine(`/status failed: ${message}`);
+        }
     }
     /**
      * α6.5 `/context` slash handler. Surfaces the three-tier context
@@ -1136,7 +1537,10 @@ export class ReplSession {
             this.appendSystemLine(capLine);
         }
         this.appendOperatorLine(brief);
-        this.patch({ briefStartedAtEpochMs: this.now() });
+        // Reset `lastCompletedOutcome` so a fresh dispatch does not
+        // inherit the prior turn's status-bar label (e.g. a stale
+        // "replied" sticking around while the next dispatch is in flight).
+        this.patch({ briefStartedAtEpochMs: this.now(), lastCompletedOutcome: null });
         // α6.9 + R3 P1 (Codex triple-review 2026-05-25): supersede the
         // prior dispatch when one is in flight. Steps in order:
         //
@@ -1463,6 +1867,7 @@ export class ReplSession {
                 apiUrl: this.options.apiUrl,
                 apiKey: this.options.apiKey,
                 workspace: this.options.workspace,
+                cyberZoo: this.options.cyberZoo,
             });
             this.patch({ sessionId, connection: 'connecting' });
             this.openStream();
@@ -1619,8 +2024,22 @@ export class ReplSession {
             }
             case 'agent.tokens': {
                 const delta = event.tokensIn + event.tokensOut;
+                // α7 cost-meter sprint — bind a client-side USD figure to this
+                // frame. The model slug rides on the event (optional for back-
+                // compat); the price ladder in `model-pricing.ts` falls back to
+                // a Sonnet-tier rate when the slug is missing, so the meter is
+                // always populated. Negative / NaN values are clamped to zero
+                // inside `computeCostUsd` so a buggy upstream never credits the
+                // meter.
+                const deltaCostUsd = computeCostUsd(event.tokensIn, event.tokensOut, event.model);
                 this.patch({
                     tokensDownstreamTotal: this.state.tokensDownstreamTotal + delta,
+                    sessionTokensIn: this.state.sessionTokensIn + event.tokensIn,
+                    sessionTokensOut: this.state.sessionTokensOut + event.tokensOut,
+                    sessionCostUsd: this.state.sessionCostUsd + deltaCostUsd,
+                    turnTokensIn: this.state.turnTokensIn + event.tokensIn,
+                    turnTokensOut: this.state.turnTokensOut + event.tokensOut,
+                    turnCostUsd: this.state.turnCostUsd + deltaCostUsd,
                     agents: this.state.agents.map((a) => a.taskId === event.taskId
                         ? {
                             ...a,
@@ -1640,10 +2059,36 @@ export class ReplSession {
                 }
                 this.askBuffer.delete(event.taskId);
                 this.askBufferPending.delete(event.taskId);
+                // Honour the work-done signal from admin-api.
+                // `outcome === 'replied'` means the turn was a pure text reply
+                // with no delegate XML and no tool call — render it as
+                // "replied" so the operator can tell the difference between
+                // "the orchestrator just talked" and "real work shipped".
+                // Older servers omit the field; default to 'shipped' so the
+                // existing wire stays back-compat.
+                const completedStatus = event.outcome === 'replied' ? 'replied' : 'shipped';
                 this.patch({
                     agents: this.state.agents.map((a) => a.taskId === event.taskId
-                        ? { ...a, status: 'shipped', detail: 'shipped' }
+                        ? { ...a, status: completedStatus, detail: completedStatus }
                         : a),
+                    // Mirror the outcome to top-level state so the status-bar
+                    // can render `replied` instead of the legacy `shipped`
+                    // label when the FSM lands in `completed`. Without this
+                    // the bottom-bar would still say "shipped" while the
+                    // agent-tree said "replied", restoring the same
+                    // contradiction this PR is fixing (Codex triple-review P2).
+                    //
+                    // r2: gate on the same stale-dispatch check that
+                    // advanceFsmOnDispatchEnd applies. If this completion
+                    // belongs to a SUPERSEDED dispatch (a newer dispatchBrief
+                    // already bumped dispatchSeq before this late terminal
+                    // arrived), don't let the status-bar label flip to the
+                    // stale outcome — the current turn is the live one.
+                    // The agent-tree row patch above is still safe because
+                    // it only updates the row keyed by taskId.
+                    ...(this.isStaleTaskEvent(event.taskId)
+                        ? {}
+                        : { lastCompletedOutcome: completedStatus }),
                 });
                 // α6.9: transition the FSM to `completed` when no other
                 // dispatch is still in flight. The check uses the agents list
@@ -1651,6 +2096,12 @@ export class ReplSession {
                 // the dispatch alive; the FSM only goes terminal when the last
                 // agent ships.
                 this.advanceFsmOnDispatchEnd('completed', 'agent_completed', event.taskId);
+                // α7 cost-meter sprint — flush the per-turn delta when the
+                // LAST agent settles. Decoupled from the FSM gate so a test
+                // fixture (or a single-agent dispatch that never reached
+                // `awaiting_response` — happens on instant SSE replay) still
+                // gets the row written into recentTurns + lastTurnDelta.
+                this.maybeFlushTurnOnAgentSettle(event.taskId);
                 if (target) {
                     // If the persona actually produced a reply via incremental
                     // agent.step events, render that reply in the transcript so
@@ -1716,6 +2167,10 @@ export class ReplSession {
                 // operator sees the bottom-bar settle back to `idle` after the
                 // last block clears.
                 this.advanceFsmOnDispatchEnd('completed', 'agent_blocked', event.taskId);
+                // α7 cost-meter sprint — flush the per-turn delta (blocked
+                // still counts as a billable turn — the operator paid for the
+                // tokens that landed before the refusal).
+                this.maybeFlushTurnOnAgentSettle(event.taskId);
                 return;
             }
             case 'agent.failed': {
@@ -1739,6 +2194,10 @@ export class ReplSession {
                 // `completed` so the bottom-bar surface tracks the dispatch
                 // collectively.
                 this.advanceFsmOnDispatchEnd('failed', 'agent_failed', event.taskId);
+                // α7 cost-meter sprint — flush the per-turn delta when the
+                // dispatch fails (the operator still paid for whatever tokens
+                // landed before the failure).
+                this.maybeFlushTurnOnAgentSettle(event.taskId);
                 return;
             }
         }
@@ -1777,13 +2236,25 @@ export class ReplSession {
      * after a manual `cancel()` finds the FSM already in `aborted` and
      * is silently dropped.
      */
+    /**
+     * 2026-05-26 — shared stale-task check used by both the FSM advance
+     * gate AND the status-bar `lastCompletedOutcome` mirror. Lifts the
+     * R2 dispatchSeq compare out of `advanceFsmOnDispatchEnd` so other
+     * agent.completed-handler side-effects (status-bar label, future
+     * metric counters) can apply the same guard without duplicating it.
+     * Returns true iff the task's stamped dispatchSeq is older than the
+     * current dispatchSeq — i.e. a newer dispatchBrief() superseded it
+     * and the late terminal event must not corrupt live-turn state.
+     */
+    isStaleTaskEvent(taskId) {
+        const taskSeq = this.taskDispatchSeq.get(taskId);
+        return taskSeq !== undefined && taskSeq < this.dispatchSeq;
+    }
     advanceFsmOnDispatchEnd(outcome, reason, taskId) {
         // R2 P1 fix (Codex triple-review 2026-05-25): a terminal event
         // for a SUPERSEDED dispatch must NOT advance the live FSM or null
-        // the live token. If the event carries a taskId and the stamped
-        // dispatchSeq for that task is older than the current dispatchSeq,
-        // the event belongs to a prior dispatch that was replaced by a
-        // newer `dispatchBrief()`. Silently drop the FSM advance.
+        // the live token. Delegates to isStaleTaskEvent so the agent.completed
+        // status-bar mirror in the handler above uses the same gate.
         if (taskId !== undefined) {
             const taskSeq = this.taskDispatchSeq.get(taskId);
             if (taskSeq !== undefined && taskSeq < this.dispatchSeq) {
@@ -1815,6 +2286,63 @@ export class ReplSession {
         this.currentDispatchToken = null;
         this.patch({ briefStartedAtEpochMs: undefined });
     }
+    /**
+     * α7 cost-meter sprint — gate the per-turn flush on "this was the
+     * LAST in-flight agent". Mirrors the `stillActive` guard inside
+     * `advanceFsmOnDispatchEnd` so a multi-agent dispatch only emits a
+     * single recentTurns row + a single lastTurnDelta flash.
+     *
+     * Idempotent: if no tokens have been billed this turn, the inner
+     * `flushTurnAccumulator` short-circuits without pushing an empty row.
+     */
+    maybeFlushTurnOnAgentSettle(taskId) {
+        const stillActive = this.state.agents.some((a) => a.status === 'queued' || a.status === 'thinking');
+        if (stillActive)
+            return;
+        this.flushTurnAccumulator(taskId);
+    }
+    /**
+     * α7 cost-meter sprint — flush the per-turn accumulator into
+     * `recentTurns` + `lastTurnDelta`. Idempotent + safe to call from any
+     * terminal-state branch (`agent.completed` / `agent.blocked` /
+     * `agent.failed`). When no tokens have been billed this turn
+     * (instant abort, cap-warning gate), the helper short-circuits
+     * without pushing an empty row.
+     */
+    flushTurnAccumulator(taskId) {
+        const turnTokensIn = this.state.turnTokensIn;
+        const turnTokensOut = this.state.turnTokensOut;
+        const turnCostUsd = this.state.turnCostUsd;
+        if (turnTokensIn === 0 && turnTokensOut === 0) {
+            // Idempotent zero-flush — never push an empty row into recentTurns.
+            return;
+        }
+        const turnId = taskId !== undefined ? taskId : `turn-${this.dispatchSeq}-${this.now()}`;
+        const newTurn = {
+            id: turnId,
+            tokensIn: turnTokensIn,
+            tokensOut: turnTokensOut,
+            costUsd: turnCostUsd,
+            completedAt: new Date(this.now()).toISOString(),
+        };
+        // Keep the buffer capped at 5 entries (oldest first). The push
+        // order matches the surface contract: `/cost` paginates oldest →
+        // newest so the operator scans top-down chronologically.
+        const recent = [...this.state.recentTurns, newTurn];
+        const trimmed = recent.length > 5 ? recent.slice(-5) : recent;
+        this.patch({
+            recentTurns: trimmed,
+            lastTurnDelta: {
+                tokensIn: turnTokensIn,
+                tokensOut: turnTokensOut,
+                costUsd: turnCostUsd,
+                completedAtEpochMs: this.now(),
+            },
+            turnTokensIn: 0,
+            turnTokensOut: 0,
+            turnCostUsd: 0,
+        });
+    }
     /* ------------- transcript helpers -------------- */
     /**
      * Look up the persona slug for a running task. Used by the tool call
@@ -1890,6 +2418,62 @@ export class ReplSession {
         //   persona  -> 'persona'
         //   system   -> 'system'
         this.persistRow(row);
+        // Leak L8 (2026-05-27): evaluate the auto-compact gate after
+        // every appendRow that produces a transcript turn. Wrapped in a
+        // setImmediate so the gate never blocks the input-handling fast
+        // path; if the threshold is tripped, the auto-trigger dispatches
+        // `/compact` in the background while the operator keeps typing.
+        if (row.source === 'operator' || row.source === 'persona') {
+            this.maybeAutoCompact();
+        }
+    }
+    /**
+     * Auto-compact gate. Cheap: builds an in-memory token estimate from
+     * the current transcript and consults `evaluateAutoCompact`. When the
+     * gate fires AND a compaction is not already in flight, we dispatch
+     * `/compact` with `trigger='auto'`. The fire-and-forget shape means
+     * the input box stays responsive while the background round-trip
+     * runs.
+     *
+     * Hysteresis: `compactionInFlight` blocks re-entry. The gate is
+     * cleared when the dispatch promise resolves regardless of outcome
+     * so a transient transport failure does not permanently disable the
+     * auto-trigger.
+     */
+    compactionInFlight = false;
+    maybeAutoCompact() {
+        if (this.compactionInFlight)
+            return;
+        if (!this.store || !this.localSessionId)
+            return;
+        if (process.env['PUGI_AUTOCOMPACT_DISABLED'] === '1')
+            return;
+        // Token estimate from the in-memory transcript. The estimate is a
+        // lower bound on actual context pressure (server-side system
+        // prompts add overhead) but the 4-char/token heuristic plus the
+        // 0.75 default threshold gives generous headroom.
+        const texts = this.state.transcript.map((r) => r.text);
+        const tokenCount = estimateTokensInMany(texts);
+        // Conservative default: assume the smallest commonly-used window
+        // (32k tokens for deepseek-v3.1). Resolving the live model slug
+        // through DispatchFSM + admin-api adds latency on a hot path; the
+        // 0.75 threshold + smallest-window assumption errs toward
+        // EARLY trigger which is the safe direction.
+        const verdict = evaluateAutoCompact({
+            tokenCount,
+            windowSize: 32_000,
+        });
+        if (verdict.kind !== 'fire')
+            return;
+        this.compactionInFlight = true;
+        void (async () => {
+            try {
+                await this.dispatchCompact('auto');
+            }
+            finally {
+                this.compactionInFlight = false;
+            }
+        })();
     }
     /**
      * Best-effort write of one transcript row into the local
@@ -1940,8 +2524,14 @@ export class ReplSession {
      * write the restored events.
      */
     restoreTranscript(events) {
+        // Leak L8 (2026-05-27): apply compact-boundary masking BEFORE the
+        // row conversion. Events strictly before the latest marker are
+        // condensed into the boundary's `keptTailTurns + marker` slice so
+        // the post-resume transcript starts at the most-recent context
+        // floor rather than re-playing the full pre-compaction history.
+        const masked = applyCompactMask(events);
         const rows = [];
-        for (const event of events) {
+        for (const event of masked) {
             const row = eventToTranscriptRow(event);
             if (row)
                 rows.push(row);
@@ -2129,6 +2719,25 @@ function eventToTranscriptRow(event) {
             timestampEpochMs: event.t,
         };
     }
+    if (event.kind === 'compaction') {
+        // Leak L8: render the marker as a system separator line on
+        // replay. The full summary text is intentionally NOT inlined here
+        // (a 2k-token summary in the transcript would defeat the purpose
+        // of compacting); the operator sees the "context compacted"
+        // banner and can run `/context` to inspect the marker payload
+        // when they want the details.
+        const compactionPayload = (event.payload ?? null);
+        const trigger = compactionPayload?.trigger === 'auto' ? 'auto' : 'manual';
+        const turns = typeof compactionPayload?.summaryTurnsBefore === 'number'
+            ? compactionPayload.summaryTurnsBefore
+            : 0;
+        return {
+            id: randomUUID(),
+            source: 'system',
+            text: `─── context compacted (${turns} turns → 1 summary, ${trigger}) ───`,
+            timestampEpochMs: event.t,
+        };
+    }
     return null;
 }
 /**
@@ -2197,6 +2806,62 @@ function formatAgeSeconds(deltaMs) {
 export function knownRoles() {
     return listRoles();
 }
+/**
+ * α7 cost-meter sprint — render a session-elapsed ms delta as the
+ * status-row's compact `XmYs` / `XhYm` shape. Distinct from
+ * `formatAgeSeconds` above because `/cost` needs minute-granularity
+ * uniformly (operator wants `2m44s`, not `2m`). Pure / branch-cheap;
+ * the TUI status row + `/cost` both call this on every render.
+ */
+function formatElapsedShort(elapsedMs) {
+    if (!Number.isFinite(elapsedMs) || elapsedMs <= 0)
+        return '0s';
+    const totalSec = Math.floor(elapsedMs / 1000);
+    if (totalSec < 60)
+        return `${totalSec}s`;
+    const min = Math.floor(totalSec / 60);
+    const sec = totalSec % 60;
+    if (min < 60)
+        return `${min}m${sec.toString().padStart(2, '0')}s`;
+    const hr = Math.floor(min / 60);
+    const restMin = min % 60;
+    return `${hr}h${restMin.toString().padStart(2, '0')}m`;
+}
+/**
+ * α7 cost-meter sprint — public-facing tier labels for the `/quota`
+ * slash. Mirrors `TIER_PRICE_LABEL` in `runtime/cli.ts` (kept in sync
+ * via `pricing.spec.ts` gate). Falls through to the raw slug when an
+ * unknown tier ships from a forward-compat admin-api build.
+ */
+const QUOTA_TIER_LABELS = Object.freeze({
+    free: 'Free',
+    founder: 'Founder ($20/mo)',
+    builder: 'Builder ($99/mo)',
+    team: 'Team ($199/mo)',
+});
+/**
+ * α7 cost-meter sprint — render the time-until-reset window for the
+ * `/quota` plan line. `resetAt` is the ISO string admin-api returns;
+ * `now` is the current epoch ms (injected for test determinism). Falls
+ * back to the raw ISO string when parsing fails so the operator never
+ * sees an empty hint.
+ */
+function formatResetWindow(resetAtIso, nowEpochMs) {
+    const resetMs = Date.parse(resetAtIso);
+    if (!Number.isFinite(resetMs))
+        return resetAtIso;
+    const deltaMs = resetMs - nowEpochMs;
+    if (deltaMs <= 0)
+        return 'now';
+    const days = Math.floor(deltaMs / (24 * 60 * 60 * 1000));
+    if (days >= 2)
+        return `in ${days}d`;
+    const hours = Math.floor(deltaMs / (60 * 60 * 1000));
+    if (hours >= 1)
+        return `in ${hours}h`;
+    const minutes = Math.max(1, Math.floor(deltaMs / (60 * 1000)));
+    return `in ${minutes}m`;
+}
 /* ------------------------------------------------------------------ */
 /* Tool call synthesiser - α6.12                                      */
 /* ------------------------------------------------------------------ */
@@ -2230,7 +2895,7 @@ export function synthesiseToolCall(input) {
     // Pattern: ToolName(args) optionally suffixed with a result hint.
     // We allow the canonical Claude Code casing AND the snake_case
     // alias `web_fetch` so the synthesiser matches what personas write.
-    const match = /^(Read|Edit|Bash|Grep|Glob|WebFetch|web_fetch)\s*\(\s*([^)]*)\s*\)\s*(.*)$/i
+    const match = /^(Read|Write|Edit|Bash|Grep|Glob|WebFetch|web_fetch)\s*\(\s*([^)]*)\s*\)\s*(.*)$/i
         .exec(detail);
     if (!match)
         return null;
@@ -2254,6 +2919,8 @@ function normaliseToolName(raw) {
         return 'web_fetch';
     if (lower === 'read')
         return 'read';
+    if (lower === 'write')
+        return 'write';
     if (lower === 'edit')
         return 'edit';
     if (lower === 'bash')
@@ -2479,7 +3146,22 @@ export function stripPersonaPrefixEcho(personaSlug, text) {
     // Escape regex specials in the display name even though THE_TEN
     // names are alpha-only today (forward-defense).
     const escaped = display.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+    // Match `<DisplayName>` (case-insensitive) followed by EITHER:
+    //   - an end-of-string, OR
+    //   - a separator (whitespace / comma / colon / dash / period+space).
+    // The `i` flag is needed so a model writing "PUGI:" or "pugi," still
+    // strips. After this match the post-fix `noSepUppercaseRe` handles
+    // the "PugiПринял" / "PugiHello" no-separator emission pattern
+    // (CEO red-alert 2026-05-27) using a SEPARATE regex without the `i`
+    // flag so the lookahead is case-strict (Pugineous must NOT strip).
     const re = new RegExp(`^${escaped}(?:[\\s,:;\\-—–]+|$)`, 'i');
+    // No-separator case-strict matcher. Display name in either of its
+    // canonical casings ("Pugi" / "PUGI") immediately followed by an
+    // uppercase Cyrillic or Latin letter. The strip is intentionally
+    // narrower than the case-insensitive `re` above because a lowercase
+    // continuation ("Pugineous") is a single word, not a display-name
+    // echo - we must not eat real content.
+    const noSepUppercaseRe = new RegExp(`^(?:${escaped}|${escaped.toUpperCase()})(?=[А-ЯЁA-Z])`);
     // Loop the strip so cascading echoes ("Pugi Pugi Pugi, координатор ...")
     // collapse to a single name. The model occasionally emits the display
     // name two or three times back-to-back when the pane prefix also
@@ -2491,10 +3173,18 @@ export function stripPersonaPrefixEcho(personaSlug, text) {
     // matches an empty string (defence-in-depth even though the current
     // pattern guarantees at least one consumed char).
     for (let i = 0; i < 3; i += 1) {
-        const m = re.exec(working);
-        if (!m || m[0].length === 0)
-            break;
-        working = working.slice(m[0].length).trimStart();
+        let m = re.exec(working);
+        if (m && m[0].length > 0) {
+            working = working.slice(m[0].length).trimStart();
+            continue;
+        }
+        // Fallback: no-separator match for "PugiПринял" / "PugiHello" shape.
+        m = noSepUppercaseRe.exec(working);
+        if (m && m[0].length > 0) {
+            working = working.slice(m[0].length);
+            continue;
+        }
+        break;
     }
     return working;
 }