pan-wizard 3.29.0 → 3.30.0

package/commands/pan/focus-auto.md

@@ -1,7 +1,7 @@
 ---
 name: focus-auto
 group: Focus
-description: Continuous scan-plan-exec loop with purpose-driven categories and 5-layer safety harness
+description: Continuous scan-plan-exec loop with purpose-driven categories and a layered safety harness
 allowed-tools:
   - Read
   - Write
@@ -107,7 +107,7 @@ Wait for the user's reply before proceeding. Do not guess or pick a default cate
 | `--stop` | — | Gracefully stop active run |
 | `--status` | — | Show current campaign progress |
 | `--dry-run` | — | Show plan without executing |
-| `--deep-review` | off | After every exec cycle, run inline OWASP security check on changed files. Verdict `block` or `review_required` stops the campaign (6th safety harness). Works with all categories. |
+| `--deep-review` | off | After every exec cycle, run inline OWASP security check on changed files. Verdict `block` or `review_required` stops the campaign (the security-gate layer of the safety harness). Works with all categories. |
 | `--parallel-research` | off | Fan out the per-item *research* stage via the Workflow tool (read-only agents). No-op fallback to sequential where the host has no Workflow tool. (ADR-0031) |
 | `--parallel-verify` | off | Fan out the per-item *verify* stage via the Workflow tool (read-only). The implement/exec stage always stays a single agent. (ADR-0031) |
 | `--clean-seal` | off | After the loop's last item, run one clean build + full verification (commands from `config.json → build`/`verification`) to catch cross-item orphans. (ADR-0031) |
@@ -185,7 +185,7 @@ Phase 2 (each cycle): Scan → Plan → Exec → Commit is strictly sequential w
 HARD STOP conditions:
 - Phase 1 fails (tests broken): Do not enter main loop — report and exit
 - Any cycle: test count drops below baseline after revert → stop campaign, preserve state
-- Context drops below 25%: stop campaign cleanly (safety harness 3)
+- Context drops below 25%: stop campaign cleanly (safety harness)
 </phase_dependencies>
 
 ### Phase 2: Main Loop
@@ -413,7 +413,7 @@ Then continue immediately to the next cycle (back to Step 2.1).
 
 4. Remove safety tag: `git tag -d focus-auto-baseline 2>/dev/null`
 
-## 6-Layer Safety Harness
+## Layered Safety Harness
 
 | Layer | Mechanism | Action |
 |-------|-----------|--------|
@@ -425,7 +425,7 @@ Then continue immediately to the next cycle (back to Step 2.1).
 | Zero-completed guard | 0 items done in a cycle | Stop — further cycles won't help |
 | Security gate (`--deep-review`) | Critical/high OWASP pattern in changed files | Revert last commit (critical) or flag for manual review (high), stop campaign |
 
-## 9 Behavioral Rules
+## Behavioral Rules
 
 1. **Read Before Write** — Read every file before editing. Understand context, callers, invariants.
 2. **Root Cause** — Fix the actual defect, not symptoms. Trace the code path.

package/hooks/dist/pan-cost-logger.js

@@ -182,7 +182,7 @@ function readActiveSessionMeta(cwd) {
 const COMMAND_TAIL_BYTES = 262144;
 
 // PAN's own command namespace, as a runtime writes it: `/pan:exec-phase` (Claude Code,
-// Gemini) or `/pan-exec-phase` (Codex, OpenCode, Copilot). Only these spawn PAN agents,
+// Gemini) or `/pan-exec-phase` (Codex, OpenCode, Copilot, and Claude Code after --unified-skills). Only these spawn PAN agents,
 // so only these are attributed — a host UI command (`/model`, `/compact`) and a plain
 // typed prompt leave the row honestly unattributed instead of borrowing a name.
 const PAN_COMMAND_RE = /<command-name>\s*\/?pan[:-]([a-z0-9][a-z0-9-]*)\s*<\/command-name>/i;

package/hooks/dist/pan-stop-guard.js

@@ -39,6 +39,19 @@
 // per stop chain, so a user who genuinely wants to stop is delayed by exactly
 // one continuation, never trapped.
 //
+// Gemini CLI (R29, 2026-09-23): the guard is registered there on AfterAgent,
+// Gemini's end-of-turn event — until this date PAN registered it under Claude's
+// `Stop` key, which Gemini skips with an "Invalid hook event name" warning, so it
+// had never run. AfterAgent honours the same {decision: 'block', reason} output
+// (a block re-prompts the agent with the reason), but its stop_hook_active is only
+// true for the AfterAgent that evaluates a continuation the block started
+// directly: a continuation that used tools reports false again (gemini-cli
+// client.ts / useGeminiStream.ts, read 2026-09-23). The flag alone would let the
+// guard block every such turn. So on AfterAgent the one-shot promise is kept with
+// a marker per session, project and target phase in the per-user 0700 hook
+// directory: a second stop aimed at the same phase is allowed. No safe marker
+// directory, or no session id, means no block (fail open, as everywhere else).
+//
 // Escape hatch: set workflow.stop_guard to false in .planning/config.json to
 // disable the guard entirely without turning off auto_advance.
 //
@@ -51,7 +64,9 @@
 // rather than only reachable via stdin (same pattern as the other PAN hooks).
 
 const fs = require('fs');
+const os = require('os');
 const path = require('path');
+const crypto = require('crypto');
 /**
  * Which planning tree this hook acts on.
  *
@@ -155,6 +170,63 @@ function readIfExists(p) {
   try { return fs.readFileSync(p, 'utf8'); } catch { return null; }
 }
 
+// Per-user hook state directory inside tmpdir, created 0700 — the same directory
+// and the same checks as bridgeDir() in pan-context-monitor.js (hooks are
+// standalone files and cannot require one another). Fail CLOSED (null) when the
+// directory is not provably ours: a shared host must not be able to pre-plant a
+// marker that silences the guard, or a symlink it writes through (M60).
+function hookStateDir() {
+  const uid = (typeof process.getuid === 'function' ? process.getuid() : process.env.USERNAME || 'win');
+  const dir = path.join(os.tmpdir(), `pan-hooks-${uid}`);
+  try {
+    fs.mkdirSync(dir, { recursive: true, mode: 0o700 });
+    const st = fs.lstatSync(dir);
+    if (st.isSymbolicLink()) return null;
+    // POSIX-only ownership and mode checks: Windows fakes mode bits (N15).
+    if (typeof process.getuid === 'function') {
+      if (st.uid !== process.getuid()) return null;
+      if ((st.mode & 0o077) !== 0) return null;
+    }
+    return dir;
+  } catch { return null; }
+}
+
+/**
+ * The file name of the one-shot marker for an AfterAgent block (Gemini CLI).
+ * Keyed by session, project and the block's reason — the reason names the target
+ * phase, so a later drop at a DIFFERENT phase in the same session is still caught.
+ * Pure; null when there is no session id to key on.
+ */
+function onceMarkerName(sessionId, projectDir, reason) {
+  if (typeof sessionId !== 'string' || !sessionId) return null;
+  const key = crypto.createHash('sha256').update(`${sessionId}\0${projectDir}\0${reason}`).digest('hex').slice(0, 32);
+  return `stop-guard-${key}.json`;
+}
+
+const MARKER_MAX_AGE_MS = 7 * 24 * 3600 * 1000;
+
+/**
+ * Apply the AfterAgent one-shot rule: returns true when this block may be issued
+ * (and records it), false when the same session already got it or no safe place
+ * to remember it exists. Best-effort pruning keeps the directory from growing.
+ */
+function claimOnceMarker(dir, name, now = Date.now()) {
+  if (!dir || !name) return false;
+  const marker = path.join(dir, name);
+  try {
+    for (const f of fs.readdirSync(dir)) {
+      if (!/^stop-guard-[0-9a-f]{32}\.json$/.test(f)) continue;
+      try { if (now - fs.statSync(path.join(dir, f)).mtimeMs > MARKER_MAX_AGE_MS) fs.unlinkSync(path.join(dir, f)); } catch { /* keep */ }
+    }
+  } catch { /* unreadable dir — the exclusive create below still decides */ }
+  try {
+    // 'wx' fails when the marker exists: the create IS the check, so two hook
+    // processes racing on the same stop cannot both block.
+    fs.writeFileSync(marker, JSON.stringify({ at: new Date(now).toISOString() }), { flag: 'wx', mode: 0o600 });
+    return true;
+  } catch { return false; }
+}
+
 function main() {
   let input = '';
   process.stdin.setEncoding('utf8');
@@ -171,13 +243,20 @@ function main() {
       let config = null;
       try { config = JSON.parse(fs.readFileSync(path.join(planningDir, 'config.json'), 'utf8')); } catch { /* no project / bad config -> allow */ }
 
-      const decision = buildStopDecision({
+      let decision = buildStopDecision({
         stopHookActive: payload.stop_hook_active === true,
         config,
         stateContent: readIfExists(path.join(planningDir, 'state.md')),
         roadmapContent: readIfExists(path.join(planningDir, 'roadmap.md')),
       });
 
+      // Gemini CLI's end-of-turn event: its stop_hook_active cannot carry the
+      // one-shot promise on its own (see the header), so a marker does.
+      if (decision && payload.hook_event_name === 'AfterAgent') {
+        const name = onceMarkerName(payload.session_id, projectDir, decision.reason);
+        if (!claimOnceMarker(hookStateDir(), name)) decision = null;
+      }
+
       if (decision) process.stdout.write(JSON.stringify(decision));
     } catch { /* fail open — never break a stop */ }
     process.exit(0);
@@ -188,4 +267,4 @@ if (require.main === module) {
   main();
 }
 
-module.exports = { buildStopDecision, UNTICKED_PHASE_RE };
+module.exports = { buildStopDecision, UNTICKED_PHASE_RE, onceMarkerName, claimOnceMarker };

package/hooks/dist/pan-trace-logger.js

@@ -175,7 +175,7 @@ function finalizeSession(cwd, sid) {
 const COMMAND_TAIL_BYTES = 262144;
 
 // PAN's own command namespace, as a runtime writes it: `/pan:exec-phase` (Claude Code,
-// Gemini) or `/pan-exec-phase` (Codex, OpenCode, Copilot). Only these spawn PAN agents,
+// Gemini) or `/pan-exec-phase` (Codex, OpenCode, Copilot, and Claude Code after --unified-skills). Only these spawn PAN agents,
 // so only these are attributed — a host UI command (`/model`, `/compact`) and a plain
 // typed prompt leave the row honestly unattributed instead of borrowing a name.
 const PAN_COMMAND_RE = /<command-name>\s*\/?pan[:-]([a-z0-9][a-z0-9-]*)\s*<\/command-name>/i;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pan-wizard",
-  "version": "3.29.0",
+  "version": "3.30.0",
   "description": "Command a bot army for your codebase: a reasoning-tier Mission Control delegates whole-project goals to specialist squads and ships behind a human merge gate. Five AI CLIs, zero context rot.",
   "bin": {
     "pan-wizard": "bin/install.js"

package/pan-wizard-core/bin/lib/agents-md.cjs

@@ -2,8 +2,9 @@
 
 // ─── AGENTS.md universal rules layer (ADR-0028 Phase 3) ─────────────────────
 //
-// AGENTS.md is the cross-runtime project-instructions standard; every PAN
-// target runtime (and Antigravity CLI) reads it natively. PAN contributes one
+// AGENTS.md is the cross-runtime project-instructions standard; Codex,
+// OpenCode and Copilot CLI read it natively, Claude Code only through the @AGENTS.md
+// import in CLAUDE.md; Gemini CLI reads GEMINI.md by default, not this file. PAN contributes one
 // marker-fenced section so agents in any runtime understand the PAN context
 // when reading the repo. User content outside the markers is never touched.
 //

package/pan-wizard-core/bin/lib/core.cjs

@@ -31,10 +31,23 @@ const { planningPath, planningRel } = require('./utils.cjs');
  * Each provider maps reasoning/mid/fast to its native model identifiers.
  * "inherit" means the host runtime uses its own top-tier model selection.
  */
+//
+// Every non-inherit value must be a model id the host accepts: a Claude Code alias
+// (sonnet, haiku) for Anthropic, otherwise a concrete API id that DEFAULT_RATES in
+// cost.cjs prices exactly (tests/documented-default-models.test.cjs pins both).
+// Until 2026-09-22 the OpenAI row held the literal strings 'mid' and 'fast', so a
+// budget-profile project on Codex or OpenCode was told to spawn a model called
+// "mid" (reality check R28).
+//   - openai: Codex's subagent docs (learn.chatgpt.com/docs/agent-configuration/
+//     subagents, read 2026-09-23): "start with gpt-6-sol. Use gpt-6-luna when you
+//     want a faster, lower-cost option". gpt-6-astra is the flagship above both.
+//   - google: the newest stable Flash and Flash-Lite on ai.google.dev's models page
+//     (read 2026-09-23). The 2.5 family these rows used is "not deprecated" but
+//     limited to users who have used it before, so a new project may not reach it.
 const PROVIDER_MODELS = {
   anthropic: { reasoning: 'inherit', mid: 'sonnet',                 fast: 'haiku' },
-  openai:    { reasoning: 'inherit', mid: 'mid',                    fast: 'fast'  },
-  google:    { reasoning: 'inherit', mid: 'gemini-2.5-flash',       fast: 'gemini-2.5-flash-lite' },
+  openai:    { reasoning: 'inherit', mid: 'gpt-6-sol',              fast: 'gpt-6-luna' },
+  google:    { reasoning: 'inherit', mid: 'gemini-3.8-flash',       fast: 'gemini-3.5-flash-lite' },
   default:   { reasoning: 'inherit', mid: 'sonnet',                 fast: 'haiku' },
 };
 
@@ -52,7 +65,7 @@ const COST_MULTIPLIERS = { reasoning: 15, mid: 3, fast: 1 };
 // isolation (each subagent runs in its own window), not a cheaper model, is what
 // keeps the main conversation clean. Cheapness is now OPT-IN: choose the `budget`
 // profile (the only column that still down-tiers) or pin a specific agent via
-// config `model_overrides`. The 3 security agents additionally pin `model: opus`
+// config `model_overrides`. The reviewer-class agents additionally pin `model: opus`
 // in their own frontmatter (a native, deliberate exception). resolve-model /
 // MODEL_PROFILES is advisory + cost-estimation; native Claude Code delegation
 // reads each agent file's static `model:` (unset → inherit).
@@ -781,7 +794,7 @@ function adjustTierForCapabilities(tier, opts) {
  * @param {string} agentType - Agent name (e.g., "pan-planner", "pan-executor")
  * @param {Object} [taskMetadata] - Optional metadata. Supports complexity fields and
  *   capability hints: {context_estimate, needs_thinking, cache_warm}.
- * @returns {string} Model identifier: "inherit", "sonnet", "haiku", "mid", "fast", etc.
+ * @returns {string} Model identifier: "inherit", a Claude Code alias ("sonnet", "haiku"), or a provider model id from PROVIDER_MODELS
  */
 function resolveModelInternal(cwd, agentType, taskMetadata) {
   const config = loadConfig(cwd);

package/pan-wizard-core/bin/lib/cost.cjs

@@ -30,8 +30,9 @@
  *   - hit rate: cache_read / (cache_read + input - cache_write) if any cache activity
  *
  * Rate table is approximate — real pricing comes from the provider's API.
- * Rates are US dollars per million tokens, indicative as of 2026-08. Users
- * can override with `.planning/config.json` → `cost.rates`.
+ * Rates are US dollars per million tokens, verified against each provider's
+ * pricing page on the date in RATES_VERIFIED_AT (below). Users can override with
+ * `.planning/config.json` → `cost.rates`.
  */
 
 const fs = require('fs');
@@ -47,63 +48,109 @@ const TOKENS_FILE = 'tokens.jsonl';
  * Override per-model in config.json → cost.rates.
  */
 const DEFAULT_RATES = {
-  // Anthropic — verified against platform pricing 2026-09-10. Opus 4.6+/Opus 5 are
-  // $5/$25 (the old $15/$75 Opus pricing ended with the 4.5 generation). Cache
-  // rates follow Anthropic's convention: read ≈ 0.1× input, write ≈ 1.25× input —
-  // EXCEPT Fable 5.1, whose cache reads bill at 0.025× input ($0.25). Fable 5.1
-  // needs its own row: without it the family-prefix fallback priced its reads at
-  // the Fable 5 rate, 4× too high on the model the `fable`/`best` aliases resolve to
-  // (model-config, read 2026-09-10: neither Fable model is any plan's default), and
-  // cached re-reads are the bulk of PAN's traffic (ADR-0044).
+  // Anthropic — platform.claude.com/docs/en/about-claude/pricing, every row read
+  // 2026-09-23. Opus 4.6+/Opus 5 are $5/$25 (the old $15/$75 Opus pricing ended with
+  // the 4.5 generation). cache_write is the 5-minute write rate (1.25× input; the
+  // 1-hour write bills 2×, which the ledger does not split yet). Cache reads follow
+  // the 0.1× convention EXCEPT on three models the page footnotes: Fable 5.1 and
+  // Mythos 5.1 at 0.025× input, Opus 5.5 at 0.05×. Each needs its own row, because
+  // the family-prefix fallback would otherwise price it at its predecessor's reads.
+  // Fable 5.1 is what the `fable`/`best` aliases resolve to (model-config: neither
+  // Fable model is any plan's default); cached re-reads are the bulk of PAN's
+  // traffic (ADR-0044).
   'claude-fable-5-1':   { input: 10.0, output: 50.0, cache_read: 0.25, cache_write: 12.5 },
   'claude-fable-5':     { input: 10.0, output: 50.0, cache_read: 1.0,  cache_write: 12.5 },
-  // Mythos 5.1 / Mythos 5 (limited availability) — platform.claude.com/docs/en/about-claude/pricing,
-  // read 2026-09-10: $10/$50; the page's cache footnote names Fable 5.1 AND Mythos 5.1 as
-  // the two models whose cache reads bill at 0.025× input; Mythos 5 follows the 0.1× rule.
-  // Added for reality check R7: resolveRate returned null for both ids.
+  // Mythos 5.1 / Mythos 5 (limited availability): $10/$50; Mythos 5.1 reads at
+  // 0.025× input like Fable 5.1, Mythos 5 follows the 0.1× rule (R7).
   'claude-mythos-5-1':  { input: 10.0, output: 50.0, cache_read: 0.25, cache_write: 12.5 },
   'claude-mythos-5':    { input: 10.0, output: 50.0, cache_read: 1.0,  cache_write: 12.5 },
+  // Opus 5.5 (claude-opus-5-5, released 2026-09-22): $4 input / $20 output, 5-minute
+  // writes $5, cache hits $0.20 — the page prices its hits at 0.05× input. Claude
+  // Code 2.1.280 made it the default model on every plan (model-config, read
+  // 2026-09-23), so PAN's inherit tier now runs on it. Without this row the family
+  // prefix priced it as Opus 5: cache reads 2.5× high, the rest a quarter high
+  // (reality check 2026-09-22, R25). The models overview lists no dated snapshot —
+  // from the 4.6 generation on, the dateless id is canonical.
+  'claude-opus-5-5':    { input: 4.0,  output: 20.0, cache_read: 0.20, cache_write: 5.0 },
   'claude-opus-5':      { input: 5.0,  output: 25.0, cache_read: 0.5,  cache_write: 6.25 },
   'claude-opus-4-8':    { input: 5.0,  output: 25.0, cache_read: 0.5,  cache_write: 6.25 },
   'claude-opus-4-7':    { input: 5.0,  output: 25.0, cache_read: 0.5,  cache_write: 6.25 },
   'claude-opus-4-6':    { input: 5.0,  output: 25.0, cache_read: 0.5,  cache_write: 6.25 },
-  // Opus 4.5 (dated id claude-opus-4-5-20251101) — same pricing page, read 2026-09-10:
-  // $5/$25/$0.50/$6.25. Without this row the dated id had no family prefix to land on
-  // and priced as null (R7).
+  // Opus 4.5 (dated id claude-opus-4-5-20251101): without this row the dated id had
+  // no family prefix to land on and priced as null (R7).
   'claude-opus-4-5':    { input: 5.0,  output: 25.0, cache_read: 0.5,  cache_write: 6.25 },
   // Sonnet 5 is $2/$10: the launch price announced as introductory through
-  // 2026-08-31 was made permanent and the scheduled rise to $3/$15 cancelled
-  // (pricing page, read 2026-09-10). Lesson: never write down a pre-announced
-  // price — this row carried the future rate for a month and over-billed by half.
+  // 2026-08-31 was made permanent and the scheduled rise to $3/$15 cancelled.
+  // Lesson: never write down a pre-announced price — this row carried the future
+  // rate for a month and over-billed by half.
   'claude-sonnet-5':    { input: 2.0,  output: 10.0, cache_read: 0.20, cache_write: 2.50 },
   'claude-sonnet-4-6':  { input: 3.0,  output: 15.0, cache_read: 0.3,  cache_write: 3.75 },
-  // Sonnet 4.5 (dated id claude-sonnet-4-5-20250929) — pricing page, read 2026-09-10:
-  // $3/$15/$0.30/$3.75 (the pre-Sonnet-5 rate; Sonnet 5 is $2/$10). R7.
+  // Sonnet 4.5 (dated id claude-sonnet-4-5-20250929): $3/$15/$0.30/$3.75 (R7).
   'claude-sonnet-4-5':  { input: 3.0,  output: 15.0, cache_read: 0.3,  cache_write: 3.75 },
+  // Haiku 4.5: API id claude-haiku-4-5-20251001, alias claude-haiku-4-5.
   'claude-haiku-4-5':   { input: 1.0,  output: 5.0,  cache_read: 0.1,  cache_write: 1.25 },
 
-  // OpenAI — verified against published pricing 2026-08. Prompt caching is a 90%
-  // input discount with no separate write charge, so cache_write bills at the
-  // plain input rate. GPT-5.6 ships in three tiers; the bare `gpt-5.6` id prices
-  // as the Sol flagship, with tier-specific keys for Terra/Luna (longest-prefix
-  // match wins in resolveRate). Luna reflects the 2026-07-30 price cut.
-  'gpt-5.6':            { input: 5.0,  output: 30.0, cache_read: 0.5,  cache_write: 5.0 },
-  'gpt-5.6-terra':      { input: 2.0,  output: 12.0, cache_read: 0.2,  cache_write: 2.0 },
-  'gpt-5.6-luna':       { input: 0.20, output: 1.20, cache_read: 0.02, cache_write: 0.20 },
+  // OpenAI — developers.openai.com/api/docs/pricing, Standard tier, short context
+  // (≤272K input tokens), read 2026-09-23. On the GPT-5.6 and GPT-6 families a cache
+  // write bills at 1.25× input IN PLACE of the input rate ("writes are not an
+  // additive fee"); GPT-5.5 charges no cache write ("-"), so its write tokens bill
+  // as plain input. Until 2026-09-22 this block said "no separate write charge" for
+  // every row, which had stopped being true — Terra and Luna writes were a fifth
+  // low (R27). Longest-prefix match lets each tier key win over a shorter one.
+  // GPT-6: Astra is the flagship; Sol and Luna were released 2026-09-22. Codex's
+  // config docs use gpt-6-sol as the default model and gpt-6-luna for lower-cost
+  // subagents (learn.chatgpt.com/docs/config-file/config-basic, /agent-configuration/subagents).
+  'gpt-6-astra':        { input: 10.0, output: 50.0, cache_read: 1.0,  cache_write: 12.5 },
+  'gpt-6-sol':          { input: 2.0,  output: 10.0, cache_read: 0.2,  cache_write: 2.5 },
+  'gpt-6-luna':         { input: 0.10, output: 0.50, cache_read: 0.01, cache_write: 0.125 },
+  // GPT-5.6. The bare `gpt-5.6` id is an alias that "routes requests to GPT-5.6 Sol"
+  // (its model page, read 2026-09-23), so it carries Sol's row — not the $5/$30 it
+  // carried before, which no page listed. Sol's price is promotional "at least
+  // through November 21, 2026": re-read it after that date.
+  'gpt-5.6':            { input: 4.0,  output: 20.0, cache_read: 0.4,  cache_write: 5.0 },
+  'gpt-5.6-sol':        { input: 4.0,  output: 20.0, cache_read: 0.4,  cache_write: 5.0 },
+  'gpt-5.6-terra':      { input: 2.0,  output: 12.0, cache_read: 0.2,  cache_write: 2.5 },
+  'gpt-5.6-luna':       { input: 0.20, output: 1.20, cache_read: 0.02, cache_write: 0.25 },
+  // Priced only in the page's separate "Cyber models" table, not in the Standard tiers.
+  'gpt-5.6-cyber':      { input: 12.5, output: 75.0, cache_read: 1.25, cache_write: 15.625 },
   'gpt-5.5':            { input: 5.0,  output: 30.0, cache_read: 0.5,  cache_write: 5.0 },
 
-  // Google Gemini — published rates (per million tokens, approximate; users can override via config.json → cost.rates).
-  // Pro tiers use the <=200K-context tier; long-context calls may be billed at ~2x. Cache rates are Google's context-cache pricing (~25% of input rate).
+  // Google Gemini — ai.google.dev/gemini-api/docs/pricing, paid tier, text input,
+  // the ≤200K-token prompt tier where the page splits by length, read 2026-09-23.
+  // cache_read is the page's "context caching" price: 0.1× input on every row, not
+  // the 0.25× this block carried until 2026-09-22 (cache reads 2.5× high, R26). The
+  // tokens that CREATE a cache bill at the standard input price (Google Cloud's
+  // context-cache overview; the Gemini API pages list only cached-token and storage
+  // charges), so cache_write = input. Cache STORAGE bills per million tokens per
+  // HOUR, which a per-token table cannot carry: PAN's Gemini totals exclude it.
+  // The 3.6/3.7/3.8 Flash prices run "through December 31, 2026" and double from
+  // January 1, 2027 (the page states both) — re-read them after that date.
+  'gemini-3.8-flash':       { input: 0.75, output: 3.75, cache_read: 0.075, cache_write: 0.75 },
+  'gemini-3.7-flash':       { input: 0.75, output: 3.75, cache_read: 0.075, cache_write: 0.75 },
+  'gemini-3.6-flash':       { input: 0.75, output: 3.75, cache_read: 0.075, cache_write: 0.75 },
+  'gemini-3.5-flash':       { input: 1.50, output: 9.00, cache_read: 0.15,  cache_write: 1.50 },
+  'gemini-3.5-flash-lite':  { input: 0.30, output: 2.50, cache_read: 0.03,  cache_write: 0.30 },
+  // Shutdown announced for 2027-05-07 (deprecations page); replacement gemini-3.5-flash-lite.
+  'gemini-3.1-flash-lite':  { input: 0.25, output: 1.50, cache_read: 0.025, cache_write: 0.25 },
+  'gemini-3-flash-preview': { input: 0.50, output: 3.00, cache_read: 0.05,  cache_write: 0.50 },
+  // The page lists 3.1 Pro only as gemini-3.1-pro-preview; the bare id keeps the
+  // preview's row so records written under it stay priced.
+  'gemini-3.1-pro':         { input: 2.00, output: 12.0, cache_read: 0.20,  cache_write: 2.00 },
+  'gemini-3.1-pro-preview': { input: 2.00, output: 12.0, cache_read: 0.20,  cache_write: 2.00 },
+  // The 2.5 family is "not deprecated", but the models page limits access to users
+  // who have used it before — new projects cannot count on it (PAN's Google routing
+  // tiers moved off it, R28). Rows kept for the ledgers that recorded it.
+  'gemini-2.5-pro':         { input: 1.25, output: 10.0, cache_read: 0.125, cache_write: 1.25 },
+  'gemini-2.5-flash':       { input: 0.30, output: 2.50, cache_read: 0.03,  cache_write: 0.30 },
+  'gemini-2.5-flash-lite':  { input: 0.10, output: 0.40, cache_read: 0.01,  cache_write: 0.10 },
   // (gemini-1.5-pro removed 2026-06: retired model; records for it fall back to tier rates.)
-  'gemini-3.1-pro':         { input: 2.00, output: 12.0, cache_read: 0.50,   cache_write: 2.00 },
-  'gemini-3.1-pro-preview': { input: 2.00, output: 12.0, cache_read: 0.50,   cache_write: 2.00 },
-  'gemini-2.5-pro':         { input: 1.25, output: 10.0, cache_read: 0.3125, cache_write: 1.25 },
-  'gemini-2.5-flash':       { input: 0.30, output: 2.50, cache_read: 0.075,  cache_write: 0.30 },
-  'gemini-2.5-flash-lite':  { input: 0.10, output: 0.40, cache_read: 0.025,  cache_write: 0.10 },
-
-  // Tier fallbacks when model id is unknown (reasoning tracks current Opus pricing)
-  'reasoning': { input: 5.0,  output: 25.0, cache_read: 0.5,  cache_write: 6.25 },
-  'mid':       { input: 3.0,  output: 15.0, cache_read: 0.3,  cache_write: 3.75 },
+
+  // Tier fallbacks when the model id is unknown: reasoning tracks the default Opus
+  // (Opus 5.5 since 2026-09-22 — the model the inherit tier runs on for most users),
+  // mid the Sonnet the `sonnet` alias resolves to on the Anthropic API (Sonnet 5),
+  // fast Haiku 4.5.
+  'reasoning': { input: 4.0,  output: 20.0, cache_read: 0.20, cache_write: 5.0 },
+  'mid':       { input: 2.0,  output: 10.0, cache_read: 0.20, cache_write: 2.50 },
   'fast':      { input: 1.0,  output: 5.0,  cache_read: 0.1,  cache_write: 1.25 },
 };
 
@@ -587,12 +634,22 @@ function cmdCostClear(cwd, raw) {
 
 // ─── Rate-table staleness ───────────────────────────────────────────────────
 
-// Date DEFAULT_RATES was last verified against published provider pricing.
-// Bump this whenever the table is re-verified; `models check` flags the table
-// once it is older than RATES_STALE_AFTER_DAYS (provider prices move faster
-// than PAN releases do).
-const RATES_VERIFIED_AT = '2026-09-10';
-const RATES_STALE_AFTER_DAYS = 180;
+// Date DEFAULT_RATES was last verified against published provider pricing — ALL
+// providers' pricing pages, not only Anthropic's (the 2026-09-10 bump covered the
+// Anthropic rows alone, and the Gemini and OpenAI rows it vouched for had never
+// matched their pages: reality check 2026-09-22, RC29). Bump it only when every
+// row has been read against its page that day.
+//
+// `models check` flags the table once it is older than RATES_STALE_AFTER_DAYS.
+// That window was half a year until 2026-09-22 (R38), which is longer than the
+// interval at which the lineup now moves: the default Claude model changed on
+// every plan inside twelve days of the last verification. Sixty days is the
+// review cadence the ecosystem reviews settled on. The calendar still cannot see
+// a default-model change inside the window — that is the job of the
+// documented-default-ids fixture test (tests/fixtures/documented-default-models.json,
+// R31), which fails the suite the day a documented default id has no exact row.
+const RATES_VERIFIED_AT = '2026-09-23';
+const RATES_STALE_AFTER_DAYS = 60;
 const RATE_TIERS = ['reasoning', 'mid', 'fast'];
 
 function checkRatesStaleness(now = new Date()) {

package/pan-wizard-core/bin/lib/memory-rebuild.cjs

@@ -10,8 +10,8 @@
  *
  * Three derived targets:
  *   1. AGENTS.md — the universal, cross-runtime tools memory. PAN owns exactly
- *      the marker-fenced `<!-- BEGIN/END PAN WIZARD -->` section (every runtime,
- *      including Copilot/.github, reads AGENTS.md natively). User content
+ *      the marker-fenced `<!-- BEGIN/END PAN WIZARD -->` section (Codex, Copilot and
+ *      OpenCode read AGENTS.md natively; Claude via the CLAUDE.md bridge; Gemini reads GEMINI.md by default). User content
  *      outside the markers is preserved byte-for-byte.
  *   2. CLAUDE.md — the Claude bridge (`@AGENTS.md` import), regenerated only
  *      when the Claude runtime is installed here.
@@ -126,7 +126,7 @@ function cmdMemoryRebuild(cwd, opts = {}, raw) {
   const targets = [];
   const warnings = [];
 
-  // 1. AGENTS.md — universal PAN section (all runtimes read it natively).
+  // 1. AGENTS.md — universal PAN section (Claude Code reads it through the bridge in step 2).
   {
     const p = path.join(cwd, 'AGENTS.md');
     const existing = safeReadFile(p);

package/pan-wizard-core/bin/lib/verify-deploy.cjs

@@ -107,7 +107,7 @@ function validateRuntimeInstall(cwd, configDir, runtime) {
         }
       }
     }
-    // Copilot/Gemini statusLine
+    // statusLine (Claude, Copilot; PAN writes none for Gemini since 2026-09-23)
     if (settings.statusLine && settings.statusLine.command) {
       hookCommands.push(settings.statusLine.command);
     }

package/pan-wizard-core/references/model-profiles.md

@@ -11,8 +11,8 @@ PAN uses three abstract tiers instead of hardcoded model names:
 | Tier | Purpose | Anthropic | OpenAI | Google |
 |------|---------|-----------|--------|--------|
 | `reasoning` | Architecture, planning, complex decisions | inherit (your session's top-tier model) | inherit | inherit |
-| `mid` | Execution, research, verification | Sonnet | mid | gemini-2.5-flash |
-| `fast` | Read-only extraction, budget tasks | Haiku | fast | gemini-2.5-flash-lite |
+| `mid` | Execution, research, verification | Sonnet | gpt-6-sol | gemini-3.8-flash |
+| `fast` | Read-only extraction, budget tasks | Haiku | gpt-6-luna | gemini-3.5-flash-lite |
 
 **Why `inherit` for reasoning?** Host runtimes map "opus" to a specific model version. PAN returns `inherit` for reasoning-tier agents, so they use whatever top-tier model the user has configured. This avoids version conflicts and silent fallbacks.
 
@@ -39,7 +39,7 @@ At install time PAN also runs a **best-effort, advisory** capability check on th
 | Class | Example model IDs | Role in PAN | Context | Relative cost | Notes |
 |-------|-------------------|-------------|---------|---------------|-------|
 | Fable / Mythos | `claude-fable-5` | **Recommended flagship** — deepest long-horizon reasoning; best for the bot army's Mission Control + planning | 1M | ~2× Opus | Runs input safety classifiers (see caveat below); requires 30-day data retention |
-| Opus | `claude-opus-5`, `claude-opus-4-8` | **Cost-conscious pick** — same 1M context + thinking, about half the cost, no cyber classifier | 1M | 1× | The safe pick when you want Opus behavior without Fable's refusal surface |
+| Opus | `claude-opus-5-5`, `claude-opus-5`, `claude-opus-4-8` | **Cost-conscious pick** — same 1M context + thinking, about half the cost, no cyber classifier | 1M | 1× | The safe pick when you want Opus behavior without Fable's refusal surface |
 
 **Why the Fable class is the recommended flagship.** It is Anthropic's deepest class for demanding, long-horizon agentic work — exactly what PAN's hierarchical bot army (Mission Control → squads → workers) asks of its reasoning tier. Select the current release in that class in your host runtime and `inherit` routes the reasoning-tier agents to it automatically.
 
@@ -117,7 +117,7 @@ PAN auto-detects the LLM provider to map tiers to the right model names:
 
 1. **Explicit config** — `routing.provider` in config.json (if not `"auto"`)
 2. **Environment variable** — `PAN_PROVIDER` env var
-3. **Runtime directory** — `.claude/` → Anthropic, `.codex/` → OpenAI, `.gemini/` → Google
+3. **Runtime directory** — `.claude/` → Anthropic, `.codex/` → OpenAI, `.gemini/` → Google, `.opencode/` → OpenAI, `.github/` → default (first match wins)
 4. **Fallback** — Default provider map (Anthropic-style names)
 
 ---

package/pan-wizard-core/references/planning-config.md

@@ -4,24 +4,22 @@ Configuration options for `.planning/` directory behavior.
 
 <config_schema>
 ```json
-"planning": {
-  "commit_docs": true,
-  "search_gitignored": false
-},
-"git": {
-  "branching_strategy": "none",
-  "phase_branch_template": "pan/phase-{phase}-{slug}",
-  "milestone_branch_template": "pan/{milestone}-{slug}"
-}
+"commit_docs": true,
+"search_gitignored": false,
+"branching_strategy": "none",
+"phase_branch_template": "pan/phase-{phase}-{slug}",
+"milestone_branch_template": "pan/{milestone}-{slug}"
 ```
 
 | Option | Default | Description |
 |--------|---------|-------------|
 | `commit_docs` | `true` | Whether to commit planning artifacts to git |
 | `search_gitignored` | `false` | Add `--no-ignore` to broad rg searches |
-| `git.branching_strategy` | `"none"` | Git branching approach: `"none"`, `"phase"`, or `"milestone"` |
-| `git.phase_branch_template` | `"pan/phase-{phase}-{slug}"` | Branch template for phase strategy |
-| `git.milestone_branch_template` | `"pan/{milestone}-{slug}"` | Branch template for milestone strategy |
+| `branching_strategy` | `"none"` | Git branching approach: `"none"`, `"phase"`, or `"milestone"` |
+| `phase_branch_template` | `"pan/phase-{phase}-{slug}"` | Branch template for phase strategy |
+| `milestone_branch_template` | `"pan/{milestone}-{slug}"` | Branch template for milestone strategy |
+
+These keys are top-level. The nested `planning.*` / `git.*` forms older versions wrote are read only when the top-level key is absent. A config `config-ensure-section` creates carries every top-level key, and one `/pan:new-project` writes carries a top-level `commit_docs` — so write the top-level key; a nested one is ignored whenever the top-level key exists.
 </config_schema>
 
 <commit_docs_behavior>
@@ -84,10 +82,8 @@ To use uncommitted mode:
 
 1. **Set config:**
    ```json
-   "planning": {
-     "commit_docs": false,
-     "search_gitignored": true
-   }
+   "commit_docs": false,
+   "search_gitignored": true
    ```
 
 2. **Add to .gitignore:**
@@ -101,7 +97,7 @@ To use uncommitted mode:
    git commit -m "chore: stop tracking planning docs"
    ```
 
-4. **Branch merges:** When using `branching_strategy: phase` or `milestone`, the `milestone-done` workflow automatically strips `.planning/` files from staging before merge commits when `commit_docs: false`.
+4. **Branch merges:** PAN never merges branches; when you merge a phase or milestone branch yourself with `commit_docs: false`, keep `.planning/` out of the merge commit.
 
 </setup_uncommitted_mode>
 
@@ -113,24 +109,24 @@ To use uncommitted mode:
 |----------|---------------------|--------------|-------------|
 | `none` | Never | N/A | N/A |
 | `phase` | At `execute-phase` start | Single phase | User merges after phase |
-| `milestone` | At first `execute-phase` of milestone | Entire milestone | At `milestone-done` |
+| `milestone` | At first `execute-phase` of milestone | Entire milestone | User merges after milestone |
 
-**When `git.branching_strategy: "none"` (default):**
+**When `branching_strategy: "none"` (default):**
 - All work commits to current branch
 - Standard PAN behavior
 
-**When `git.branching_strategy: "phase"`:**
+**When `branching_strategy: "phase"`:**
 - `execute-phase` creates/switches to a branch before execution
 - Branch name from `phase_branch_template` (e.g., `pan/phase-03-authentication`)
 - All plan commits go to that branch
 - User merges branches manually after phase completion
-- `milestone-done` offers to merge all phase branches
+- `milestone-done` does not merge — merge the phase branches yourself
 
-**When `git.branching_strategy: "milestone"`:**
+**When `branching_strategy: "milestone"`:**
 - First `execute-phase` of milestone creates the milestone branch
 - Branch name from `milestone_branch_template` (e.g., `pan/v1.0-mvp`)
 - All phases in milestone commit to same branch
-- `milestone-done` offers to merge milestone branch to main
+- `milestone-done` archives and tags but does not merge — merge the milestone branch yourself
 
 **Template variables:**
 

package/pan-wizard-core/workflows/settings.md

@@ -31,7 +31,7 @@ Parse current values (default to `true` if not present):
 - `workflow.nyquist_validation` — validation architecture research during plan-phase
 - `model_profile` — which model each agent uses (default: `balanced`)
 - `routing.strategy` — how model tiers are adjusted at runtime (default: `static`)
-- `git.branching_strategy` — branching approach (default: `"none"`)
+- `branching_strategy` — branching approach (default: `"none"`; a top-level key — a nested `git.branching_strategy` is ignored when the top-level key exists, as it does in every config `config-ensure-section` creates — write the top-level key)
 </step>
 
 <step name="present_settings">
@@ -131,9 +131,7 @@ Merge new settings into existing config.json:
     "auto_advance": true/false,
     "nyquist_validation": true/false
   },
-  "git": {
-    "branching_strategy": "none" | "phase" | "milestone"
-  },
+  "branching_strategy": "none" | "phase" | "milestone",
   "routing": {
     "strategy": "static" | "complexity"
   }