fix(pi-dev): port pi-watch to pi 1.x, fail on provider errors, refresh alias ladders

BUGS.md

@@ -108,3 +108,11 @@ entries' IDs; manual edits to fix typos are fine.
 - **Severity**: medium
 - **Proposed fix**: Have fold carry notes into the requirement (e.g. reviewNote) or print them on stderr so Claude must act on them; and either let the page set certainty directly or document that notes are advisory. Omit stage-1 items from the stage-2 record, or mark them as pinned.
 
+## BUG-12: pi-watch dispatch prints nothing when the provider request fails
+- **Observed**: 2026-10-15
+- **File**: skills/pi-dev/scripts/pi-watch/pi-watch.mjs:442
+- **Description**: With an expired anthropic OAuth token, pi-watch --alias sonnet --tools read --require-trailer STATUS wrote empty stdout and exited 6 (missing trailer). Without --require-trailer it exits 0 with empty output. Raw pi -p --provider anthropic --model claude-sonnet-4-6 prints 'OAuth refresh failed for anthropic: ... invalid_grant'. session.prompt() does not throw on a provider error, so pi-watch's catch never runs; the failure sits on the final assistant message (stopReason/errorMessage), which pi-watch never inspects. Callers such as adversarial-review read the empty reply as a reviewer result. This is the dispatch side of the expired-token case in BUG-9. Seen while verifying the 1.x SDK port on zpyoung/fix-pi-watch; BUG-9 shows the same empty-reply symptom on the pre-port SDK.
+- **Severity**: high
+- **Proposed fix**: After session.prompt(), check the last assistant message's stopReason; on error, write errorMessage to stderr and exit non-zero.
+- **Resolved**: 2026-10-15. After session.prompt(), pi-watch reads the last assistant message the way pi's print mode does: stopReason error or aborted exits 1 with errorMessage on stderr, and length exits 1 as truncated. The check runs before --require-trailer, so a provider failure no longer surfaces as exit 6. BUG-9 (--check passing an expired token) stays open.
+

DEFERRED.md

@@ -140,3 +140,20 @@ Reviewed every sprint planning. Use `/quirk:artifacts:defer` to append.
 - **Estimated effort**: M
 - **Priority**: P2
 - **Proposed owner**: unassigned
+- **Resolved**: 2026-10-15. pi-watch.mjs now builds a ModelRuntime and passes it to createAgentSession. package.json pins ^1.2.0, not ^1.0.0, because pi-coding-agent 1.0.0 and 1.1.0 fail on a fresh install: their caret range on pi-codemode resolves to 1.2.0, which dropped an export they import. The launcher reinstalls deps older than that floor, and a committed pnpm-workspace.yaml declines the two dependency build scripts so pnpm 12's install doesn't exit 1.
+
+## DEFER-17: Refresh pi-watch alias ladders for the pi 1.x model catalog
+- **Deferred**: 2026-10-15
+- **Session context**: pi-watch 1.x SDK port on zpyoung/fix-pi-watch
+- **Why deferred**: pi 1.1.0 --list-models lists anthropic/claude-sonnet-5-5, anthropic/claude-haiku-5-5 and openai/gpt-6.1-sol, which no ladder in skills/pi-dev/scripts/pi-watch/pi-watch.mjs ALIASES includes, so sonnet/haiku/codex resolve to older models. The launcher's --where staleness probe greps for gpt-5.6 and needs the same update. Left out of the 1.x SDK port because a ladder refresh is a separate model-choice decision.
+- **Estimated effort**: S
+- **Priority**: P2
+- **Resolved**: 2026-10-15. The ladders now lead with GPT-6.1/6 Sol, GPT-6 Luna, and Claude 5.5. A new opt-in `fable` alias covers Claude Fable, and grok routes to grok-4.7/4.6 via xai, then Copilot. GPT-6 Astra is excluded as pro-priced, and terra stays on GPT-5.6 Terra because GPT-6 has no Terra tier. Entries the pinned pi-ai 1.2.0 catalog doesn't ship were removed. The --where probe now looks for gpt-6.1-sol, and adversarial-review's ALIAS_LADDER heads match.
+
+## DEFER-18: Replace the per-provider model dumps in pi-dev's models.md
+- **Deferred**: 2026-10-15
+- **Session context**: DEFER-17 alias ladder refresh on zpyoung/fix-pi-watch
+- **Why deferred**: skills/pi-dev/reference/models.md has a hand-copied pi --list-models table for each provider (openai-codex, openai, anthropic, google, github-copilot). It's a pre-1.x snapshot that lacks GPT-6, Claude 5.x and Grok 4.x and still lists grok-code-fast-1. Either regenerate it from the pi-ai catalog or replace it with a pointer to pi --list-models <pattern>. Left out of the DEFER-17 ladder refresh because it's a doc restructure, not a ladder change.
+- **Estimated effort**: S
+- **Priority**: P3
+

skills/adversarial-review/scripts/adversarial-review

@@ -851,11 +851,11 @@ def run_prepass(
 # within each family; the ladder is walked in this order once the author's family
 # is filtered out.
 ALIAS_LADDER = (
-    ("codex", "openai", "openai-codex/gpt-5.6-sol", "high"),
-    ("gemini", "google", "google/gemini-3.2-pro-preview", "high"),
+    ("codex", "openai", "openai-codex/gpt-6.1-sol", "high"),
+    ("gemini", "google", "google/gemini-3.1-pro-preview", "high"),
     ("terra", "openai", "openai-codex/gpt-5.6-terra", "high"),
-    ("opus", "anthropic", "anthropic/claude-opus-4-7", "high"),
-    ("sonnet", "anthropic", "anthropic/claude-sonnet-4-7", "high"),
+    ("opus", "anthropic", "anthropic/claude-opus-5-5", "high"),
+    ("sonnet", "anthropic", "anthropic/claude-sonnet-5-5", "high"),
     ("flash", "google", "google/gemini-flash-latest", "medium"),
 )
 

skills/pi-dev/SKILL.md

@@ -42,21 +42,22 @@ The ✓/✗ report prints to **stderr** (like pi-watch's other progress output)
 
 | Alias | Default thinking | Routes to |
 |---|---|---|
-| `codex` (default coding) | `medium` | gpt-5.6-sol → 5.5 → 5.4 → 5.3-codex (across `openai-codex`/`openai`/`github-copilot`) |
-| `codex-max` | `max` | gpt-5.6-sol → 5.5-codex-max → 5.4-codex-max → 5.1-codex-max |
-| `codex-mini` | `medium` | gpt-5.6-luna → 5.4-mini → 5.1-codex-mini |
-| `codex-spark` | `high` | gpt-5.4-codex-spark → 5.3-codex-spark |
+| `codex` (default coding) | `medium` | gpt-6.1-sol → 6-sol → 5.6-sol → 5.5 (across `openai-codex`/`openai`/`github-copilot`) |
+| `codex-max` | `max` | gpt-6.1-sol → 6-sol → 5.6-sol, at `max` effort |
+| `codex-mini` | `medium` | gpt-6-luna → 5.6-luna |
+| `codex-spark` | `high` | gpt-5.3-codex-spark |
 | `terra` (balanced) | `high` | gpt-5.6-terra → 5.4 |
-| `sonnet` | `high` | claude-sonnet-4-7 → 4-6 (anthropic + copilot) |
-| `opus` | `high` | claude-opus-4-7 → 4-6 |
-| `haiku` | `medium` | claude-haiku-4-6 → 4-5 |
-| `gemini` | `high` | gemini-3.2-pro-preview → 3.1 → 3 |
-| `flash` | `medium` | gemini-flash-latest → 3-flash-preview |
-| `grok` | `medium` | github-copilot/grok-code-fast-1 |
+| `sonnet` | `high` | claude-sonnet-5-5 → 5 → 4-6 (anthropic + copilot) |
+| `opus` | `high` | claude-opus-5-5 → 5 → 4-8 → 4-7 |
+| `haiku` | `medium` | claude-haiku-5-5 → 4-5 |
+| `fable` (opt-in) | `high` | claude-fable-5-1 → 5 |
+| `gemini` | `high` | gemini-3.1-pro-preview |
+| `flash` | `medium` | gemini-flash-latest → 3.8-flash → 3-flash-preview |
+| `grok` | `medium` | grok-4.7 → 4.6 (xai, then copilot) |
 
-GPT-5.6 replaced the suffix ladder with three capability tiers: **Sol** (flagship, $5/$30, the only tier that unlocks `max` reasoning effort), **Terra** (balanced, $2.50/$15), **Luna** (fast/cheapest, $1/$6). Sol takes over `codex`/`codex-max`, Luna takes over `codex-mini`, and Terra gets its own `terra` alias.
+GPT-6 ships three tiers: **Sol** (flagship, $2/$10) backs `codex` and `codex-max`, **Luna** (fastest and cheapest, $0.10/$0.50) backs `codex-mini`, and **Astra** ($10/$50) is excluded like a pro tier. GPT-6 has no Terra tier, so `terra` stays on GPT-5.6 Terra ($2/$12), which now costs more than GPT-6 Sol. Claude **Fable** ($10/$50, 2.5× Opus 5.5) gets the opt-in `fable` alias and never appears in the `opus` ladder.
 
-Aliases auto-upgrade as new models ship — newest entries at the top of each ladder are speculative and skipped if not yet in `pi --list-models`. **OpenAI pro / deep-research tiers are deliberately excluded** from alias ladders: `gpt-5-pro`, `gpt-5.x-pro`, `o1-pro`, `o3-pro`, `o3-deep-research`, `o4-mini-deep-research` (~10–30× cost, multi-minute latency). GPT-5.6's **Sol Pro** and **Ultra mode** are the same lever — they're request-time params (`reasoning.mode`), not separate model IDs, so there's no alias for them. Dispatch pro-tier only on explicit user request via `--provider`/`--model`. *(Note: `gemini-*-pro-preview` is Google's standard tier, not a pro tier — it routes via the `gemini` alias.)*
+Aliases auto-upgrade as new models ship — newest entries at the top of each ladder are speculative and skipped if not yet in `pi --list-models`. **OpenAI pro / deep-research tiers are deliberately excluded** from alias ladders: `gpt-5-pro`, `gpt-5.x-pro`, `o1-pro`, `o3-pro`, `o3-deep-research`, `o4-mini-deep-research` (~10–30× cost, multi-minute latency), and `gpt-6-astra` (5× GPT-6 Sol). GPT-5.6's **Sol Pro** and **Ultra mode** are the same lever — they're request-time params (`reasoning.mode`), not separate model IDs, so there's no alias for them. Dispatch pro-tier only on explicit user request via `--provider`/`--model`. *(Note: `gemini-*-pro-preview` is Google's standard tier, not a pro tier — it routes via the `gemini` alias.)*
 
 To see the full ladders: `pi-watch --list-aliases`.
 
@@ -120,7 +121,7 @@ QUIRK="$(node -e 'const fs=require("fs"),os=require("os"),p=require("path");cons
 install -m755 "$QUIRK/skills/pi-dev/scripts/pi-watch/pi-watch" ~/.local/bin/pi-watch  # or any PATH dir
 ```
 
-Verify: `pi --version` → expect ≥ 0.65.1. `pi /login` for any subscription provider you want (Claude, ChatGPT, Copilot). Then `pi-watch --check` to confirm which aliases resolve to an authed model — the first run also auto-installs the current version's Node deps if they're missing.
+Verify: `pi --version` → expect ≥ 0.65.1, and `node --version` → expect ≥ 22.19 (the pi SDK's floor). `pi /login` for any subscription provider you want (Claude, ChatGPT, Copilot). Then `pi-watch --check` to confirm which aliases resolve to an authed model — the first run also auto-installs the current version's Node deps if they're missing or outdated.
 
 **Updating:** `claude plugin update quirk` is enough — `pi-watch` auto-follows. Run `pi-watch --where` to see which version/dir it resolves to (and whether that copy has the current alias ladder). Re-run the `install` line above only if the launcher script itself changes (rare); mjs/alias updates are picked up automatically.