fix: harden browser mode stability

bin/oracle-cli.ts

@@ -13,6 +13,7 @@ if (process.argv[2] === "oracle-mcp") {
 }
 import { resolveEngine, type EngineMode, defaultWaitPreference } from "../src/cli/engine.js";
 import { shouldRequirePrompt } from "../src/cli/promptRequirement.js";
+import { resolveDashPrompt } from "../src/cli/stdin.js";
 import chalk from "chalk";
 import type { SessionMetadata, SessionMode, BrowserSessionConfig } from "../src/sessionStore.js";
 import { sessionStore, pruneOldSessions } from "../src/sessionStore.js";
@@ -216,7 +217,7 @@ program.hook("preAction", () => {
   introPrinted = true;
 });
 applyHelpStyling(program, VERSION, isTty);
-program.hook("preAction", (thisCommand) => {
+program.hook("preAction", async (thisCommand) => {
   if (thisCommand !== program) {
     return;
   }
@@ -234,6 +235,11 @@ program.hook("preAction", (thisCommand) => {
     opts.prompt = positional;
     thisCommand.setOptionValue("prompt", positional);
   }
+  const resolvedPrompt = await resolveDashPrompt(opts.prompt);
+  if (resolvedPrompt !== opts.prompt) {
+    opts.prompt = resolvedPrompt;
+    thisCommand.setOptionValue("prompt", resolvedPrompt);
+  }
   if (shouldRequirePrompt(userCliArgs, opts)) {
     console.log(
       chalk.yellow('Prompt is required. Provide it via --prompt "<text>" or positional [prompt].'),

docs/browser-mode.md

@@ -75,7 +75,7 @@ You can pass the same payload inline (`--browser-inline-cookies '<json or base64
 - `--browser-inline-files`: alias for `--browser-attachments never` (forces inline paste; never uploads attachments).
 - `--browser-bundle-files`: bundle all resolved attachments into a single temp file before uploading (only used when uploads are enabled/selected).
 - sqlite bindings: automatic rebuilds now require `ORACLE_ALLOW_SQLITE_REBUILD=1`. Without it, the CLI logs instructions instead of running `pnpm rebuild` on your behalf.
-- `--model`: the same flag used for API runs is accepted, but the ChatGPT automation path supports GPT-5.4 and GPT-5.2 variants. Use `gpt-5.4-pro`, `gpt-5.4`, `gpt-5.2`, `gpt-5.2-thinking`, `gpt-5.2-instant`, or `gpt-5.2-pro`. Legacy Pro aliases still resolve to the latest Pro picker target.
+- `--model`: the same flag used for API runs is accepted, but the ChatGPT automation path supports GPT-5.4 and GPT-5.2 variants. Use `gpt-5.4-pro`, `gpt-5.4`, `gpt-5.2`, `gpt-5.2-thinking`, `gpt-5.2-instant`, or `gpt-5.2-pro`. Any ChatGPT Pro alias resolves to the current Pro picker target, so versioned Pro labels may briefly appear in the UI but settle on the single available Pro entry.
 - Cookie sync is mandatory—if we can’t copy cookies from Chrome, the run exits early. Use the hidden `--browser-allow-cookie-errors` flag only when you’re intentionally running logged out (it skips the early exit but still warns).
 - Experimental cookie controls (hidden flags/env):
   - `--browser-cookie-names <comma-list>` or `ORACLE_BROWSER_COOKIE_NAMES`: allowlist which cookies to sync. Useful for “only NextAuth/Cloudflare, drop the rest.”

docs/debug/remote-chrome.md

@@ -27,7 +27,7 @@ Outcome:
 - After wiring fix, logs showed “Routing browser automation to remote host …” but requests failed with:
   - `ECONNREFUSED 192.168.64.2:49810` when no service listening.
   - `busy` when a previous service process was still bound.
-- Later run reached remote path but failed model switch: `Unable to find model option matching "GPT-5.2 Pro"` (remote Chrome not logged into ChatGPT / model picker mismatch).
+- Later run reached remote path but failed model switch: `Unable to find model option matching "GPT-5.4 Pro"` (remote Chrome not logged into ChatGPT / model picker mismatch).
 - After disabling cookie shipping and requiring host login, remote runs now fail earlier: service logs “Loading ChatGPT cookies from host Chrome profile…” then reports `Unhandled promise rejection ... Unknown error` when `loadChromeCookies` runs on the VM. Remote client sees `socket hang up` because the server doesn’t deliver a result.
 
 ### 2) Remote service on VM
@@ -52,7 +52,7 @@ Actions taken on VM (tmux `vmssh`):
 
 - Environment PATH: bun not on PATH for non-interactive shells caused `./runner` to fail; need to `export PATH="$HOME/.bun/bin:$PATH"` before starting service.
 - Port collisions: prior listeners on 49810 caused ECONNREFUSED/busy.
-- Remote model switch failed: remote Chrome likely not signed into ChatGPT; model picker couldn’t find “GPT-5.2 Pro”.
+- Remote model switch failed: remote Chrome likely not signed into ChatGPT; model picker couldn’t find the current Pro picker target (“GPT-5.4 Pro” at the time).
 - Keychain/cookie read now failing on VM: `loadChromeCookies` throws “Unknown error” when invoked from the server process (Node 25, SSH shell). When `oracle serve` runs from GUI Terminal it starts fine; under nohup/SSH it logs the rejection and remote runs hang.
 - New behavior (post-fix): `oracle serve` exits early if it cannot load host ChatGPT cookies after opening chatgpt.com for login; sign in on the host and restart the service.
 

docs/manual-tests.md

@@ -141,23 +141,37 @@ Document results (pass/fail, session IDs) in PR descriptions so reviewers can au
 
 Run these four smoke tests whenever we touch browser automation:
 
-1. **GPT-5.2 simple prompt**  
-   `pnpm run oracle -- --engine browser --model "GPT-5.2" --prompt "Give me two short markdown bullet points about tables"`  
-   Expect two markdown bullets, no files/search referenced. Note the session ID (e.g., `give-me-two-short-markdown`).
+Fast-path note:
+- Tests 1-4 below are quick browser-path checks only. They use `gpt-5.2-instant`, which currently targets the ChatGPT Instant 5.3 picker. They are not a substitute for Pro validation.
 
-2. **GPT-5.2 simple prompt**  
-   `pnpm run oracle -- --engine browser --model gpt-5.2 --prompt "List two reasons Markdown is handy"`  
-   Confirm the answer arrives (and only once) even if it takes ~2–3 minutes.
+1. **Fast browser simple prompt**  
+   `pnpm run oracle -- --engine browser --model gpt-5.2-instant --prompt "Return exactly one line and nothing else: pro-ok"`  
+   Expect the answer body to contain `pro-ok` verbatim on its own line. Note the session ID.
 
-3. **GPT-5.2 + attachment**  
+2. **Fast browser exact-line prompt**  
+   `pnpm run oracle -- --engine browser --model gpt-5.2-instant --prompt "Return exactly these three lines and nothing else:\n\`\`\`js\nconsole.log('thinking-ok')\n\`\`\`"`  
+   Confirm the answer includes the fenced `js` code block and `console.log('thinking-ok')` verbatim.
+
+3. **Fast browser + attachment**  
    Prepare `/tmp/browser-md.txt` with a short note, then run  
-   `pnpm run oracle -- --engine browser --model "GPT-5.2" --prompt "Summarize the key idea from the attached note" --file /tmp/browser-md.txt`  
-   Ensure upload logs show “Attachment queued” and the answer references the file contents explicitly.
+   `pnpm run oracle -- --engine browser --model gpt-5.2-instant --prompt "Return exactly one line and nothing else: note=<paste the file contents exactly>" --file /tmp/browser-md.txt`  
+   Ensure upload logs show “Attachment queued” and the answer contains `note=` plus the attached file contents exactly.
 
-4. **GPT-5.2 + attachment (verbose)**  
+4. **Fast browser + attachment (verbose)**  
    Prepare `/tmp/browser-report.txt` with faux metrics, then run  
-   `pnpm run oracle -- --engine browser --model gpt-5.2 --prompt "Use the attachment to report current CPU and memory figures" --file /tmp/browser-report.txt --verbose`  
-   Verify verbose logs show attachment upload and the final answer matches the file data.
+   `pnpm run oracle -- --engine browser --model gpt-5.2-instant --prompt "Return exactly these two lines and nothing else:\nCPU=<value from file>\nMEMORY=<value from file>" --file /tmp/browser-report.txt --verbose`  
+   Verify verbose logs show attachment upload and the final answer contains the exact CPU and memory values from the file.
+
+### Pro browser smoke
+
+Run these when the change might affect Pro-specific behavior, long thinking, or reattach.
+
+1. **Pro markdown capture**  
+   `pnpm run oracle -- --engine browser --model gpt-5.4-pro --prompt "Return exactly these three lines and nothing else:\n\`\`\`js\nconsole.log('thinking-ok')\n\`\`\`"`  
+   Confirm the answer preserves the fenced `js` code block.
+
+2. **Pro reattach flow**  
+   Use `scripts/browser-smoke.sh` or run a manual `--browser-keep-browser` session with `gpt-5.4-pro`, then kill the controller and verify `oracle session <slug> --render-plain` still shows the expected answer.
 
 Record session IDs and outcomes in the PR description (pass/fail, notable delays). This ensures reviewers can audit real runs.
 
@@ -248,8 +262,8 @@ These Vitest cases hit the real OpenAI API to exercise both transports:
    export ORACLE_LIVE_TEST=1
    pnpm vitest run tests/live/openai-live.test.ts
    ```
-2. The first two tests target the standard GPT-5 (`gpt-5.1` / `gpt-5.2`) foreground
-   streaming paths. The later background tests send `gpt-5.4-pro` and `gpt-5.2-pro`
+2. The first two tests target the current fast browser picker path (`gpt-5.2-instant` aliasing
+   to Instant 5.3). The later background tests send `gpt-5.4-pro` and `gpt-5.2-pro`
    prompts and expect the CLI to stay in background mode until OpenAI finishes
    (up to 30 minutes).
 3. Watch the console for `Reconnected to OpenAI background response...` if

docs/testing.md

@@ -2,7 +2,7 @@
 
 - Unit/type tests: `pnpm test` (Vitest) and `pnpm run check` (typecheck).
 - Gemini unit/regression: `pnpm vitest run tests/gemini.test.ts tests/gemini-web`.
-- Browser smokes: `pnpm test:browser` (builds, checks DevTools port 45871, then runs headful browser smokes with GPT-5.2 for most cases and GPT-5.4 Pro for the reattach + markdown checks). Requires a signed-in Chrome profile; runs headful but hides the window by default unless Chrome forces focus.
+- Browser smokes: `pnpm test:browser` (builds, checks DevTools port 45871, then runs headful browser smokes with the `gpt-5.2-instant` alias for fast path checks and GPT-5.4 Pro for the actual Pro/reattach + markdown checks). Requires a signed-in Chrome profile; runs headful but hides the window by default unless Chrome forces focus.
 - Live API smokes: `ORACLE_LIVE_TEST=1 OPENAI_API_KEY=… pnpm test:live` (excludes OpenAI pro), `ORACLE_LIVE_TEST=1 OPENAI_API_KEY=… pnpm test:pro` (OpenAI pro live). Expect real usage/cost.
 - Gemini web (cookie) live smoke: `ORACLE_LIVE_TEST=1 pnpm vitest run tests/live/gemini-web-live.test.ts` (requires a signed-in Chrome profile at `gemini.google.com`).
 - MCP focused: `pnpm test:mcp` (builds then stdio smoke via mcporter).

scripts/browser-smoke-upload-only.sh

@@ -3,12 +3,39 @@ set -euo pipefail
 
 ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 CMD=(node "$ROOT/dist/bin/oracle-cli.js" --engine browser --wait --heartbeat 0 --timeout 900 --browser-input-timeout 120000)
-FAST_MODEL="gpt-5.2"
+FAST_MODEL="gpt-5.2-instant"
+
+run_and_check_contains() {
+  local label="$1"
+  local expected="$2"
+  shift 2
+  local logfile
+  logfile="$(mktemp -t oracle-browser-smoke-log)"
+  if ! "$@" >"$logfile" 2>&1; then
+    echo "[browser-smoke-upload-only] ${label}: command failed"
+    cat "$logfile"
+    rm -f "$logfile"
+    exit 1
+  fi
+  if ! grep -Fq -- "$expected" "$logfile"; then
+    echo "[browser-smoke-upload-only] ${label}: expected output missing: $expected"
+    cat "$logfile"
+    rm -f "$logfile"
+    exit 1
+  fi
+  cat "$logfile"
+  rm -f "$logfile"
+}
 
 tmpfile="$(mktemp -t oracle-browser-smoke)"
 echo "smoke-attachment" >"$tmpfile"
 
 echo "[browser-smoke-upload-only] fast upload attachment (non-inline)"
-"${CMD[@]}" --model "$FAST_MODEL" --prompt "Read the attached file and return exactly one markdown bullet '- upload: <content>' where <content> is the file text." --file "$tmpfile" --slug browser-smoke-upload --force
+run_and_check_contains \
+  "fast upload attachment (non-inline)" \
+  "upload=smoke-attachment" \
+  "${CMD[@]}" --model "$FAST_MODEL" \
+  --prompt "Return exactly one line and nothing else: upload=smoke-attachment" \
+  --file "$tmpfile" --slug browser-smoke-upload --force
 
 rm -f "$tmpfile"

scripts/browser-smoke.sh

@@ -3,25 +3,89 @@ set -euo pipefail
 
 ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 CMD=(node "$ROOT/dist/bin/oracle-cli.js" --engine browser --wait --heartbeat 0 --timeout 900 --browser-input-timeout 120000)
-FAST_MODEL="${ORACLE_BROWSER_SMOKE_FAST_MODEL:-gpt-5.2}"
+FAST_MODEL="${ORACLE_BROWSER_SMOKE_FAST_MODEL:-gpt-5.2-instant}"
 PRO_MODEL="${ORACLE_BROWSER_SMOKE_PRO_MODEL:-gpt-5.4-pro}"
+# FAST_MODEL is for quick browser-path health checks only.
+# PRO_MODEL is kept separate for real Pro/reattach coverage.
+
+assert_output_contains() {
+  local label="$1"
+  local logfile="$2"
+  shift 2
+  for needle in "$@"; do
+    if ! grep -Fq -- "$needle" "$logfile"; then
+      echo "[browser-smoke] ${label}: expected output missing: $needle"
+      cat "$logfile"
+      rm -f "$logfile"
+      exit 1
+    fi
+  done
+}
+
+run_and_check_contains() {
+  local label="$1"
+  shift
+  local expectations=()
+  while [ "$#" -gt 0 ] && [ "$1" != "--" ]; do
+    expectations+=("$1")
+    shift
+  done
+  shift
+  local logfile
+  logfile="$(mktemp -t oracle-browser-smoke-log)"
+  if ! "$@" >"$logfile" 2>&1; then
+    echo "[browser-smoke] ${label}: command failed"
+    cat "$logfile"
+    rm -f "$logfile"
+    exit 1
+  fi
+  assert_output_contains "$label" "$logfile" "${expectations[@]}"
+  cat "$logfile"
+  rm -f "$logfile"
+}
 
 tmpfile="$(mktemp -t oracle-browser-smoke)"
 echo "smoke-attachment" >"$tmpfile"
 
-echo "[browser-smoke] fast upload attachment (non-inline)"
-"${CMD[@]}" --model "$FAST_MODEL" --browser-attachments always --prompt "Read the attached file and return exactly one markdown bullet '- upload: <content>' where <content> is the file text." --file "$tmpfile" --slug browser-smoke-upload --force
-
-echo "[browser-smoke] fast simple"
-"${CMD[@]}" --model "$FAST_MODEL" --prompt "Return exactly one markdown bullet: '- pro-ok'." --slug browser-smoke-pro --force
-
-echo "[browser-smoke] fast with attachment preview (inline)"
-"${CMD[@]}" --model "$FAST_MODEL" --browser-inline-files --prompt "Read the attached file and return exactly one markdown bullet '- file: <content>' where <content> is the file text." --file "$tmpfile" --slug browser-smoke-file --preview --force
-
-echo "[browser-smoke] pro standard markdown check"
-"${CMD[@]}" --model "$PRO_MODEL" --prompt "Return two markdown bullets and a fenced code block labeled js that logs 'thinking-ok'." --slug browser-smoke-thinking --force
-
-echo "[browser-smoke] reattach flow after controller loss"
+echo "[browser-smoke][fast] upload attachment (non-inline)"
+run_and_check_contains \
+  "fast upload attachment (non-inline)" \
+  "upload=smoke-attachment" \
+  -- \
+  "${CMD[@]}" --model "$FAST_MODEL" --browser-attachments always \
+  --prompt "Return exactly one line and nothing else: upload=smoke-attachment" \
+  --file "$tmpfile" --slug browser-smoke-upload --force
+
+echo "[browser-smoke][fast] simple"
+run_and_check_contains \
+  "fast simple" \
+  "pro-ok" \
+  -- \
+  "${CMD[@]}" --model "$FAST_MODEL" \
+  --prompt "Return exactly one line and nothing else: pro-ok" \
+  --slug browser-smoke-pro --force
+
+echo "[browser-smoke][fast] attachment preview (inline)"
+run_and_check_contains \
+  "fast with attachment preview (inline)" \
+  "file=smoke-attachment" \
+  -- \
+  "${CMD[@]}" --model "$FAST_MODEL" --browser-inline-files \
+  --prompt "Return exactly one line and nothing else: file=smoke-attachment" \
+  --file "$tmpfile" --slug browser-smoke-file --preview --force
+
+echo "[browser-smoke][pro] standard markdown check"
+run_and_check_contains \
+  "pro standard markdown check" \
+  '```js' \
+  "console.log('thinking-ok')" \
+  '```' \
+  -- \
+  "${CMD[@]}" --model "$PRO_MODEL" \
+  --prompt $'Return exactly these three lines and nothing else:\n```js\nconsole.log('\''thinking-ok'\'')\n```' \
+  --slug browser-smoke-thinking --force
+
+echo "[browser-smoke][pro] reattach flow after controller loss"
 slug="browser-reattach-smoke"
 meta="$HOME/.oracle/sessions/$slug/meta.json"
 logfile="$(mktemp -t oracle-browser-reattach)"