docs: gradio 6.18 migration done + DEV RUNNING; SSR fix superseded

Records the real fix for the "Starting" hang (gradio 5.49 -> 6.18; SSR was a
red herring), the show_api follow-up, the HF rollout wedge cleared by
pause/unpause, and dev confirmed RUNNING on 5ba3350. Flags prod migration as
the remaining open item (quota-gated). SHOWCASE_STATUS.md updated separately
(it lives outside this repo).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

TODO.md

@@ -12,6 +12,15 @@ pdac-analysis-orchestrator.
 
 ## Open
 
+### Deploy / infra
+
+- **High — Migrate PROD specialist to gradio 6.18.** Local `main` (`5ba3350`) has
+  the verified gradio-6 migration; dev is RUNNING on it. Prod (`origin`) is still
+  gradio 5.49 (`d38496a`) and PAUSED on the cpu-basic quota (3/3). To ship: push
+  `main` → `origin`, free a slot (e.g. pause `bcc-lit-agent`), then restart/unpause
+  the prod Space. Until then prod stays on 5.49 and would hit the same "Starting"
+  hang if it ever comes off pause without the migration.
+
 ### Cross-dataset integration (ADR-0001)
 
 Combine ≥2 datasets — early (pool + batch covariate) / late (meta-analysis) /
@@ -107,10 +116,18 @@ refuse, driven by registry metadata. ADR + plan + per-task spin-off prompts in
 
 ## Done (recent)
 
-- 2026-06-21 — Fixed HF Space stuck on "Starting" (`gradio==5.49.0` auto-enables
-  experimental SSR; HF readiness check hangs). `ssr_mode=False` default in
-  `GradioAgentUI.launch()` (`d38496a`); pushed to dev (rebuilding) + prod
-  (staged, still quota-paused). Workaround: Space var `GRADIO_SSR_MODE=false`.
+- 2026-06-22 — **REALLY fixed HF Space stuck on "Starting": gradio 5.49 → 6.18
+  migration** (commits `b4e7ac0` + `5ba3350`, branch `fix/gradio-6-migration` →
+  merged to `main`). The 2026-06-21 SSR fix was a red herring (the working
+  orchestrator runs 6.18 *with SSR on*); real cause is gradio 5.49 no longer
+  completing HF's readiness handshake. Cascade: `mcp` 1.10.1→1.28.0,
+  `langchain-mcp-adapters` 0.2.2→0.3.0, `fastmcp` pinned `==3.2.3`; `gradio_ui.py`
+  ported to the 6.0 API; `ssr_mode=False` removed. Verified: 851 tests, 49 tools,
+  HTTP 200. **DEV RUNNING** on `5ba3350` (a pause→unpause cleared an HF-side
+  rollout wedge). Prod migration still pending (see Open → Deploy/infra).
+- 2026-06-21 — [SUPERSEDED by the gradio-6 migration above] Attempted SSR fix for
+  the "Starting" hang: `ssr_mode=False` default in `GradioAgentUI.launch()`
+  (`d38496a`). Did not actually fix it; removed during the migration.
 - 2026-06-17 — UI download buttons fixed: exports no longer truncate the
   solution (the old path dropped everything after the first blank line); TXT
   downloads again; added a **Full log (.pdf)** export of the full generated

memory.md

@@ -4,7 +4,73 @@ This file tracks current status, recent decisions, and next steps.
 Update it whenever meaningful work is completed or the direction changes.
 Stable architectural facts belong in `CLAUDE.md`, not here.
 
-Last updated: 2026-06-21
+Last updated: 2026-06-22
+
+---
+
+## 2026-06-21/22 session — REAL fix for Space stuck on "Starting": gradio 5.49 → 6.18 (SSR was a red herring) — DEV RUNNING ✅
+
+The dev specialist Space (`Paper2Agent_decoupleRpy_dev`) was still hanging on
+**APP_STARTING** even after the SSR fix (`d38496a`) was deployed — the app bound
+7860, served, loaded 49 tools, but HF never flipped the badge to Running (0
+ready replicas, public URL 404).
+
+**Diagnosis (via HF runtime API, git keychain token):** the Space had been
+wedged building an *old* commit (`c54b550`, pre-SSR-fix); the d38496a/81fb4aa
+pushes updated the repo but never rebuilt the container. Forced a factory
+rebuild → it built `81fb4aa` (SSR fix live) and **still hung on APP_STARTING**.
+Comparing run logs against the known-good `pdac-analysis-orchestrator-dev`
+(RUNNING) showed that Space runs **gradio 6.18 WITH SSR enabled** and reaches
+Running, while this one on **5.49.0** hangs regardless of `ssr_mode`. → Root
+cause: **gradio 5.49.0 no longer completes HF's current Spaces readiness
+handshake.** The SSR commit chased the wrong thing.
+
+**Fix (commit `b4e7ac0`, branch `fix/gradio-6-migration`):** migrate to gradio
+6.18.0.
+- README `sdk_version` 5.49.0 → 6.18.0.
+- requirements: `gradio[oauth,mcp]` 5.49.0 → 6.18.0 lifts the old `mcp==1.10.1`
+  hard-pin → `langchain-mcp-adapters` 0.2.2 → 0.3.0, `mcp` → 1.28.0. **`fastmcp`
+  pinned ==3.2.3**: 2.11.2 is incompatible with mcp≥1.28 (`@mcp.tool` returns a
+  non-callable FunctionTool → 142 test failures + breaks direct tool calls);
+  3.4.x breaks `from fastmcp import FastMCP`.
+- `gradio_ui.py` ported to the 6.0 API: theme+css moved off `Blocks()` onto
+  `launch()`; removed `Chatbot(type=, show_copy_button=)` and `Textbox(
+  show_copy_button=)`; dropped the obsolete `ssr_mode=False` override (6.x runs
+  with SSR on, like the working Space).
+- **Verified locally** (pinned set): UI builds, 49 MCP tools discover (no
+  errors), full suite **851 passed / 55 skipped**.
+- **Second fix needed (`5ba3350`):** the first dev build crash-looped with
+  `TypeError: Blocks.launch() got an unexpected keyword argument 'show_api'` —
+  gradio 6.0 removed `show_api` from `launch()`. Local checks had only built
+  `create_app()` + inspected the launch signature, never actually *called*
+  `launch()`; fixed and re-verified by really launching (HTTP 200).
+- **HF rollout wedge (infra, not code):** after the `5ba3350` push the dev Space
+  got stuck `APP_STARTING`/`rep=0` for hours regardless of config — ruled out
+  gradio version (6.18 = the RUNNING orchestrator), SSR (tried on AND off via
+  `GRADIO_SSR_MODE` var), code (startup log byte-identical to the working
+  orchestrator, incl. the benign `__del__` `ValueError`), and quota (had its
+  slot). Other cpu-basic Spaces stayed RUNNING → space-specific. A **pause →
+  unpause** cycle finally cleared it (promoted after the watch window).
+- **DEV CONFIRMED RUNNING** ✅ (`hf-dev` main = `5ba3350`): runtime RUNNING,
+  rep=1, authed `https://anne-voigt-paper2agent-decouplerpy-dev.hf.space/` →
+  HTTP 200 (404 unauthenticated is just the private-Space gate), 49 tools
+  loaded, SSR on (Node proxy → Python :7861, same as the orchestrator).
+- **Local `main` fast-forwarded to `5ba3350`.** Branch `fix/gradio-6-migration`
+  merged.
+- **Not yet on prod** (`origin`): prod is PAUSED on the cpu-basic quota (3/3:
+  dev specialist + dev orchestrator + bcc-lit-agent all RUNNING) and needs a
+  freed slot + manual unpause; push `main` → `origin` once a slot is free. The
+  two SSR-era prod commits (`d38496a`/`81fb4aa`) are superseded by this
+  migration. **`ssr_mode=False` is no longer set in code** — 6.x runs with SSR
+  on like the working Space.
+
+---
+
+## 2026-06-21 session — Fix HF Space stuck on "Starting" (SSR) — SUPERSEDED, see entry above
+
+NOTE: the SSR theory below was a **red herring** — the real cause was the gradio
+5.49 ↔ HF readiness incompatibility (see the gradio 6.18 migration entry above).
+Kept for history.
 
 ---