[jerry78424]

Problem and result

_note_unreadable_tool_images - the method that promises "a picture a tool returned that this server cannot read becomes a short note in its place instead of a 400" - encodes the picture without starting the elastic encoder first. Once the idle sweep has stopped the encoder (elastic_idle_s without a picture, default 120 s), the tool's picture writes to a pipe that no longer exists: the AttributeError on the None proc escapes the except (ValueError, OSError) filter and the route's body-shape filter, and the request dies as a 400 "the request is malformed (AttributeError...)" - the opposite of what the method exists to do. And because clients re-send the whole history every turn, every later request of the conversation dies the same way at this method's own line: the conversation stays broken until a picture from another conversation (or a restart) starts the encoder.

Reproduce: "elastic": true + "vram_elastic": true; send one picture, let the encoder idle-stop (log: [strata] images: the image encoder stopped after N s ...); then send a tool message whose content carries an image block (Claude Code's Read of a picture - the exact input 3709c15 added this method for). Expected per the method's own docstring: a [image omitted: ...] note. Actual: AttributeError: 'NoneType' object has no attribute 'stdin' (inside Vision.encode's self.proc.stdin.write) surfaces as a 400.

After the change the tool picture starts the encoder first, the note is written as documented, and the conversation survives.

Related discussion

Follows #2641 (same branch and author; closed by the 0.2.0-beta tracker reset, and GitHub would not reopen it after the rebase force-push). Fixes #3171. The prepare() path does not have this problem - its encode block calls _vision_up() inside the FIFO turn; the tool path (3709c15, wrapped into _turn by 873a0d6) never got the wake, and the elastic idle stop made the missing wake fatal. Related: #2435 (the sibling: a start that fails on the prepare path), #2343 (MCP picture results will send more traffic through this exact path). Independent of the #2455 successor (its cached() skip lands in this same method later - textual overlap only). #3174 is a parallel PR with the core wake alone; this one is the same fix plus the two sides that adding a wake to this path brings.

Approach

• serve/server.py: inside the tool path's turn, _vision_up() before the encode (the caller holds the FIFO, as prepare does); the turn's end updates last_image_at in a finally - without it the idle sweep judges the just-started encoder by the previous picture's timestamp and reclaims it at once, and a tool-picture-only conversation never moves the timer; start()'s RuntimeError joins the filter, so a failing start becomes the documented note, not a dead request.
• serve/test_vision_elastic.py: two cases - a tool picture wakes a stopped encoder, and the turn holds it across the idle check; a start that raises becomes the note.
• Compatibility: the widened filter matches what can still raise (encode()'s own pipe death already becomes ValueError/OSError internally - serve: an untimed read of the engine/vision pipe can hold the request turn forever (intermittent hang; the #481 lost-step class) #1317's read timeout raises ValueError, a write to a dead pipe is OSError - and ClientGone is a bare Exception, so a client that left still propagates). No native files touched.

Validation

Passed:

• Both new cases fail on the base tree at the fake's "a picture reached an encoder that was not started" assert (the bug's exact signature), and a real Vision with the proc unloaded raises the AttributeError verbatim.
• python -m unittest serve.test_vision_elastic serve.test_vision_busy (CPU-only, no engine): 13 OK; python -m unittest serve.test_images: 19 OK - at e88ee39d (v0.2.0-beta-12), rebased 2026-10-20, Windows 11 26200, Python 3.12 venv. On the previous base f57c0adc the same files were 9/9 + 32 OK.
• With [Other] serve: an all-cached image turn never wakes the elastic encoder #2455 on top (cherry-picked, auto-merge clean): the three files stay 19 + 19 OK - the two PRs conflict in neither order.

Failed: none observed.

Not run: CUDA/HIP/SYCL builds (no native files touched); the live GPU end-to-end is #3171's reporter's log, reproduced here by the two CPU tests' fail-on-base/pass-after signature.

Limits

• The wake reuses the same FIFO turn the encode already runs in, so a tool picture now costs the encoder start (~seconds, logged) on an idle-stopped encoder - the same cost the user-picture path already pays; it is what the note path cannot afford to skip.
• No native files touched; nothing here changes non-elastic behavior.