[netlogics]

Recognize llama-server's context-overflow error

Summary

Auto compaction isn't triggered when a llama.cpp model's context window is exceeded.

Problem

When a provider rejects a prompt that exceeds the model's context window, ECA classifies the error with eca.llm-providers.errors/classify-error and, on :context-overflow, runs the automatic recovery path (prune-tool-results! auto-compact retry), as described in docs/config/context-management.md.

If no pattern matches, the error falls through to :unknown, so recovery never runs and the raw 400 is shown to the user instead.

llama.cpp's llama-server phrases the rejection as:

{"error":{"code":400,"message":"request (68183 tokens) exceeds the available context size (65536 tokens), try increasing it","type":"exceed_context_size_error","n_prompt_tokens":68183,"n_ctx":65536}}

The existing catch-all regex requires max/maximum/model right after exceeds the, so available doesn't match. Every llama-server wording since the message was introduced uses "exceeds the available context size", so none of them were recognized.

Changes

src/eca/llm_providers/errors.clj context-overflow-patterns:

• Add available to the catch-all's alternation: exceeds? the? ?(?:available|max(?:imum)?|model).{0,20}(?:token|context).
• Also match the error's type field, exceed_context_size_error, a stable identifier, so detection keeps working if llama.cpp rewords the message.

test/eca/llm_providers/errors_test.clj:

• llama-server's full 400 body.
• The type field on its own.
• llama-server's earlier wording ("the request exceeds the available context size, try increasing it") with no body, i.e. the message-only path.
• A guard against loosening the catch-all: Azure's rate-limit text ("exceeded token rate limit") must not classify as an overflow.

CHANGELOG.md:

• Add an entry under ## Unreleased.

Testing

clojure -M:test --focus eca.llm-providers.errors-test
# 15 tests, 137 assertions, 0 failures.

Also checked against a live llama-server: an over-long prompt returns the 400 above on /v1/chat/completions and /v1/responses, streaming or not. The wording and type are the same in current ggml-org/llama.cpp master. That body classifies as :unknown on master and :context-overflow with this change.

Checklist

• I added a entry in changelog under unreleased section.
• This is not an AI slop.

🤖 Generated with ECA (Claude Code)

[zikajk]

Looks good!
Please, do the rebase and feel free to merge.

[netlogics]

Looks good! Please, do the rebase and feel free to merge.

Thanks for the review and approval @zikajk . I only have read access so I can't merge it myself. I reviewed the dev docs but apologies if I missed something.

[zikajk]

Sorry, my bad @netlogics. Merged 🙂

[netlogics]

Thanks