revert(proxy): drop overfit Caido reconnect/HTTPQL band-aids, keep serialization lock

[devin-ai-integration]

Summary

Followup to #794. Most of what #794 added to strix/tools/proxy/{tools,caido_api}.py overfit to specific past error messages rather than fixing anything: a reconnect path that matched on substrings like "transport is already connected" / "connector is closed", and an injected HTTPQL "hint" gated on "httpql" in str(exc). This reverts both files to their pre-#794 shape and keeps only the one real fix — serializing access to the non-concurrency-safe Caido GraphQL transport.

Kept (real concurrency fix): the asyncio lock. Two in-flight requests on the shared client race and raise "Transport is already connected", so all proxy calls run one at a time.

• caido_api.call_with_client stays, reduced to pure serialization:

  async def call_with_client[T](fn) -> T:
      async with _CLIENT_LOCK:
          client = _CLIENT_CACHE.get("default") or await _new_client()
          _CLIENT_CACHE["default"] = client
          return await fn(client)   # no reconnect, no idempotent=

• host-side tools.py serializes through a module-level _CAIDO_CALL_LOCK via a small _call(client, fn) helper.

Dropped (overfit / redundant):

• _CONNECTION_ERROR_MARKERS, _is_connection_error, _safe_aclose, the reconnect + idempotent= retry, and SharedCaidoClient.
• reconnect_caido and the project-id plumbing — bootstrap_caido returns just the Client again; session_manager/runner store the raw client.
• _HTTPQL_HINT / _is_httpql_error / _httpql_error. The HTTPQL syntax it injected is already documented in the list_requests docstring.

Moved to the prompt (where it belongs): a concise HTTPQL syntax note added to the PROXY section of system_prompt.jinja (quote strings, unquoted ints, AND/OR with no NOT, operator-by-field-type), instead of surfacing it as a runtime error hint.

Tests: tests/test_proxy_client.py trimmed to the surviving guarantees — client caching, error propagation, and serialization on both the call_with_client and host _call paths.

Net: +86 / −438 across 7 files.

Link to Devin session: https://app.devin.ai/sessions/5453b8e1865246cca776e31159dd9a02
Requested by: @0xallam

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Greptile Summary

This PR removes the Caido reconnect and runtime HTTPQL hint workarounds while keeping request serialization. The main changes are:

• Serializes sandbox-side and host-side Caido calls with async locks.
• Restores raw Caido client ownership in the runtime session bundle.
• Closes the connected client when project setup fails.
• Moves HTTPQL syntax guidance into the system prompt.
• Narrows proxy tests to caching, serialization, and error propagation.

Confidence Score: 5/5

This looks safe to merge.

• Project setup failures now close the connected client before propagating.
• Successful bootstrap ownership moves into the session cache without an intervening async boundary.
• No blocking issues remain in the changed code.

Important Files Changed

Filename	Overview
strix/runtime/caido_bootstrap.py	Closes the connected Caido client when project creation or selection fails.
strix/runtime/session_manager.py	Stores and cleans up the raw Caido client in the session bundle.
strix/tools/proxy/caido_api.py	Removes message-based reconnection and retries while retaining serialized client access.
strix/tools/proxy/tools.py	Serializes host-side proxy calls and removes runtime HTTPQL error rewriting.
strix/agents/prompts/system_prompt.jinja	Adds concise HTTPQL syntax guidance to the proxy prompt.
tests/test_proxy_client.py	Focuses proxy client tests on caching, serialization, and error propagation.

_{Reviews (2): Last reviewed commit: "fix(proxy): close Caido client if projec..." | Re-trigger Greptile}

[devin-ai-integration]

Thanks for the review.

P1 (bootstrap client leak) — fixed in 4b07166: bootstrap_caido now closes the connected client (best-effort) if project.create/select fails, so a failed init no longer leaks the transport.

Teardown race (close_client outside the serialization lock) — leaving as-is. This is pre-existing behavior, not introduced by this PR (this is a revert of #794 back to the prior shape), and coordinating session teardown with in-flight proxy calls is a separate, more involved change that's out of scope for this revert. Happy to track it separately if wanted.

[devin-ai-integration]

@greptile

[devin-ai-integration]

✅ Live proxy-function testing passed

Exercised every caido_api proxy function end-to-end against a real caido-cli v0.56.0 instance (same version/flags as containers/docker-entrypoint.sh) with real proxied HTTP/HTTPS traffic. Shell/script-based, no recording.

Function coverage — all PASS

• get_client() connects
• list_requests() valid HTTPQL → 10 edges
• view_request() request (151B) + response (651B)
• repeat_request() replay → DONE, 651B, 1269ms
• list_sitemap() → example.com + httpbin.org
• view_sitemap_entry() → related requests
• scope_rules() create/list/get/update/delete round-trip

Revert behavior verified

• Module-level invalid HTTPQL raises OperationUserError cleanly; loop stays healthy.
• Host-side tools.list_requests invalid filter → generic {"success": false, "error": "list_requests failed: ...Invalid HTTPQL query..."} with none of the removed structured-hint phrases. Runtime hint injection confirmed gone.

Concurrency — the serialization fix works

• 40× concurrent list_requests → 40/40 ok, 0 transport races.
• 40× concurrent mixed call_with_client → 40/40 ok, 0 races.
• Control (bypass the lock): 40/40 fail, 20 exactly "Transport is already connected" — proves the lock is the fix and the test genuinely detects the race.

uv run pytest tests/test_proxy_client.py -q → 8 passed.

Testing via Devin: https://app.devin.ai/sessions/5453b8e1865246cca776e31159dd9a02