Repository navigation
Controller fixes from code review: queue timeout, LF line model, rejection reasons - #115
Conversation
kudama
left a comment
There was a problem hiding this comment.
Reviewed 2026-11-18 against f2d59c9. Independently ran the full suite: 538 passed, 2 subtests passed; GitHub Tests and Checkmk compatibility are green. LF-only source numbering and the execution/queue budget separation are sensible changes, but I recommend addressing the queue-expiry gap before merge; details inline. Also reject nonfinite queue-timeout settings.
Independent expiry probe: with execution timeout=5s, queue_timeout=0.05s and capacity=2, a second task remains queued after 0.15s behind blocked inference; a third submission is rejected with 503. Only after releasing the first task is the second marked queue_expired, with no additional inference. Thus the new setting is currently a dispatch-time eligibility test, not an actual bound on queued lifetime/capacity.
600s is a reasonable provisional configurable default, not a measured scheduling recommendation. No controller configuration, model default, deployed source, or service state was changed during review. This PR targets forge-v2, not main; the draft v2 specification is an additional branch commit rather than another new open PR.
The execution deadline previously started at enqueue time while a single worker drains the queue, so a task could be killed (or expire before it ever ran) purely because of work ahead of it. - Execution deadline (FORGE_TASK_TIMEOUT) now starts at dispatch. - New FORGE_QUEUE_TIMEOUT (default 600s) bounds queue wait; expiry is the distinct terminal error 'queue_expired', decided before the task is marked running and without consuming inference. - Docs updated; this intentionally changes documented v1 semantics to meet docs/v2 SRV-04 / MCP-07 (queue wait separate from execution time).
…xplicitly Line numbers were produced with str.splitlines(), which also splits on \x0b, \x0c, \x1c-\x1e, \x85 and U+2028/9. Any such character shifted every later line number away from git/editor numbering (8 of 8 separators reproduced), while evidence validation used the same wrong numbering, so a citation could pass validation yet point a reviewer at the wrong line. - Add tools.split_lines (LF only) and use it for reads, snapshot reads, result-evidence bounds and the evaluator. - CRLF sources can never satisfy exact multi-line replacement (the model is shown text the raw source does not contain). Preflight now rejects any source containing CR with the explicit reason 'unsupported_line_endings' instead of failing every edit later as 'replacement_not_exact'. - Tests compare snapshot numbering with 'git grep -n' as the oracle.
Every preflight failure (about a dozen distinct HandoffError codes such as stale_revision, runtime_limits_mismatch, source_hash_mismatch, scope_denied) collapsed to a bare 422 'invalid_work_order', and the MCP bridge collapsed it again to 'invalid_request', so callers could not tell what to fix (#76, docs/v2 MCP-07). - Controller: 422 body now carries {'detail','reason'} when the failure is a HandoffError whose code matches ^[a-z0-9_-]{1,64}$. KeyError/ValueError and malformed codes still return the bare 'invalid_work_order'. - Bridge: relays 'invalid_request: <reason>' only for a bounded, well-formed reason on a 422 invalid_work_order; never relays other upstream text. The existing 403 source_reads_required behaviour is unchanged. - Docs list the closed reason vocabulary.
f2d59c9 to
2711efa
Compare
|
Review fixes pushed in 2711efa (2026-11-18). Queued expiry now runs independently, settles/reclaims waiting entries while inference continues, preserves FIFO and Queue accounting, and cancels timers at dispatch/shutdown. NaN/infinity are rejected for both direct and environment configuration. Original review regressions failed before the fix; full suite now passes: 548 tests and 2 subtests. Per Mike’s correction, rebased the controller-only commits onto main and retargeted this PR to main. The draft v2 docs remain untouched on forge-v2 and are excluded from this diff. Verified the tested runtime/scripts/tests and relevant v1 docs are byte-identical across the rebase. Remote update used an expected-head lease to preserve any concurrent edits. Nothing merged or deployed; next gate is refreshed CI/review. |
|
Refreshed CI is green for 2711efa: Tests (including fresh Ubuntu/macOS environments) and Checkmk compatibility both passed. Both review findings are resolved. PR #115 targets main and excludes docs/v2; the v2 specification remains on its separate branch. Ready for merge review; not merged or deployed. |
Fixes three issues in the existing Forge controller:
Review follow-up adds regression coverage for expiry while inference remains blocked, repeated slot reuse without physical queue buildup, surviving FIFO order, queue completion accounting, cancellation/shutdown, running-task protection, and nonfinite direct/environment configuration. The 600-second default remains configurable.
Validation: full suite passes, 548 tests and 2 subtests. New regressions reproduced the failures on the original PR. Runtime/source/test files are unchanged by the rebase onto main.
Targets main because these are existing-controller fixes. Draft Forge v2 specification documents remain separately on forge-v2; this PR does not include them. No deployed service or model configuration changed.