saddle-Clanker-Saddle
|
6e569e25df
Advisor refinements: hidden tool bubbles, chattier prompt, reports in reach, parking guard
|
4 săptămâni în urmă |
clanker
|
4eccca4faa
M0003 event indexes, batched chunks ingestion, LLM usage surfacing, resume failure signatures
|
1 lună în urmă |
clanker
|
7007d4b94c
protocol-native integration pass: goal-era arc end to end against the real daemon — the harness speaks tool calls (mock script per §7 grammar: narrated probes, question via run_command's saddle block, submit_plan/create_task/enqueue+suspend, the worker's five identical probes tripping the loop warning, finish, the coalesced merge/push/submit_report resume) and asserts the ledger world (turn_text/tool_call/tool_result/usage events, §7.2 kind set, seq exactly-once on chunks, the event log page + SSE replay, tool results as role=tool turns, the retired results-entries array consumed silently); web fixes surfaced by the pass: a review-kind block push on a goal run routes to submit_completion (the daemon's submit_report done-park now reaches completion_review with its goal-scoped narrative row), the narrative derives block-raised from the run task's question/RFC raises, the U2 mechanical rows (session context restored, parked-until) persist as system comment events on resume/park with the goal thread filtered like the task page, and the replay bundle's remark query pins the model-actor row; integration 125/0, 13 bundles sequential 966/0
|
1 lună în urmă |
clanker
|
4b1ee65655
Plan 17 M4: migrate the harness and bundles onto the ledger
|
1 lună în urmă |
clanker
|
db546c35a7
r23: D4 web half — parks close their session rows 'parked'
|
1 lună în urmă |
clanker
|
2d0e09d1d9
r5: harness split (arch 18) — the monolith becomes a package. test/harness.py: the FakeSaddle daemon, every shared wire helper (http/act/login/slot_public/wait_for/quiet_window/db/check, task_state/wait_task_done/create_queued/enqueue_task/saddle_api/saddle_card/http_get_page/dump_threads and friends), the per-shard fixtures (Ctx/start_bundle), and the runner (run_bundle + run_all_parallel + main) — the parallel spawn now targets the entry shim so 'harness' stays one module instance per process, and run_bundle imports the bundle module lazily (after that shard's environment exists, so star-imported runtime globals are the shard's own) then re-syncs WORK/DB_PATH/ports onto it (the sequential debug path reuses modules across shards). test/bundles/<name>.py ×11: the scenario bodies moved verbatim (bundle_X -> run(ctx)); the x7 duplicated goal_state closures collapse onto harness.goal_state, late-bound through the runner's fixture ctx so the answers/replay restart scenarios' re-logins keep working (the first cut snapped the auth at shard start and the replay redelivery scenario caught it). test/web_stack_test.py stays the meson entry (args unchanged) and the shard spawn target. The S-workstream handoff survivors verified byte-intact after the move: the goalui S2 verdict-goal pins (resolveReviewBlock/goal.open_blocks/feedback posting/extraText gating on the goal page) and the goals S3 park-second waits (strictly-newer completed_at vs suspended_at seconds under SADDLE_GOAL_RESUME_S=1). Both execution modes green: parallel shards 851/0, WEBSTACK_SEQUENTIAL 851/0. 851 checks.
|
1 lună în urmă |