-
6766e2f78c
refactor: rename pod_spec param to user_pod_spec in build_job_spec
yuqyang
2026-04-01 01:14:07 -07:00
-
f8159ba872
refactor: remove all default template paths — manifest path always explicit
yuqyang
2026-04-01 00:31:15 -07:00
-
f074c1cd33
refactor: remove ControllerConfig, move poll/queue settings to k8s.yaml
yuqyang
2026-04-01 00:13:36 -07:00
-
3482dfde10
feat: Jinja2 job manifest template + PodPatcher for controller
yuqyang
2026-03-31 23:40:19 -07:00
-
17dffd7e1d
feat(examples): add rollout archive operation at the end of each iteration
yuqyang
2026-03-31 18:58:03 -07:00
-
87ff4ba4e2
test: align math-poc with new deploy schema and separate experiment config
yuqyang
2026-03-31 18:38:12 -07:00
-
672a710662
update docs to clearly tell user's requirement
yuqyang
2026-03-30 21:35:40 -07:00
-
d3b3fa43c0
small updates
yuqyang
2026-03-30 21:27:50 -07:00
-
081b093e6c
refactor: align deploy flow and config with mode-based schema
yuqyang
2026-03-30 21:21:09 -07:00
-
a2a28b4fc9
refactor: adopt AGL_BASE_URL naming and add deploy guide
yuqyang
2026-03-30 09:24:17 -07:00
-
ac0f5f5615
refactor: switch deploy config to typed YAML schema
yuqyang
2026-03-30 08:49:28 -07:00
-
e037ddbba2
feat: add Python deploy entrypoint with --config support
yuqyang
2026-03-30 08:21:49 -07:00
-
805f70eca8
refactor: split pod-facing vs host-facing agl-lite URLs in deploy.sh
yuqyang
2026-03-30 05:05:44 -07:00
-
ae53bfe125
fix: make in-k8s serve honor hooks/gateway config and restore math-poc mock e2e
yuqyang
2026-03-30 03:48:01 -07:00
-
254fa161c4
refactor: refine deploy.sh modes with explicit external/host behavior
yuqyang
2026-03-30 03:00:15 -07:00
-
3e28829410
feat: add explicit deploy modes --agl-in-k8s/--agl-in-host and auto-launch host service
yuqyang
2026-03-30 02:06:35 -07:00
-
462f23fe26
feat: add calc_x migration baseline and prioritize it for training parity
yuqyang
2026-03-30 01:54:42 -07:00
-
ad26645c6d
refactor: split math-verl training example from math-poc pipeline example
yuqyang
2026-03-30 00:57:13 -07:00
-
cff633f309
refactor: make math-poc VERL training path vllm-only
yuqyang
2026-03-29 22:41:00 -07:00
-
29033ea20c
feat: add math-poc VERL training entry script and wire resources_id
yuqyang
2026-03-29 20:34:43 -07:00
-
a85b4d7993
refactor: delete async_server.py, update __init__.py
yuqyang
2026-03-29 20:11:20 -07:00
-
88d90755c2
refactor: adapt entrypoint.py + config.yaml — remove agentlightning deps
yuqyang
2026-03-29 20:10:35 -07:00
-
79c026d2ca
refactor: adapt dataset.py — replace agentlightning.types.Dataset with Sequence
yuqyang
2026-03-29 20:08:42 -07:00
-
ab6ca6a574
refactor: adapt trainer.py — use AglLiteDaemon internally
yuqyang
2026-03-29 20:02:18 -07:00
-
3700d9c3a1
refactor: adapt daemon.py — replace AgentModeDaemon with AglLiteDaemon
yuqyang
2026-03-29 19:06:08 -07:00
-
7dd323d0b3
update for documents
yuqyang
2026-03-28 00:04:06 -07:00
-
fe72b2019d
chore: copy verl module from agent-lightning verbatim
yuqyang
2026-03-27 23:59:47 -07:00
-
d2888593e0
run.sh: remove unnecessary pip-cache setup from mount section
yuqyang
2026-03-27 07:59:16 -07:00
-
a5c3cf7aba
swe_bench: minikube mount for host-accessible artifacts
yuqyang
2026-03-27 07:57:23 -07:00
-
c3e9e476fe
swe_bench: add pip cache hostPath volume for faster container startup
yuqyang
2026-03-27 07:33:15 -07:00
-
06cba884c4
add AGL_ATTEMPT_ID env var (logical alias for AGL_POD_UID)
yuqyang
2026-03-27 07:27:36 -07:00
-
80255e4548
swe_bench: move all event posting to grade.py, pure bash entrypoint
yuqyang
2026-03-27 07:24:14 -07:00
-
deabe53444
swe_bench: remove patch_size from grade.py and reward event
yuqyang
2026-03-27 07:17:44 -07:00
-
73617f82ba
swe_bench: reorganize output flow, grade.py takes path argument
yuqyang
2026-03-27 07:14:57 -07:00
-
1f33e3c245
swe_bench: install swebench explicitly in entrypoint via python3 -m pip
yuqyang
2026-03-27 05:04:15 -07:00
-
be511e7be0
swe_bench: extract grading into grade.py, mounted via ConfigMap
yuqyang
2026-03-27 04:55:01 -07:00
-
94b709af23
todo: mark completed SWE-bench refactoring items
yuqyang
2026-03-27 04:40:11 -07:00
-
313a537b31
swe_bench: move grading into container, hostPath for artifacts
yuqyang
2026-03-27 04:40:01 -07:00
-
c89247e6ed
todo: update SWE-bench design — grading in container, hostPath for artifacts
yuqyang
2026-03-27 04:31:11 -07:00
-
bbcb514caf
fix: pipe large artifact payload to curl via stdin
yuqyang
2026-03-27 00:52:04 -07:00
-
349ae38c64
start_vllm.sh: enable tool use by default (hermes parser)
yuqyang
2026-03-26 23:13:27 -07:00
-
a803cc64f0
start_vllm.sh: make max-model-len optional (use model default)
yuqyang
2026-03-26 23:12:19 -07:00
-
b78d8a6bc2
start_vllm.sh: add --tool-call-parser flag for tool use support
yuqyang
2026-03-26 23:09:43 -07:00
-
e3099b42fa
todo: mark SWE-bench as [ongoing]
yuqyang
2026-03-26 23:05:38 -07:00
-
5248d7fe86
swe_bench run.sh: add cleanup step before each run
yuqyang
2026-03-26 22:58:23 -07:00
-
3fc35e6ea3
revert: remove blanket controller restart from deploy.sh
yuqyang
2026-03-26 22:58:05 -07:00
-
b8fdd094d8
Revert "deploy.sh: use config hash annotation instead of blanket restart"
yuqyang
2026-03-26 22:57:52 -07:00
-
fc1b15aaf7
deploy.sh: use config hash annotation instead of blanket restart
yuqyang
2026-03-26 22:55:45 -07:00
-
66093a4968
fix: add auth header to event POSTs, note vLLM tool-choice requirement
yuqyang
2026-03-26 22:49:22 -07:00
-
b71252ef98
fix: ANTHROPIC_BASE_URL without /v1, restart controller on redeploy
yuqyang
2026-03-26 22:44:46 -07:00
-
eee55e2d0c
fix: add swebench/ Docker Hub prefix to SWE-bench image names
yuqyang
2026-03-26 20:13:36 -07:00
-
4479cb036e
chore: default artifact dir to ./artifacts (project-local, gitignored)
yuqyang
2026-03-26 19:57:22 -07:00
-
9a5f046448
gitignore: add artifacts/
yuqyang
2026-03-26 19:56:49 -07:00
-
392e09efbc
swe_bench: default model to Qwen2.5-Coder-3B-Instruct for quick testing
yuqyang
2026-03-26 19:44:35 -07:00
-
74d4c3576c
swe_bench: check vLLM model availability in run.sh
yuqyang
2026-03-26 19:36:51 -07:00
-
b89d105599
feat: env var substitution in gateway config (${VAR_NAME})
yuqyang
2026-03-26 19:33:02 -07:00
-
e7aa70bacf
refactor: remove mini_swe_agent, use claude_code only
yuqyang
2026-03-26 08:33:01 -07:00
-
307d5e6056
todo: mark SWE-bench example [completed]
yuqyang
2026-03-26 08:02:02 -07:00
-
7b6f7b690a
feat: SWE-bench example — hooks, agents, rl_loop, deploy config
yuqyang
2026-03-26 08:01:56 -07:00
-
4e7e02853c
feat: artifact event type — large files written to disk
yuqyang
2026-03-26 07:53:05 -07:00
-
2e9ac2a4be
todo: resolve Q3 — ConfigMap for agent scripts
yuqyang
2026-03-26 07:50:46 -07:00
-
cfa09ae395
todo: mark SWE-bench example [ready], resolve open questions 2 and 5
yuqyang
2026-03-26 07:45:10 -07:00
-
9b3163a4df
deploy.sh: only patch CoreDNS when AGL_LITE_URL uses host.minikube.internal
yuqyang
2026-03-26 07:31:35 -07:00
-
58dc0cea12
update SWE-bench proposal + deploy.sh improvements
yuqyang
2026-03-26 07:29:20 -07:00
-
bac897a316
todo: rewrite SWE-bench proposal for hooks architecture
yuqyang
2026-03-26 01:35:54 -07:00
-
6606e926ea
remove legacy code
yuqyang
2026-03-26 00:35:14 -07:00
-
6aea2f6ce7
cleanup: remove legacy math-poc files, rename v2 to final
yuqyang
2026-03-26 00:30:26 -07:00
-
abc6f44526
todo: confirm and remove Store Hooks + Math-poc restructuring
yuqyang
2026-03-26 00:29:39 -07:00
-
e09180e8f9
fix: vllm hooks inject AGL_MODEL_NAME, job-template adds hostAliases
yuqyang
2026-03-26 00:24:59 -07:00
-
c351368305
fix: math-poc hooks respect job-template defaults, read ground_truth from input
yuqyang
2026-03-25 23:55:32 -07:00
-
b899c2f1b7
todo: mark math-poc restructuring as completed
yuqyang
2026-03-25 23:01:30 -07:00
-
6e875a1dc1
examples: math-poc restructured with hooks and mode subfolders
yuqyang
2026-03-25 23:01:23 -07:00
-
b9bd98b850
todo: add math-poc restructuring with hooks [ongoing]
yuqyang
2026-03-25 22:56:50 -07:00
-
9dc3b80115
schemas: remove data field from RolloutMetadata
yuqyang
2026-03-25 22:55:58 -07:00
-
ec76bb7fd4
schemas: typed RolloutMetadata with algorithm control indexes
yuqyang
2026-03-25 22:52:35 -07:00
-
bd723cc86b
refactor: remove AGL_TASK_INPUT auto-injection from job_builder
yuqyang
2026-03-25 22:50:18 -07:00
-
c690432db6
todo: mark Store Hooks as completed
yuqyang
2026-03-25 22:19:31 -07:00
-
163606fc99
feat: implement Store Hooks — rollout lifecycle hooks in server
yuqyang
2026-03-25 22:19:25 -07:00
-
6dfa1ce93b
schemas: add metadata field to Rollout and EnqueueRolloutRequest
yuqyang
2026-03-25 22:14:25 -07:00
-
b8bdbb97b5
todo: finalize Store Hooks design — hooks as sync pre-processors in server
yuqyang
2026-03-25 20:54:51 -07:00
-
6212bc25f3
todo: Task Controller architecture — task-agnostic algorithm layer [discuss]
yuqyang
2026-03-25 20:18:55 -07:00
-
2ce355d1ef
todo: SWE-bench — single-container design (agent + eval in one rollout)
yuqyang
2026-03-25 19:01:09 -07:00
-
67089804a4
todo: update SWE-bench design — resolve (A)(B)(C)(D)(E) per discussion
yuqyang
2026-03-25 00:50:36 -07:00
-
17bd0d5def
todo: add SWE-bench example design [discuss]
yuqyang
2026-03-25 00:11:57 -07:00
-
ed3dd4a00d
slides update
yuqyang
2026-03-22 22:36:34 -07:00
-
1966c8714c
small changes
yuqyang
2026-03-22 08:48:37 -07:00
-
77392d321d
docs: slides — add subtitle to Three Design Choices, split job-template into two pages
yuqyang
2026-03-22 08:41:07 -07:00
-
e6859a0da1
docs: add job-template slide to Part 4 (store + controller)
yuqyang
2026-03-22 08:36:37 -07:00
-
a41979a06a
docs: slides v3 — no animations, agent_output event, deployment demo, restructured
yuqyang
2026-03-22 08:08:48 -07:00
-
ee35412ed9
docs: rewrite refactor review slides — positive framing, no AL criticism
yuqyang
2026-03-22 07:52:23 -07:00
-
16756d9569
docs: add refactor review slides for Agent Lightning developers
yuqyang
2026-03-22 07:20:43 -07:00
-
51787311ce
docs: update todo — Phase 5a+5b complete, 5c (full training loop) next
yuqyang
2026-03-22 07:05:45 -07:00
-
167e1ae9f6
feat: AglLiteDaemon — VERL trainer bridge via AglLiteClient
yuqyang
2026-03-22 07:04:34 -07:00
-
7901c3b39a
feat: triplet format on GET /api/events?format=triplet
yuqyang
2026-03-22 06:55:44 -07:00
-
805d1454c7
docs: fix Phase 5b framing — agent-lightning depends on agl-lite HTTP, not vice versa
yuqyang
2026-03-22 06:36:20 -07:00
-
7255d65220
docs: Phase 5 design — triplet API + VERL integration options A/B
yuqyang
2026-03-22 06:21:46 -07:00
-
9fa477335c
docs: update todo — Phase 4b with colocated topology, gateway injection, new decisions
yuqyang
2026-03-22 05:17:31 -07:00
-
f6215e2eb5
refactor: colocate agl-lite with algorithm on host in vLLM mode
yuqyang
2026-03-22 05:08:22 -07:00
-
f6119deac9
feat: gateway param injection — return_token_ids for RL training
yuqyang
2026-03-22 03:40:57 -07:00
-
46f7214e7e
docs: update todo — Phase 4b complete, performance baseline to backlog
yuqyang
2026-03-22 03:23:34 -07:00