[coding-agent-rl] Refactor coding-agent RL: turn-node TrajectoryManager + pluggable harness layer#2005

Merged

zhuzilin merged 66 commits into

jingshenghang:refactor_trajectory_manager

Jun 17, 2026

Commits on Jun 8, 2026

refactor(agent): replace segment-based trajectory with turn-node TrajectoryTree
jingshenghang
committed
refactor(agent): drop dead code left by trajectory-manager refactor
jingshenghang
committed
refactor(agent): port anthropic + trajectory_manager from trajectory-manager-migration-v2
jingshenghang
committed
refactor(agent): rewrite openai.py for Codex CLI + TrajectoryManager
jingshenghang
committed
refactor(agent): centralize snapshot-threshold default + filter access log
jingshenghang
committed
feat(agent): add fork-merge rescue for short assistant rewrites
jingshenghang
committed
fix(agent): replace sib.messages on fork-merge rescue
jingshenghang
committed
refactor(agent): drop billing-header scrub now cc emits no header
jingshenghang
committed
refactor(agent): migrate TrajectoryManager and adapters (v4)
jingshenghang
committed
refactor(agent): drop drift fork/merge params, strict exact-prefix linearization
jingshenghang
committed
feat(agent): assistant-rewrite merge to de-dilute reward
jingshenghang
committed
feat(agent): TrajectoryManager re-accepts fork_merge_max_response_tokens
jingshenghang
committed
docs(test): spec for TrajectoryManager e2e test script
jingshenghang
committed
test(agent): end-to-end TrajectoryManager test matrix (append_turn/get_trajectory)
jingshenghang
committed
test(agent): dump raw append_turn inputs in e2e readable output
jingshenghang
committed
test(agent): 1.7 now shows token drift's effect on the linearized sample
jingshenghang
committed
test(agent): every case prints [samples] + mask info
jingshenghang
committed
test(agent): make reward-split explicit in dump + conservation assert
jingshenghang
committed
test(agent): set every case's input reward to 1.0
jingshenghang
committed
test(agent): render whitespace in token labels as visible ␣
jingshenghang
committed
test(agent): add Group 4 (boundary/defensive/feature) -> 98% coverage
jingshenghang
committed
test(agent): assert full output via golden token+loss strings
jingshenghang
committed
refactor(agent): drift-tolerant trajectory linearization
jingshenghang
committed
chore(agent): untrack e2e test design doc and trajectory_manager tests
jingshenghang
committed
chore(agent): drop comments/docstrings in generate.py and TurnRecord
jingshenghang
committed
docs(agent): tighten trajectory_manager comments to why-not-what
jingshenghang
committed

Commits on Jun 9, 2026

refactor(agent): slim adapters and trajectory_manager, add e2e test
jingshenghang
committed
refactor(agent): replace segment-based trajectory with turn-node TrajectoryTree
jingshenghang
committed
refactor(agent): drop dead code left by trajectory-manager refactor
jingshenghang
committed
refactor(agent): port anthropic + trajectory_manager from trajectory-manager-migration-v2
jingshenghang
committed
refactor(agent): rewrite openai.py for Codex CLI + TrajectoryManager
jingshenghang
committed
refactor(agent): centralize snapshot-threshold default + filter access log
jingshenghang
committed
feat(agent): add fork-merge rescue for short assistant rewrites
jingshenghang
committed
fix(agent): replace sib.messages on fork-merge rescue
jingshenghang
committed
refactor(agent): drop billing-header scrub now cc emits no header
jingshenghang
committed
fix(agent): use rollout_id after upstream group_id revert
jingshenghang
committed
refactor(agent): assert base_sample in get_trajectory instead of defaulting
jingshenghang
committed
Merge branch 'trajectory-manager-migration-v4' into refactor_trajectory_manager
jingshenghang
committed
fix(agent): mask entire drifted response span in B1 replace
jingshenghang
committed
refactor(agent): extract shared adapter pipeline; move TurnRecord + aiohttp_threaded into slime.agent
jingshenghang
committed
refactor(agent): symmetric adapters + de-scaffold common
jingshenghang
committed

Commits on Jun 10, 2026

refactor(agent): drop tools param; dict== mount-point matching
jingshenghang
committed
refactor(agent): collapse Node turn_* into turn; rename Node->MessageNode; drop node_match_key
jingshenghang
committed
perf(agent): chunked common-prefix; rename _lcp_len -> _common_prefix_len
jingshenghang
committed

Commits on Jun 11, 2026

refactor(agent): trajectory manager cleanup + adapter tweaks
jingshenghang
committed
refactor(agent): fold classify_drift into _SampleBuilder as classify_token_drift
jingshenghang
committed
chore(agent): untrack test_trajectory_manager.py
jingshenghang
committed

Commits on Jun 12, 2026

refactor(agent): gate fork on full output_ids length
jingshenghang
committed
Refactor coding-agent RL harness: pluggable agent harness layer + SWE eval
jingshenghang
committed
Merge pull request #4 from jingshenghang/refactor_harness
jingshenghang
authored

Commits on Jun 15, 2026

refactor(agent): unify env vars under SLIME_AGENT_*/ADAPTER_*, tidy harness + adapter
jingshenghang
committed
refactor(agent): make harnesses SingletonMeta-backed, drop module-level singletons
jingshenghang
committed
refactor(agent): rename debug hook to debug_callback, pass TurnRecord whole
jingshenghang
committed

Commits on Jun 16, 2026

refactor(agent): tidy adapters, trajectory_manager, harness and SWE example
jingshenghang
committed
refactor(agent): drop optional sandbox metadata file/json, keep image key only
jingshenghang
committed
refactor(agent): reorganize agent tests under tests/test_agent/, tidy SWE example and anthropic adapter
jingshenghang
committed
refactor(agent): simplify adapters, harness and sandbox
jingshenghang
committed
docs(swe): move fan-out/sandbox notes from launcher into README
jingshenghang
committed
Merge remote-tracking branch 'origin/main' into refactor_trajectory_manager
jingshenghang
committed
fix(agent): lazy-import load_tokenizer so CPU agent test needs no transformers
jingshenghang
committed
Revert "fix(agent): lazy-import load_tokenizer so CPU agent test needs no transformers"
jingshenghang
committed
test(agent): stub transformers before importing generate in CPU rollout test
jingshenghang
committed
test(agent): shim asyncio.timeout for py3.10 CI in CPU rollout test
jingshenghang
committed
ci: rename agent-adapter-test job to agent-test (covers more than adapters)
jingshenghang
committed
refactor(agent): rename trajectory_manager.py back to trajectory.py
jingshenghang
committed
refactor(agent): drop duplicate trajectory_manager.py left by rename
jingshenghang
committed