Repository navigation

feat(loop): derive the prompt context budget from the model's advertised window - #8053

Open

henrypark133 wants to merge 29 commits into

mainfrom

context-length

Commits on Sep 1, 2026
- feat(loop-contracts): derive a prompt context budget from an advertised window
  
  henrypark133
  and
  claude
  committed
  4bede65
- feat(loop-contracts): carry a resolved context budget on LoopRunContext
  
  henrypark133
  and
  claude
  committed
  198974f
- refactor(agent-loop): pass the compaction budget as an argument
  
  henrypark133
  and
  claude
  committed
  328ffaa
- feat(agent-loop): honor a run-resolved context budget when compacting
  
  henrypark133
  and
  claude
  committed
  267b0b5
- feat(loop-host): expose the provider-advertised context window on the gateway
  
  henrypark133
  and
  claude
  committed
  c98383f
- refactor(loop-host): thread a prompt context budget to the model port
  
  henrypark133
  and
  claude
  committed
  2100238
- feat(turn-runner): resolve the prompt context budget from the run's model
  
  henrypark133
  and
  claude
  committed
  b963204
- test(support): let the scripted model advertise a context window
  
  henrypark133
  and
  claude
  committed
  1c714d2
Commits on Sep 2, 2026
- test(loop-host): pin that the gateway wrapper applies its prompt context budget
  
  henrypark133
  and
  claude
  committed
  2888785
- test(turn-runner): pin that the derived budget sizes the host's outbound request
  
  henrypark133
  and
  claude
  committed
  55ff1c9
- test(integration): prove the model-derived context budget through a real turn
  
  henrypark133
  and
  claude
  committed
  d293fe1
- docs: record the context_length consumer and tidy review leftovers
  
  henrypark133
  and
  claude
  committed
  adfc802
- refactor(loop-contracts): move context_budget tests out of line
  
  henrypark133
  and
  claude
  committed
  c842ed0
- chore(architecture-tests): re-pin ironclaw_loop_contracts ceiling at 13773
  
  henrypark133
  and
  claude
  committed
  b687623
- docs(internal): design note and plan for the model-derived context budget
  
  henrypark133
  and
  claude
  committed
  ca68ac4
- fix(llm): keep model_metadata() free of token refresh I/O
  
  henrypark133
  and
  claude
  committed
  120402e
Commits on Sep 3, 2026
- refactor(agent-loop): move compaction strategy tests out of line
  
  henrypark133
  and
  claude
  committed
  ec9ead5
- ci(reborn-plan): register tests/trace_llm_tests.rs as a root test partition
  
  henrypark133
  and
  claude
  committed
  3ad5c72
- fix(loop-contracts): treat a window too small for any transcript as unknown
  
  henrypark133
  and
  claude
  committed
  2ced021
- test(integration): read the last captured model request, not index 5
  
  henrypark133
  and
  claude
  committed
  2b35d1e
- fix(llm): never advertise a guessed or too-large context window
  
  henrypark133
  and
  claude
  committed
  b1f172f
- fix(agent-loop): bound the compaction preserve tail by the run's visible budget
  
  henrypark133
  and
  claude
  committed
  0eb4701
- Merge origin/main into context-length
  
  henrypark133
  and
  claude
  committed
  03bf75c
- test(webui): use the live extension id in the notification-setup boundary test
  
  henrypark133
  and
  claude
  committed
  a2699f9
- fix(llm): FailoverProvider advertises the smallest window across its chain
  
  henrypark133
  and
  claude
  committed
  cd277c0
- test(loop-host): pin the window probe never refreshes a token, through the gateway
  
  henrypark133
  and
  claude
  committed
  c7af1f1
- refactor(llm): query each failover member once when advertising the window
  
  henrypark133
  and
  claude
  committed
  1f870b7
- ci(nextest): give the whole-tree architecture scans real timeout headroom
  
  henrypark133
  and
  claude
  committed
  f0cdace
- test(loop-host): pin the window probe's snapshot-identity and metadata-error paths
  
  henrypark133
  and
  claude
  committed
  16c51cc

Navigation Menu

Repository navigation

feat(loop): derive the prompt context budget from the model's advertised window - #8053

feat(loop): derive the prompt context budget from the model's advertised window#8053
henrypark133 wants to merge 29 commits into
mainfrom
context-length

Commits on Sep 1, 2026

feat(loop-contracts): derive a prompt context budget from an advertised window

feat(loop-contracts): carry a resolved context budget on LoopRunContext

refactor(agent-loop): pass the compaction budget as an argument

feat(agent-loop): honor a run-resolved context budget when compacting

feat(loop-host): expose the provider-advertised context window on the gateway

refactor(loop-host): thread a prompt context budget to the model port

feat(turn-runner): resolve the prompt context budget from the run's model

test(support): let the scripted model advertise a context window

Commits on Sep 2, 2026

test(loop-host): pin that the gateway wrapper applies its prompt context budget

test(turn-runner): pin that the derived budget sizes the host's outbound request

test(integration): prove the model-derived context budget through a real turn

docs: record the context_length consumer and tidy review leftovers

refactor(loop-contracts): move context_budget tests out of line

chore(architecture-tests): re-pin ironclaw_loop_contracts ceiling at 13773

docs(internal): design note and plan for the model-derived context budget

fix(llm): keep model_metadata() free of token refresh I/O

Commits on Sep 3, 2026

refactor(agent-loop): move compaction strategy tests out of line

ci(reborn-plan): register tests/trace_llm_tests.rs as a root test partition

fix(loop-contracts): treat a window too small for any transcript as unknown

test(integration): read the last captured model request, not index 5

fix(llm): never advertise a guessed or too-large context window

fix(agent-loop): bound the compaction preserve tail by the run's visible budget

Merge origin/main into context-length

test(webui): use the live extension id in the notification-setup boundary test

fix(llm): FailoverProvider advertises the smallest window across its chain

test(loop-host): pin the window probe never refreshes a token, through the gateway

refactor(llm): query each failover member once when advertising the window

ci(nextest): give the whole-tree architecture scans real timeout headroom

test(loop-host): pin the window probe's snapshot-identity and metadata-error paths