fix(pinch): cap outsized tool results even inside the protected window #44

alee · 2026-09-06T23:10:45Z

alee commented

2026-09-06 23:10:45 +00:00

keep_last_turns has no size limit inside it: an entire autonomous tool-call loop with no new user message counts as one protected turn, so one outsized tool result inside it -- a full verbose test run, a huge file read -- shipped verbatim regardless of size.

Measured live 2026-09-06: a 324k-token conversation only shrank ~8% because nearly all of it sat inside the protected window, and even after pruning was still ~6x pinch.budget_tokens. This matters beyond raw token cost too -- the pruned size is what feeds required_context_tokens (dispatcher.py measures it post-prune before tier/candidate selection), so a poorly-pruned conversation can also keep a request above a smaller, cheaper model's context ceiling that a properly-pruned one would have dropped below.

New pinch.protected_max_chars (default 20000, null to disable): any tool result inside the protected window over this many characters still gets the same head/tail elision candidates already get. Deliberately a much higher bar than max_summarize_chars (4000) -- recent results are more likely to still matter -- so this only catches true outliers, never ordinary recent tool output.

The pure prune_context function's own default stays None, so every existing caller/test is unaffected unless it opts in; PinchConfig supplies the real production default.

New regression tests confirmed failing (TypeError, param didn't exist) on the unfixed code, passing after. Full suite: 1473 passed, 0 failures.

keep_last_turns has no size limit inside it: an entire autonomous tool-call loop with no new user message counts as one protected turn, so one outsized tool result inside it -- a full verbose test run, a huge file read -- shipped verbatim regardless of size. **Measured live 2026-09-06**: a 324k-token conversation only shrank ~8% because nearly all of it sat inside the protected window, and even after pruning was still ~6x `pinch.budget_tokens`. This matters beyond raw token cost too -- the pruned size is what feeds `required_context_tokens` (dispatcher.py measures it post-prune before tier/candidate selection), so a poorly-pruned conversation can also keep a request above a smaller, cheaper model's context ceiling that a properly-pruned one would have dropped below. New `pinch.protected_max_chars` (default 20000, null to disable): any tool result inside the protected window over this many characters still gets the same head/tail elision candidates already get. Deliberately a much higher bar than `max_summarize_chars` (4000) -- recent results are more likely to still matter -- so this only catches true outliers, never ordinary recent tool output. The pure `prune_context` function's own default stays `None`, so every existing caller/test is unaffected unless it opts in; `PinchConfig` supplies the real production default. New regression tests confirmed failing (TypeError, param didn't exist) on the unfixed code, passing after. Full suite: 1473 passed, 0 failures.

alee added 1 commit 2026-09-06 23:10:45 +00:00

fix(pinch): cap outsized tool results even inside the protected window 06193b9144

keep_last_turns has no size limit inside it: an entire autonomous
tool-call loop with no new user message counts as one protected turn,
so one outsized tool result inside it -- a full verbose test run, a
huge file read -- shipped verbatim regardless of size.

Measured live 2026-09-06: a 324k-token conversation only shrank ~8%
because nearly all of it sat inside the protected window, and even
after pruning was still ~6x pinch.budget_tokens. This matters beyond
raw token cost too -- the pruned size is what feeds
required_context_tokens (dispatcher.py measures it post-prune before
tier/candidate selection), so a poorly-pruned conversation can also
keep a request above a smaller, cheaper model's context ceiling that
a properly-pruned one would have dropped below.

New pinch.protected_max_chars (default 20000, null to disable): any
tool result inside the protected window over this many characters
still gets the same head/tail elision candidates already get.
Deliberately a much higher bar than max_summarize_chars (4000) --
recent results are more likely to still matter -- so this only
catches true outliers, never ordinary recent tool output. The pure
function's own default stays None, so every existing caller/test is
unaffected unless it opts in; PinchConfig supplies the real default
so production gets the fix without a signature change elsewhere.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VRQXz5SYZYVWscxS1QqF6U

alee merged commit d2c5235c08 into main

2026-09-06 23:11:09 +00:00

alee referenced this issue from a commit

2026-09-06 23:11:11 +00:00

Merge pull request 'fix(pinch): cap outsized tool results even inside the protected window' (#44) from fix/pinch-protected-window-size-cap into main

alee referenced this issue from a commit

2026-09-07 02:15:15 +00:00

docs: refresh CLAUDE.md and companion docs for PRs #40-#47

alee referenced this pull request

2026-09-07 02:15:30 +00:00

docs: refresh CLAUDE.md and companion docs for PRs #40-#50 #51

Sign in to join this conversation.