fix(agent): reserve output tokens in compression threshold

[lEWFkRAD]

What does this PR do?

Makes proactive context compression reserve the response-token budget that the outgoing request will send. When a provider/profile default like provider: custom reserves max_tokens=65536, the compressor now compares prompt estimates against (context_length - output_reservation) * compression.threshold instead of the raw context window.

Fixes #43547

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Test-only change
• Refactor

Changes Made

• agent/context_compressor.py derives threshold/tail budgets from an output-reservation-aware input budget when a response cap is known.
• run_agent.py resolves the same output cap precedence used by request construction: request overrides, ephemeral retry caps, user model.max_tokens, Anthropic defaults, provider profile defaults, and Claude aggregator defaults.
• agent/agent_init.py, agent/turn_context.py, agent/agent_runtime_helpers.py, and agent/chat_completion_helpers.py sync the reservation at init, preflight, model switch, and fallback activation.
• tests/test_ctx_halving_fix.py adds regression coverage for the custom provider 65536 default against a 131072 context window.

How to Test

1. scripts/run_tests.sh tests/test_ctx_halving_fix.py
2. scripts/run_tests.sh tests/agent/test_context_compressor.py
3. scripts/run_tests.sh tests/run_agent/test_switch_model_context.py tests/run_agent/test_compressor_fallback_update.py
4. ruff check agent/context_compressor.py run_agent.py agent/agent_init.py agent/turn_context.py agent/agent_runtime_helpers.py agent/chat_completion_helpers.py tests/test_ctx_halving_fix.py
5. git diff --check

Checklist

• I have read the Contributing Guide
• My changes follow the project's conventional commit format
• I have searched existing issues/PRs for duplicates
• This PR is focused and minimal
• I have run tests with scripts/run_tests.sh
• I have added or updated tests for behavior changes
• I have tested on the relevant platform (WSL/Linux runner on Windows host)

Documentation / Housekeeping

• N/A - no documentation changes needed

For New Skills

• N/A - no skill changes

Screenshots / Logs

N/A

[liuhao1024]

Verified — clean.

Reviewed the full diff across context_compressor.py, run_agent.py, agent_init.py, agent_runtime_helpers.py, chat_completion_helpers.py, turn_context.py, and tests.

What was checked:

• Output reservation correctly subtracted from context_length before threshold percentage: input_budget = max(context_length - output_reservation, 1) → threshold = max(input_budget * threshold_percent, 1)
• When output_reservation_tokens = 0, behavior reverts to original context_length * threshold_percent path (no minimum floor applied — this is intentional for the reservation-aware path)
• _resolved_output_reservation_tokens() cascades correctly: override_cap → ephemeral_cap → user_cap → anthropic/provider profile → fallback 0
• Anthropic cap guard: if cap >= context_length: cap = max(context_length - 1, 1) prevents reserving the entire window
• All 4 call sites (agent_init, switch_model, try_activate_fallback, turn_context) call _sync_context_compressor_output_reservation after compressor creation/update
• _coerce_output_reservation_tokens handles bool, overflow, None, negative values
• Tests validate: reservation reduces threshold correctly, custom provider default caps propagate to compressor

Why this matters: Without output reservation, proactive compression triggers at 50% of context_length (e.g. 64K of 128K). If the model reserves 32K for output, the effective input budget is only 96K — compression should trigger at 48K, not 64K. This fix prevents context overflow on models with large output caps.

Good change. No issues found.

[tonydwb]

Code Review Summary

Verdict: Approved

Clean, well-scoped fix/feature with comprehensive tests. No issues found.

• Logic is correct and focused
• Tests cover the new behavior
• No security concerns
• Good error handling

---

Reviewed by Hermes Agent

[alt-glitch]

Duplicate of #43651 — same fix (reserve the output-token budget in the proactive compaction threshold, comparing against (context_length - output_reservation) * threshold) for #43547. #43651 is the earlier-open canonical.

[lEWFkRAD]

Flagged as duplicate of #43651 by @alt-glitch. Closing this one in favor of the earlier PR.

[lEWFkRAD]

Closing in favor of #43651 which contains the same fix.