feat(context): bound per-tool output before it enters agent history

[devin-ai-integration]

Summary

Restores the pre-v1 per-tool-output truncation that was lost in the OpenAI-Agents-SDK migration. Today a single verbose tool result (a recursive find, a noisy scanner, a full page dump) is appended raw to the SQLiteSession and replayed on every subsequent turn — pinning the conversation near the model's context window for the rest of the scan (and, once the window is exceeded, causing a hard failure since there's no compaction). This PR caps each result before it enters history.

This is layer 1 of a two-layer context-management effort; a follow-up will add provider-agnostic conversation compaction (the summarizer removed in 369fa56). This PR is independently useful and low-risk: caps are generous, so normal outputs are untouched.

Three parts:

1. New ContextSettings config group (strix/config/settings.py) — all env-tunable, with headroom for the upcoming compaction layer:

tool_output_max_tokens = 8_000   # STRIX_TOOL_OUTPUT_MAX_TOKENS  (SDK shell cap)
tool_output_max_lines  = 2_000   # STRIX_TOOL_OUTPUT_MAX_LINES
tool_output_max_bytes  = 50*1024 # STRIX_TOOL_OUTPUT_MAX_BYTES
# (+ auto_compact / buffer / keep / fallback / summary knobs reserved for the follow-up)

2. Default the SDK shell tools' own cap (factory.py). exec_command / write_stdin already accept max_output_tokens and truncate head+tail internally, but Strix never set it (⇒ None ⇒ unbounded). We now inject the configured default when the model omits it, and respect an explicit model-supplied value:

def _apply_shell_output_cap(parsed):
    if parsed.get("max_output_tokens") is None:
        parsed["max_output_tokens"] = load_settings().context.tool_output_max_tokens

3. Bound Strix's own tool results (strix/tools/output_store.py + factory.py). Every FunctionTool/CustomTool result (and capped error strings) passes through bound_text, which returns the text unchanged when small, else a head+tail preview joined by a notice recording what was dropped:

<first ~half of the lines/bytes>

[... 1234 lines (56789 bytes) truncated ...]

<last ~half of the lines/bytes>

Truncation triggers on whichever limit (line count or UTF-8 byte size) is hit first, and the byte split is multibyte-safe (never cuts a character in half). The base-tool wrapper is idempotent (base tools are shared singletons reused across every agent build).

Test plan

• New tests/test_output_store.py: passthrough for small output, head+tail on line limit, byte-limit enforcement on a single long line, multibyte safety, dropped-count notice.
• tests/test_agent_factory_shell.py: asserts the injected max_output_tokens default and that an explicit value is preserved.
• uv run pytest green (2 pre-existing unrelated failures in test_runner_root_prompt.py exist on main too — a broken SimpleNamespace settings mock). ruff + mypy + bandit pass via pre-commit.

Link to Devin session: https://app.devin.ai/sessions/3461e7dce476402aa2091e4af3f55c00
Requested by: @0xallam

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Greptile Summary

The PR restores bounded per-tool output before results enter persistent agent history.

• Adds validated, environment-configurable context and tool-output limits.
• Clamps SDK shell output tokens and bounds FunctionTool and native CustomTool results.
• Preserves head and tail output within the configured byte ceiling and reports dropped content.
• Adds coverage for shell caps, native filesystem tools, configuration validation, byte safety, and truncation accounting.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
strix/agents/factory.py	Applies configured output bounds across shell, FunctionTool, and native filesystem CustomTool paths; the previously reported bypasses are addressed.
strix/config/settings.py	Introduces context-management settings with positive-value validation and a safe minimum byte ceiling.
strix/tools/output_store.py	Implements UTF-8-safe head-and-tail truncation while accounting for metadata within the byte cap and reporting final retained-line counts.
tests/test_agent_factory_shell.py	Covers default and oversized shell caps plus native and chat-completions filesystem tool handling.
tests/test_config_loader.py	Verifies rejection of byte ceilings too small to contain truncation metadata.
tests/test_output_store.py	Covers passthrough, line and byte truncation, multibyte safety, byte-cap enforcement, and dropped-line accounting.

_{Reviews (6): Last reviewed commit: "fix(context): reject tool-output byte ce..." | Re-trigger Greptile}

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile