feat(context): spill oversized tool output to disk with a retrieval tool

[devin-ai-integration]

Summary

Completes the per-tool-output story from #880. Today an oversized tool result is truncated to a head+tail preview and the middle is discarded — which can drop the one line that matters (a buried grep hit, a stack frame, a leaked credential). This persists the full output and gives the agent a way to page back to it, so history stays bounded but nothing is lost.

Stacked on #881 → #880. Review/merge those first; the base retargets automatically.

strix/tools/output_store.py

def bound_and_store(text, *, max_lines, max_bytes) -> str:
    parts = _head_tail(text, max_lines, max_bytes)
    if parts is None: return text          # small: unchanged
    output_id = store_full_output(text)     # spill FULL text to the store
    # preview notice now carries the id:
    # [... N lines (M bytes) truncated — full output saved as output_id="ab12…";
    #  read it with read_tool_output(output_id="ab12…") ...]
    return _join(head, tail, notice)         # falls back to a plain preview if the write fails

def read_stored_output(output_id, *, offset=0, limit=2000) -> str: ...

output_id is a 32-char hex token, validated with ^[0-9a-f]{32}$ before touching the filesystem, so a crafted id can't escape the store dir. read_stored_output serves validated, paginated chunks and appends a "call again with offset=…" hint when more remains.

read_tool_output tool (strix/tools/tool_output/)

A new @function_tool registered for every scan agent. The agent calls read_tool_output(output_id, offset, limit) to retrieve any elided output by id and page through it.

Wiring

• strix/agents/factory.py: _bound_result now spills via bound_and_store (was bound_text); read_tool_output added to _BASE_TOOLS. Error strings still use the non-spilling bound_text (they're short).
• strix/core/runner.py: configure_output_store(state_dir / "tool-output") so spilled files live beside the rest of the run's .state and are cleaned up with it. Defaults to ~/.strix/tool-output if unconfigured.

The store dir is held in a module-level dict so it can be swapped per scan without a global rebind.

Test plan

tests/test_output_store.py adds: small output not spilled, oversized output spilled + fully retrievable (incl. a middle line absent from the preview but present in the retrieved full text), pagination hint/offset, path-traversal rejection, and missing-id handling. uv run pytest green (same 2 pre-existing unrelated test_runner_root_prompt.py failures as main); ruff + mypy + bandit pass.

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

This revision adds durable, bounded retrieval for oversized tool results and addresses the previously reported pagination defects.

• Spills complete oversized tool output into the scan’s state directory while retaining a bounded preview in agent history.
• Registers read_tool_output for all agents and exempts retrieval pages from generic result rebounding.
• Uses bounded byte-range reads with UTF-8 boundary handling, continuation-hint reservation, and explicit rejection of undersized page limits.
• Adds regression coverage for lossless long-line and multibyte pagination, response ceilings, arbitrary offsets, traversal rejection, and missing output IDs.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains; the current code fixes the previously reported retrieval re-truncation, full-file pagination, skipped content, oversized pages, UTF-8 corruption, and stalled-offset behaviors.

Important Files Changed

Filename	Overview
strix/tools/output_store.py	Implements per-scan output persistence and bounded byte-offset retrieval; the current pagination logic resolves the previously reported content-loss, UTF-8, progress, and response-ceiling issues.
strix/agents/factory.py	Registers the retrieval tool and safely exempts its pages from generic output bounding, resolving the prior re-truncation issue.
strix/core/runner.py	Configures the output store beneath the active run’s state directory.
strix/tools/tool_output/tool.py	Exposes the stored-output reader as a paginated agent-callable tool with byte-based offset and limit parameters.
tests/test_output_store.py	Adds focused regression tests covering spill persistence, bounded and lossless pagination, UTF-8 boundaries, continuation hints, and invalid identifiers.
tests/test_agent_factory_shell.py	Verifies retrieval results bypass the generic bounding wrapper while ordinary function tools remain bounded.

_{Reviews (12): Last reviewed commit: "fix(context): honor the caller's retriev..." | Re-trigger Greptile}

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[5hy7xz92nd-oss]

Deep research REALITY INGESTION LAYER GitHub Manus Base44 Supabase Notion APIs Financial Systems CRM Banking Data | v ENTITY GRAPH "What exists?" Company Person Product Customer Repository Document Financial Record AI Agent | v BEHAVIOR GRAPH "What is happening?" Events Actions Patterns Velocity Trends Changes Interactions Outcomes | v EVIDENCE GRAPH "Why believe it?" Claims Sources Documents Transactions Verification Provenance | v CONFIDENCE ENGINE "How certain are we?" Source Reliability Data Freshness Conflict Detection Historical Accuracy Confidence Score | v AI REASONING ENGINE "What does it mean?" Context Analysis Risk Detection Scenario Modeling Forecasting Opportunity Discovery | v DECISION ENGINE Funding Investment Credit M&A Risk Operations Automation | v HUMAN GOVERNANCE Review Approval Override Compliance | v OUTCOME LEARNING Repayment Growth Failure Performance Decision Accuracy | ↺ UPDATED INTELLIGENCE MODEL@Manus @datacamp @call-e REALITY INGESTION LAYER GitHub Manus Base44 Supabase Notion APIs Financial Systems CRM Banking Data | v ENTITY GRAPH "What exists?" Company Person Product Customer Repository Document Financial Record AI Agent | v BEHAVIOR GRAPH "What is happening?" Events Actions Patterns Velocity Trends Changes Interactions Outcomes | v EVIDENCE GRAPH "Why believe it?" Claims Sources Documents Transactions Verification Provenance | v CONFIDENCE ENGINE "How certain are we?" Source Reliability Data Freshness Conflict Detection Historical Accuracy Confidence Score | v AI REASONING ENGINE "What does it mean?" Context Analysis Risk Detection Scenario Modeling Forecasting Opportunity Discovery | v DECISION ENGINE Funding Investment Credit M&A Risk Operations Automation | v HUMAN GOVERNANCE Review Approval Override Compliance | v OUTCOME LEARNING Repayment Growth Failure Performance Decision Accuracy | ↺ UPDATED INTELLIGENCE MODEL@GitHub @Malwarebytes @heygen