title: Stateless JIT Context Injection for Agentic Git Workflows date: 2026-02-24 tags:

• context_management

• git

• workflow options: token_size: 1070 id: 26030 status: proposed superseded_by: null type: adr version: 1.0.0 birth: ‘2026-02-24’ authors:

• name: Vadim Rudakov email: rudakow.wadim@gmail.com description: 2026-02-24.

ADR-26030: Stateless JIT Context Injection for Agentic Git Workflows¶

Date¶

2026-02-24

Status¶

proposed

Context¶

Standard agentic workflows rely on stateful “sessions” where chat history is preserved across turns. In the context of local SLM deployment (1B–14B parameters) on Fedora/Debian systems, this “Recursive Context Accumulation” creates two critical failures:

1. Computational Debt: Accumulating history consumes VRAM and CPU cycles linearly per turn, leading to performance degradation and context window saturation for models like qwen2.5-coder or ministral.

2. State Drift: LLMs often “hallucinate” based on code discussed in previous turns that may have since been reverted or modified in the actual filesystem.

Furthermore, commercial API usage incurs a “Context Tax”—paying repeatedly for the same history tokens. To satisfy the requirements of ADR-26024 for a “computational/traceable artifact” (the CHANGELOG), the agent must have a perfectly accurate, high-fidelity view of the current staging area without the noise of conversational history.

Decision¶

We will implement the Stateless JIT (Just-In-Time) Injection pattern for all Git-interfacing agents. The agent shall be treated as a “Stateless Observer”—a pure function that maps current system state to a structured command.

1. Zero-Context Execution¶

• The agent session (memory buffer) is purged before every invocation.

• No user/assistant chat history is carried over between discrete Git operations.

2. Just-In-Time (JIT) State Assembly¶

The prompt payload is constructed at the moment of execution from the Single Source of Truth (SSoT):

• Scope: git status --short provides the high-level modified file list.

• Substance: git diff --cached provides the hunk-level changes to be summarized.

• Configuration: Domain constants (valid types, section mappings) are pulled from pyproject.toml [tool.commit-convention] per turn to ensure compliance with ADR-26029.

3. Format Enforcement¶

• The agent is prompted to output only a valid Git command or a JSON object containing the commit body.

• The output MUST follow the structured format: - Verb: file-path — what_and_why.

Consequences¶

Positive¶

• Token Economy: Reduces token consumption by ~70–90% in multi-commit workflows by eliminating history redundancy.

• Model Reliability: Small LMs (1B–14B) exhibit higher adherence to structure when the context window is focused solely on the current diff.

• Traceability: Each commit message is guaranteed to be derived from the actual git diff present at the time of generation, eliminating “narrative hallucinations”.

• Portability: Configuration is tied to the repository’s pyproject.toml, supporting the hub-and-spoke model.

Negative / Risks¶

• Loss of Implicit Intent: The agent does not “remember” the user’s verbal reasoning from five minutes ago. Mitigation: The wrapper script shall support an optional --intent or -m hint to pass high-level motivation into the JIT payload.

• Diff Bloat: Large refactors may exceed the context window. Mitigation: Wrapper script must detect token count and truncate/summarize the diff via git diff --stat if it exceeds model limits (e.g., 4k tokens).

Alternatives¶

• Option A: Stateful Sessions. Rejection Reason: Inherently scales poorly on local hardware; introduces VRAM exhaustion and non-deterministic “drift” in long sessions.

• Option B: Manual Release Note Synthesis. Rejection Reason: Superseded by ADR-26024; manual curation is non-scalable and lacks deterministic traceability.

• Option C: Vector Store (RAG) for History. Rejection Reason: Excessive complexity for local CLI tools; introduces a persistent state-tracking layer that violates the simplicity principle.

References¶

• ADR-26024: Structured Commit Bodies for Automated CHANGELOG Generation

• ADR-26029: pyproject.toml as Tool Configuration Hub

• ADR-26025: RFC→ADR Workflow Formalization

Participants¶

1. Vadim Rudakov

2. Gemini (Principal Systems Architect)