EQUIPA boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run examines EQUIPA as a shipped agent orchestration system, including its retained context and its separate improvement programs. The intended use is to explain its design and operating responsibilities, not to establish production effectiveness. The source is the acquired checkout at the full commit above. HEAD matches that commit and the checkout reports no working-tree changes. The analysis cutoff is 2026-10-15; later revisions, live deployments and linked external pages supply no evidence.

The target class is enclosing runtime: EQUIPA selects and dispatches agents, controls task iteration, admits results through review and merge paths, and records outcomes for later consumption. Claude Code supplies an inner agent execution engine; its internals are not EQUIPA's own implementation. EQUIPA also ships an Ollama tool-calling loop. This is not merely an MCP adapter or a lesson-retrieval API. The boundary is whole-system, including package entry points, root-level drivers, interactive session skills, prompts, standing orders, maintenance, evaluation and improvement tooling. These same-repository consumers are not treated as external hosts. Evidence: SRC-1, equipa/cli.py, equipa/agent_runner.py, ollama_agent.py, forgesmith.py, .claude/skills/forge-end/SKILL.md.

The entry-point trace starts at forge_orchestrator.py and equipa/__main__.py, which import the package CLI. Its mode registry includes single and parallel tasks, goals, automatic dispatch, initiatives, configuration and MCP serving. The MCP dispatch handler launches python -m equipa.cli; the arena driver launches the root orchestrator. Goal planning/evaluation surrounds task execution in equipa/manager.py; initiative waves reuse parallel task dispatch in equipa/initiative_runner.py. These are included shipped drivers, not excluded enclosing applications. Evidence: SRC-1, forge_orchestrator.py, equipa/__main__.py, equipa/cli.py, equipa/mcp_server.py, equipa/manager.py, equipa/initiative_runner.py, tools/forge_arena.py.

cmd = [sys.executable, "-m", "equipa.cli", "--task", str(task_id), "--role", role, "--yes"] --- equipa/mcp_server.py

The inspected operating path covers production, admission and later consumption of state. Task dispatch creates worktrees and routes execution toward security review and gated merge. The dev-test loop restores session/checkpoint context and constructs developer and tester calls. Post-task telemetry records status and runs, invokes scoring and reflection, and updates the values of injected episodes. Episode recording sanitizes agent-authored reflection before SQLite insertion; prompt construction retrieves lessons and episodes and tracks which were shown. Project-context loading also reads session notes, questions and decisions. Interactive session-end instructions write those same categories. Evidence: SRC-1, equipa/dispatch.py, equipa/loops.py, equipa/cli.py, equipa/lessons.py, equipa/prompts.py, equipa/sessions.py, equipa/tasks.py, .claude/skills/forge-end/SKILL.md.

update_task_status(task["id"], outcome, output=output, merged_sha=merged_sha) record_agent_run(task, result, outcome, role=role, model=model, max_turns=max_turns, cycle_number=cycle_number) --- equipa/cli.py

static_prefix = common_text + "\n\n---\n\n" + role_text + standing_orders --- equipa/prompts.py:508-508

The improvement plane is inside the boundary. forgesmith.py collects/analyzes runs, evaluates previous changes, extracts lessons, calls SIMBA, scores runs, applies changes, and conditionally invokes OPRO and GEPA; it also calls the LITM audit. Prompt construction imports GEPA's A/B prompt selector. Benchmarking and sweep programs are included as consumers of task execution and its recorded results. Their presence and source logic do not establish that a scheduled job ran or that a proposed prompt improved results. Evidence: SRC-1, forgesmith.py, equipa/prompts.py, forgesmith_gepa.py, forgesmith_litm.py, scripts/forgesmith_simba.py, scripts/autoresearch_loop.py, scripts/equipa_harness_sweep.py, tools/forge_arena.py.

findings, runs, blocked = run_analysis(cfg) --- forgesmith.py:3209-3209

Tracked top-level coverage

Every top-level tracked entry is classified below. Included means eligible within this system boundary, not that every file has been read or its behavior verified. Initial inspection traced the paths described above and the named source ranges; tree/function/search inspection established the remaining layout. Detailed internals not yet inspected remain available at the same frozen commit. No material shipped path was inaccessible. All rows refer to SRC-1.

External dependencies and evidence limits

The overall tier is code-grounded. Only source text was inspected; no target, tests, examples, services or model calls were executed. Target-authored claims of production use and improvement in SRC-1, README.md, remain reported operation. There is no supplied run trace or causal experiment, so this boundary cannot establish reliability, security isolation in deployment, realized cost, retrieval quality or prompt-improvement effects.

External dependencies include Claude Code and model services, Ollama model/embedding services, Python/SQLite, Git, configured MCP servers, project repositories and their build/test systems, and optional optimizer dependencies. Their implementation and deployed state are excluded: this prevents attributing their internal scheduling, inference, sandboxing, persistence or test validity to EQUIPA. EQUIPA's own launch, configuration, guard and result-handling logic stays included. Evidence: SRC-1, README.md, equipa/agent_runner.py, ollama_agent.py, equipa/loops.py, equipa_setup.py.

Operator databases, credentials, ignored local configuration, project worktrees, checkpoint/session files and remote deployment hosts are not supplied frozen evidence. Their absence prevents claims about actual tasks, stored memories, enabled features, installed plugins, live scheduling and success rates. This does not exclude shipped producers, maintainers or consumers of that state. Available but uninspected source files above are not access gaps; later jobs may inspect them at this same registered commit.

Source register