Ephemeral Buffer MCP Server (ephemeral-buffer)

An ephemeral in-memory command output capture and hybrid search engine (BM25 + Semantic Embeddings) for AI coding assistants (Claude Code, Antigravity, Cursor, etc.).

🎯 The Problem This Solves

When coding agents run commands that generate large outputs (thousands of lines of build logs, test runs, stack traces, JSON dumps), agents face two failure modes:

1. Context Pollution: Ingesting megabytes of raw text blows out token limits and degrades model reasoning.

2. Blind Bash Filtering: Agents waste multiple turns running head, tail, grep, and awk trying to guess error patterns.

Related MCP server: Qurio MCP Server

💡 The Solution

ephemeral-buffer provides a transient in-memory ring buffer with Dual Hybrid Indexing and Content-Aware Structure Parsing:

• BM25 Lexical Search (SQLite FTS5): For exact matches on error codes (NullPointerException, ECONNREFUSED, exit 137, HTTP 502). Builds without SQLite FTS5 use a complete token-based Python fallback with lower ranking performance.

• Dense Semantic Vector Search (FastEmbed ONNX): For fuzzy conceptual queries ("Where did the DB connection pool fail?" or "Why did authentication fail?").

• Unified Diff Structural Mapping: Automatically detects git diffs and PR diffs (gh pr diff, git show, git diff), parses modified files, additions/deletions, and generates a line-indexed file map in the summary.

• Smart Signal Filtering: Scans command/build/test logs for diagnostic keywords, suppresses false positives in diffs and source code, and accurately captures test runner failures, unhandled exceptions, and merge conflicts. Use content_type='log' when a plain-text capture should be signal-scanned.

• Successful test-run summaries such as OK or 25 passed suppress fixture-only error and failure keywords while retaining the original output for search.

• Reciprocal Rank Fusion (RRF): Blends lexical and semantic ranking for high precision retrieval.

• LRU Capture Eviction: Holds up to 25 captures and 50 MiB of captured content by default, evicting the least recently used captures when either limit is reached.

• Thread-Safe Shared Engine: Serializes ingestion, search, LRU updates, eviction, and cleanup across MCP requests and CLI socket clients.

📦 Installation

Requirements

• Python 3.10 or newer

• A supported MCP client if you want to use the server from an AI coding assistant

• Network access on first use if FastEmbed needs to download its embedding model

The package is installed from PyPI as ephemeral-buffer-mcp. It provides both the MCP server and the ephbuf command-line client.

Recommended: install from PyPI

Create an isolated virtual environment and install the latest published package:

python3 -m venv .venv
.venv/bin/python -m pip install --upgrade pip
.venv/bin/python -m pip install ephemeral-buffer-mcp

Verify that the CLI is available:

.venv/bin/ephbuf --help

On Windows, use the equivalent commands from the virtual environment's Scripts directory:

py -3 -m venv .venv
.venv\Scripts\python.exe -m pip install --upgrade pip
.venv\Scripts\python.exe -m pip install ephemeral-buffer-mcp
.venv\Scripts\ephbuf.exe --help

To upgrade or remove the package:

.venv/bin/python -m pip install --upgrade ephemeral-buffer-mcp
.venv/bin/python -m pip uninstall ephemeral-buffer-mcp

Install from a source checkout

Use an editable install when developing or testing local changes:

git clone https://github.com/k-rister/ephemeral-buffer-mcp.git
cd ephemeral-buffer-mcp
python3 -m venv .venv
.venv/bin/python -m pip install --upgrade pip
.venv/bin/python -m pip install -e .

The editable install exposes the same ephbuf command and MCP server as the PyPI installation. The reproducible, locked contributor environment is documented in Testing the Server.

When launching from a source checkout with run.sh, the launcher prefers .venv/bin/python, then supports the legacy venv/bin/python layout, before using an explicit PYTHON override or python3. An invalid PYTHON override fails with an actionable error.

Start the MCP server manually

The MCP server uses stdio for communication with the MCP host. Start it with the Python interpreter from the environment where the package was installed:

.venv/bin/python -m server

Normally you should let your MCP client start this process automatically. Do not start a separate server for every shell command: the ephbuf CLI sends captures to the running server over its local Unix socket. The private CLI socket uses versioned length-prefixed request and response frames, so fragmented reads do not depend on half-closing the connection.

Start an isolated Codex session

When using the global Codex MCP configuration, start Codex through the installed codex-ephemeral launcher. From a source checkout, use ./codex-ephemeral. The launcher creates a unique EPHEMERAL_SESSION_ID and exports it to Codex, the MCP server, and ephbuf. It also passes the ID explicitly through Codex's MCP configuration so it is available even when Codex sanitizes the MCP child environment. The global configuration requires this identity, so starting Codex directly will fail closed instead of attaching to another session's socket.

The launcher also enables EPHEMERAL_ALLOW_STDIO_WITHOUT_SOCKET=1 because some Codex execution environments deny Unix-socket creation. In that case MCP over stdio remains available and runtime diagnostics report the socket failure; ephbuf CLI support remains available when the environment permits socket creation.

Configure other coding agents

The server is not Codex-specific. Any MCP client that launches a local stdio server can use the same command and environment settings:

{
  "mcpServers": {
    "ephemeral-buffer": {
      "command": "/absolute/path/to/ephemeral-buffer-mcp/run.sh",
      "args": [],
      "env": {
        "EPHEMERAL_REQUIRE_ISOLATION": "1",
        "EPHEMERAL_SESSION_ID": "agent-session-unique-value",
        "EPHEMERAL_ALLOW_STDIO_WITHOUT_SOCKET": "1"
      }
    }
  }
}

This shape applies to clients such as Claude-style .mcp.json and Gemini/Antigravity-style mcp_config.json. Use a different EPHEMERAL_SESSION_ID for every concurrent agent session. If the client supports environment interpolation, use its per-session identifier; otherwise generate an ID when starting the agent and place that same value in the agent's shell environment:

export EPHEMERAL_SESSION_ID="agent-$(date +%s)-$$"
export EPHEMERAL_REQUIRE_ISOLATION=1
export EPHEMERAL_ALLOW_STDIO_WITHOUT_SOCKET=1

The MCP server and ephbuf CLI must inherit the same ID. An MCP client that does not pass environment variables to child processes can still use the MCP tools, but its separate ephbuf shell commands will need an equivalent session environment configured independently.

For other shell-launched agents, use the installed generic launcher. From a source checkout, use ./ephemeral-agent:

ephemeral-agent claude
ephemeral-agent gemini

Or source the environment into an existing shell before launching the agent:

source ephemeral-session-env
claude   # or another coding-agent command

GUI-launched agents still need their MCP configuration to provide a unique session ID per conversation; these scripts cannot modify an already-running GUI process.

Configure an MCP client

MCP clients generally need the command, arguments, and environment used to start a stdio server. A portable configuration looks like this:

{
  "mcpServers": {
    "ephemeral-buffer": {
      "command": "/absolute/path/to/ephemeral-buffer-mcp/.venv/bin/python",
      "args": ["-m", "server"]
    }
  }
}

Replace the command with the absolute path to the virtual-environment Python interpreter. Using an absolute path avoids differences between GUI-launched clients and interactive shells. On Windows, use a path such as C:\\path\\to\\ephemeral-buffer-mcp\\.venv\\Scripts\\python.exe.

For clients that provide a command-line registration command, register the same executable and arguments conceptually as:

command: /absolute/path/to/.venv/bin/python
arguments: -m server

After saving the configuration, restart or reload the MCP client and confirm that the ephemeral-buffer tools appear. The server exposes capture, search, summary, slice, consolidation, diagnostics, and cleanup tools; the ephbuf CLI is a separate convenience client for shell output.

Isolate concurrent agent sessions

For a coding agent that may run alongside another agent, configure the same session identity for the MCP server and the ephbuf CLI:

"env": {
  "EPHEMERAL_REQUIRE_ISOLATION": "1",
  "EPHEMERAL_SESSION_ID": "agent-session-1"
}

The server and CLI then derive the same session-specific socket, while a missing identity fails closed instead of falling back to the shared legacy socket. EPHEMERAL_SOCKET_PATH may be used instead when the launcher assigns the socket path directly. The MCP initialization instructions describe this policy to the client, but the environment checks enforce it independently.

The CLI bounds each socket connect, send, and receive operation to 10 seconds by default. Set EPHEMERAL_SOCKET_TIMEOUT_SECONDS to a positive number of seconds when a different limit is appropriate; timeout failures return a nonzero CLI result.

Background model warm-up

After socket startup succeeds, the server loads FastEmbed and runs a small deterministic embedding in a background thread. MCP startup and BM25 search do not wait for this work. A semantic request arriving during warm-up waits on the same model lock instead of starting a duplicate load. Subsequent operations use the local model cache. To disable warm-up for lexical-only or memory-constrained deployments, or to select a compatible model and cache location, set:

export EPHEMERAL_EMBEDDING_WARMUP=0
export EPHEMERAL_EMBEDDING_MODEL="BAAI/bge-small-en-v1.5-fp32"
export EPHEMERAL_FASTEMBED_CACHE_DIR="$HOME/.cache/ephemeral-buffer"

The default model is BAAI/bge-small-en-v1.5-fp32, the upstream fp32 ONNX export of bge-small-en-v1.5 (about 130 MB), which the server registers with FastEmbed itself. FastEmbed's own catalogue entry, selectable as BAAI/bge-small-en-v1.5, is a smaller reduced-precision export that produces identical vectors but whose matrix kernels do not parallelize on common CPU hosts. In a release-preparation comparison on a 16-vCPU Linux x86_64 host, the fp32 export reached 69.1 semantic chunks per second with 1.85 seconds median indexing at 1,024 lines, versus 3.1 chunks per second and 41.15 seconds for the catalogue file. That roughly 22x difference is host-specific evidence, not a universal performance guarantee. Embedding inference uses ONNX Runtime's default thread count; bound it with EPHEMERAL_EMBEDDING_THREADS when the server shares a host with an interactive coding agent, and check both settings on the deployment host with benchmark_semantic_index.py. get_buffer_stats reports the active thread setting.

export EPHEMERAL_EMBEDDING_THREADS=4

Warm-up failures do not stop the server. BM25 remains available, and hybrid search returns lexical results with a semantic_fallback error-class field when semantic initialization fails. get_buffer_stats and get_runtime_diagnostics report warm-up as not-started, loading, ready, failed, or disabled; failures expose only the exception class.

Lexical and semantic search index different chunk grids. BM25 keeps four-line sliding windows with two-line overlap so exact line ranges stay tight. Semantic embeddings use separate windows packed from consecutive lines, closed at eight lines or 1,024 UTF-8 bytes, whichever comes first, with no overlap. That keeps every window inside the model's token limit and, because embedding cost tracks total tokens, roughly halves the work by not embedding the overlap twice. Hybrid ranking fuses the two grids in line space: a semantic window boosts the lexical hits it overlaps and appears on its own only when nothing lexical matched inside it. Each match reports chunk_index as lexical or semantic alongside its line range. Tune the windows when output has very long or very short lines:

export EPHEMERAL_SEMANTIC_CHUNK_LINES=8
export EPHEMERAL_SEMANTIC_CHUNK_BYTES=1024
export EPHEMERAL_SEMANTIC_CHUNK_OVERLAP=0

Larger windows embed fewer chunks but dilute single-line signals and cost more per token; overlap improves recall across window boundaries at proportional embedding cost. Compare strategies on the deployment host with benchmark_semantic_index.py --semantic-chunk-lines ... --semantic-chunk-overlap ....

Semantic indexing is prefetched after ingestion by default, so a search that arrives after the background job finishes pays no indexing cost. Disable it for lexical-only or CPU-constrained hosts, or bound the worker pool:

export EPHEMERAL_SEMANTIC_PREFETCH=0
export EPHEMERAL_SEMANTIC_PREFETCH_WORKERS=1

Every eligible capture is queued at ingestion and a bounded worker pool drains the queue newest-first, since the latest capture is the most likely search target; a burst of captures is never silently skipped. A semantic or hybrid search waits for a job that is already running, and pulls a still-queued capture out of the queue to index it on a separate on-demand pool so it does not wait behind older work. That pool is bounded by the same worker count, so searches that give up waiting on captures that are then evicted cannot accumulate indexing threads. Failed jobs retry through the normal lazy path, and evicted or cleared captures drop their queued prefetch and on-demand work. Embedding inference is serialized by one model lock, so extra workers only overlap bookkeeping; tune EPHEMERAL_EMBEDDING_THREADS instead of the worker count for throughput. get_buffer_stats and get_runtime_diagnostics expose only aggregate pending, queued, running, and failed counts.

Indexing cost grows linearly with capture size, so a hybrid search that arrives before a large capture's index is ready waits at most a configurable budget:

export EPHEMERAL_SEMANTIC_WAIT_SECONDS=10

When the budget expires, hybrid search returns the BM25 results immediately with semantic_coverage set to pending (the tool response says semantic pending (lexical only)), and indexing continues in the background so repeating the search returns full hybrid ranking. Every other hybrid response reports semantic_coverage as complete, or unavailable when the semantic backend failed and semantic_fallback names the exception class; BM25 results and exact line ranges are identical either way. The default of 10 seconds kept a 2,048-line capture fully hybrid in the release-preparation Linux run (about 4.0 seconds to index), while 8,192- and 16,384-line captures took about 17.0 and 32.9 seconds to index and therefore answered lexical-first within the budget. Measured first-search p95 was about 4.0 seconds at 2,048 lines (complete) and 10.002 seconds at the larger sizes (pending); every needle in the benchmark fixture still ranked first. These figures are host-specific: Apple silicon and other Linux hosts can differ materially, so run benchmark_semantic_index.py on the deployment host before choosing a wait budget. Set 0 to always answer lexical-first while the index builds, or inf to wait for the index unconditionally. Semantic mode has no lexical result to fall back on, so it always waits for the index. An empty capture has nothing to index, so it reports complete coverage for semantic and hybrid searches. get_buffer_stats reports the budget and the number of on-demand index jobs running and queued.

The exact model and cache location can also be supplied in the MCP client's env configuration. Keep the model cache writable by the user running the MCP client.

Installation choices at a glance

If the MCP client cannot find the server, check the absolute interpreter path, the selected Python environment, and the client's server logs. For socket, capture-limit, logging, and deployment troubleshooting, see OPERATIONS.md.

🏗 Architecture & Flow

flowchart TD
    subgraph Ingestion["1. Ingestion & Diagnostics"]
        A["CLI Pipe: command 2>&1 | ephbuf"] --> D["Unix Socket (platform temp dir)"]
        B["Agent Tool: execute_and_capture(cmd)"] --> E["Ephemeral Ring Buffer Engine"]
        C["Agent Tool: capture_text / capture_file"] --> E
        D --> E
        P["Agent Tool: preflight_command(cmd, cwd)"] --> Q["Path, executable & repository diagnostics"]
    end

    subgraph Execution["2. Durable Phase Execution"]
        X["Agent Tools: start_execution / resume_execution"] --> Y["Phase Execution Manager"]
        Y --> Z["Persistent Execution State & Bounded Phase Output"]
        Y --> E
        Z --> AA["get_execution / get_execution_output / list_executions"]
    end

    subgraph Indexing["3. Classification & Search Indexing"]
        E --> F["SQLite FTS5 (BM25 Lexical) or Python lexical fallback"]
        E --> G["Optional/background FastEmbed ONNX (Dense Vectors)"]
        E --> K["Diff & Signal Parser (File Maps & Conflict Detection)"]
        E --> M["consolidate_captures: bounded source-aware JSON"]
        M --> E
    end

    subgraph Querying["4. Agent Query & Retrieval"]
        F & G --> H["Reciprocal Rank Fusion (RRF)"]
        H --> I["search_capture(query, mode='hybrid')"]
        K --> L["get_capture_summary: diff stats & bounded file map"]
        L --> N["get_capture_slice: exact lines"]
        I --> J["Precise Context Chunk + Line Numbers"]
    end

🚀 How to Use It

1. From the Terminal (CLI Pipe via ephbuf)

You can pipe command output directly into the running MCP server:

# Pipe any command output into the buffer
pytest -v 2>&1 | ephbuf --label "pytest run"

# Pipe git diffs directly
git diff HEAD~3 | ephbuf --label "feature diff" --type diff

# Or wrap command execution
ephbuf --label "backend build" -- cargo build --verbose

The optional --type/-t hint accepts auto (the default), diff, log, or text. Use diff for unified patches when automatic detection is ambiguous; otherwise auto classifies diffs, build/test logs, and plain text from the content and label.

ephbuf also bounds wrapped-command and piped-stdin capture with --max-output-bytes; it defaults to EPHEMERAL_MAX_BUFFER_BYTES or 50 MiB and retains the beginning and end of oversized output. Use --timeout-seconds to stop a wrapped command after a bounded runtime; timed out commands retain the output collected so far and exit with status 124. Requested max_output_bytes and capture_file max_bytes values may not exceed the configured buffer byte limit; the tools return a validation error instead of silently clamping them.

2. From the AI Agent via MCP Tools

The agent has access to the following tools:

Capture tools return a compact JSON summary so an agent can decide whether it needs the full output before spending context on retrieval. The summary uses schema_version: 1 and includes status (captured, success, failed, or timed_out), duration_ms, retained and original byte sizes, approximate token counts, truncation and partial-execution flags, typed warning/error signals, and bounded caller-provided structured_metrics. The token values are deterministic planning estimates based on four UTF-8 bytes per token; they are not provider billing counts. Use get_capture_summary for the compact form, set include_previews=True only when a head/tail sample is useful, and use get_capture_slice or search_capture for complete or targeted content. Diff file maps are also bounded and report omitted entries; use get_capture_slice for the complete diff. Raw retained captures are unchanged by summary generation.

The deterministic summary benchmark measures the initial agent-prompt reduction for representative successful, failed, noisy, truncated, and timed-out captures. It exercises the public capture API, compares a parent-contract formatted-text response with the current compact summary-first decision prompt, and verifies through the public slice API that retained output can still be retrieved unchanged. The parent response is reconstructed from the prior execute_and_capture contract; the current path uses the public compact response. The proxy divides UTF-8 prompt bytes by four and is intended for regression comparison, not provider billing or exact token accounting. Execution summaries also bound command and label metadata so unusually long shell commands cannot re-expand the initial response.

To generate the machine-readable benchmark record:

.venv/bin/python benchmark_effectiveness.py --summary --output /tmp/capture-summary.json

Choose the execution path based on the output and inspection goal:

• Use direct command execution for a small, targeted inspection where the output is already bounded and immediate terminal feedback is sufficient.

• Use execute_and_capture once output may be noisy, large, or uncertain—such as tests, builds, and logs—because it bounds context and makes later search and exact retrieval available. This is an advisory heuristic, not a hard line-count policy.

• Use capture_text when output is already in hand, or capture_file for a file that has been checked and intentionally selected for ingestion.

Use preflight_command when repository identity or path resolution is uncertain before a sensitive command. It reports resolved facts and explicit unavailable states without running the requested command or exposing command output. It cannot predict shell expansion, aliases, pipelines, redirections, environment changes, or arbitrary shell logic, so normal command validation and user intent checks remain necessary.

Resumable phase execution

Durable phase execution and its process-group recovery currently require Linux with file-locking support and /proc process identities. The package's Windows installations remain usable for MCP stdio and text/file capture. Bounded subprocess capture requires POSIX pipe and process-group support, and start_execution additionally requires Linux leases, /proc process identities, pidfd signaling, and selector support; startup rejects the request with a clear platform error when those recovery backends are unavailable.

Use start_execution when a long-running workflow has meaningful checkpoints:

start_execution(
  execution_id="release-checks",
  phases=[
    {"name": "tests", "command": "python -m unittest", "timeout_seconds": 900},
    {"name": "publish", "command": "./publish.sh", "side_effects": "unsafe"},
  ],
)

Each phase is persisted as pending, started, completed, failed, interrupted, or timed_out. Output, exit status, duration, truncation, and caller metrics are written after every finished phase. If the server restarts while a phase is started, the next inspection records it as interrupted. resume_execution skips every completed phase and continues at the first incomplete phase. Failed and timed-out phases require retry_failed=True; safe phases recovered as interrupted resume automatically. A timed-out phase whose process-group cleanup is not confirmed remains fence-pending and blocks retry. An interrupted phase marked side_effects: "unsafe" additionally requires confirm_unsafe=True unless the execution was created with the explicit resume_policy="allow-unsafe". The persisted idempotency_key is an audit boundary for an external operation; it does not replace confirmation or provide an external deduplication guarantee. The compatibility alias unsafe_side_effects: true is normalized to side_effects: "unsafe"; if both fields are supplied, they must agree. The command is run beneath a Linux subreaper supervisor that adopts and terminates descendants which escape the original process group. The supervisor identity is checkpointed while a phase runs and is signalled through a pidfd before restart recovery permits that phase to resume. Linux process start and boot identities are revalidated while the pidfd is pinned, protecting recovery from signalling a reused process ID; resume stays blocked if supervisor termination or group absence cannot be confirmed. A phase marks its launch fence before spawning the command, so a crash before process identity is persisted also fails closed. Older in-progress records without these identity fields are recovered as fence-pending and must be inspected or retired rather than being retried automatically.

Execution responses include a human-readable summary, machine-readable execution_status and partial fields, per-phase event history, and the next resumable phase. If detailed metadata would exceed the 64 KiB tool response budget, the server returns a compact response with the durable execution_id and response_truncated: true; call get_execution or the bounded output tool to retrieve details. Durable JSON state defaults to a temporary local process-local directory created securely with owner-only permissions; set EPHEMERAL_SESSION_ID, EPHEMERAL_SOCKET_PATH, or EPHEMERAL_EXECUTION_STATE_DIR to persist and share state across server restarts. State can otherwise be placed elsewhere with EPHEMERAL_EXECUTION_STATE_DIR. State contains the commands and bounded outputs, so keep any explicitly configured directory protected when commands or results are sensitive.

Execution metadata is bounded to 64 MiB per record, 64 phases, 32 attempts per phase, 16 KiB of structured metrics, and 1,000 records per state directory; list results are paginated with a maximum page size of 100. There is no automatic expiry: when the record cap is reached, stop the server and archive or rotate the state directory, or remove completed records together with their matching summary files before restarting. State and execution leases are isolated by the explicit execution-state directory, session ID, or socket path; without one, each server process receives a fresh private state directory that is removed during normal shutdown on POSIX platforms. Windows may retain that temporary directory because secure owner-identity cleanup is not available there.

Before any repository-sensitive command or file capture, verify the intended working directory and target path. Prefer an explicit cwd, confirm the repository identity, and resolve symlinks when path identity matters. Shell expansion, inherited working directories, and symlinks can target a different location than the spelling suggests. Capture limits protect context size; they do not validate command intent, path identity, or filesystem safety.

For diff captures, get_capture_summary reports the detected file map, addition/deletion statistics, line ranges, and merge-conflict signals. Use get_capture_slice with those ranges to retrieve the complete file context.

Consolidating multi-result workflows

When a workflow produces several captures—for example, one command per repository—use consolidate_captures to give the agent one bounded overview:

consolidate_captures(
  capture_ids=["cap_1", "cap_2", "cap_3"],
  label="organization activity",
  max_captures=25,
  max_bytes=20000
)

The resulting capture is JSON with source metadata and records containing the original capture_id and source_line. It can be searched normally with search_capture, and exact consolidated context can be retrieved with get_capture_slice. The response reports omitted records, missing IDs, and the original source IDs; use those original IDs to retrieve complete detail when the consolidated byte budget is reached. Calling the tool without capture_ids consolidates the currently active captures, up to max_captures.

This workflow keeps the server responsible for bounded execution, storage, and retrieval while leaving prioritization and interpretation to the coding agent.

3. Capture Hygiene

Keep captures focused so search results remain useful and the agent receives only the context it needs:

• Capture one command or related output stream at a time, using a descriptive label.

• Start with get_capture_summary, then use search_capture or get_capture_slice for targeted retrieval instead of repeatedly recapturing the same output.

• Use clear_captures(capture_id) when a capture is no longer needed; use clear_captures("all") between unrelated investigations.

The buffer is intentionally transient and bounded by the LRU capture limit, but explicit cleanup prevents recent investigations from obscuring the active one before automatic eviction occurs. Its memory metrics separate captured content and embedding bytes from process RSS; the unaccounted RSS value includes model, index, and Python object overhead and is approximate.

The server defaults can be overridden with EPHEMERAL_MAX_CAPTURES and EPHEMERAL_MAX_BUFFER_BYTES. Session-aware launchers can set EPHEMERAL_SESSION_ID so each server/CLI pair automatically derives a unique socket path; EPHEMERAL_SOCKET_PATH remains an explicit override. The byte limit accounts for captured UTF-8 content plus its label; a capture is rejected when their combined size exceeds the limit. get_buffer_stats also reports embedding model readiness, embedding/cache settings, and process memory separately.

Indexed chunks are bounded separately by EPHEMERAL_MAX_INDEXED_CHUNKS, which defaults to 32,768 total chunks across retained captures. LRU eviction makes room for a new capture when possible. A capture that exceeds the entire index budget is rejected rather than partially indexed, so accepted captures remain fully searchable and cannot produce silent semantic false negatives. The optional set_semantic_index_budget tool changes this limit for the current session only and is disabled unless EPHEMERAL_ALLOW_RUNTIME_INDEX_BUDGET=1 is set at startup. Decreases use the same deterministic LRU order and may evict multiple captures; deployment defaults are not changed or persisted. execute_and_capture retains the beginning and end of oversized command output and marks the capture with its original byte count. Each returned head or tail preview is independently capped at 4 KiB of UTF-8 data; a truncation marker directs the agent to get_capture_slice for the complete content.

Call get_runtime_diagnostics() when reporting a field observation. It is explicitly opt-in and returns operational metadata only; captured content, labels, command arguments, and session ID values are excluded. Sanitize any additional output before sharing it.

Operational events are written as privacy-safe JSON lines to stderr. Warnings and errors are enabled by default; set EPHEMERAL_LOG_LEVEL=INFO to include normal readiness, eviction, process lifecycle, and MCP tool start/completion events. Tool lifecycle events contain only a local call ID, tool name, duration, success state, and error class. Logs never include captured content, labels, query text, or command text. A start event without a matching completion or failure event identifies a stalled tool/session boundary. The Codex A/B adapter can persist these events per MCP run with --diagnostic-log-dir /path/to/logs; keep that directory outside the repository. The files contain lifecycle metadata only.

Optional local usage metrics

Set EPHEMERAL_METRICS=1 to collect content-free, in-process usage metrics. The metrics include per-tool call counts, success/failure counts, duration totals, aggregate capture/search/retrieval, empty-search, eviction, and cleanup events, plus session-scoped data-path byte counters. The byte counters cover input, retained, and original capture bytes; tool/search/retrieval response bytes; and framed socket request/response bytes. They are disabled by default, are never sent anywhere, and do not retain captured content, labels, commands, or query text. When enabled, both get_runtime_diagnostics() and get_buffer_stats() include the same aggregate metrics snapshot. The event keys are stable and zero-filled when no event has occurred, as are the byte-counter keys. Wire counts include framing headers and payload bytes actually consumed, including partial malformed requests; payload bytes rejected from an oversized frame before reading are not counted. MCP tool-response counts measure UTF-8 response content and exclude transport-envelope overhead. Metrics are process-lifetime state: restarting the server clears them, while capture-associated correlation state is released when a capture is evicted or explicitly cleared.

Effectiveness metrics and privacy

The built-in metrics are local, opt-in operational telemetry. Set EPHEMERAL_METRICS=1 only when you want measurements for the current server process; nothing is uploaded or shared by the server. The metrics contain counts, durations, byte sizes, and bounded lifecycle outcomes, but do not retain captured content, labels, command arguments, or query text. Runtime logs follow the same privacy model. Treat any captured output or diagnostic excerpt as potentially sensitive and sanitize it before sharing.

Interpret the measurements in two separate layers:

The effectiveness harness measures server-side behavior with deterministic fixtures and does not invoke a coding-agent model. A successful targeted retrieval means only that the fixture's expected marker was found. It is not a measure of answer quality, search relevance in a real repository, token cost, or end-to-end task completion. Use representative, privacy-reviewed tasks for those questions and report the fixture, seed, repetition count, success rate, useful-search rate, byte measurements, and local timing separately.

For a reproducible local diagnostic, start the server with EPHEMERAL_METRICS=1, exercise the workflow, then request get_runtime_diagnostics() and get_buffer_stats(). A safe bug report includes the version/commit, Python/platform, configuration limits, operation name, reproduction steps, and sanitized metric output; it excludes captures, credentials, tokens, private paths, source code, user data, and raw queries.

See OPERATIONS.md for deployment settings, troubleshooting, release verification, and repository maintenance procedures.

🛠 Testing the Server

Set up a local development environment from a fresh checkout:

python3.12 -m venv .venv
.venv/bin/python -m pip install --require-hashes -r requirements-dev-lock-py312.txt

The committed requirements-dev-lock-py312.txt file is the reproducible Python 3.12 development and release environment. Python 3.10 remains supported through the direct requirements and tested constraints.txt file; the CI matrix exercises both paths. Keep requirements.txt and requirements-dev.txt as the reviewable dependency inputs, and regenerate the Python 3.12 locks with pip-tools after an intentional dependency update:

.venv/bin/python -m pip install pip-tools
.venv/bin/pip-compile --generate-hashes --output-file=requirements-lock-py312.txt requirements.txt
.venv/bin/pip-compile --generate-hashes --output-file=requirements-dev-lock-py312.txt requirements-dev.txt

Review the resulting changes, run the full test matrix, and run pip-audit before merging. Downstream users install the package normally; its compatible dependency ranges in pyproject.toml are intentionally not replaced by the development locks.

Run the test suite:

.venv/bin/python -m unittest test_benchmark_warmup.py test_engine.py test_capture_utils.py test_config.py test_cli.py test_server.py test_execution.py test_execution_server.py
.venv/bin/python -m unittest test_e2e_pipe.py

Measure focused-test coverage locally:

.venv/bin/python -m coverage run --source=. --omit='test_*.py,setup.py,benchmark_concurrency.py,benchmark_effectiveness.py,benchmark_latency.py,benchmark_warmup.py,benchmark_agent_ab_repository_fixture.py,release_checks.py' -m unittest test_benchmark_concurrency.py test_benchmark_effectiveness.py test_benchmark_warmup.py test_release_checks.py test_benchmark_agent_ab_repository_fixture.py test_engine.py test_capture_utils.py test_config.py test_cli.py test_server.py test_execution.py test_execution_server.py
.venv/bin/python -m coverage report

CI requires 100% coverage for application runtime modules and excludes test, benchmark, release-check, and packaging-metadata files from that gate. Coverage reports are uploaded for inspection, and new runtime paths should include targeted tests. The release guardrail utility is measured separately because it is a workflow utility rather than application runtime code:

COVERAGE_FILE=.coverage.release .venv/bin/python -m coverage run --source=. -m unittest test_release_checks.py
COVERAGE_FILE=.coverage.release .venv/bin/python -m coverage report --include='release_checks.py'

GitHub Actions runs the compile check, focused tests, and end-to-end test on Python 3.10 and 3.12 for pushes to main and pull requests. The FastEmbed model is loaded on the first capture or semantic search rather than during server import. Set EPHEMERAL_EMBEDDING_MODEL to select a compatible model and EPHEMERAL_FASTEMBED_CACHE_DIR to control its cache directory. The model cache is retained between CI runs to reduce startup time. CI unit and end-to-end tests set the internal EPHEMERAL_TEST_EMBEDDINGS=1 flag, which uses a small deterministic embedding substitute so test execution does not depend on a model download; release and benchmark jobs continue to exercise FastEmbed. It also builds the wheel and verifies the installed ephbuf entry point. CI audits the declared dependencies with pip-audit and fails if known vulnerabilities are found. CI installs the hashed Python 3.12 development/runtime locks and uses the tested constraints.txt path for Python 3.10. The direct requirements and constraints are updated only after the full test matrix passes; lock updates must be reviewed together with their resolver output and audit results.

Pushing a version tag such as v0.1.1 runs the release workflow, which first verifies that the tag is valid SemVer, points to a commit contained in the default branch, and starts from a clean checkout. It also requires the tag, pyproject.toml, and a dated matching CHANGELOG.md section to agree. The workflow then builds wheel and source distributions, validates their metadata, verifies the installed package, and uploads the artifacts for review. A failed guardrail reports the mismatched value or source-state problem before building. The workflow creates a GitHub Release using the matching changelog section, attaches the wheel, source distribution, and SHA256SUMS, and links back to the workflow run containing the build-provenance attestation. Verify a downloaded artifact with sha256sum --check SHA256SUMS from the directory containing the files. The same verified distributions are then published to PyPI through trusted publishing. After the repository's pypi environment is configured with a PyPI trusted publisher, the workflow publishes the distributions to PyPI automatically.

Run the concurrency benchmark:

.venv/bin/python benchmark_concurrency.py --captures 32 --workers 8

The benchmark accepts --min-ingest-per-second and --min-reads-per-second thresholds for direct checks. For repeatable regression checks, pass --baseline benchmark_baseline.json --output benchmark-concurrency.json. The checked-in baseline uses a 20% tolerance: a run fails only when ingest or read throughput drops below 80% of its baseline. Each scheduled or manually dispatched GitHub Actions run records the raw JSON result as an artifact and adds the measurements and regression status to the workflow summary. This benchmark remains optional and is not part of the required pull-request checks; update the baseline deliberately when the runner or benchmark workload changes.

Measure command-capture latency by output size and pipeline phase:

EPHEMERAL_TEST_EMBEDDINGS=1 .venv/bin/python benchmark_latency.py \
  --samples 5 --output benchmark-latency.json

The latency harness reports cold-start time plus warm median and p95 timings for command execution, BM25 capture/indexing, deferred semantic indexing, summary generation, and the combined pipeline. Semantic embeddings are now materialized when semantic or hybrid search first needs them, so the warmup explicitly materializes the warm engine's embeddings before timed samples begin. Use the separate semantic-index phase when evaluating end-to-end costs; cold start includes the first engine's model and embedding setup, while warm samples reuse the configured model cache and engine. The benchmark is diagnostic and optional, not a required pull-request check.

Compare lazy semantic indexing with the default asynchronous prefetch:

EPHEMERAL_TEST_EMBEDDINGS=1 .venv/bin/python benchmark_prefetch.py \
  --line-count 256 --samples 5 --output benchmark-prefetch.json

The prefetch harness reports ingestion, first semantic search, and subsequent semantic search medians for both modes. Prefetch can reduce first-query latency when work completes during ingestion, while adding bounded background resource use. In the release-preparation 256-line Linux run, first-search median changed from 0.435 seconds without prefetch to 0.425 seconds with it, while ingestion changed from 1.19 ms to 1.63 ms; treat this as deployment-specific diagnostic evidence rather than a guaranteed improvement.

Measure semantic indexing cost and first hybrid-search latency by capture size:

.venv/bin/python benchmark_semantic_index.py \
  --samples 3 --output benchmark-semantic-index.json

The semantic-index harness ingests a fresh deterministic log-like capture per sample, then reports median and p95 timings for ingestion, lazy embedding materialization, the first hybrid (or --mode semantic) search that triggers it, and a subsequent search over the complete index, plus chunk count and indexing throughput. The first search runs against the semantic wait budget (--semantic-wait-seconds, default EPHEMERAL_SEMANTIC_WAIT_SECONDS or 10), so the report shows how often it answered with pending semantic coverage (first_search_pending_rate) and the needle rank it achieved (first_search_needle_mrr) next to the rank over the complete index (needle_mrr); pass --semantic-wait-seconds inf to time the full lazy indexing cost inside the first search instead. Default sizes are 16, 256, 2,048, and 8,192 lines. Run it without EPHEMERAL_TEST_EMBEDDINGS=1 to measure the configured FastEmbed model; with deterministic test embeddings it only validates the harness. Prefetch and startup warm-up are disabled inside the harness so the lazy cost is visible. Results are host-specific diagnostic evidence, not a required CI gate.

For a release comparison, record the same workload under each model and compare the selected run with the versioned result tool:

.venv/bin/python benchmark_semantic_index.py \
  --embedding-model BAAI/bge-small-en-v1.5-fp32 --line-counts 1024 --samples 3 \
  --result results/fp32.result.json --experiment embedding-model \
  --metadata variant=fp32 --metadata host_class=linux-x86_64
.venv/bin/python benchmark_semantic_index.py \
  --embedding-model BAAI/bge-small-en-v1.5 --line-counts 1024 --samples 3 \
  --result results/catalogue.result.json --experiment embedding-model \
  --metadata variant=catalogue --metadata host_class=linux-x86_64
.venv/bin/python compare_workload_results.py \
  results/fp32.result.json#lines-1024 results/catalogue.result.json#lines-1024 \
  --statistic median --metric semantic_index --metric throughput_per_second

Keep result documents together with the release benchmark artifacts. Compare results only within the same host class and record model, thread, chunking, prefetch, warm-up, and wait-budget settings; do not treat a different host as a regression.

Compare lazy model loading with background startup warm-up:

EPHEMERAL_TEST_EMBEDDINGS=1 .venv/bin/python benchmark_warmup.py \
  --samples 5 --output benchmark-warmup.json

The warm-up harness reports engine initialization, time to embedding readiness, first semantic-search latency, and process RSS change for both policies. For the lazy policy, embedding readiness is measured through completion of the first semantic search rather than reported as instantaneous. Run it without deterministic test embeddings to measure the configured FastEmbed model and host cache; each policy is measured in a fresh worker process so model pages retained by the allocator do not contaminate the other policy. In the release-preparation Linux run, first-search median was 0.335 seconds without warm-up versus 0.0122 seconds with warm-up, with approximately 186 MB RSS increase in both modes. Results are deployment-specific and the benchmark is optional.

Measure direct-versus-captured routing tradeoffs with synthetic output profiles:

EPHEMERAL_TEST_EMBEDDINGS=1 .venv/bin/python benchmark_routing.py \
  --samples 5 --output benchmark-routing.json

The routing harness compares direct command completion with bounded capture and summary generation for targeted (16 lines), test-like (256 lines), and build/log-like (2048 lines) output. Use the medians and p95 values to keep the heuristic honest: direct execution generally has lower latency for small, bounded output, while capture adds searchable context and bounded response size for noisy or uncertain output. These synthetic measurements are guidance, not universal thresholds or a required CI gate. Both benchmark summaries use the nearest-rank p95 convention: for n samples, p95 is the value at sorted rank ceil(0.95 * n), with ranks starting at one.

Measure command-output handling effectiveness with deterministic synthetic data:

EPHEMERAL_TEST_EMBEDDINGS=1 .venv/bin/python benchmark_effectiveness.py \
  --mode both --output benchmark-effectiveness.json

The effectiveness harness compares a full-output baseline with an engine-backed MCP workflow across large output, failure logs, diffs, and follow-up searches. It reports per-scenario success, search usefulness, retrievals, bytes examined, resolution time, and an aggregate comparison as machine-readable JSON. Token usage is explicitly marked unavailable because this harness does not invoke a model. Fixtures contain no project content and are generated in code, so runs are reproducible. This is a server-side smoke evaluation, not a claim about any particular coding agent or model.

Evaluate search relevance across supported modes with deterministic synthetic fixtures:

EPHEMERAL_TEST_EMBEDDINGS=1 .venv/bin/python benchmark_relevance.py \
  --top-k 3 --baseline benchmark_relevance_baseline.json \
  --output search-relevance.json

The relevance benchmark covers exact errors, punctuation-heavy queries, conceptual semantic queries, and lexical/semantic conflicts. It reports hit@1, hit@k, and mean reciprocal rank (MRR) for BM25, semantic, and hybrid search. The fixtures define expected retrieval markers, so this measures retrieval relevance only—not agent answer quality, token usage, or universal performance. CI compares the deterministic scores with the checked-in benchmark_relevance_baseline.json and fails only when a metric drops by more than the recorded tolerance. Baseline changes must be deliberate, reviewable, and accompanied by a fixture or intended-behavior explanation. The evaluation uploads its machine-readable JSON result with the other benchmark artifacts.

The baseline stores only schema and fixture versions, deterministic embedding mode, aggregate per-mode scores, query counts, and explicit tolerances. It does not store captures, raw command output, or user queries. Environment-specific latency and effectiveness results remain CI artifacts and summaries rather than exact cross-run gates; retention follows the workflow artifact policy.

Example benchmark results

The reproducible paired evaluation was run on 2026-09-11 with five repetitions, seed 20260907, and deterministic test embeddings. Across four synthetic scenarios, both the direct-output baseline and the MCP workflow completed all 20 tasks, and every MCP search was useful. The MCP workflow examined 30–85% fewer bytes than the baseline per scenario (68% on average):

The consolidation evaluation, using the same seed and five repetitions, reduced the initial multi-result overview from an average of 4,298 bytes to 213 bytes (95% fewer overview bytes), while preserving a 100% targeted retrieval success rate. The consolidated workflow retrieved 3.6% fewer bytes overall and took 2.2 times as long locally as sequential processing in this run. These measurements quantify the MCP data path: fewer bytes need to be returned to the agent before it asks for targeted detail.

They are not universal performance guarantees. The fixtures are synthetic, the harness does not invoke a model, and the byte reduction is not a direct token-savings measurement. In this run, consolidated processing took about 2.2 times longer locally than sequential processing, while retrieving similar detail. The benchmark therefore demonstrates context-size and workflow-shaping benefits, not that every workload will be faster or that search results will be relevant for arbitrary repositories. Re-run the commands below with representative, privacy-reviewed tasks before making project-specific claims.

Run a controlled local A/B evaluation with repeated paired measurements:

EPHEMERAL_TEST_EMBEDDINGS=1 .venv/bin/python benchmark_effectiveness.py \
  --ab-runs 5 --seed 20260907 --output benchmark-effectiveness-ab.json

The A/B report uses the same deterministic fixtures in both modes, seeded task and mode ordering, and local-only measurements. It reports completion rate, mean/min/max time, standard deviation, repeated commands, search usefulness, byte reduction, and local MCP processing overhead for each scenario. The timing ratio covers only local capture, indexing, search, and retrieval; it is not an agent-level performance measurement. Since no model is invoked, token usage is unavailable. Treat the recommendations as synthetic benchmark guidance and repeat the evaluation with representative agent tasks before generalizing the results.

For an end-to-end coding-agent evaluation, generate a counterbalanced, privacy-safe schedule:

.venv/bin/python benchmark_agent_ab.py \
  --schedule-output agent-ab-schedule.json --repetitions 5 --seed 20260909

Run each scheduled task with MCP disabled (control) and enabled (mcp) using the same model configuration, repository fixture, environment, and reset policy. Have an external agent adapter write a records envelope containing only the schedule, non-secret protocol identifiers, and per-run fields: completion, signal retrieval, duration, tool calls, repeated commands, context proxy bytes (total plus prompt/output components), provider usage samples, and peak RSS bytes sampled from that invocation's process. On hosts without a supported per-process RSS interface, peak RSS is recorded as zero and should be treated as unavailable. Summarize it with:

.venv/bin/python benchmark_agent_ab.py \
  --records agent-ab-records.json --output agent-ab-summary.json

The analyzer validates balanced paired runs, reports per-mode aggregates and MCP-minus-control deltas with uncertainty, and emits recommendations. It does not invoke a model, require credentials, or accept raw prompts, transcripts, commands, captures, or user content. Do not commit agent records or generated captures; share only the aggregate summary after privacy review.

The repository includes a Codex CLI adapter for executing that protocol. It uses the requested model explicitly, creates a fresh fixture copy for every run, isolates control and MCP configuration, and writes metadata-only records. Create a private task manifest (do not commit it) with one prompt and optional signal marker for each scheduled task:

{
  "tasks": {
    "targeted-inspection": {"prompt": "Inspect the fixture and report the marker.", "signal_marker": "MARKER"},
    "noisy-test-failure": {"prompt": "Run the fixture test and report the failure marker.", "signal_marker": "MARKER"},
    "build-log-search": {"prompt": "Inspect the build log and report the marker.", "signal_marker": "MARKER"},
    "follow-up-context": {"prompt": "Find the earlier marker and report it.", "signal_marker": "MARKER"}
  }
}

Run the adapter from the repository checkout:

.venv/bin/python run_codex_agent_ab.py \
  --schedule agent-ab-schedule.json \
  --tasks /path/to/private-agent-tasks.json \
  --repository /path/to/privacy-reviewed-fixture \
  --model gpt-5.6-luna \
  --output /tmp/agent-ab-records.json
.venv/bin/python benchmark_agent_ab.py \
  --records /tmp/agent-ab-records.json \
  --output /tmp/agent-ab-summary.json

The runner requires a locally authenticated codex CLI. It uses codex exec --json --ephemeral, uses a read-only sandbox by default. context_bytes_proxy is an observable prompt/event-envelope proxy because the CLI does not expose the model's internal context size. When the fixture does not contain an importable server module, pass --mcp-server-script /absolute/path/to/server.py. For MCP experiments where the client must be allowed to call the configured server, add --allow-mcp-approvals. This uses Codex automatic review with a workspace-write sandbox, and should only be used with a disposable, privacy-reviewed fixture. Control runs continue to use the read-only sandbox without MCP approval routing; the records protocol identifies the selected policy. Add --require-mcp-calls when the MCP arm must exercise at least one MCP tool; runs that bypass MCP are then marked incomplete with reason mcp_not_used. Review prompts, fixtures, and generated records for privacy before sharing; the runner does not persist transcripts in its records output.

Runner records use version 5 and add exit code, failure reason, MCP-specific tool-call count, provider-reported input/output token counts, and every provider usage sample when Codex emits them. They also break the observable context proxy into prompt and output byte components. Version-1 through version-4 records remain readable; missing provider metrics are reported as unavailable rather than zero. Version-5 MCP records additionally include content-free session data-path byte counters for capture input/retention, tool/search/retrieval responses, and framed socket traffic. The adapter enables local metrics for MCP runs and collects the server snapshot after each run; missing snapshots are represented as zero counters and should be treated as unavailable when diagnosing a failed run. Summaries also report usage sample counts, monotonicity observations, and first-to-last deltas. Monotonic samples are explicitly inconclusive: they may be cumulative or per-turn values and require a controlled calibration matrix.

Generate the reviewed synthetic EB-heavy fixture and its task manifest with:

.venv/bin/python benchmark_agent_ab_fixtures.py \
  --fixture-output /tmp/agent-ab-fixture \
  --manifest-output /tmp/agent-ab-tasks.json

The fixture generates large test and build output at runtime, plus a small targeted-output task and a follow-up retrieval task. The manifest includes output-size bands, expected signals, and objective success criteria. It is synthetic and contains no user logs or captured output.

For a repository-shaped evaluation, set AGENT_AB_FIXTURE_PROFILE:

AGENT_AB_FIXTURE_PROFILE=repository-shaped-v1 \
  CODEX_HOME=/path/to/writable/authenticated-codex-home \
  ./run_agent_ab_experiment.sh

This profile contains source, tests, configuration, and repository workflow tooling. Its test and build commands inspect that layout while producing deterministic noisy signals. It is still synthetic and privacy-safe; it does not contain a production repository or user data.

Repeat the complete five-repetition Codex A/B run with the repository script. Set CODEX_HOME to a writable, authenticated Codex home; generated fixtures, records, and lifecycle logs remain under /tmp by default:

CODEX_HOME=/path/to/writable/authenticated-codex-home \
  ./run_agent_ab_experiment.sh

Override AGENT_AB_RUN_DIR, AGENT_AB_MODEL, AGENT_AB_REPETITIONS, AGENT_AB_SEED, AGENT_AB_TIMEOUT_SECONDS, or AGENT_AB_TEST_EMBEDDINGS to repeat a different experiment. The script also writes records.result.json and summary.result.json in the common workload result format; set AGENT_AB_EXPERIMENT and AGENT_AB_VARIANT to assign them to an experiment group (see "Organizing experiments"). The test embedding setting defaults to 1 for deterministic, offline runs; set it to 0 to exercise the configured FastEmbed model and measure real model startup and first-query behavior. Keep the model, embedding cache, and other environment settings identical across paired runs. The script uses the synthetic fixture by default; set AGENT_AB_FIXTURE_PROFILE=repository-shaped-v1 for the repository-shaped synthetic profile. Use the lower-level runner commands above for a privacy-reviewed production-repository fixture and private task manifest.

Create and compare an aggregate agent A/B baseline after a privacy review:

.venv/bin/python benchmark_agent_ab_baseline.py \
  --summary /tmp/agent-ab-summary.json \
  --create-baseline \
  --output benchmark_agent_ab_baseline.json
.venv/bin/python benchmark_agent_ab_baseline.py \
  --summary /tmp/agent-ab-summary.json \
  --baseline benchmark_agent_ab_baseline.json \
  --fail-on-regression \
  --output /tmp/agent-ab-comparison.json

The baseline stores agent and embedding model configuration, embedding mode and cache, fixture, seed, repetition, aggregate metrics, and explicit tolerances only. It never stores prompts, transcripts, commands, captures, or user content. This comparison is a documented manual workflow; live model calls are not part of required pull-request CI. Update a checked-in baseline only when fixture or model changes are explained in review.

The current agent-level evaluation supports a provisional routing policy: prefer MCP for large noisy output and follow-up retrieval, and prefer direct execution for small targeted inspections. Do not treat this as a universal default or convert it into hard numeric thresholds yet. The repository-shaped profile is synthetic, so production-repository behavior should be validated separately before generalizing the result. Provider-reported token deltas are diagnostic until a calibration matrix establishes whether usage samples are cumulative or per-turn.

Compare sequential per-capture retrieval with the consolidated workflow:

EPHEMERAL_TEST_EMBEDDINGS=1 .venv/bin/python benchmark_effectiveness.py \
  --consolidation-runs 5 --seed 20260907 \
  --output benchmark-effectiveness-consolidation.json

This report measures overview and retrieval response bytes, targeted retrieval success, search/retrieval counts, local processing time, and omitted records. It models each synthetic scenario as a repository result and does not invoke a coding-agent model; response-byte reductions therefore describe the MCP data path, not end-to-end agent performance.

When sharing a result, include the command, seed, repetitions, benchmark evaluation name, success rate, useful-search rate, byte reduction, and timing scope. Do not attach generated captures or paste raw command output. Check any surrounding report or wrapper for repository-specific content before sharing the benchmark JSON.

Machine-readable workload results

Every benchmark, evaluation, and the Codex A/B runner can emit one common, versioned JSON document in addition to its own report and --output record. Pass --result PATH to write it to a file, or --result - to print it on stdout; the human-readable report then moves to stderr so stdout stays valid JSON:

EPHEMERAL_TEST_EMBEDDINGS=1 .venv/bin/python benchmark_latency.py \
  --samples 3 --result benchmark-latency.result.json
.venv/bin/python benchmark_semantic_index.py --samples 3 --result - > semantic-index.result.json
.venv/bin/python workload_results.py benchmark-latency.result.json semantic-index.result.json

The document is a coding-agent-workload-result (format version 1). It is tool- and task-agnostic: it records what was measured, never how, so a consumer does not need to know about embeddings, BM25, or any other producer mechanism. The reference validator is workload_results.py (also usable as a CLI, as above) and the same contract is published as JSON Schema in workload_result.schema.json.

{
  "format": "coding-agent-workload-result",
  "format_version": 1,
  "workload": {"name": "capture-latency", "kind": "benchmark",
               "producer": "benchmark_latency.py",
               "parameters": {"line_counts": [16, 256, 2048], "samples": 3}},
  "environment": {"python_version": "3.12.14", "platform": "...",
                  "cpu_count": 8, "recorded_at": "2026-09-18T15:00:00+00:00",
                  "tool": {"name": "ephemeral-buffer-mcp", "version": "0.4.0"},
                  "source_revision": "..."},
  "status": "success",
  "errors": [],
  "measurements": {},
  "runs": [
    {"id": "lines-256",
     "labels": {"cache_state": "warm", "line_count": 256},
     "status": "success",
     "measurements": {
       "output_bytes": {"unit": "bytes", "value": 3840},
       "wall_time_seconds": {"unit": "seconds", "median": 0.012, "p95": 0.015, "samples": 3}},
     "phases": [
       {"name": "command", "unit": "seconds", "median": 0.004, "p95": 0.005, "samples": 3},
       {"name": "ingest", "unit": "seconds", "median": 0.006, "p95": 0.008, "samples": 3}],
     "errors": []}
  ],
  "experiment": {"group": "chunk-sweep",
                 "metadata": {"variant": "chunk-8", "model": "bge-small-fp32"}},
  "details": {"...": "the producer's own record, for humans"}
}

• workload.name identifies the measurement and workload.parameters holds everything needed to repeat it (sizes, seeds, modes, option overrides); environment explains why two results may legitimately differ.

• Each run is one comparable unit of work with a stable id, descriptive labels (cache_state, mode, task_id, repetition, line_count, profile, ...), a status of success, failure, timeout, error, or partial, and its own errors. Failed, timed-out, and partial runs are kept with their status rather than dropped.

• A measurement has a unit (seconds, bytes, count, tokens, ratio, per_second, or score) and one or more statistics (value, sum, mean, median, min, max, p95, stdev), optionally with the number of samples and a note explaining a proxy. null means the statistic is unavailable, never zero. phases is an ordered timeline of seconds measurements.

• Canonical names such as wall_time_seconds, queue_wait_seconds, tool_calls, output_bytes, context_bytes, estimated_tokens, retained_summary_tokens, input_tokens, output_tokens, peak_rss_bytes, rss_delta_bytes, success_rate, and throughput_per_second pin their unit so results from different producers line up; producers add their own names beside them. The JSON Schema pins those units too.

• Run ids are unique within a document. JSON Schema cannot express that rule (its uniqueItems only rejects fully identical runs), so a consumer that validates with the schema alone must check ids itself or run workload_results.py on the document first.

• details carries the producer's native record for people who need it; its shape is producer-specific and versioned separately by workload.producer_schema_version.

• The optional experiment block assigns the document to a named group of related runs and carries flat metadata (scalar values only) describing what varied; see "Organizing experiments" below. Every producer accepts --experiment GROUP, --metadata KEY=VALUE (repeatable; values parse as JSON when possible), and --redact KEY.

OPERATIONS.md describes how comparison tooling and regression checks should consume these documents. The shared format is not a privacy exemption: apply the same review to a result file as to any other benchmark output before sharing it.

Comparing workload results

compare_workload_results.py compares two or more result documents without rerunning anything. The first reference is the baseline and every later one is compared against it. For each run and measurement the documents share it prints the absolute delta, the percentage change, and an outcome; runs pair by id, measurements and phases by name, and every statistic is compared only with the same statistic (median against median, never against mean). The example below records a baseline, halves the semantic chunk size, and compares the two:

EPHEMERAL_TEST_EMBEDDINGS=1 .venv/bin/python benchmark_latency.py \
  --samples 3 --line-counts 256 2048 --result before.result.json
EPHEMERAL_TEST_EMBEDDINGS=1 EPHEMERAL_SEMANTIC_CHUNK_LINES=2 .venv/bin/python benchmark_latency.py \
  --samples 3 --line-counts 256 2048 --result after.result.json
.venv/bin/python compare_workload_results.py before.result.json after.result.json \
  --tolerance 5 --statistic median --metric wall_time_seconds --metric ingest --metric semantic_index

workload: capture-latency
  baseline: before.result.json  producer=benchmark_latency.py  kind=benchmark  status=success  runs=3  recorded=2026-09-18T18:16:11+00:00  python=3.12.14  tool=ephemeral-buffer-mcp 0.4.0  revision=2b58f2490571
  candidate: after.result.json  producer=benchmark_latency.py  kind=benchmark  status=success  runs=3  recorded=2026-09-18T18:16:28+00:00  python=3.12.14  tool=ephemeral-buffer-mcp 0.4.0  revision=2b58f2490571
  statistics=median tolerance=5% metrics=ingest,semantic_index,wall_time_seconds

before.result.json -> after.result.json
  run         metric                stat    baseline     candidate    delta         change   outcome
  lines-256   wall_time_seconds     median  0.009854 s   0.01054 s    +0.0006824 s  +6.9%    regressed
  lines-256   phase:ingest          median  0.0007052 s  0.0007507 s  +4.546e-05 s  +6.4%    regressed
  lines-256   phase:semantic_index  median  0.0008512 s  0.002113 s   +0.001262 s   +148.3%  regressed
  lines-2048  wall_time_seconds     median  0.02084 s    0.02982 s    +0.008981 s   +43.1%   regressed
  lines-2048  phase:ingest          median  0.002724 s   0.002792 s   +6.775e-05 s  +2.5%    unchanged
  lines-2048  phase:semantic_index  median  0.006279 s   0.01706 s    +0.01078 s    +171.6%  regressed
  summary: 0 improved, 5 regressed, 0 changed, 1 unchanged, 0 missing, 0 incompatible; runs compared=3 missing=0; status success -> success

The phase rows attribute the wall-time regression to semantic indexing rather than ingestion. Outcomes are:

• improved or regressed: the value moved beyond --tolerance PERCENT (default 0) in the metric's better or worse direction. Canonical measurements know their direction (time, bytes, tokens, tool calls, and memory are better lower; success_rate and throughput are better higher); other seconds, bytes, tokens, and per_second metrics follow their unit, and --direction NAME=lower|higher declares the rest.

• changed: the value moved but the metric has no known direction, such as a producer-specific count, ratio, or score. A change from a zero baseline has no percentage and counts as beyond any tolerance.

• unchanged: within the tolerance.

• missing: the run, measurement, or statistic is absent, null, or has samples: 0 on at least one side. These are listed with the side that lacks them, never silently skipped.

• incompatible: both sides report the metric with different units.

Documents must describe the same workload.name unless --allow-workload-mismatch is given, and each comparison lists the workload.parameters and environment fields that differ (CPU count, tool version, source revision) so a different setup is not mistaken for a regression. Non-success runs appear with their status and errors.

PATH#RUN_ID selects one run from a document. When the baseline and a candidate each select a single run, those two runs pair even though their ids differ, which compares two configurations recorded in the same document:

# Agent configurations: control versus MCP in one A/B summary, or two
# agent-run documents recorded under different models or policies.
.venv/bin/python benchmark_agent_ab.py --records runs.json --result summary.result.json
.venv/bin/python compare_workload_results.py summary.result.json#control summary.result.json#mcp --statistic mean
.venv/bin/python compare_workload_results.py codex-gpt5.result.json codex-candidate.result.json \
  --select mode=mcp --metric input_tokens --metric output_tokens --metric wall_time_seconds --metric tool_calls

# Tool-output policies: the same latency workload under two server settings.
EPHEMERAL_SEMANTIC_PREFETCH=0 .venv/bin/python benchmark_latency.py --result lazy.result.json
.venv/bin/python benchmark_latency.py --result prefetch.result.json
.venv/bin/python compare_workload_results.py lazy.result.json prefetch.result.json --metric wall_time_seconds

# Summarization strategies: retained-summary size and prompt-token proxies per
# task; the reduction ratios need an explicit direction.
.venv/bin/python benchmark_effectiveness.py --summary --result summary-a.result.json
.venv/bin/python compare_workload_results.py summary-a.result.json summary-b.result.json \
  --metric retained_summary_tokens --metric estimated_tokens --metric summary_token_reduction \
  --direction summary_token_reduction=higher

--select KEY=VALUE keeps only runs whose label matches (values parse as JSON when possible, so line_count=256 compares a number), --metric NAME limits the report to named measurements or phases, and --statistic NAME chooses the statistics (default value, median, mean, and p95; all adds sum, min, max, and stdev; stdev describes spread rather than level, so its changes are reported as changed and never judged). --format json prints, and --output PATH writes, a coding-agent-workload-comparison document (format version 1) with the same entries plus each document's workload and environment blocks. --check exits with status 2 when any metric regressed or any document has a non-success status, which OPERATIONS.md uses for regression checks. A document narrowed with PATH#RUN_ID or --select is judged by its selected runs, and the report shows the whole file's document_status beside it when the two differ.

Organizing experiments

A performance or token-efficiency study is rarely one comparison: a chunk-size sweep, a model change, or a prompt-policy A/B produces several result documents whose relationship is otherwise only in their file names. Every producer therefore accepts --experiment GROUP to assign its result document to an experiment or run group, and --metadata KEY=VALUE to record what varied. The workflow is: tag each run when it is recorded, list the group to see what exists and which runs failed, and compare documents by group and metadata instead of by path.

Record. Give every run of one study the same group and describe the variable under test in metadata. Conventional keys are task_type, repository_revision, agent_configuration, model, tool_version, variant, environment, workload_size, and started_at; any other snake_case key is allowed. Values are identifiers (at most 256 characters), never prompts or captured content, and started_at must be an ISO 8601 timestamp with a UTC offset (2026-09-18T10:00:00+00:00 or a trailing Z) so documents order by instant. A value that breaks these rules is rejected when the arguments are parsed, before the workload runs:

EPHEMERAL_TEST_EMBEDDINGS=1 .venv/bin/python benchmark_latency.py --samples 3 --line-counts 256 \
  --result results/chunk-8.result.json --experiment chunk-sweep \
  --metadata variant=chunk-8 --metadata model=bge-small-fp32
EPHEMERAL_TEST_EMBEDDINGS=1 EPHEMERAL_SEMANTIC_CHUNK_LINES=4 .venv/bin/python benchmark_latency.py --samples 3 --line-counts 256 \
  --result results/chunk-4.result.json --experiment chunk-sweep \
  --metadata variant=chunk-4 --metadata model=bge-small-fp32
EPHEMERAL_TEST_EMBEDDINGS=1 EPHEMERAL_SEMANTIC_CHUNK_LINES=16 .venv/bin/python benchmark_latency.py --samples 3 --line-counts 256 \
  --result results/chunk-16.result.json --experiment chunk-sweep \
  --metadata variant=chunk-16 --metadata model=bge-small-fp32

For coding-agent runs, run_agent_ab_experiment.sh tags both of its result documents (records.result.json and summary.result.json) with the model, fixture profile, repetition count, embedding environment, and start time, and takes the group from AGENT_AB_EXPERIMENT and the variant from AGENT_AB_VARIANT:

AGENT_AB_EXPERIMENT=policy-ab AGENT_AB_VARIANT=summarize-first \
  CODEX_HOME=/path/to/writable/authenticated-codex-home ./run_agent_ab_experiment.sh

List. list_workload_results.py searches files and directories (recursively) for result documents, ignores other JSON files such as records and schedules, and prints one row per document with its group, status, run count, time, and metadata. --group NAME and --where KEY=VALUE filter by group and metadata, --workload NAME and --status STATUS narrow further, --field KEY shows chosen metadata keys as columns, and --runs lists every run with its labels and status so failed, timed-out, and partial runs inside a document are visible:

.venv/bin/python list_workload_results.py results --field variant --field model

path                          group        workload           status   runs  time                       variant   model           errors
results/chunk-8.result.json   chunk-sweep  capture-latency    success  2     2026-09-18T18:52:02+00:00  chunk-8   bge-small-fp32
results/chunk-16.result.json  chunk-sweep  capture-latency    success  2     2026-09-18T18:52:03+00:00  chunk-16  bge-small-fp32
results/chunk-4.result.json   chunk-sweep  capture-latency    success  2     2026-09-18T18:52:03+00:00  chunk-4   bge-small-fp32
results/prefetch.result.json  -            semantic-prefetch  success  2     2026-09-18T18:52:04+00:00  -         bge-small-fp32

Rows are ordered by group, then by started_at metadata (or the recording time when it is absent, both normalised to UTC), then by path. Documents that do not satisfy the format are listed as invalid with the reason and make the command exit with status 1, so a broken file is never mistaken for an absent one. --format json prints a coding-agent-workload-listing document whose entries carry each document's group, metadata, status, errors, and per-run status, and --format paths prints only the selected paths for shell substitution.

Compare. compare_workload_results.py accepts DIR@GROUP references beside file references: the reference expands to every document under the directory that belongs to the group, in the same order as the listing, and DIR@GROUP,KEY=VALUE keeps only documents whose metadata matches. A lone results@chunk-sweep compares every later document of the sweep against the earliest; naming two selectors picks the baseline explicitly, and #RUN_ID still selects one run from each document:

.venv/bin/python compare_workload_results.py results@chunk-sweep,variant=chunk-8 results@chunk-sweep,variant=chunk-4 \
  --statistic median --metric wall_time_seconds --metric semantic_index

workload: capture-latency
  baseline: results/chunk-8.result.json  group=chunk-sweep  producer=benchmark_latency.py  kind=benchmark  status=success  runs=2  recorded=2026-09-18T18:52:02+00:00  python=3.12.14  tool=ephemeral-buffer-mcp 0.4.0  revision=0a9e8da36b46
  candidate: results/chunk-4.result.json  group=chunk-sweep  producer=benchmark_latency.py  kind=benchmark  status=success  runs=2  recorded=2026-09-18T18:52:03+00:00  python=3.12.14  tool=ephemeral-buffer-mcp 0.4.0  revision=0a9e8da36b46
  statistics=median tolerance=0% metrics=semantic_index,wall_time_seconds

results/chunk-8.result.json -> results/chunk-4.result.json
  experiment differences: variant: "chunk-8" -> "chunk-4"
  run        metric                stat    baseline    candidate   delta         change  outcome
  lines-256  wall_time_seconds     median  0.0159 s    0.01414 s   -0.00176 s    -11.1%  improved
  lines-256  phase:semantic_index  median  0.001104 s  0.001814 s  +0.0007096 s  +64.3%  regressed
  summary: 1 improved, 1 regressed, 0 changed, 0 unchanged, 0 missing, 0 incompatible; runs compared=2 missing=0; status success -> success

The report and the JSON comparison show the group beside each document and list the metadata that differs (experiment differences) next to the parameter and environment differences, so a delta can be read together with the variable that caused it. Selector values may not contain commas; a value containing @ is fine because only the first @ separates the directory from the group. A group reference fails on an invalid document only when that document claims the requested group; stale or broken files in other groups do not block the comparison (the listing still reports them). Several group references into one directory scan it once.

Sensitive metadata. Metadata keys that name credentials (token, key, password, secret, credentials, authorization, bearer, or any *_token or *_key) are stored as [redacted] by every producer, the validator rejects a document that carries a real value under such a key, and --redact KEY stores [redacted] for any other key whose value should not leave the machine (the key stays visible so readers know it was set). The listing tool's --redact KEY masks a value in its output without changing the file. Metadata that should not be recorded at all is simply not passed; the result format is still not a privacy exemption, so review documents before sharing them as with any other benchmark output.