Mastra #27148: small local regression-data packet

Target: Mastra issue #27148.

Free, synthetic reproduction data, not a source fix or an upstream PR. The packet adds boundary and non-digit controls to the four reported strings. No accounts, API keys, providers, storage, model calls, or customer data are needed.

Replay

Use Node 24.14.1 (.node-version); do not update your global Node installation. The manifest requires >=22.22.0 because of a locked transitive dependency, but the recorded run is Node 24.14.1 / ICU 78.2 / Unicode 17.0, Darwin arm64, en-US Intl. Other Node/ICU/platform combinations have not been qualified.

Run from this directory. Set NODE_BIN to the directory containing your isolated Node 24.14.1 node and npm binaries. Installation needs public npm-registry network access. Lifecycle scripts, audit and funding requests are disabled:

NODE_BIN="/path/to/node-24.14.1/bin"
SETUP_HOME="$(mktemp -d)"
env -i PATH="$NODE_BIN:/usr/bin:/bin" HOME="$SETUP_HOME" \
  TMPDIR="$SETUP_HOME" XDG_CONFIG_HOME="$SETUP_HOME" \
  NPM_CONFIG_USERCONFIG="$SETUP_HOME/user.conf" \
  NPM_CONFIG_GLOBALCONFIG="$SETUP_HOME/global.conf" \
  NPM_CONFIG_CACHE="$SETUP_HOME/cache" LANG=C LC_ALL=C TZ=UTC NO_COLOR=1 \
  "$NODE_BIN/npm" ci --ignore-scripts --no-audit --no-fund
rm -rf "$SETUP_HOME"

env -i PATH="$NODE_BIN:/usr/bin:/bin" LANG=C LC_ALL=C TZ=UTC \
  "$NODE_BIN/node" reproduce.mjs > observed.actual.json
env -i PATH="$NODE_BIN:/usr/bin:/bin" LANG=C LC_ALL=C TZ=UTC \
  "$NODE_BIN/node" reproduce.mjs --csv > observed.actual.csv
env -i PATH="$NODE_BIN:/usr/bin:/bin" LANG=C LC_ALL=C TZ=UTC \
  "$NODE_BIN/node" --test test-reproduce.mjs

Check each command's exit status; do not treat a created output file as success. node reproduce.mjs emits stable JSON by default; --csv emits CSV, and --self-test runs deliberate guard probes in a separate process, not the measurement. Errors, invalid fixtures/counts, failed imports, guarded attempts, or a worker timeout produce nonzero exit status. Use the explicit env -i commands above: unlike npm run, they also exclude inherited startup options such as NODE_OPTIONS before Node launches.

What was observed

The real published @mastra/memory/processors and alias mastra-memory-alpha/processors exports are each instantiated with new TokenCounter() and called through countString(string). Direct tokenx calls use estimateTokenCount. No wrapper implementation substitutes for the SDK.

Input	memory 1.37.1	memory 1.38.0-alpha.0	tokenx 1.3.0	tokenx 1.6.0
"0".repeat(28000)	1	1	1	9334
"12345678901234567890"	1	1	1	7
"1700000000000"	1	1	1	5
"hello world this is text"	5	5	5	5
spaced numbers	3	3	3	12
mixed JSON text	17	17	17	20
decimal text	3	3	3	9
CJK control	13	13	13	12
emoji control	3	3	3	8

All 16 cases / 64 estimates are in observed.json and observed.csv. Controls are observations, not assumptions that non-digit behavior is unchanged. The tests compare two fresh JSON runs and two fresh CSV runs byte-for-byte, confirm the original four counts, check fixture hashes/lengths, and compare the case matrix with the checked-in observations. Recorded results came after a fresh npm ci --ignore-scripts --no-audit --no-fund (172 installed packages).

Interpretation and boundaries

• One installation, not a before/after experiment. Root tokenx 1.6.0 coexists with the SDK counts above. This does not prove that adding/upgrading it had no effect, nor identify which tokenx version/code a published SDK bundles. Equal output or equal countString function hashes are not proof of bundled identity.
• tokenx 1.6.0 is only a versioned estimator comparator, never model-token ground truth. ascii_digit_heuristic_ceil_n_over_3 is the issue's heuristic for nonempty ASCII-digit-only strings. It is null for all other strings, not a claim about all models or all text. No actual model usage is measured.
• comparator_over_stable is an estimator-count ratio; a zero denominator is JSON null / an empty CSV cell. It is not an accuracy or billing metric.
• No claims about OM triggers, buffers, heap/OOM, context limits, end-to-end memory behavior, or the full Mastra test suite. No proposed fix is evaluated.
• The previously inspected current-main source at 03646c587b915ced4c9fad3189391909a296f441 is an API reference only, not the executed npm build. observed.json records actual installed package versions, lock integrities, selected ESM entry paths resolved with import.meta.resolve, entry/package/function SHA-256s, lock/fixture hashes, and runtime details. Entry hashes do not represent a complete transitive module graph.

Isolation and data handling

The launcher creates a temporary empty HOME and starts the measurement child with only HOME, TMPDIR, XDG_CONFIG_HOME, LANG, LC_ALL, TZ, and NO_COLOR. macOS's injected __CF_USER_TEXT_ENCODING is removed before validation/imports. Tests add fake API keys, proxy and OTEL headers to the launcher environment and verify the child rejects inherited ENV and does not report those sentinels. No real HOME/config is passed to third-party imports.

Before any dynamic third-party import, the child guards native fetch, HTTP(S), HTTP/2 connection, net/socket, TLS, datagram, DNS (including promises and resolvers), subprocess and worker APIs. Attempts are counted before throwing; every non-self-test denial immediately sets a failing process exit status even if caught. Reports are written synchronously at exit, after ordinary deferred beforeExit work. A denial in a later exit listener cannot revise already-written stdout but still fails the process, and the launcher propagates that failure: exit status is required for acceptance, not just a zero-attempt JSON snapshot. Lifecycle regressions exercise both listener orderings with the actual worker. The measured runs recorded zero attempts through guarded APIs during imports and cases; 23 deliberate self-test probes were separately blocked. This is process-local instrumentation, not an OS sandbox or comprehensive proof of offline enforcement. Native addons, uncovered APIs, or malicious code are not a supported threat model. Runtime counting requires no registry access; setup does.

fixtures.json and the JSON observation are authoritative. Inputs remain strings: leading zeros and long numeric identifiers are never converted to Number. The 28,000-zero input uses a bounded declarative repeat generator. Every expanded input has UTF-8 byte length, Unicode code-point length, UTF-16 length and SHA-256. CSV contains only validated labels, lengths, hashes and numeric observations—not raw identifiers/inputs that spreadsheets could round or interpret as formulas. Reports contain no timestamps, absolute paths, headers or private-trace content.

Independent verification

A fresh whitelist-only installation completed npm ci --ignore-scripts --no-audit --no-fund with 172 packages and no engine warning on the recorded Node runtime. After a focused runner correction, an independent recheck in that same locked installation passed 7 tests, reproduced two identical JSON runs and two identical CSV runs, and confirmed all 16 inputs / 64 estimates. The 23 deliberate guard probes remain separate from normal observations with zero guarded attempts.

The recheck also exercised caught calls deferred by beforeExit and calls in a later exit listener: those processes and the launcher return nonzero, even when stdout was already written. Exit status is authoritative. These checks do not make the process-local guard an OS sandbox.

Recorded SHA-256 values:

observed.json  8a2d202728389fe5322547e3da76991b0208813567d6b1ae58534f89778d4537
observed.csv   21dae204914173075c8163b0051cc8cfaf164646695f3d43eaf7daf85e8efdb8

Reuse

Original probe code and synthetic data: Apache-2.0, see LICENSE. Installed third-party packages retain their own licenses (including Apache-2.0 and MIT); this packet does not relicense or redistribute node_modules. The issue workflow requires waiting for triage labels before opening a source PR; this standalone data packet is not a source fix or a claim of upstream approval. Development environments, dependencies, internal handoff drafts and private verification records are not part of the reusable delivery.