Repository navigation
failproofai 1.0.14
Changes since 1.0.13 (v1.0.13..main at 18a79c7, 2026-11-13).
Highlight: telemetry survives delivery outages
In 1.0.13, the failproofaid collector handled a long network or server outage the same way it handled a rejected batch. After three failed passes the batch was marked .poison and never sent again, so events recorded during an outage could stay stuck even after the server came back. In 1.0.14, network errors and HTTP 408, 429 and 5xx responses never use up a batch's retries. The batch stays on disk, unchanged, and is sent automatically once delivery works again. This applies to both NDJSON event batches and OTLP exports. (#997)
Fixes
Outages no longer poison telemetry. Network failures and 408, 429 and 5xx responses are always retried, however long the outage lasts. (#997)
An interrupted acknowledgement keeps the batch. If the connection drops while the server's response is being read, the batch is kept and retried. It is no longer counted as delivered or as rejected. (#997)
Permanent rejections are kept and not resent. A batch the server rejects for good, such as 400 or event_too_large, is kept under state/failed/ for inspection. It is not sent again automatically and is never deleted. (#997)
Partial acknowledgements have their own budget. A 200 that stores some events and skips others is now counted separately from outage passes. A recovered endpoint can finish delivering the batch, and a skipped event is counted only once in metrics. (#997)
Credential refusals stay recoverable after an outage. A batch refused with 401/403 after an outage is still redriven when a working key arrives. Refusals have their own bounded budget, so repeated refusals can't hold up other batches. (#997)
failproofai reads the new retry state. The CLI's delivery status understands the new .sN (skipped acks) and .rN (client rejections) suffixes on parked batch filenames. (#997)
macOS ARM releases build again. The darwin-arm64 daemon is now built and install-tested on macos-15 runners. GitHub retired the macos-14 runner image on 2026-11-02, which would have blocked this release. (#999, #997)
Upgrade notes
Old .poison files are not resent. Batches marked .poison before 1.0.14 stay where they are. Recover them by hand.
OTLP exports refused with 401/403 need manual recovery, even after the key is fixed:
Stop the daemon.
Confirm you have a working events:add key for the same Cloud origin and organisation.
Back up the affected files under <FAILPROOFAI_HOME>/state/failed/otlp/.
Move only those batches into state/spool-otlp/ with fresh otlp-.jsonl names. Don't overwrite any file, and don't change the metadata or body bytes.
Restart the daemon, confirm delivery, then delete the backup.
Never move batches that org_boundary archived into another organisation's spool.
Security and dependencies
Daemon: rustix 1.1.4 → 1.1.6 for RUSTSEC-2026-0486 (#999).
Daemon: rusqlite 0.40 → 0.41 (bundled SQLite via libsqlite3-sys 0.39) and notify 8.2.0 → 8.2.1 (#975, #999).
CLI/dashboard: posthog-node 5.61.1 and sql.js 1.15.2 (#976, #979, #974).
Docs
New pages for the MCP server, the Slack app and product analytics (#982).
The ingestion retry docs, OTLP credential-refusal recovery steps and troubleshooting pages now match the new delivery behaviour. Every public OpenAPI operation now has a unique ID. (#997)
Not in this package
fp login browser sign-in (#981) ships in fp-cloud-cli on PyPI, not in the npm package, and the PR is still open.
Bumps to the @failproofai/sdk integration fixtures and to the Python SDK's test dependencies (oauthlib 4.0.1) don't affect failproofai users (#966, #999, #1000–#1003).