Agent ETA — DeepSeek reliability preview

Makes the DeepSeek panel and prospective comparisons more reliable without changing either estimator.

• Show only a forecast that was successfully written to the journal, with its timestamp. Withdraw stale numbers after 60 seconds until a new event arrives.
• Show cold-start history progress, preserve collapse preference, and keep both estimates visible in the compact row.
• Keep last-write time correct regardless of journal filename order; disclose incomplete history lines.
• Enforce the existing restart exclusion: a partial turn cannot borrow its start or prediction from another host mount.

Validation: 326 local tests, the controlled subprocess failure/rework fixture, and all four GitHub CI jobs (Windows/Linux, Node 22/24) passed. Native macOS DeepSeek Harness panel and compact comparison checked through computer use. No added model requests; synthetic UI examples remain excluded from real data. No claimed accuracy improvement. Estimator parameters and evaluation windows are unchanged.

Install and configure · Protocol and limitations

Agent ETA — DeepSeek Harness panel preview

Adds an experimental native DeepSeek Harness panel above the conversation input.

• Original remaining-time estimator retained; candidate forecasts shown alongside it.
• Step, test-failure and rework signals, recording counts and last-write time.
• Local-only observation with no additional model calls; explicit cold-start and error states.
• Labelled UI-only demo, excluded from real observations.

Install using DeepSeek Harness's bundled CLI:

dsh plugin --profile desktop add 'github:maxi-max-dev/agent-eta#v0.12.0-deepseek.5' --ignore-scripts

Setup and private log directory · Collection protocol

Validation: 318 local tests passed; native macOS DeepSeek Harness 0.2.1-rc.3 panel checked through computer use; isolated real subprocess fail→edit→pass fixture passed. The fixture is not an autonomous coding accuracy benchmark. No demonstrated accuracy improvement yet. Restart currently resets estimator warm-up while retaining journals and recording counts.

MIT licensed. GitHub preview only, not an npm registry publication. The portable v0.11.0 release remains available.

Agent ETA v0.11.0 — Portable synthetic walkthrough

A new npm run demo:portable walkthrough makes the portable lifecycle and evidence pipeline inspectable without an existing tracker database. It shows cold start, experimental estimates, a ten-minute pause that adds no active time, terminal status, stale observation and recovery with an observation gap. A three-minute presentation guide explains each state and its limits.

Every run, history, duration and outcome is explicitly synthetic. The example uses a fresh temporary store with synthetic-demo-only, rejects existing-database arguments and ignores configured stores. It retains labeled artifacts, the original export, verified file integrity, evaluation/comparison reports and absolute argument arrays for read-only reproduction. It does not measure a real agent, establish accuracy, certify novice usability or submit to a competition.

Validation: 302 local tests, eight targeted checks, clean installed-package CLI/SDK/HTTP checks, original-store byte preservation, and saved-demo report reproduction across working directories. Four public Linux/Windows × Node 22/24 CI jobs passed before release. Same-input synthetic prediction and evaluation outputs remain identical to v0.10.0. The 106 packaged files match the reviewed source and contain no private workspace data.

The model, frozen evaluation protocols, lifecycle semantics and retention policy remain unchanged. Independent future real prediction evidence is still pending. This is an experimental GitHub release, not an npm registry publication.

Agent ETA v0.10.0 — Saved export verification

Saved journal exports now have a built-in acceptance check: agent-eta verify-export FILE. It streams the existing file, checks its envelope, run/forecast ordering, manifest/end/actual counts and original-byte SHA-256, and exits nonzero for truncated, appended or inconsistent output. The old agentwhen alias also supports it.

Verification reads only the named file, creates no database and preserves every input byte. A complete export can still contain damaged original forecasts or orphan receipts; these remain evidence, with forecast validity explicitly unchecked. The report does not establish the original export process exit status, equality with a source database, authentic prospective collection or prediction accuracy.

Validation: 300 local tests; clean installed-package CLI/SDK/HTTP checks; unchanged same-sample synthetic prediction, evaluation and holdout output versus v0.9.2; four public Linux/Windows × Node 22/24 CI jobs. Packaged files were checked against the reviewed source and scanned for private database/workspace artifacts.

The default model, frozen evaluation protocols and original records remain unchanged. Independent real future-holdout evidence is still pending. This is an experimental GitHub release, not an npm registry publication.

Agent ETA v0.9.2 — Isolate SDK example tracking failures

Agent ETA v0.9.2 repairs the runnable JavaScript SDK example so a tracking failure does not abort the operation it is observing. Previously, a failed timer-driven heartbeat could escape as an uncaught exception, and a failed terminal write could replace the operation's original error.

• The copyable example preserves the operation's exact return value or original thrown value.
• A heartbeat failure emits an incomplete-tracking notice and stops that timer; a final write remains best effort. Failed final writes are never reported as saved completion.
• Tracking startup failures still allow the operation to execute. Status/warning callback exceptions and rejected promises are contained, and the timer and constructed tracker are cleaned up when the operation settles.
• Document the boundaries: this is example code, not a new package SDK export. Lower-level SDK methods, the CLI wrapper, default estimator and retention are unchanged. CPU-blocking work, human waiting, cancellation and remote work need an appropriate explicit lifecycle or process boundary.

Validation: 294 local tests, including isolated SQLite write failures and error-preservation regressions. Clean-package checks cover the runnable installed example, actual heartbeat/final-write failure, CLI/SDK/wrapper/HTTP behavior, and identical forecasts/evaluation reports against v0.9.1. Public Linux/Windows × Node 22/24 CI passed on the release commit.

This is an integration-reliability fix backed by synthetic fault-injection checks. Independent real-user adoption and future prediction accuracy remain pending. No private databases, raw conversations or provider logs are published.

Copyable example · Failure semantics and limits · Roadmap

Agent ETA v0.9.1 — Faster latest-run lists as history grows

Agent ETA v0.9.1 reduces the work needed to list the latest 50 runs as the local database grows. A start-time index replaces a full-table scan and sort, while preserving newest-insertion ordering when start times tie.

• Existing stores gain the index on their next writable tracker open. Run rows, original forecast receipts, run IDs and the latest-50 cutoff are preserved.
• Default predictions, history selection, heartbeat semantics, refresh frequency and retention are unchanged. Read-only doctor/export/usage do not migrate the database.
• Include a reproducible synthetic benchmark with raw samples, environment, query plans, upgrade cost and storage cost.

On an Apple M4 / Node 22.23.1, the complete list's median duration with 100,000 synthetic terminal runs fell from 28.52 ms to 12.13 ms (about 57%). First open including index construction took 18.42 ms once and logical storage grew by about 1.41 MiB. Small databases showed smaller benefits. This is local synthetic list performance, not end-to-end browser latency, real-agent throughput or improved forecast accuracy.

Validation: 290 local tests, including upgrade preservation and state/ordering regressions; clean-package CLI/SDK/wrapper/HTTP checks; identical same-input forecasts and evaluation reports against v0.9.0. Public Linux/Windows × Node 22/24 CI passed on the release commit.

The release remains experimental. Independent future prediction validation remains pending, and no model is promoted. No private databases, raw conversations or provider logs are published.

Benchmark and reproduction · Raw synthetic samples · Roadmap

Agent ETA v0.9.0 — First-run dashboard commands that work from any directory

Agent ETA v0.9.0 fixes the first-run dashboard instructions for installed packages. The empty page now generates a wrapper template using the running server's actual database and absolute executable/CLI paths, so users can start from another working directory instead of relying on a checkout-relative command.

• Show the local database and a copyable POSIX-shell or PowerShell command template. Replace the trailing COMMAND with real work; copying executes nothing.
• Escape paths with spaces and apostrophes, retain clipboard fallback feedback, and hide setup instructions when a run is selected or the server disconnects.
• Keep the existing local-origin/host/cross-site/GET-only guard around setup metadata. No URL parameter selects the local file. Disabled and in-memory trackers offer no persistent-store command.
• Preserve the default estimator, forecast semantics and existing CLI/SDK compatibility.

Validation: 288 local tests passed; clean-package CLI/SDK/wrapper/HTTP checks passed. Browser acceptance checked the empty page, matching clipboard content, shell switching, a 390 px layout without horizontal overflow, cold start, pause and disconnection. Public Linux/Windows × Node 22/24 CI passed on the release commit.

The current Codex task also used the published v0.8.0 wrapper for the actual project test process: 18.305 seconds, successful exit, four real-clock cold-start receipts, zero numeric forecasts. This is one controlled process-boundary integration case, not a native vendor hook, independent user study or accuracy holdout. The desktop build was not obtained. Private databases, raw session logs and screenshots containing local paths are not published.

The release remains experimental. Real numeric accuracy and broader host compatibility remain unproven; no model is promoted.

Onboarding details and evidence limits · Roadmap

Agent ETA v0.8.0 — Read-only onboarding diagnostics

Agent ETA v0.8.0 makes portable onboarding easier to diagnose. When an agent changes working directories, doctor shows the actual database path and selection source, the matching history count, and absolute argument arrays it can reuse to start tracking and open the dashboard.

• doctor --db DATABASE --profile PROFILE --class CLASS is read-only: no database/folder creation, forecast capture, heartbeat, provider-log import or server startup.
• Distinguish an uninitialized path, a readable tracker schema and an unreadable/foreign/corrupt store. Blocked stores are never repaired or overwritten.
• Explain cold start using the same eligible-history selection and 200-run cap as the predictor. Keep the experimental, uncalibrated boundary explicit.
• Return argument arrays that work across directories and paths with spaces. Report the dashboard address as untested until the caller starts it.
• Update the bilingual setup instructions and portable Skill. Retain the legacy CLI alias and database-path precedence.

Validation: 283 local tests passed, including seven diagnostic tests. Clean-package acceptance exercised cross-directory CLI lifecycle calls, a synchronous process wrapper, the JavaScript SDK and the same-database loopback HTTP dashboard. Database/WAL bytes remain unchanged during diagnosis. Identical synthetic inputs retain v0.7.0 forecasts and evaluation/comparison output. Skill validation passed. All four Linux/Windows × Node 22/24 CI jobs passed on the release commit (CI evidence).

This remains experimental. Wiring runs and controlled clocks are not accuracy evidence. This release does not establish vendor-native integrations, independent novice-user usability, write permission, an unattended collection service, or improved real-world prediction accuracy. The default estimator is unchanged, and private conversations/databases are not included.

Onboarding guide · Roadmap

Agent ETA v0.7.0 — Frozen review targets and guarded holdout checks

Agent ETA v0.7.0 adds a reproducible review step for the existing empirical shadow candidate. Declare a fresh future window and exact workflow with gate-plan, then use gate to check saved paired forecasts against targets frozen before implementation.

• Freeze scope, primary/baseline/candidate versions, and fixed error, severe-underestimate, coverage, interval-width, availability and sample-count targets.
• Withhold accuracy metrics until the start window has ended and all in-scope tasks have closed. Reject invalid or future-dated evidence and mixed model versions.
• Report every observed value and threshold. A numerically favorable result still requires independent review; no command promotes the candidate or changes the default estimator.
• Preserve v0.6.0 evaluation/comparison output and the legacy CLI alias. Normalize macOS disconnected export pipes to the existing closed-output error.

Validation: 276 local tests passed; a fresh package installation verified the CLI/API, read-only behavior, fixed-clock determinism, unchanged v0.6.0 reports, missing-file behavior and package contents. All four Linux/Windows × Node 22/24 CI jobs passed on the published commit (CI evidence).

All new acceptance samples use synthetic data and controlled clocks. This is an experimental tooling release: it does not establish improved real-world accuracy, an independently collected holdout, or sufficient statistical power. The review targets are engineering screening criteria. Private databases and conversations are not included.

Protocols: candidate review targets · roadmap

Agent ETA v0.6.0 — Frozen cohorts and future validation windows

Agent ETA v0.6.0 adds reproducible separation of old tuning runs from a predeclared future run window. It keeps the displayed ETA and default model unchanged.

• cohort-plan reads an existing local database, freezes completed tuning runs and their saved receipt IDs, and emits a manifest with future UTC start/end bounds.
• compare --plan FILE --cohort tuning|holdout assigns complete runs by start time, preserves pending and cross-window long runs, rejects edited/deleted frozen rows, and ignores predictions added later to tuning runs.
• Reports include profile/class and fixed active-duration strata, sparse-history abstentions, missing/invalid records, and all cohort exclusions. Unplanned compare and evaluate remain compatible with v0.5.0.
• The time-split protocol was frozen in commit 5f2c702 before implementation or inspection of a target cohort's results. Usage is documented in both READMEs and docs/TIME-SPLIT.md.

Validation: 263 local tests, 12 cohort-specific tests, clean tarball installation and CLI checks, byte-identical unplanned reports versus v0.5.0 on the same synthetic fixture, unchanged database/WAL checks, and Linux/Windows × Node 22/24 public CI.

This is an experimental release of evaluation tooling. Synthetic fixtures and controlled clocks verify behavior, not real forecast accuracy. A manifest/hash does not prove that a holdout is real or untouched. Promotion gates and an independently documented future real cohort remain pending; promotionEligible stays false. Keep local manifests, reports and databases private. No data migration, upload, deletion, new background service or model promotion is introduced.

中文：本版把旧样本与未来新运行按事先约定的时间切开，补齐任务长度和类型分层，防止事后补预测或只挑已经结束的好看样本。测试通过不等于预测更准；默认模型保持原样，下一步是先冻结晋升门槛，再收集独立未来实测。