v1.0.20

Codex discovery client update

Updated the Codex model-discovery client version from 0.155.0 to 0.159.2.

Model IDs, authenticated account filtering, authentication, provider routing and request budgets are unchanged. No daemon protocol or schema changes.

Source: 65a1a411a49207d54ce97515325535e3ff25e2ce.

The focused discovery regression and full project check passed, including normal commit hooks. Fresh local installed-release qualification passed 12/12 (installer and npm), including the selected model catalog, 17 Python imports and SDK/RPC checks. Actual installed Python: 3.13.12; runtime: 1.0.20. Exact-commit CI passed. Fresh unchanged default-installer qualification passed 9/9 with no version or source overrides. Installed Node 22.23.3, actual Python 3.13.15 and runtime 1.0.20; exact clean source commit, 17 runtime imports and six selected model/provider entries. No model requests or provider account queries. All qualification profiles were shut down last, checked for remaining owned processes, and frozen. No live research sessions were upgraded or restarted.

v1.0.19

Fixed

• Fixed intermittent Native session catalog stale failures during child creation, family messaging, and child listing. Live local and supervisor-owned sessions use current owner metadata instead of racing against their on-disk indexes.
• Preserved inactive-session validation, family visibility, session-name collision checks, and local/remote message routing. No source-consistency or authorization guard was relaxed.

Source qualification

• Exact source: 7def07724e24391c377aa62348717165f9c3347b; clean source build using Node 22.12.0.
• 180 targeted TypeScript tests across 17 files passed; no failures or skips. Full project check and normal commit hooks passed.
• Native regression cases reproduce the original stale-index errors and verify that truly stale inactive catalogs still fail.
• Existing fixture mismatches were reproduced on unchanged 1.0.18 before correction. Kernel test setup was corrected without changing bootstrap behavior or assertions.

Installed qualification

• Fresh local installer and npm qualification: 12/12 passed.
• Actual installed Python 3.13.12, runtime 1.0.19, 17 runtime imports, SDK/RPC smoke, and six API/Codex catalog entries. No model requests or account queries.
• Shutdown ran last for both isolated profiles; no owned/controller processes remained. Profiles are frozen.

Exact-commit CI passed. Fresh unchanged default-installer qualification passed 9/9 in one fresh profile, with no version or source overrides. Installed Node 22.23.3, actual Python 3.13.15 and runtime 1.0.19; exact clean commit, 17 runtime imports and six selected model/provider entries. No model requests or account queries. Shutdown ran last, no owned/controller processes remained, and the profile is frozen. No live research session was upgraded or restarted.

Base Context 1.0.18

Fixed

• Added gpt-6.1-sol for OpenAI API keys (openai) and Codex subscriptions (openai-codex), with documented pricing, context limits, and mandatory low-through-max reasoning.
• Updated Codex discovery to include account-authorized GPT-6 Sol and Luna. Discovery still rejects unknown or unauthorized models.

Qualification

• 154 targeted tests passed, including API/subscription request routing, model-picker registration, and authorized discovery. Full project checks passed.
• No real model inference was used. Subscription access remains subject to the account's model availability.
• Clean source fabbb3fa90a47d6d87931fdcc67886385821ef9f; native CI passed: https://github.com/BaseModelAI/base-context/actions/runs/36611040317.
• Fresh local qualification passed 12/12, including six installed model/provider entries. An initial QA-only ESM checker error was corrected before the fresh run. All test profiles were shut down and frozen.
• Public default installation passed 9/9 in nine attempts with no version overrides: Node22.23.3, Python3.13.15, runtime1.0.18, all17 Python imports and all six model/provider entries. Shutdown and owned-process cleanup completed; the profile is frozen.
• All four public npm latest tags resolve1.0.18. No existing live session was upgraded or restarted.

Base Context 1.0.17

Fixed

• Compaction can summarize old recovered tool output instead of keeping it permanently in active context. Original evidence remains recoverable, with source ownership and acknowledgement checks preserved.
• Queued job-watch status snapshots keep the latest observation per registration. Important events remain visible; already-processing input is unchanged. Unregister removes queued routine updates without stopping the job.
• Session attachment labels native context checkpoints accurately rather than counting them as summary calls.

Qualification

• 471 targeted TypeScript tests and 3 Python tests passed; full project checks and normal commit hooks passed.
• Clean Node 22.12 build from 29e8dd35411c44d7716b0dc73c988634a94c9d36.
• Fresh local installer and npm qualification: 12/12 passed, with zero inference. Both isolated profiles were shut down and frozen.
• Native CI passed for the exact source commit: https://github.com/BaseModelAI/base-context/actions/runs/36601207729
• Fresh public-default installer qualification: 9/9 passed without selection overrides or inference. Selected 1.0.17, managed Node 22.23.3 and Python 3.13.15; all 17 runtime imports passed. The isolated profile was shut down and frozen.

Base Context 1.0.16

Base Context 1.0.16

• Added job-watch for native job monitoring, prepared reports, and scoped waiting in goal and autonomous modes.
• Added bounded-inspect for selected text and JSON evidence under shared source/output budgets, with lossless numeric text.
• Improved skill recovery, child-result capsule guidance, and reuse of small project helpers.
• Fixed background child-usage updates stopping an unfinished goal during request preparation. Source/branch ownership and native epoch acknowledgement remain strict.
• Kept failed goals visible as Goal error in the bottom-right status tray.
• Fixed temporary memory retention for a final multi-chunk line without a trailing newline.

Validation

429 distinct TypeScript cases and 39 Python cases passed across named selections. Autonomous-only and combined goal/autonomous paths were tested with mocked providers and real local kernels. Full repository checks, normal commit hooks and the clean source build passed. Fresh prepublication installation qualification passed 12/12, including packaged skills, 17 runtime imports, SDK/RPC and clean shutdowns, without inference. Fresh public-default installer qualification also passed 9/9, including both skills and all 17 runtime imports, using managed Node 22.23.3 and Python 3.13.15. Its isolated profile shut down cleanly.

Source: 2ec583d3c870d042a46a5ecc1e6ffa83b8edd95e.

Native CI passed.

Important limits

Watch state is bounded. A prolonged undelivered backlog can explicitly stop monitoring with unknown/refusal while the job keeps running. Recovery does not claim exactly-once delivery after an ambiguous crash. Unsafe structured probe numbers require strings.

Inspector cursors use stat identity, JSON has an explicit source ceiling, and the outer transport envelope is separate. No model-cost or training-throughput saving is promised.

Publishing does not upgrade running agents. Check ownership of background jobs before stopping or restarting workers.

Base Context 1.0.15

Base Context 1.0.15

• Fixed compaction of long-lived sessions by selecting the active canonical context before source hydration. This covers manual, agent-requested, threshold, and overflow-recovery compaction without increasing source limits.
• Fixed Codex HTTP retries losing native context acknowledgement after retryable failures. Only an unchanged prepared request can reuse its acknowledged epoch.
• Changed /goal stop and /goal clear to acknowledge the goal that was cleared instead of reporting only the resulting idle state.
• Fixed daemon worker connection retries retaining failed socket and channel references.
• Handled Python kernel stderr stream errors through the existing kernel error path.
• Fixed the goal status display being overwritten by delayed state reads and streamed catch-up snapshots.

Upstream review: screened 222 Prime Agent commits through 2026-09-27. The selected changes and deferred/rejected adaptations are documented in packages/coding-agent/docs/upstream-1.0.15.md. No PrimeRL provider, subscription, or model changes are included. Native capsule, source, replay, ownership, and acknowledgement contracts remain in place.

Base Context 1.0.14

Base Context 1.0.14

Fixes

• Fixed compaction recovery and saved-session resume after failed or interrupted partial tool calls, without replaying failed provider output or weakening source checks.
• Enforced autonomous turn, token, and time limits inside native tool loops at completed-turn boundaries. Running tools finish normally; default limits and token accounting are unchanged.
• Fixed JSON print mode returning success for terminal assistant, session-command, or compaction failures. Unrun autonomous quality gates are no longer treated as passed.
• Fixed daemon admission for custom --session-dir transcript directories while preserving supervisor ownership checks.
• Rejected unknown explicit CLI models before saved-model fallback or runtime creation.
• Preserved nested Codex error messages and codes instead of reducing them to an unknown error.

Models

• Added GPT-6 Sol and GPT-6 Luna to OpenAI API and Codex subscription model pickers, with provider-specific reasoning and context metadata.
• OpenAI API context capacity is 1,050,000 tokens. Codex uses its documented configurable maximum of 872,000 tokens; its upstream default of 272,000 is a separate setting. Output capacity is 128,000 tokens.
• Catalog cost estimates do not represent Codex subscription invoices. Model availability still depends on provider/account access.

Documentation

Clarified headless exit status and stderr handling, client-owned versus resident session lifetimes, stable session identity, persistent goal budgets, autonomous token accounting, and existing delegation controls. No change to the core operating model or default autonomy limits.

Packages

All four npm packages and the bundled Python runtime are version 1.0.14. External dependency versions are unchanged.

Qualification

• 500 focused tests passed, including native autonomous tool loops, daemon lifecycles, goal persistence, cold recovery, model pickers and offline Responses routing.
• Formatting, type, installer and browser checks passed (1,060 files).
• Prepared from clean source commit 1c9651df8abd812cdba9e323517067363acc621c.
• 11 isolated prepublication checks passed for the separate npm and GitHub artifacts: installation, selected build, CLI, Python runtime, SDK/RPC, and shutdown.
• All four npm packages are public at 1.0.14; downloaded public npm and GitHub archives match the qualified artifacts.
• 8/8 fresh public-default checks passed: normal installation, installed Node/build/CLI, release-local Python runtime, and shutdown. The default installer selected the clean release commit above.
• Live account acceptance and paid inference were not tested. Existing research installations were not activated or changed.

Base Context 1.0.13

Base Context 1.0.13

• Changed the default automatic compaction target to 90% of the model context window for root sessions and recursive subagents.
• Retained required-context headroom and the model-window output reserve. Explicit numeric targets and "model-limit" remain unchanged.
• Preserved earlier agent-requested compaction through await compact.run(...), including Astra, and documented its settings and turn-boundary behavior.
• Preserved the 1.0.12 child recovery, cache stability, status coalescing, and usage-accounting fixes.

No live research sessions or default installation were activated or restarted during release qualification.

Source qualification:127 focused tests passed;2 credential-gated live tests skipped. Native threshold regression covers depths0,1,2,3. Root lint/type/installer/browser checks and clean source build passed. No paid inference was used.

All 11 isolated prepublication checks passed for the separate GitHub and npm archives, including CLI, Python runtime, SDK, RPC, and test-owned shutdown.

Base Context 1.0.12

Base Context 1.0.12

• Fixed primary-child compaction and recovery, resumed child usage forwarding, nested attribution, and terminal error/cancellation status.
• Added receipt-based per-purpose and captured-family usage to /context, with explicit missing data, goal scope, and catalog-cost labels.
• Reduced dashboard inference with unchanged-input reuse and one-minute refresh coalescing.
• Kept refinement advice fresh without rewriting the system prefix. Added stable cache identities for auxiliary requests.
• Added earlier model-aware compaction targets with required-context headroom and a model-limit opt-out.
• Fixed SDK abort-listener retention and Azure API-version loss.
• Added guidance for bounded output paging and nonurgent update batching.

All four npm packages use version 1.0.12. Usage estimates are not provider invoices, and cache savings are not guaranteed.

Validation: 915 source tests passed (11 skipped). Separate npm and GitHub archives passed 11 isolated installation, CLI, SDK/RPC, and Python runtime checks.

v1.0.11

Base Context 1.0.11

Fixes

• Restored recovery compaction for long-running sessions that use prime_context without a request-budget profile.
• Deferred pre-turn recovery until a real containing request establishes replay coverage, without reporting a false compaction failure.
• Preserved exact recovery coverage across over-budget requests and repeated summaries.
• Kept valid full-native payload-hook metadata compatible with optional replay-contract capture.
• Fixed large-session status, listing, and attachment by reading captured usage totals instead of the entire journal.
• Fixed /tree, /context, fork-message selection, and refinement history reads over large unrelated archives.
• Fixed small selected forks over large histories while preserving required updates, invocation records, ancestry, and copy limits.

Checks

• 220 tests passed across the final focused files. Eight existing runtime fork cases and the existing refinement controls also passed.
• Full source, installer, and browser checks passed on Node 22.12.
• A captured 197 MB session passed real isolated daemon open, detach, same-session reattach, tree/context views, message selection, and small-fork creation.
• Native automatic recovery and a fresh-process resume with another summary retained all three recovery records with no failed compaction outcomes. The following main request shrank from about 7.25 MB to 69 KB. These probes used local fake responses, not paid inference.

Replay acknowledgements, source identity, cancellation, dependencies, and selected-output limits remain enforced. Full-history exports/copies and provider requests remain capacity-bounded. This release does not update or restart any running session, sub-agent, kernel, schedule, or external job.