v0.26.0

0.26.0 (2026-10-04)

Features

• admit signed measured fleet data with owner permission (#205) (dcfb49b)

v0.25.0

0.25.0 (2026-10-04)

Features

• add local fleet data hub and launcher view (#204) (feca029)
• telemetry: record safe decision evidence and read-only timelines (#202) (2e01297)

Bug Fixes

• git: report bounded probe startup diagnostics (#199) (6ea0200)
• runtime: bind rewards and promotion to verified evidence (#198) (49aeaac)

v0.24.0

0.24.0 (2026-09-29)

Features

• package: admit optional model, dataset, knowledge and report groups (#191) (f719ee1)
• web: drive browser and Three.js games through the contract (5291135)

Bug Fixes

• web: sink the validation seed into every episode reset (961b817)

Documentation

• decisions: record the episode termination contract as ADR-0047 (cb43aa4)
• package: align skills and guides with the implemented package contract (#194) (f62dd2f)

v0.23.0

0.23.0 (2026-09-28)

Features

• entry-point: pin one entry point per project with launch attestation (9452131), closes #159
• episode: make an indeterminate outcome absorbing (GLR #155) (226f1b9)
• episode: require a termination reason for every episode (GLR #156) (af2f010)
• glr-cli: bind a default goal per project (#151) (8422166)
• glr-cli: record the trainer window during glr train (fd652f1)
• glr-cli: report offline conformance for materialized packages (#163) (859b4f4)
• glr-recording: add PID-scoped Windows Graphics Capture recording (345a5fe)
• governance: add agent onboarding, modular tools, watchdog and fork gate (852332f)
• hooks: register lifecycle hooks per environment and dimension (#175) (5cacd40)
• learnability: bound state-action cardinality against the step budget (1a3d233)
• learnability: record the declared scale as a first-class metric (f43ccd8)
• metrics: fail closed on a declared metric that is never emitted (f19edb9)
• project: let a project declare the environment its roles receive (#190) (b37d4cc), closes #169
• safety: bound negative shaping with optional budgets (5aa7796), closes #161

Bug Fixes

• align release metadata and formatting with the recording crate (0bb5d2e)
• cli: name the real terminal state in the capture warning (bb94a74)
• cli: publish one learnability shape from train and runs show (#180) (c580983)
• cli: publish the absent learnability verdict with null fields (824ed59)
• cli: record entry point candidates the walk cannot inspect (#185) (2cba0d4)
• entry-point: fail closed on a directory the walk cannot list (#176) (97d15d6), closes #159
• episode: bind the termination sink with the run (6b8c5cb)
• episode: land every termination in the run store (b58d437)
• goal: revalidate stored bindings and record the context source (6ff1985)
• governance: harden the watchdog contract and the layout gate (e6d132e)
• learnability: land the coverage floor a game-backed run can reach (c697862)
• learnability: let coverage efficiency decay with every step (90a356f)
• learnability: re-export the state-action-cells metric from the package (22705d2)
• package: share one per-component local override predicate (#181) (db37185)
• recording: resolve review findings on PID window recording (78d003d)
• safety: fail closed on unclassified reward terms (#173) (ff39105), closes #160
• tests: pin utf-8 on locale-bound read_text and write_text calls (404f67b)

Documentation

• add portable training package ADR and phased plan (d4844d4)
• decisions: promote the remote admission contract to ADR-0045 (b6a86c5)
• docs/guides/learnability-budget.md and ADR-0039. (1a3d233)
• document window recording configuration and outputs (cb3e42e)
• entry-point: exit code 2 on a run is not retryable; pin the unlistable root (#177) (36c4110), closes #159
• planning: anchor remote admission on symbols and align with ADR-0041 (71b6681)
• planning: propose authenticated remote role admission contract (e52cd49)
• planning: record the design review rulings for remote role admission (d54832d)

v0.22.0

0.22.0 (2026-09-17)

Features

• define module boundaries and build trained-stage installers (48b37c4)

v0.21.0

0.21.0 (2026-09-17)

Features

• bound the runtime start readiness window (1a54aa7)
• give each workbench server an instance identity (#143) (182fb69), closes #141
• host an externally driven loop inside a GLR run (0d0ece8)
• standardize downstream Python quality and migrations (f809476)

Bug Fixes

• bind training roles to one trial identity (00a260a)
• include quality templates in skill publication (15279c1)
• own the GLR environment namespace for every child process (221a155)
• pass a hosted command line through without expansion (9b276c3)

v0.20.0

0.20.0 (2026-09-16)

Features

• add adaptable agent workbench and process trace (2252bb7)
• add bridge telemetry ingestion and dashboard state views (332b815)
• add dashboard video and evidence workspace (a59d797)
• add embedded training dashboard and durable observability (f13cdfb)
• embed React dashboard assets through CI (52c1e26)

Bug Fixes

• render structured process logs and document data collection (481e6c1)

Documentation

• clarify bridge telemetry ingestion contract (2c8a2a1)
• format workbench integration example (e25a286)
• sync data collection skills into agent plugin (7fc3770)

v0.19.0

0.19.0 (2026-09-15)

Features

• add offline source package handoff (6b77e9b)
• align recording skills and update notifications (b9e6e36)

Bug Fixes

• preserve Python run store compatibility (58d24e4)

Documentation

• add interface-first game training showcase (35673d4)
• improve training media playback (30b724e)

v0.18.0

0.18.0 (2026-09-15)

Features

• add declarative plugin extension system (#131) (e2bd5df)
• normalize managed project output layout (8cbf92d)

Bug Fixes

• sync skills on current release (84d48f5)

v0.17.0

0.17.0 (2026-09-14)

Features

• apply updates and persist staged checkpoints (33ada85)