Repository navigation
Releases
Releases · ilanbm/kpopper
kpopper 0.18.0
Changed
- New projects automatically keep one private local record under
~/.kpopper/projects/<project>-<uuid>/, or an absoluteKPOPPER_PROJECTS_DIR.
The first valid record-creating write selects its location without a mode-choice
prompt or announcement. Reads and previews do not create a record or binding. - Existing records and explicit configurations retain their location and behavior.
Related worktrees share a stable project identity; copied projects require explicit
recovery before writing another owner's record. No existing record is migrated
automatically. Central storage does not provide remote synchronization or publication
permission. See project knowledge storage for recovery,
source resolution and remote-environment requirements.
Experimental
- Advanced branch-specific storage is explicitly selected with
kpop experimental advanced.config --mode advancedremains a compatibility
alias. Existing Advanced projects continue to work; their record merge limitations
remain documented separately.
Included changes
- Fix cross-platform Git path fixtures
- Default new projects to Simple storage; make Advanced experimental (#322)
Decisions recorded: d.simple_default_project_storage
kpopper 0.17.3
Fixed
- Core references can read a judgment's authored
verdictas a premise. This
removes incorrectmissing_referencefindings when one judgment depends on
another, preserves change detection when the verdict changes, and keeps the
conclusion's truth and acceptance separate from reading its text. Existing
records require no migration.
Included changes
- Fix core references to authored judgment verdicts (#320)
kpopper 0.17.2
Fixed
- Ordinary record reads evaluate recorded structured conditions together and reuse
their results within the same captured world. This removes repeated full-record
evaluator launches from opening, checks and attention on larger records, while
preserving the existing per-condition evaluator when a batch cannot be used. - Skills, host adapters and hook recovery guidance direct agents and delegated
workers through the bounded kpopper reader. An unavailable reader is reported
explicitly instead of prompting direct loading or searching of the record YAML.
Existing records require no migration.
Included changes
- Batch ordinary predicates and require bounded record reads (#319)
- Trim routine platform tests and migrate Apple Silicon CI to macOS 15 (#317)
Decisions recorded: d.record_yaml_is_reader_boundary, d.structured_predicates_batch_once
kpopper 0.17.1
- Claude and Codex receive a bounded knowledge-handoff reminder during the current turn,
including for research and evidence saved outside the repository. It asks for verified
retention or an explicit unresolved outcome without creating a Stop continuation.
Recording guidance distinguishes accepted writes, existing coverage, drafts and failures;
successful routine recording stays quiet. Prompt identity survives record parse failures,
and context compaction re-arms the reminder without creating another user request. - Report writes acquire private-state and record locks in canonical order, including when
the state directory sorts before the record. Interrupted writes recover under the same order. - Report batches retain ownership for each written entry, so the writing session can correct
its new entries while another session remains refused, including after an interrupted
write is recovered. Older unbound journals do not grant ownership to the retrying session.
Legacy history receipts preserve YAML dates using the existing typed JSON format instead
of reporting an error after a successful write. Invalid undeclared predicate values are
refused before publication with guidance to use an executable predicate orreopened_by. - Routine recording no longer requires an unrelated maintenance proposal or health paragraph.
Readback and hook advice use the same scope. Relevant declared reliability limits and
requests for ongoing checking remain visible, within the existing opening budgets.
Included changes
- Fix task knowledge retention and report recovery
- Give the Intel macOS test job three hours (#313)
- Name the missing record or unknown id when same, distinct, migrate or assess refuse (#309)
Decisions recorded: d.maintenance_advice_stays_on_task, d.task_knowledge_handoff
kpopper 0.17.0
Ongoing knowledge maintenance
- Ordinary work can propose source or clock checks for knowledge that will be
revisited, even before followups are configured. Decline, snooze, pause and
guidance-off suppress repeated offers while relevant missing, failed or overdue
evidence remains visible. - Reader guidance directs agents to complete task-required file comparisons and
read other project materials within the existing authorization. - Explicit declarations select source scope, evidence requirements and cadence in
whole local calendar days. Recurring checks share the existing followup ledger,
ownership and daily task budget. Proposals do not install a schedule or grant
source access, record changes or publication authority. - Keep successful source observations separate from failed attempts, evidence
expiry, the next check's due time and reported host execution. Current-use
assessment also checks declared dependencies and source-to-record alignment;
a recent observation does not automatically validate an old conclusion. - Inspection receipts are host-attested, not authenticated source truth or proof
that future scheduled runs will occur. New maintenance capture uses the version
2 private ledger; older runtimes must be upgraded before using that ledger.
Existing ordinary followups keep their behavior. See
clock and source maintenance.
Included changes
- Add recurring source and clock maintenance (#296)
- Make the native crate pass its documented clippy check (#298)
- Refuse a hub brief that is named but missing (#300)
- Tell a record that is not valid YAML the same way on remeasure (#302)
- Space the numbers in context and search limit refusals (#308)
- Show an accepted session id in the no-intent refusals (#306)
- Check every example record's verdict in a native test (#305)
- Check that every hook manifest route reaches the native command (#304)
- Match the local check recipe to CI and keep formatters off the record (#307)
- Describe every option of export, history, answer, pending and board (#303)
Decisions recorded: d.maintenance_notice_keeps_source_access, d.maintenance_reuses_followups
kpopper 0.16.1
Fixed
setpreserves long text values when adding their firstofdate. The date
follows the complete wrapped value instead of interrupting it and causing
invalid_history_yaml; existing dates, source citations and comments are preserved.
Included changes
- Keep inserted dates after wrapped set values (#299)
kpopper 0.16.0
Native model plugins
- Added an opt-in
modelskill and nativekpop-modelcompanion for packaging
a committed knowledge model, its source/history closure and domain skill as
an installable plugin. Read/query models use the shared reader without custom
application code; dated case applications can provide a native adapter. - Generated bundles include a pinned complete engine, validate the model before
publishing the output directory, and prepare private verified generations on
explicit setup. Updates, rollback and damaged-cache recovery preserve model
identity and keep case writes out of the canonical model and project record. - Initial support is local Codex on macOS ARM64 with kpopper 0.15.1. Consumers
need no Python or compiler. Source checkouts still need the qualified prebuilt
companion and engine to generate a bundle; project storage and hooks remain
unchanged. See Create a reusable model plugin.
Documentation
- The README separates using knowledge to support project work from building a
reusable knowledge model as the product. Both paths share installation and link
to their next steps; the guide defines a living knowledge model and distinguishes
complete models from reading excerpts. - Expanded the living knowledge model guide to
cover an application-owned model bundle alongside an optional project record,
explicit project/model selection and ephemeral case data. It describes an
application pattern using existing operations, with the optional native plugin
companion providing packaging. The public flight example is a reading excerpt with synthetic
results; it is not a standalone runnable model or a complete account of rights.
Conversation freshness
- Managed Codex conversations refresh previously received evidence and its
declared support before the next answer when captured record inputs change.
Unchanged inputs add no context. Deleted entries are named; unavailable,
out-of-scope or oversized refreshes require an explicit current read. - Received context frames provide a short, session-bound
revision_reffor
answer metadata and follow-upsession viewcalls. The runtime verifies the
retained frame and full revision; unknown or expired references are refused. - Refresh preserves the original source, read mode and navigation profile.
Stop remains passive, delta remains off by default, and the existing session
opt-out controls apply. No record migration is required. Refresh checks
captured record inputs, not changes in the external world; the bounded hook
deadline and complete-body limits still apply. - Source tracking survives interrupted turns, compaction, failed transcript
reads and same-session restarts. Failed refreshes are retried, and complete
retained managed reads can acknowledge recovery. Raw stdout alone does not
prove host reception. A disabled or terminated hook cannot guarantee a warning. - Revision references include a random session epoch so restarting a session
cannot reuse an old reference for new evidence. Managed views reuse their
captured source for freshness; unrelated resolver settings do not trigger a
refresh. - Preserve tracked evidence when a resumed session has a missing record or an
ordinary opening fallback. Honor workspace opt-out settings for external
records, and measure the full refresh envelope against its configured budget. - Source-handover warnings identify the selected record and explicitly keep
complete reads of a previous record from serving as current evidence. Bounded
openings preserve the warning and pending selection when paths are long. - Reuse the immutable snapshot's input index across node evaluations to avoid
rebuilding it for every record. See the conversation lifecycle
and canonical view reference.
Fixed
- Skill guidance uses kpopper in ordinary conversation. Fully qualified skill
names such askpopper:groundare reserved for invocation instructions,
technical documentation, debugging and explanations of a specific skill.
Included changes
- Finish Windows release checks and bind reused workflow evidence (#301)
- Verify release test reuse after a Windows fixture correction (#295)
- Stabilize Windows transient-publisher test scheduling (#294)
- Add native packaging for reusable knowledge models (#281)
- Clarify project work and reusable knowledge models in the README (#279)
- Warn when formatters may rewrite the record (#292)
- Record evidence for release renumbering (#282)
- Name checked recovery for conflicted records (#291)
- Show pending knowledge in ordinary branch comparisons (#289)
- Report unavailable record paths consistently (#283)
- Add focused CI for reviewed reader changes (#293)
- Pin check and release workflow actions to commit SHAs (#290)
- Keep Gemini and CI example versions aligned with releases (#285)
- Check CLI references in skills and documentation (#280)
- Cover the launchers' Windows, Linux, hosted and unknown-hook paths (#288)
- Give the greenhouse example's page brief a tab that serves its sessions (#287)
- Add Codex listing metadata for every canonical skill (#286)
- Describe the question report the ingestion writer retains (#284)
- Refresh the flight example source after history cleanup (#278)
- Document separate decision models and project knowledge (#277)
- Refresh changed conversation evidence before the next answer (#271)
- Use kpopper as the product name in conversation (#274)
Decisions recorded: decision.source_refresh_before_answer
kpopper 0.15.1
Fixed
- Ordinary checked reads and
kpop session openavoid repeated record
conversions while assessing judgments and building session data. Source
freshness checks, judgment results and record formats are unchanged.
Included changes
- Reuse ordinary reader conversion within immutable projections (#272)
kpopper 0.15.0
Conversation context
- Codex now opens a canonical graph view by default. Selected evidence keeps its
complete original bodies, source IDs and typed relations; the rest of the record
stays reachable through folded navigation groups. Large openings provide a bound
read route instead of a clipped graph. - Follow-up queries automatically use eligible source IDs cited in the last two
completed answers as ranking hints. Global discovery still runs. Explicit
--anchorand--no-auto-anchorscontrols are available. kpop session viewsupports question focus, exact IDs, group and membership
expansion, coarse navigation for large records, and independently enforced token
and byte budgets. JSON and checked-text encodings represent the same selected
evidence; Codex defaults to tagged checked text.- Managed Codex reads deliver complete context frames through PostToolUse. Stop
checks retained frames locally, without an extra model request. Missing or lost
context can be recovered with a full read using--view-transport stdout.
Experimental
- Delta delivery is available with
KPOPPER_VIEW_DELTA=1and remains off by
default. It requires an exact retained base and must be smaller in both bytes
and reference tokens. Source changes, compaction, rollback, interruption and
invalid state require a full checkpoint. End-to-end savings are not established.
Compatibility and controls
- This context upgrade requires no record migration. Existing writers and readers
remain available; standalone view calls return the full selected view and other
hosts keep their existing opening unless explicitly opted in. - Use
KPOPPER_CANONICAL_VIEW=0orkpop session disableto opt out. Restore the
canonical allowance withkpop session enable --tokens 16000. The hard override
KPOPPER_SESSION_DISABLE=1takes precedence over explicit opt-in. - Navigation labels, queued markers and saved files do not count as source-body
reads. Complete retained frames establish delivery under the supported host
contract; they do not prove source truth or model understanding.
Documentation and releases
- Added an upgrade guide, a complete conversation walkthrough, a runnable shipping
policy example and a canonical view reference, covering follow-ups, changed
sources, recovery and optional delta delivery. - Release notes now preserve curated
Unreleasedexplanations and list included
PRs for traceability. Historical entries remain unchanged. Authors can describe
behavior, settings, compatibility and limits before the release bot chooses the
next version.
Included changes
- Add canonical conversation context and curated release notes (#269)
Decisions recorded: decision.canonical_context_default, decision.context_delta_opt_in, decision.curated_release_notes
kpopper 0.14.0
- Normalize alternate compact record paths on Windows (#268) — patch
- Fix Windows compact history paths and stress validation (#267) — patch
- Fix compact history release acceptance (#266) — patch
- Fix Windows contribution test compilation (#265) — patch
- Use compact node history by default (#264) — minor
- Put source sections after findings in writing guidance and examples (#263) — patch
- Fix historical source citations and Git pin checks (#259) — patch
- Preserve history review requirements and branch evidence (#260) — minor
- Read candidate metadata in trusted crate verification (#258) — patch
Decisions recorded: d.compact_history_is_default, d.history_review_follows_lineage, d.source_sections_last, d.workspace_operations, d.write_path_returns_reach, d.written_layer_drawn_by_build, d.young_sections_read_by_shape