v0.1.6 — context selection by audience and destination (experimental)

v0.1.6 — experimental prerelease

Bundles context-docs 0.1.6, adopt-context-docs 0.1.7 and standard 0.1.1.
Release preparation and validation completed October 10, 2026.

Release commit: b10a3058e43f9af278cbb7a6b2b70d41df16ea7d. Remote CI passed.

Changes since v0.1.5

Core guidance now selects retained information by its intended readers, sharing
destination and lasting usefulness. The
standard
covers public contributor docs, restricted team procedures, personal workspace
notes and mixed audiences. It preserves relevant operational constraints and
failed evidence while separating current context from detailed history and
incidental local state. Public derivatives must retain provenance and meaningful
diagnostic relationships; cleanup does not authorize discarding source evidence.
The standard adds a local working state role and applies audience selection to
optional journal payloads without enabling logging. The core guidance excludes
secrets and unnecessary personal or customer data from context. The starter
context template follows the same guidance.

Adoption now explicitly preserves existing equivalent maintenance rules as
written. Newly added wording adapts to local terminology without client-specific
invocation syntax.

Repository updates add bounded adoption and audience evaluations and clarify
future marker-format grading. The revised evaluator accepts supported fenced
markers only in future freezes; the completed adoption study keeps its original
failure and scores. The
evidence cleanup record documents normalized
public copies and unchanged historical scores; it does not sanitize older Git
history. See the changelog for per-change validation.

Installation and upgrade

Follow the installation instructions for the
chosen client and scope. Install or update both complete skill folders together;
adoption requires the sibling core package. Review existing files, back up local
customizations and preserve them when applying the package changes.

Existing projects need no directory or document migration. Preserve their loaded
maintenance rules, adoption dates, entry paths and logging settings. Installing
or updating the skills does not enable logging. The existing opt-in auto preference
and explicit-setting precedence remain unchanged, as do the journal helper and
schema. Ordinary Markdown maintenance has no runtime dependency; the optional
helper requires Python 3.9+, and auto detection also requires Git.

Validation and limits

The local release validation passed all 21 CI command
groups using Python 3.12.14 on macOS: 146 unit tests,
study and adoption grader controls, static checks, and four unchanged regenerated
reports. Both packages were copied outside the repository; their metadata and
packaged links passed, and all 14 packaged files matched the candidate source.
Commands used repository-root paths rather than CI subdirectory working
directories. These are local CI-equivalent checks, not a remote Linux CI result.
The candidate skills match commit 3de7d304d4c4a4cd46dffc8ab39c86d21d4869b3.

GPT 6 Astra High drafted these notes. The Claude Opus 5.5 review
found no blocking issue; its package-identity evidence gap was resolved by hash
comparison. All 14 candidate files match every frozen audience-pilot package copy;
all four adoption files also match the earlier adoption-study source. The
validation record retains those comparisons.

The October 9 public-source adoption regression
ran four native Claude Opus 5.5 sessions with core 0.1.5 and adoption 0.1.7.
Execution passed 4/4; frozen mechanical checks passed 3/4, retaining the
fenced-marker failure. Both reviewers accepted the target preservation behavior
in all four sessions. Full-claim semantic acceptance remained Astra 1/4 versus
Claude 3/4; both found the repeat's unsupported verification claim. These
assessments remain separate.

The October 10 audience/destination pilot
used this package bundle in five native Claude Opus 5.5 sessions.
Execution passed 5/5. GPT 6 Astra High and Claude Opus 5.5 each accepted
19/19 target criteria and 15/15 general criteria. The mixed case completed
its API documentation edit and asked the permitted audience/destination question;
it did not produce a handoff artifact. The audit attempted no writes or
history access. Neither study retried subjects or repaired their outputs.

Mechanical checks are separate from actual agent evaluations. These small,
authored studies used explicit invocation and restricted tools on one client and
responding model, with unblinded counterpart review. They establish neither
general reliability nor comparative benefit, automatic skill selection or
unrestricted-client behavior. Earlier factual, invocation and journal-access
limitations remain recorded in the linked reports and
project context.