# Changelog All notable changes to this project are documented here. The format is [Keep a Changelog](https://keepachangelog.com/en/2.0.0/) 2.0.0, and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). ## [Unreleased] ## [1.2.2] - 2026-10-24 ### Added - `wiki.drift` — the linter now compares numbers restated on wiki pages with the prompt itself: an invariants count or a gate width that disagrees is an error. Verified on two fixtures (`Nine rules` against ten, `ten-item gate` against twelve), each producing exactly the expected finding. - The bilingual wiki is published: 10 pages plus `_Sidebar` and `_Footer` at , with navigation for both languages. ### Fixed - `wiki/02-Modules.md` said "Nine rules never change" while the prompt had ten, and the Russian page claimed ten while listing nine: the delivery invariant was missing from both. Both now state ten and name it. - The wiki module table still described M10 as "point it at a new domain"; it now describes the session-to-environment procedure, and M8 names the twelve-item gate. ## [1.2.1] - 2026-10-24 ### Fixed - `SKILL.md` held the Russian text and `SKILL.ru.md` the English one: the two files were published swapped. The contents are exchanged and English is canonical again. - The linter gained the rule that would have caught it: `language.direction` measures the Cyrillic share of prose, so a `.ru.md` file holding English and a canonical file holding Russian both fail. Verified on two fixtures — a swapped `SKILL` pair and a swapped `PROMPT` pair — each producing exactly the expected errors. - The external check verifies the same property over the **published** bytes, not the local tree, so a swap that reaches GitHub is caught by reading GitHub. - The old pairing rule only proved that two files both existed. Existence is not direction: this defect passed the linter, the documentation contract and the audit, and was found by diffing language proportions. ## [1.2.0] - 2026-10-24 ### Added - Root `SKILL.md` / `SKILL.ru.md` — the portable entry users install: five steps from a finished session to a specialist environment, per-runtime registration, a handoff prompt, and the session-persona mode demoted to a side mode. - `tools/install.py` — installs the skill and the `/agent-prompt-kit` command into AutoClaw2 or HARNESS, backs up whatever it replaces, and `--check` reports files plus registration state. For SagaAI it **refuses to guess** and prints the two questions to answer instead. - `docs/INSTALL.md` / `.ru.md` — measured mechanics per runtime: skill roots and command file format in AutoClaw2, the workspace registry and session index in HARNESS, plus a trap table where every row maps to an observed symptom. - Related sessions as a first-class source: M3 gains the session-index row and two rules (a past session is evidence of what happened; index noise is not evidence), and M6 gains the `session-search` and `workspace-create` capabilities. - The quality gate grew to twelve items: item 12 checks **delivery** — a directory at a named path, a truthful registration step, a handoff prompt. - Three failure modes (17–19): produced-but-not-installed, environment substituted by a persona, reasoning from index noise. - A tenth invariant: a reusable deliverable counts only when it is registered, or the exact registration step is named. ### Changed - `M0` now states two outputs — an answer, and a specialist environment when the work repeats — and names the failure it prevents: a good answer that dies with the chat. - `M10` rewritten: it was "fill a card" with no destination; it is now the five-step procedure ending in registration and handoff, with a runtime table that says what the model can do and what a human must do. - README (both languages) leads with the purpose instead of the prompt mechanics, and gains a Runtimes section. - The linter and the external check track ten invariants, a twelve-item gate, and the presence of `SKILL.md` and `docs/INSTALL.md`. - Numbers in the README are measured, not estimated: the lint badge reports 27 distinct rule ids because that is what the tool actually contains. ### Fixed - Documentation no longer describes the kit as "paste this into your system prompt" as the primary route; that route is one of three, and the other two are commands. - A claim of "34 contract checks" was invented in the previous version and replaced by the measured count. - `docs/QA.md` and its translation had a duplicated row after an automated count sweep; removed. ## [1.1.0] - 2026-10-24 ### Added - `M13 — Experience reuse`: before planning a non-trivial task, search a store of prior work; record it after verification. Experience is a hypothesis, never authority. - Four experience types — task record, recipe, artifact, failure record — with a fixed metadata header, so the store is queryable without reading every file. - A six-question applicability check, because blind copying of an old recipe is the cheapest way to reproduce an old bug. - `tools/experience.py` — the store and its one-command retrieval: `init`, `add`, `search`, `show`, `check`, `stats`, `deprecate`, `selftest`. Lexical ranking prints the reason for each position; `add` refuses a record with a token-shaped literal, an absolute machine path, or a recipe missing `Rollback`. - `skills/experience-reuse/` — the portable skill: workflow, retrieval rules, applicability checklist, four record templates, three sanitised examples, and a security/retention policy. Every file in English and Russian. - `memory/` — the store's home, with the rules for what belongs there and what never does. - The quality gate grew from ten tests to eleven: retrieval ran, every reused record has a verdict, nothing sensitive entered the store. - Four new failure modes (13–16): blind recipe copy, stale experience as truth, store as a dump, two stores. - `prompt_lint.py` now enforces the new contracts: M13 markers in both languages, eleven gate rows, the store, the tool and the skill sections. ### Changed - `M3` — the Experience Store joins the source table at priority 5.5, with rule 7: experience never overrides the current repository, attached material or an explicit instruction. - `M5` — the loop gained a **Retrieve** step between Frame and Plan, and Record now decides whether a record is warranted. - `M6` — capabilities `memory-search` and `memory-write`, plus rule 8: experience writes are reviewable and never contain sensitive values. - `M7` — an optional `## Reused experience` section, and rule 7: reused work is named, with what was adapted. - `M10` — a domain may add record types, but not change the M13 invariants. - Ten invariants: prior experience is retrieved before planning, validated before reuse, sanitised before storage. ## [1.0.0] - 2026-10-24 ### Added - `PROMPT.md` — the portable system prompt: thirteen modules (`M0`–`M12`), each with a self-check line, plus eight invariants that hold in every domain. - Task Card and Domain Card as mandatory pre-work outputs: the subject, the practitioner, the skill set, the facts that must be verified and a testable definition of done. - Context ingestion with a priority order and a five-column Context Ledger, so no claim enters an answer unsourced. - Three evidence labels — `[fact]`, `[inference]`, `[assumption]` — with a rule that an assumption names its consequence. - A capability-class tool policy with an explicit degradation column and seven safety rules, including the sandbox-before-irreversible rule and a ban on displaying secrets. - A five-section output contract, a ten-item quality gate reported as one line, and a table of twelve failure modes with counters. - A specialization protocol: a new domain costs one card, one examples file and one acceptance line. - `examples/EXAMPLES.md` — three unrelated domains (PHP panel, Windows service, documentation product) using one prompt with zero edits. - `tools/prompt_lint.py` — 28 structural contract checks: module sequence, self-check lines, invariant count, translation pairing, placeholders, links, secrets, absolute paths, dangling module references, version agreement. - `docs/DESIGN.md`, `docs/ADAPTATION.md`, `docs/QA.md`, `docs/VALIDATION.md` — rationale, the fifteen-minute procedure, the human review checklist, and the objective checked requirement by requirement. - Russian translations for every document, following the `.ru.md` convention. ### Security - The kit ships without tokens, absolute machine paths or personal addresses; the linter fails the build if one appears, which is how the repository enforces on itself the rule the prompt states. [Unreleased]: https://github.com/paulmann/agent-prompt-kit/compare/v1.2.2...HEAD [1.2.2]: https://github.com/paulmann/agent-prompt-kit/compare/v1.2.1...v1.2.2 [1.2.1]: https://github.com/paulmann/agent-prompt-kit/compare/v1.2.0...v1.2.1 [1.2.0]: https://github.com/paulmann/agent-prompt-kit/compare/v1.1.0...v1.2.0 [1.1.0]: https://github.com/paulmann/agent-prompt-kit/compare/v1.0.0...v1.1.0 [1.0.0]: https://github.com/paulmann/agent-prompt-kit/releases/tag/v1.0.0