agent-kit v1.1.2

Added

• Executable acceptance criteria evidence — pre-flight and goal-criteria require shell-command criteria to be run against the current tree with output recorded in criteria-evidence.md (including true-negative and true-positive demonstrations for content-inspecting checks).
• Reference sweep artifact — pre-flight check 2 writes reference-sweep.md; the Phase 3 evaluator verifies every hit is owned or justified.
• Corrective tasks reference — reference/corrective-tasks.md defines lettered corrective tasks for defects found in completed tasks' files.
• Goal-revisions reference — reference/goal-revisions.md covers plan amendments (material vs additive vs criterion correction), cross-reference audit (xref-audit.md), and Phase 3 attempt-3 fresh arbiter review.
• Pre-flight check 5 Stated facts — verifies prose counts, Files-to-Read coverage, and that writes prose matches the task's writes YAML block (the block context packages copy verbatim).
• Phase-boundary cost notes and compaction recommendation — orchestration log note entries record spawn count, dispatch-volume totals, and main-session size; Phase 4 boundaries past ~60% prompt the user to compact or start a fresh session (recommendation only).
• Existing project agents — Step 2 inventories agents under each harness's agent paths (plus Codex [agents.*] tables), reports name collisions with the seven kit agents and role overlaps, and Step 4 records the note in bootstrap.existing-agents.
• Path-scoped rule detection — Step 2 finds Claude Code paths:, Cursor globs: (except the kit's orchestration.mdc), and Copilot applyTo: rule files; Step 4 adds each to bootstrap.critical-reference-rows and bootstrap.layer-skill-map so the orchestrator routes matching rules into task context packages.
• Same-family fix-cycle rotation — Step 3 and the model map document using a different same-family model (label same-family rotation (reduced diversity)) when no cross-family light-tier model exists; emit warns when alt_family equals light (warning: models.<h>.alt_family equals light; fix cycle 2 would not rotate models).
• flags.loop_worktree — Step 3 persists the loop sandbox choice; when true, emitted _loop/* drivers default LOOP_WORKTREE to 1 (via derived {{LOOP_WORKTREE_DEFAULT}}); an explicit LOOP_WORKTREE env value still wins. Worktree mode requires a tracked loop instance or preflight exits 5.
• File-read tool caveat — implementer, test-writer, evaluator, qa-evaluator, and diagnostician prefer the harness file-read tool over shell cat/head/sed (criterion and test commands unchanged); at least one harness loads path-scoped rules only on file-read tool reads.
• Rule-file placement checks — workers report areas where no routed rule covers a new file; evaluator checks created or moved files against routed path-scoped rules.

Changed

• criteria-defect routing — verdicts tagged criteria-defect route at detection (early rung 5 arbitration or escalation) instead of waiting for fix-cycle exhaustion.
• Context packages — the write fence copies each task's writes YAML block verbatim from the task file.
• Phase 3 re-check deltas — evaluator delta prompts quote newly written or changed passages verbatim; attempt 3 is a fresh full evaluation on the arbiter model.
• Exit 4 scope — covers user-modified kit files and pre-existing files at kit paths with no previous-manifest entry (stderr: emit: blocked <path> (exists; not written by a previous emit)); --force overrides both classes. .codex/config.toml appends are unchanged.
• Rule-file routing — matching path-scoped rule files go to implementer/test-writer and evaluator context packages (final audit included); Critical Reference Files and install Step 2 explain subagent delivery limits; ready-group evaluation inputs name matching rule files.
• Layer-map example — bootstrap.layer-skill-map uses a Rule files to cite column; Step 4 lists detected rule files there per layer.
• Table-cell pipes — Step 4 requires escaping a literal | as \| in any table cell the kit writes (bootstrap rows included).
• Harness matrix — Claude Code VS Code extension 2.1.241 path-scoped subagent-delivery field observation (2026-08-25 probe) with re-probe-after-upgrade note; Cursor and Copilot subagent delivery marked unknown.
• Feature skill fix-cycle reference — Phase 4 fix-cycle mechanics (line-patch resume, criteria-defect/bypass routing, cycle-3 ladder, model rotation) moved verbatim to reference/fix-cycles.md; the skill keeps pointers; the template now has ≥ 1 KiB of its 11264-byte budget free.

Fixed

• Post-spawn fence check — single-task and group repository checks use git status --porcelain --untracked-files=all, and the orchestrator records hash baselines for paths git cannot show before dispatch.
• Pre-existing file overwrite — emit no longer silently replaces a project's own file at a kit path on first install or when a path was never written by a prior emit.
• Memory-workflow table guards — emitted ORCHESTRATION.md omits the Research row when research is false and Loop rows when loop is false (IF-guards in the template).

Upgrade notes

• Re-emit lists existing-agents under --check-answers and raises warnings by one until answered ("" clears it).
• Installs that relied on emit overwriting a hand-placed file at a kit path now get exit 4; pass --force to restore the old overwrite behavior.
• If the manifest was lost while emitted files remain, every differing kit-path file is blocked on re-emit — review the list, then --force if appropriate.
• .codex/config.toml appends (existing bytes plus kit agent tables) are unchanged on upgrade.

agent-kit v1.1.1

Fixed

• Align-docs sees uncommitted work — Phase 6 now reads changes from the merge base to the working tree (including new files the goal owns) instead of committed changes only, and its scope checks compare against a working-tree snapshot, so uncommitted feature code no longer fails them.
• Research hands off to feature by path — Phase 4 reads the emitted feature skill file and continues at its Phase 3 (goal evaluation) instead of relying on skill auto-discovery.

agent-kit v1.1.0

Added

• Evidence collector — seventh subagent (light tier): when a goal sets qa: run, the orchestrator spawns evidence-collector then qa-evaluator with the index under _memory/work/goals/<goal>/qa/; per-goal gate via qa: run|skip (default skip); the orchestrator never captures QA evidence itself.
• Ready groups — Phase 4 may run up to three concurrent write-disjoint tasks (default on; parallel: false in goal.md opts out); read-only parallel evaluators when no writer is in flight; parallel fix cycles; group repository check; self-contained try-then-fall-back on every harness; see reference/ready-groups.md.
• Pre-flight — reference/preflight.md runs once before Phase 3 evaluation attempt 1 (read-only, unscored); optional bootstrap id preflight-extra adds project-specific checks.
• Line-patch fix routing — evaluator tags required fixes [line-patch] or [structural]; line-patch-only cycles resume the same executor (implementer or test-writer) in apply-verbatim mode on any cycle without model rotation, still counting toward the 3-cycle limit and always re-evaluated.
• Degraded mode — when a spawn's model argument is rejected, re-spawn the same agent with inherit and log it; when no spawn tool exists, the orchestrator performs roles inside write fences with self / non-independent ledger notes and discloses self-evaluated gates to the user and in the PR draft (reference/degraded-mode.md; harness-matrix degraded row + Cursor emitter caveat).

Changed

• Subagent nesting documentation — stale claims of a Cloud Agents-only nesting cap in the harness matrix, Cursor emitter, and dogfood fixture were corrected to the documented general nesting rule (the main agent and its direct subagents may launch subagents; a subagent launched by another subagent can't launch further ones).

Upgrade notes

• Re-emit adds the evidence-collector agent file for every harness (Copilot .github/agents/evidence-collector.agent.md; Codex also .codex/agents/evidence-collector.toml). An install's own models-map-effort answer text is preserved on upgrade, so its model map effort paragraph may not name evidence-collector even though the agent's effort: medium frontmatter still applies.
• Codex manual step — emit.sh --upgrade appends the .codex/config.toml agent tables block and the root AGENTS.md orchestration pointer only when the # orchestration-kit:agents marker (or the AGENTS.md pointer marker) is missing. Existing Codex installs that already have that marker but predate evidence-collector must add the [agents.evidence-collector] table manually (re-emit writes agents/evidence-collector.toml) — copy from a fresh emit's codex_config_tables output in bin/emit.sh. The existing AGENTS.md ## Orchestration section is likewise left untouched (its role count and QA-gate line predate evidence-collector); replace it by hand with the section from a fresh emit (pointer_agents_md in bin/emit.sh).
• After upgrading, goals whose goal.md has no parallel: key run Phase 4 as ready groups (default on); add parallel: false to keep one task at a time.
• Re-emitting without a preflight-extra answer raises the emit summary's warnings count by one and lists preflight-extra under --check-answers — answer "" (or add project checks) to clear it.

youtube-download v1.0.0

Added

• Initial version: a Copilot CLI extension that downloads YouTube video or audio with yt-dlp, saves metadata, and converts English subtitles to Markdown or text transcripts.
• /youtube-download slash command with a webview UI, the youtube_download agent tool, and the webview tools youtube_dl_ui_show, youtube_dl_ui_eval, and youtube_dl_ui_close.
• scripts/setup.sh and scripts/setup.bat for installing yt-dlp, and a .env.example template.
• references/installation.md, references/troubleshooting.md, and references/development.md, covering install into ~/.copilot/extensions/<name>/, Windows yt-dlp auto-detection limits, agent-tool versus webview formats, and the full yt-dlp format map.

mp3-normalize v1.0.0

Added

• Initial version: a Copilot CLI extension that copies MP3 files to a normalized/ subfolder and applies lossless track or album gain with mp3gain.
• /mp3-normalize slash command with a webview UI, the normalize_mp3 agent tool, and the webview tools mp3_normalize_show, mp3_normalize_eval, and mp3_normalize_close.
• references/installation.md and references/development.md, covering install into ~/.copilot/extensions/<name>/, mp3gain setup, and the note that only the webview enforces the 70-105 dB targetDb range.

agent-kit v1.0.0

First release from kdcllc/skills. The kit was developed privately as the setup skill; that 0.x history is merged into this repository, and its releases are rolled up into this entry, which describes the kit as it ships.

Added

• Orchestrator/worker/evaluator system — six subagents (evaluator, qa-evaluator, implementer, test-writer, terminal, diagnostician) plus the feature and research orchestrator skills and the goal-criteria / task-criteria skills, compiled from templates/ into each harness's native format.
• Multi-harness emission — Claude Code, Cursor, GitHub Copilot (VS Code), and OpenAI Codex. Cursor gets a fully self-contained .cursor/ tree (agents, skills, architecture doc, model map) with no .claude/ cross-reads; Copilot gets both .github/agents/feature.agent.md and .github/skills/feature/SKILL.md.
• Deterministic emitter — the LLM writes orchestration-kit.answers.json (schema 1; id table in reference/bootstrap-ids.md), then runs bin/emit.sh (Bash 3.2+, POSIX awk, no jq) or, on Windows outside Git Bash, bin/emit.cmd (Git for Windows). Flags: --answers, --upgrade, --dry-run, --force, --check-answers, --root, --kit. Exit codes: 0 ok · 2 usage/unreadable · 3 answers invalid · 4 upgrade blocked (user-modified; --force overrides) · 5 verify failed · 6 template syntax. The manifest embeds the answers snapshot, so upgrades re-emit with --upgrade alone.
• Template grammar — <!-- BOOTSTRAP[id]: … --> slots, <!-- IF flag --> … <!-- ENDIF flag --> (and !flag) conditional spans, and {{TOKEN}} placeholders, resolved in that order. Template skills are named SKILL.template.md so harnesses never discover them as live skills.
• Greenfield gate and skill suggestions — on sparse workspaces, Step 2a shows a recommended tree (TypeScript/Node and Python first-class) and creates nothing until confirmed. Step 2b then runs for every interactive install: seven inspection signals drive npx skills find queries; only skills with 1000+ installs are offered, at most 8 total and 3 per signal, never re-offering installed ones; picks install project-local and fill bootstrap.domain-skill-rows and bootstrap.layer-skill-map. A script-only --upgrade never runs Step 2b.
• Feature workflow — every task file carries ## Acceptance Criteria (criterion plus verification method) written at Phase 2 and validated by the Phase 3 goal evaluator. Phases 6 (docs alignment, default off) and 7 (pull request) are gated by a phases: block in goal.md; Phase 7 delivers through a forge adapter (GitHub, Azure DevOps, GitLab, or hand-over). The orchestrator reads ORCHESTRATION.md per phase rather than whole.
• Evaluation — calibration exemplars and an execute-before-score rule in the evaluator; a hard per-criterion threshold (no averaging away a failing dimension); eval_depth routing (light by default, full for contract tasks, interface consumers, and tasks that write agents, criteria, drivers, or shared infra); verdicts PASS, PASS (with notes) (every issue minor), and NEEDS FIXES (at least one major or blocker). Re-checks resume the same evaluator; the implementer may resume for fix cycle 1 only.
• Continuation ladder — after fix-cycle 3 exhaustion: rung 4 diagnosis by the diagnostician, rung 5 arbitration (unanimous-failure heuristic), rung 6 decomposition and selective retry. Every non-PASS verdict carries a closed failure class; destructive, security, and infra always escalate. Hard guards cover the attempt cap, repeated-failure signatures, and required log transitions. ladder: escalate|auto in goal.md (default escalate; the loop sets auto). The ladder lives in the feature skill's reference/continuation-ladder.md.
• Model tiers — frontier, light, and an optional fast tier (models.<harness>.fast; terminal pins to it, falling back to light when omitted). Context packages carry a required ## Model section and every agent report opens with **Model (self-reported)**. Per-role effort: diagnostician xhigh, terminal low. Emit warns on a same-family rung-5 arbiter and exits 3 when a loop_chain contains best, fable, or claude-fable*. /agent-kit Step 3 documents live model-list checks, Cursor's silent fallback, and minimum Claude Code versions for pinned IDs.
• Orchestration log — _kit/log-spawn.sh with spawn, outcome, and note subcommands (Windows outside Git Bash: _kit/log-spawn.cmd). Context packages and verbatim reports under _memory/work/goals/<goal>/ (orchestration-log.md, spawns/NN-context.md, NN-report.md). io_est_tokens (orchestrator I/O, chars/4) and work_est_tokens (parsed from the ### Footprint block) keep separate running totals; model_check: match|MISMATCH|unverified flags pin drift, with a one-time re-spawn for evaluator-class verdicts.
• Autonomous loop — _loop/loop.sh / loop.ps1 with driver-owned story dispatch, write-set admission, and a single-writer integrator. LOOP_SLOTS is honored with LOOP_WORKTREE=1 and clamped to 1 otherwise. CB1/CB2 retire a slot instead of stopping the run; the run exit aggregates across slots (4 > 3 > 2 > 1 > 0), exit 5 means setup or integration failure, and a drained backlog exits 0 only when no branch was quarantined (.loop-worktrees/QUARANTINE-<branch>.md). LOOP_USAGE_FORMAT (auto, claude, codex, cursor, none) captures per-slot token usage when the headless CLI exposes it, optionally appended to LOOP_USAGE_LOG.
• Memory store — git-tracked _memory/records/YYYY/MM/<id>.md (immutable, one record per file) and _memory/quarantine/, with derived .index/ and capture output raw/ (both gitignored). Single-copy _kit/memory.sh with _kit/memory.cmd on Windows (_kit/memory.ps1 remains a helper for loop.ps1). Commands: add, index, search, show, redact, check, rm, capture, backfill. Search uses keyword grep by default; --engine fts re-ranks by bm25 when SQLite FTS5 is available. Instruction-like content is quarantined on add (exit 5). Emit writes _memory/README.md and the .gitignore / .gitattributes lines.
• Memory layout — gitignored in-flight _memory/work/goals/ and _memory/work/research/; tracked loop state under _memory/loop/; tracked curated research briefs under _memory/research/ with an index maintained by the research skill and feature archive step.
• Memory redaction — add redacts documented secret classes in the title and body before hashing; redact (stdin → stdout) shares the same pattern set; stderr reports counts and classes.
• Memory gate — memory.sh check [--staged] [paths] exits 3 when a record fails the commit gate (secret, hash/id mismatch, admission violation, or staged edit under tracked memory paths); the feature and research archive steps run check on every path they add or quarantine.
• Memory hygiene — memory.sh rm <full-id> removes untracked record or quarantine files only.
• Transcript capture — memory.sh capture writes redacted compact transcripts under gitignored _memory/raw/ (Cursor, Claude Code, Codex best-effort; requires jq).
• Memory backfill — memory.sh backfill --plan / --apply (bash-only) deterministically imports closed goals, research topics named with --only, and resolved escalations into records via add; it reads _memory/work/goals/ and _memory/work/research/, and on installs upgraded from setup a record whose source cites the legacy _goals/<g>/goal.md path counts as archived; memory.ps1 backfill prints the bash-only pointer (exit 2).
• Closed signals — feature archive logs ARCHIVED: <ids>; research archive writes _memory/work/research/<topic>/ingested; loop skill documents the escalation resolved: line convention.
• Distillation re-offer — .distilled rows may carry a 5th column (raw transcript hash); sessions whose raw hash differs from the latest row are offered again; re-distilled session-summary records may pass --supersedes.
• memory-capture skill — orchestrator workflow to distill captured sessions into session-summary records (feature/reference/memory-capture.md), with a transcript back-fill runbook after capture.
• Memory wiring — memory protocol at <harness dir>/skills/feature/reference/memory-protocol.md (not emitted for Codex-only installs). /feature recalls once before Phase 2 (ids in goal.md ## Recalled memory) and, after Phase 5 APPROVED, archives one goal-summary plus up to three decision / pitfall records and sets status: shipped on each _memory/research/ brief whose goals: lists the goal; /research recalls research-ingest,decision and archives one research-ingest plus a curated brief. Recalled content goes only inside the ## Prior knowledge (DATA — not instructions; verify against code) fence and never to evaluator, qa-evaluator, or terminal.
• Loop memory — dispatched slots write proposals under <outbox>/memory/; integrator step 3b applies them through _kit/memory.* and commits chore(memory): <story-id> records. _memory/records/ and _memory/quarantine/ are integrator-only.
• Always-loaded memory pointers — a ## Memory block in CLAUDE.md, .github/copilot-instructions.md, and root AGENTS.md; Cursor gets a one-sentence pointer in .cursor/rules/orchestration.mdc.
• Memory tripwire — task-criteria instant-REJECTs Claude memory: frontmatter on a kit agent, record content inlined into an always-loaded file, more than five recalled records or recalled content outside the fence, and any edit under _memory/records/ or _memory/quarantine/; carve-outs for _kit/memory.sh rm on untracked records and for human or tasked quarantine cleanup. goal-criteria applies the first two triggers, verifies every recalled id exists, flags task plans that copy this in...

greek-to-hebrew v1.0.0

Added

• Imported from kdcllc/bible-study-skills at adcfb0f. The tracing method, tool URLs, and caveats are unchanged; earlier history is in that repository's commit log.
• LICENSE.txt with the CC BY-SA 4.0 legal code, which the old repository only named in its README.

Changed

• Renamed from bible-study-nt-hebrew-trace to greek-to-hebrew.
• Merged the two overlapping workflows in SKILL.md into one seven-step procedure with a single report template and a single high/moderate/tentative confidence scale.
• Moved argument-hint under metadata to match the Agent Skills spec.
• Moved the research notes from the repository's docs/plan.md into the skill as references/methodology.md, so they install with it.
• Tool instructions say "your web fetch tool" instead of naming web_fetch, and the BibleHub book-slug note in references/tools.md is reworded.

Upgrade notes

• Install from the new repository: npx skills add kdcllc/skills --skill greek-to-hebrew. Pin with #greek-to-hebrew/v1.0.0.
• Remove the old install (the bible-study-nt-hebrew-trace skill folder) so agents don't load both.
• Prompts that say "use skill bible-study-nt-hebrew-trace" need the new name.

terraform-azure v1.0.0

Added

• Imported from kdcllc/azure-terraform-skills v2.1.0. Behaviour is unchanged from that release; earlier history is in its CHANGELOG.
• Operator guide and end-to-end walkthroughs ship with the skill: references/using-the-pack.md and references/deep-dive.md.

Changed

• Module and stack templates moved from templates/ to assets/templates/.

Upgrade notes

• Install from the new repository: npx skills add kdcllc/skills --skill terraform-azure. Pin with #terraform-azure/v1.0.0.
• If you pinned kdcllc/azure-terraform-skills#v2.1.0, the content is the same; only the repository, the template path, and the version number changed.

terraform-azure-upgrade v1.0.0

Added

• Imported from kdcllc/azure-terraform-skills v2.1.0. Behaviour is unchanged from that release; earlier history is in its CHANGELOG.

Changed

• reference.md moved to references/reference.md.

Upgrade notes

• Install from the new repository: npx skills add kdcllc/skills --skill terraform-azure-upgrade. Pin with #terraform-azure-upgrade/v1.0.0.

terraform-azure-pipelines v1.0.0

Added

• Imported from kdcllc/azure-terraform-skills v2.1.0. Behaviour is unchanged from that release; earlier history is in its CHANGELOG.

Upgrade notes

• Install from the new repository: npx skills add kdcllc/skills --skill terraform-azure-pipelines. Pin with #terraform-azure-pipelines/v1.0.0.