feat(aidd-context): add Kilo-native generation targets - #992

🎯 What & why

Adds native Kilo support to aidd-context, enabling Kilo projects to onboard AIDD and generate discoverable project memory, rules, skills, agents, and workflows while preserving compatibility with other supported tools.

🛠️ How it works

• Detection and memory: Recognizes six Kilo signals: .kilo/, root kilo.json / kilo.jsonc, .kilo/kilo.json / .kilo/kilo.jsonc, and legacy .kilocode/. Neither AGENTS.md alone nor OpenCode-only configuration triggers Kilo detection. Shared AGENTS.md project memory preserves user-owned content.
• Rules: Publishes categorized .kilo/rules/<category>/<slug>.md files and registers their exact paths in Kilo instructions. Legacy .kilocode/ projects write new rules only to canonical .kilo/ paths. Unsupported path-scoped rules are rejected before any mutation rather than silently converted to non-enforceable prose.
• Configuration and safety: Handles JSON/JSONC configuration candidates, explicit selection, ambiguous ownership, preflight checks, staged writes, and best-effort rollback. Tests cover configuration preservation, hardlinks, non-regular paths, late failures, and temporary-file cleanup.
• Skills: Supports native .kilo/skills/ generation and explicitly selected portable .agents/skills/ placement.
• Agents and workflows: Generates native Markdown artifacts under .kilo/agents/ and .kilo/commands/.
• Hooks: Provides Kilo JavaScript/TypeScript plugin guidance without generating unsupported declarative hooks.
• Cross-tool improvements: Hardens shared project-memory filesystem operations, rule publication, and hook compatibility checks.
• Skill evaluation: Extends skill-eval with Claude/Codex support, JSON checks, project-tree snapshots, and explicit PASS / FAIL / INCOMPLETE reporting. Invalid --jobs and --repeat values are rejected before workers or evaluations start.

The rule-generation guidance keeps its detection contract within the skill, preserving standalone installation compatibility.

Implementation details and host contracts are documented in the [aidd-context README](https://github.com/ai-driven-dev/framework/blob/3c5d574545da673a61110d435bf42ec1892aab40/plugins/aidd-context/README.md).

🧪 How to verify

Run the focused Kilo CLI tests:

cd cli
pnpm test:e2e:kilo
pnpm test:e2e:kilo:native

Validation evidence:

• All applicable GitHub Actions checks passed on PR HEAD 3c5d574; conditional release jobs were skipped.
• Linux Kilo CLI 7.7.5 runtime tests use a deterministic local provider.
• Authenticated Codex CLI 0.162.1 successfully generated a native Kilo skill, action, and synthetic payload using AIDD.
• Linux Kilo 7.8.8 runtime validation confirmed skill discovery, loading, action reading, and payload consumption.
• Targeted publication tests cover legacy detection, JSON/JSONC selection, ambiguity refusal, idempotence, hardlinks, file identity, preflight, and rollback.
• Targeted tests also cover invalid skill-eval execution counts and standalone skill-link boundaries.

These are scoped behavioral and runtime validations, not a claim that the full skill-eval --host=codex suite passed.

⚠️ Heads-up

The AC references below correspond to the acceptance criteria defined in #914.

Maintainer decisions requested

1. PR scope — cross-tool changes

Besides native Kilo generation, this PR includes shared project-memory and rule-publishing hardening, hook compatibility improvements, and skill-eval extensions.

Should these changes remain together in #992, or should some independent improvements be extracted into follow-up PRs?

2. AC16 — Official documentation traceability

Official Kilo paths, formats, sources, and verification provenance are centralized in [official-sources.md](https://github.com/ai-driven-dev/framework/blob/3c5d574545da673a61110d435bf42ec1892aab40/aidd_docs/tasks/2026_10/2026_10_10_kilo-native-generation-914/official-sources.md).

Issue #914 records September 25, 2026, as the requested verification date; the documentation was independently rechecked on October 13, 2026.

Does this centralized evidence satisfy AC16, or is per-claim traceability required?

3. Authenticated Claude validation

CONTRIBUTING.md requests testing with Claude and another tool. Authenticated Codex generation and Kilo runtime validations are available, but no authenticated Claude Code generation session has been performed.

Can the available Codex/Kilo evidence be accepted, or could a maintainer perform the missing Claude validation?

Remaining acceptance-criteria coverage

The following acceptance criteria from #914 remain partially demonstrated. These gaps remain tracked under the original issue unless explicitly associated with a separate follow-up below.

Maintainer guidance on whether additional evidence is required before merging would be appreciated.

• AC1–4 — Detection: The six Kilo signals, negative OpenCode/AGENTS.md cases, and legacy rule publication have targeted coverage. Complete behavioral validation across every applicable generator is not established.

• AC7 — Configuration precedence: JSON/JSONC candidates, ownership, explicit selection, and ambiguity handling are tested. Complete Kilo runtime precedence is not established.

• AC13 — Idempotence: Deterministic publication paths have replay tests. Idempotence of an entire LLM conversation is not guaranteed. No separate issue has been created for this broader interpretation of the acceptance criterion.

• AC14 — Write safety: Preflight, refusal paths, hardlinks, and best-effort rollback are tested. The separate project-memory corruption defect involving I/O write failures is tracked in [#1018](fix(aidd-context): preserve project-memory files on write failure #1018). This PR does not claim to resolve that issue or guarantee crash recovery, concurrent-mutation safety, or atomic multi-file updates.

• AC15 — Runtime coverage: Kilo runtime validation remains Linux-based. Native Kilo runtime validation on macOS and Windows is outstanding.

AC16 and authenticated Claude validation are addressed in the maintainer decisions above.

The remaining gaps do not necessarily indicate confirmed implementation defects. Some concern the scope of required guarantees or the completeness of the available evidence.

Related PR and independent follow-ups

Related integration:

• [#1069](feat(cli): emit rules in Kilo flat builds #1069) — Flat Kilo rule builds: The Kilo-native publisher rejects unsupported path-scoped rules, consistent with the policy adopted in feat(cli): emit rules in Kilo flat builds #1069. The implementations address separate generation paths. feat(cli): emit rules in Kilo flat builds #1069 has not been merged or cherry-picked into feat(aidd-context): add Kilo-native generation targets #992; integration tests on a common base remain appropriate.

Independent follow-up issues:

• [#994](feat(framework): resolve resources shipped with installed plugins #994) — Installed-plugin resource resolution: The manual memory-sync path does not provide every protection available in the executable hook.
• [#1018](fix(aidd-context): preserve project-memory files on write failure #1018) — Project-memory write failures: Tracks the independently reproduced partial-write corruption defect, related to AC14.
• [#1050](fix(cli): isolate ambient session IDs in telemetry lifecycle tests #1050) — Telemetry test isolation: Independent session-ID fixture issue, now closed.
• [#1112](fix(scripts): exclude generated AIDD cache from Markdown link checks #1112) — Markdown validator/cache: Generated .aidd/cache content causes unrelated Markdown link-validation failures.

These issues are tracked independently. This PR does not claim to resolve them, and their existence does not automatically mean that every corresponding acceptance-criteria gap requires a separate implementation.

Maintainer review and the outstanding decisions remain necessary before merging.

🔗 Linked issue

Refs #914

✅ I certify

• I DO CERTIFY I READ EACH LINE OF THE PULL REQUEST BECAUSE I AM A SOFTWARE ENGINEER, NOT A AI PUPPY.