# PLN-0003 — observed delivery evidence

Date: 2026-11-13. Actual build/test platform: macOS ARM64, Darwin 27.2.0.
Toolchain: GraalVM CE 25.5.5.0.3 / JDK 25.0.5.0.3, SDKMAN
25.5.5.0+3-graalce; Gradle Wrapper 9.3.1; Java release target 25, no previews.

## Scope, authority and source identity

WRK-0001 / PLN-0003, based on accepted DEC-0002 and the two actual human answers
preserved in SVY-0001 and DSC-0045: local Codex/Claude CLIs, Discussion/synthesis
proposals validated by AIE. This does not authorize source repository execution,
formal decision approval or human-reviewed Work closure.

Git HEAD c84bb8d36f88ef249a249cdfe7b476fd148717af, with uncommitted changes.
HEAD is the baseline, not a claim that this implementation was committed there.
ART-0033 captures 95 source/test/resource/build/documentation hashes, including
untracked files and excluding mutable Work bookkeeping.
Manifest SHA-256: 7f2363afe9f1f11608c726aa893b345ccc44b91cb60908c63659e044203e4b35.

## Implementation and technical boundaries

- agent/CliAgentAdapter normalizes versioned Codex and Claude wire responses into
  existing typed AgentProposal records. The framework assigns identity/fixture=false
  and validates references/policy before domain mutations.
- agent/internal/LocalProcess uses ProcessBuilder vectors, private directories,
  concurrent virtual-thread I/O, byte/time limits, descendant termination and cleanup.
  No shell command text or source repository working directory is supplied.
- Optional Invocation/Provenance and timeout metadata preserve old serialized/digest
  behavior when absent. Previous RUN outputs enter context as derived interpretations.
- CLI offers doctor, portable configuration, explicit/natural activity routing and
  guided init with console or prepared input. Setup validates before publishing.
- State/conformance/EPIC mechanics never call a model. No framework, database,
  mandatory Node dependency or generic execution engine was added. Ten modules remain.

Profiles are AIE role instructions; vendor configuration profiles are not loaded.
Configured default model is not represented as a known actual server-selected model.
Credentials remain in provider stores/environment, outside canonical artifacts.

## Executed build/test commands

```sh
./gradlew test --tests '*LocalCliAcceptanceTest' --console=plain
./gradlew test --tests '*CliTransportTest' --tests '*LocalCliAcceptanceTest' --console=plain
./gradlew test --console=plain
./gradlew clean test bootJar nativeCompile --console=plain
./gradlew nativeDistZip nativeDistTar -PtargetPlatform=macos-arm64 --console=plain
```

Four initial tests compiled and failed because capabilities were absent (ART-0021).
Then 13 targeted scenarios and the intermediate 101-test suite passed. After adding
identity/setup regressions, final clean result was 104 tests across 17 suites:
0 failures, 0 errors, 0 skipped (ART-0029/0030). Modulith verified ten modules.
Spring AOT, JAR, nativeCompile and both native distribution tasks succeeded.
Seven upstream deprecation warnings and build-time SLF4J no-provider notices were
observed. No other native platform or CI run is claimed.

## Actual provider probes, separate from automated fixtures

Doctor observed Codex 0.169.0 and Claude Code 2.1.311 with required controls (ART-0023).
All inference probes used a tiny synthetic project and 60-second provider timeouts:

1. Codex RUN-0001 summarized its Work goal and DSC-0001 (ART-0024).
2. Claude RUN-0002 resumed that persisted summary and distinguished it from canonical
   synthetic evidence (ART-0025).
3. RUN-0003 omitted --activity: Codex classified the request as summarize, then
   returned a synthesis. Both invocations have provenance (ART-0026).

Initial Claude inference failed authentication. Normal CLI auth reported a login,
while the filtered process environment did not. USER had been omitted. Preserving
standard user/identity/XDG values restored the provider's private login lookup and
retry succeeded without AIE copying credentials or initiating login (ART-0028).
An identity-forwarding regression now covers this boundary.

The restricted development tool sandbox also blocked a JVM sysctl query during
doctor. Repeating the local probe outside that tool sandbox succeeded. These failed
attempts were not recorded as successful RUNs. Automated tests require no real CLI,
account, credential, network service or LLM.

Retained synthetic live fixture: /tmp/aie-cli-live.nFOJGv.
No private conversation history, source repository or Zelda data was supplied.
Three successful RUNs persisted and can be inspected independently of the providers.
These probes demonstrate transport/continuity, not product research or test results.

## Native/runtime and distribution observations

The executable alone was copied outside the build directory and exercised with:

```sh
JAVA_HOME=/nonexistent GRAALVM_HOME=/nonexistent \
PATH=/usr/bin:/bin:/usr/sbin:/sbin AIE_KEEP_FIXTURE=1 \
bash scripts/native-smoke.sh /temporary/path/ai-engineering
```

Complete native smoke passed (ART-0031). The local-CLI section used controlled shell
executables, no JVM/model, and no Git/tools on PATH: guided setup, codecs, classification,
specialization, provenance, replay and nested reload all passed. Previous discussion,
EPIC, multi-Git and operational-cycle scenarios passed as well. No adjacent Native
Image helper libraries were copied. The smoke ran outside the development tool
sandbox to permit normal process/sysctl operations.

The native executable loaded the three real RUNs produced by the JAR; its complete
JSON context matched fresh JVM output byte-for-byte, and project validation returned
valid=true with no issues (ART-0032). Actual inference was verified on the JVM;
native transport was verified with controlled provider executables. Those are
distinct observations, not a claim of native real-provider inference.

ZIP and TAR were extracted and retained executable permissions. Both contained the
exact native binary exercised by the smoke and current docs. ZIP validated the live
fixture without JVM; TAR executed --version without JVM. Onboarding updated the
source project's guide/manifest and preserved root AGENTS.md bytes.

Native retained fixture:
/var/folders/0t/tfsqc14j5wbgh9h4s4dytwb00000gq/T/ai-engineering-smoke.drBqKs
Its root also contains independent runtime/studio Git repositories used by the
multi-repository continuity regression. All such fixtures are explicitly synthetic.

## Artifact SHA-256

ai-engineering:
bbc97798b8e83f07df15b5847314ca3a678ba51d8ae19349b96d67a986914c70

ai-engineering-0.1.0-macos-arm64.zip:
92c32d70383fd06cdeeccc3897447ed29319532b893ea57810f3bef3f9675260

ai-engineering-0.1.0-macos-arm64.tar.gz:
bc81fc88b25c6fef86c7c0570b3537f36fa5ada341c1a116f51403f5c7df6d8c

## Remaining limits and next experiment

Local adapters are proposal-only. Source editing/execution, direct APIs, vendor-native
profile configuration, automatic decomposition and legacy import require separate work.
Provider/OS controls are not an AIE OS sandbox. Authentication and model availability
remain external; no silent retry or executor fallback is performed. Model classification
can cost a separate call; explicit --activity skips it. Payloads are references unless
recorded as Discussion text. Pre-publication failures leave canonical Work unchanged;
cross-file application still has the existing explicit interruption/repair limits.

No real executor was automatically enabled in this source project. No Zelda files
were changed. Work completion remains subject to actual human review; implementation
authority is not fabricated delivery approval.

Next experiment: use one configured CLI for a bounded real discussion, answer its
survey and review a resulting DEC draft before planning any repository execution.
