English · Français

EGX_Terminal

EGX_Terminal is a research and governance workspace for designing inspectable, token-conscious terminal environments that remain LLM-, harness-, and provider-agnostic. Its V1 boundary preserves the evidence, decisions, and validation methods behind that work; it does not contain a universal agent runtime.

Why this exists

Terminal-agent environments can become structurally dependent on one provider, model, or harness convention. Instructions, hooks, skills, and memory can also accumulate until their always-on cost and behavioral effect are difficult to inspect. A plausible mechanism is not necessarily a demonstrated one.

This repository uses a bounded progression: audit the available evidence, state a hypothesis, isolate a prototype, pre-register the protocol, test it, record a decision, and preserve both positive and negative results. Complexity is promoted only after its value has been shown under explicit criteria.

What V1 actually ships

V1 contains:

• a project charter and current status;
• repository audits, cross-repository research, and an architecture synthesis;
• decision records with evidence, consequences, and revision conditions;
• immutable archives of two rejected kernel experiments;
• experimental protocols, fixtures, graders, and compatibility reports;
• local verification scripts and their applicable tests; and
• a neutral repository root.

A clone exposes none of the known active Codex, Claude Code, or OpenCode project surfaces recorded in neutral-root-surfaces.json. No kernel, adapter, or behavioral payload is active at the root. This is a bounded registry of current conventions, not a claim about every future harness convention.

The Git history has been audited and remediated. After GitHub selected a personal author address for the first squash merge, the affected repository was made private and the canonical repository was recreated from a content-identical clean history. PR #2 is complete and was squash-merged as commit c708bc6af88b5e98a08fc801e476d4c16248e701, establishing a 40-commit linear-history checkpoint. All commits in that checkpoint use approved GitHub noreply metadata; the affected object was not imported. The active policy continues to accept attributable GitHub noreply contributors and the exact GitHub web committer used for squash merges, without accepting personal email addresses or undeclared bots; see the identity remediation report and run python scripts/check_git_history.py --fail-on-review to verify the gate.

At the 2026-07-22 pretransition checkpoint 23cd5c596159fda6866e0fdc6ef0ba7bcf0d2515, main contained 41 commits and the canonical repository had no Git tag or published release. This count is a historical checkpoint, not a current counter. v1.0.0 is retained only as a historical release record withdrawn during the privacy remediation; its terminal governance conclusions remain in force, while Decision 0014 records the limits of the preserved micro-kernel runtime evidence. Its former release metadata is preserved only in verified private evidence outside this repository. GitHub's immutable-release reservation prevents reuse of that tag name in the recreated repository. Decision 0016 records the first controlled public transition and its rollback. The corrected final publication was subsequently applied and verified: the repository is public, Levels A, B, and C passed, and public run 30002915548 succeeded. No tag or release is active; v1.0.1 and its SSH-signing gate remain a separate later mission.

Key V1 decision

V1 is maintained with a neutral root. Useful principles remain available as documentation, while any future capability or adapter must be explicit and opt-in. Hidden instruction injection is outside the V1 boundary.

The final reasoning is recorded in the micro-kernel evaluation (0004), its runtime-evidence erratum (0014), and the neutral-root decision (0005). For the micro-kernel, the payload, protocol, fixtures, graders, manifest, and recorded hashes of preserved artifacts remain directly auditable. The micro-kernel final results preserve the published historical claims, but the original aggregate and source runtime data are irrecoverable. The balanced-kernel challenge results remain a separate historical record.

Experimental results

The micro-kernel scores, ties, token totals, and overhead shown below are values published in the historical report. Their internal arithmetic remains checkable, but the irrecoverable original aggregate and source runtime data prevent independent verification of the underlying runtime observations; see Decision 0014.

Experiment	Verdict	Functional success (baseline / candidate)	Primary wins (baseline / candidate)	Ties	False completions (baseline / candidate)	Token overhead	Pre-registered ceiling
Balanced kernel	REJECTED_AS_BALANCED	One baseline cell unavailable after scoring	2 / 0	3 complete pairs	0/5 / 2/6	+19.166% on the five complete pairs	n/a
Micro-kernel	REJECT_MICRO	5/5 / 5/5	0 / 0	5	0 / 0	+7.85% (exact value: 7.850019354305572%)	5%

According to the historical micro-kernel report, all five pairs were behavioral ties and the published token overhead exceeded the pre-registered budget. The terminal governance verdict remains REJECT_MICRO; this summary does not revalidate the unavailable runtime sources. The balanced result's +19.166% is a descriptive total for the five complete pairs; the six-pair overhead is unavailable because one baseline observation was lost after scoring. Neither rejection is presented as a promotion.

Compatibility evidence

The archived experiments established the following bounded evidence:

• deterministic static distribution into disposable targets;
• instruction discovery by Codex CLI 0.144.6 in the measured fixtures;
• mock OpenCode 1.17.9 discovery and single injection of the expected kernel;
• a direct Ollama 0.20.2 inference with a precisely identified Qwen model; and
• real OpenCode 1.17.9 → Ollama 0.20.2 → qwen3.6:27b transport.

The resulting compatibility classification is SUPPORTED — EXPERIMENTAL for the measured cells. The runtime compatibility summary separates static distribution, instruction discovery, runtime transport, formal output adherence, and behavioral effectiveness. Qwen is a model served by a runtime; OpenCode is a harness. Validating one combination does not establish universal compatibility, and transport success does not establish behavioral effectiveness. Claude Code was not installed and was not validated.

Repository map

Path	Purpose
docs/CHARTER.md	Defines the mission, constraints, and current non-objectives.
docs/STATUS.md	Records the current phase, closed decisions, and next gate.
docs/research/	Contains the external audits, evidence ledgers, and synthesis.
docs/decisions/	Contains consequential project decisions and their evidence.
docs/publication/	Defines the inspected public boundary and documented exceptions.
experiments/kernel-v1/	Archives the rejected balanced-kernel prototype and evidence.
experiments/kernel-micro-v1/	Archives the rejected micro-kernel and final campaign.
scripts/	Provides lightweight, deterministic V1 checks.
tests/	Tests the checks applicable to the neutral-root distribution.

Quickstart

The canonical local validation path requires Python 3.11+, a full Git clone, and REUSE 6.2.0 installed from the repository's hashed locks as described in the detailed quickstart. It uses the standard-library tomllib module and is officially verified from Python 3.11 onward; Python 3.9 and 3.10 are not supported by this path. The current environment was verified with Python 3.11.9 and Git 2.54.0; Git 2.54.0 is a verified environment, not a declared minimum. The active project checks use only the Python standard library. Installing the separate REUSE validator requires package-index access; validation itself requires no network, model, or agent runtime.

From the repository root:

python scripts/check_neutral_root.py
python scripts/check_public_surface.py
python scripts/check_licensing.py
python scripts/check_git_history.py --fail-on-review
python scripts/check_markdown_links.py
python scripts/check_github_governance.py
python -m unittest discover -s tests -v
reuse lint

Then read the charter, status, decision register, experimental archives, and publication boundary, in that order. The detailed quickstart distinguishes the complete clone audit from the smaller content-only sequence available in a GitHub source archive, and explains how to inspect evidence without relaunching a benchmark.

Community and publication

See contribution guidelines, governance, and the security policy. The SHA-pinned, read-only validation workflow defines the public checks. The GitHub publication record and release policy distinguish repository publication, immutable releases, tag signatures, and future release gates.

What is intentionally deferred

V1 does not implement memory, hooks, skills, routing, compression, active adapters, packaging, model distribution, cloud integrations, telemetry, or automated publication. These remain deferred until a concrete need and new evidence justify their ongoing cost and maintenance surface.

V1 is also not an installable universal agent, a complete orchestration framework, a model distribution, a memory system, a large skills collection, a hooks or routing system, a guarantee across all harnesses and models, or an automatically injected behavioral doctrine.

Design constraints

• Prefer explicit mechanisms over invisible ones.
• Make capabilities opt-in rather than always-on.
• Keep evidence local, inspectable, and attributable.
• Degrade cleanly when an optional mechanism is unavailable.
• Keep abstractions reconstructible rather than structurally provider-bound.
• Treat providers, runtimes, models, and harnesses as distinct roles.
• Keep complexity proportional to demonstrated value.
• Preserve existing user work.
• Scale verification with risk and reversibility.

These are documented project principles, not an automatically injected payload.

Current status

The sole canonical repository is ElGrandeXu/EGX_Terminal. Its durable phase is PUBLIC_REPOSITORY_VERIFIED. The final publication was validated at checkpoint 1d79ea37a1c614728cc7651c4d611218eaca174a with 43 linear commits: public workflow_dispatch run 30002915548 passed repository / ubuntu, repository / windows, and licensing / reuse, and Levels A, B, and C succeeded. PVR, secret scanning, push protection, vulnerability alerts, and the no-bypass main-protection ruleset were observed active. The merge gate now requires the PR branch to be up to date with main and the three required checks to pass for that state. Anonymous REST log downloads remained HTTP 403 and were classified INFO / PLATFORM_AMBIGUITY, consistently with the earlier diagnosis.

The first public transition and rollback remain historical facts, as do the identity remediation and repository recreation. The dated remote observations are not a claim that the offline checker queries GitHub in real time. No tag or release is active, and v1.0.1 preparation remains outside this mission.

License

Licensing is determined per file: original code and functional artifacts use Apache-2.0, while original documentation and research use CC-BY-4.0. This is not a choice between two licenses for the same file. See the short license summary, machine-readable REUSE metadata, and detailed licensing policy.

Validate the complete local mapping offline with:

python scripts/check_licensing.py