Add localhost-3000 hackathon submission

submissions/localhost-3000/.mutagent/config.yaml

@@ -0,0 +1,23 @@
+config_version: "0.3.0"
+
+global:
+  workspace:
+    repo: "chandu-jannu/fun"
+    path: "."
+  models:
+    default: "gpt-5.6-terra"
+  targets:
+    - name: "local-codex-workspace"
+      platform: "local-codex"
+      subject: "agent"
+      artifact_format: "code"
+      default: true
+      mode: "local"
+      root: "."
+      apply:
+        kind: "code-pr"
+        pr: false
+
+lifecycle:
+  agentspec:
+    spec_dir: ".mutagent/specs"

submissions/localhost-3000/.mutagent/specs/orbix-first-pass-candidate-evaluator/agentspec.decisions.md

@@ -0,0 +1,31 @@
+# Decisions — Orbix First-Pass Candidate Evaluation Agent
+
+## Human decision authority
+
+- Decision: The agent creates evidence-cited draft evaluations only; a recruiter is the decision of record for advancement, rejection, annotations, and overrides.
+- Rejected alternative: Automatic pass/reject thresholds or ATS handoff.
+- Rationale: Hiring outcomes, compensation, role level, and any process-ending decision require human review.
+
+## Evidence model
+
+- Decision: Skill assessment uses `strong`, `weak`, or `insufficient` evidence; direct resume-answer contradictions are separate conflict flags.
+- Rejected alternative: Fold contradictions and insufficient evidence into a single numeric score.
+- Rationale: Recruiters must see uncertainty and conflicting evidence rather than a non-explainable aggregate.
+
+## Interview failure behavior
+
+- Decision: Consent decline, timeout, abandonment, and technical failure stop scoring and report generation, with distinct recorded states.
+- Rejected alternative: Score partial interviews or treat failures as low performance.
+- Rationale: Partial or technically compromised evidence is not a valid assessment.
+
+## Audit preservation
+
+- Decision: Recruiter annotations and overrides append a separate audit event and preserve the agent's original draft.
+- Rejected alternative: Allow human review to overwrite the generated draft.
+- Rationale: The evidence trail must remain reviewable and attributable.
+
+## Implementation target
+
+- Decision: Build a code artifact for the local Codex harness in the Orbix workspace.
+- Rejected alternative: A framework SDK or multiple decision-making agents.
+- Rationale: The first release needs one bounded workflow with reusable services and human downstream review.

submissions/localhost-3000/.mutagentrc.json

@@ -0,0 +1,5 @@
+{
+  "endpoint": "https://api.mutagent.io",
+  "defaultWorkspace": "4a096d90-88c5-449e-87be-e0d297241f02",
+  "defaultOrganization": "org_q9EaDbbsdJcmO2KzPFm2n"
+}

submissions/localhost-3000/.vercel/README.txt

@@ -0,0 +1,11 @@
+> Why do I have a folder named ".vercel" in my project?
+The ".vercel" folder is created when you link a directory to a Vercel project.
+
+> What does the "project.json" file contain?
+The "project.json" file contains:
+- The ID of the Vercel project that you linked ("projectId")
+- The ID of the user or team your Vercel project is owned by ("orgId")
+
+> Should I commit the ".vercel" folder?
+No, you should not share the ".vercel" folder with anyone.
+Upon creation, it will be automatically added to your ".gitignore" file.

submissions/localhost-3000/.vercel/project.json

@@ -0,0 +1 @@
+{"projectId":"prj_pbEaGuboguHRSY4fkFw923ob74wV","orgId":"team_QI8FTPW2kmY5q5VQDZ3BZM88","projectName":"fun"}

submissions/localhost-3000/.vscode/settings.json

@@ -0,0 +1,3 @@
+{
+    "github.copilot.chat.claudeAgent.allowAutoPermissions": true
+}

submissions/localhost-3000/AGENTS.md

@@ -0,0 +1,16 @@
+<!-- @mutagent/helix boot -->
+# Helix — MutagenT ADL conductor
+
+This project has the Helix orchestrator installed. To boot it, read and adopt the agent
+definition at `.agents/skills/mutagent-helix/orchestrator.md` (run its activation-instructions: persona → system index →
+ADL dashboard), then await a `*command`.
+
+Trigger: `*mutagent` · `/mutagent-helix` · `boot`.
+
+DASHBOARD RENDERING — HARD RULE (Codex): on `*mutagent`/`boot`/`*help`/`*status`, output the
+orchestrator's `help-display-template` VERBATIM inside a fenced `text` code block. Preserve its
+EXACT shape — the boxed MUTAGENT header (box-drawing chars), every panel (lifecycle · system index ·
+setup/onboarding · state), and the command roster. Replace ONLY the `{placeholder}` tokens with
+live values; change NOTHING else. Do NOT summarize, shorten, paraphrase, drop panels, or convert it
+to Markdown headings/tables/bullets unless the operator explicitly asks for a condensed view.
+<!-- @mutagent/helix boot -->

submissions/localhost-3000/CLAUDE.md

@@ -0,0 +1,9 @@
+<!-- @mutagent/helix boot -->
+# Helix — MutagenT ADL conductor
+
+This project has the Helix orchestrator installed. To boot it, read and adopt the agent
+definition at `.claude/skills/mutagent-helix/orchestrator.md` (run its activation-instructions: persona → system index →
+ADL dashboard), then await a `*command`.
+
+Trigger: `*mutagent` · `/mutagent-helix` · `boot`.
+<!-- @mutagent/helix boot -->

submissions/localhost-3000/README.md

@@ -0,0 +1,188 @@
+# Orbix — First-Pass Candidate Evaluation Agent
+
+**Team:** localhost:3000
+
+## What it does
+
+Orbix is an AI-powered hiring platform. This submission specifies and implements the agent behind its core AI-interview flow: a bounded, auditable agent that takes a candidate through resume parsing → consent → a structured interview → an evidence-cited skill assessment, and then produces a **draft-only report** for a human recruiter.
+
+The agent **never makes a final hire or reject decision**.
+
+Every skill rating (`strong` / `weak` / `insufficient`) is grounded in specific resume and interview evidence. Direct contradictions between a candidate's resume and interview answers are flagged **separately** from the skill rating rather than being silently converted into a lower score.
+
+Consent is required and recorded before interview data is captured. Incomplete interviews — including abandoned, timed-out, or technically failed sessions — do not produce a score.
+
+## How it was built
+
+### 1. Spec
+
+We ran the full `*spec` interview through Helix (Codex harness), covering:
+
+- Candidate journey
+- Privacy and explainability requirements
+- Access boundaries
+- What the agent may read and write
+- Human-approval boundaries
+- Evaluation scenarios
+The specification included 10 representative scenarios, including:
+
+- Consent refusal
+- Resume/interview conflicts
+- Missing role requirements
+- Incomplete interviews
+The resulting specification files are:
+
+```text
+agentspec.yaml
+agentspec.decisions.md
+```
+
+Both files document the approved agent contract and design decisions.
+
+### 2. Build
+
+The Codex session reached its monthly usage limit during `*build`, before Helix's builder agent could run.
+
+Rather than block the submission, the agent was implemented directly against the approved specification.
+
+The implementation follows the same contract and constraints defined by the specification and was verified through the evaluation suite.
+
+To be transparent: the submitted implementation was not generated by `*build`; it is a direct implementation of the approved specification.
+
+### 3. Evaluate
+
+An evaluation suite was created covering all 10 scenarios defined in the specification.
+
+**Result: 23/23 checks passed.**
+
+The evaluation suite is:
+
+```text
+agents/orbix-first-pass-candidate-evaluator.eval.ts
+```
+
+## Important safety finding
+
+While grounding the specification against Orbix's existing `app.js`, we identified a potentially unsafe automated hiring decision:
+
+```js
+c.status = score >= 75 ? 'Passed' : score >= 50 ? 'Review' : 'Rejected';
+```
+
+This automatically assigns a candidate status based on a score.
+
+That conflicts with the agent's safety boundary because the AI system should not make the final hiring decision.
+
+Instead of hiding this issue, we explicitly identified it during the spec-first process.
+
+The intended recruiter workflow is:
+
+```text
+AI evaluation
+      ↓
+Draft verification report
+      ↓
+Human recruiter review
+      ↓
+Human decision
+```
+
+This demonstrates how the specification can identify safety issues in the surrounding product rather than simply documenting the existing implementation.
+
+## Mutagent Agent
+
+The main agent is:
+
+```text
+orbix-first-pass-candidate-evaluator
+```
+
+| Item | Path |
+|---|---|
+| Agent implementation | `agents/orbix-first-pass-candidate-evaluator.ts` |
+| Agent specification | `agentspec.yaml` |
+| Specification decisions | `agentspec.decisions.md` |
+| Evaluation suite | `agents/orbix-first-pass-candidate-evaluator.eval.ts` |
+
+### Key capabilities
+
+- Candidate information analysis
+- Resume and project claim evaluation
+- Structured interview evaluation
+- Evidence-grounded skill assessment
+- Resume/interview contradiction detection
+- Consent-aware interview workflow
+- Incomplete-interview handling
+- Structured verification results
+- Human-review boundary
+- Auditable evaluation process
+## Run the evaluation
+
+From the submission directory:
+
+```bash
+npm install -g tsx
+npx tsx agents/orbix-first-pass-candidate-evaluator.eval.ts
+```
+
+The evaluation suite contains 10 scenarios and 23 checks.
+
+Expected result:
+
+```text
+23/23 checks pass
+```
+
+## Project Structure
+
+```text
+localhost-3000/
+├── README.md
+├── agentspec.yaml
+├── agentspec.decisions.md
+│
+├── agents/
+│   ├── orbix-first-pass-candidate-evaluator.ts
+│   └── orbix-first-pass-candidate-evaluator.eval.ts
+│
+├── transcripts/
+│   ├── Claude Code session transcripts
+│   └── Codex session transcripts
+│
+├── app.js
+├── styles.css
+├── ai-interview-agent.html
+├── ai-verification.html
+├── analytics.html
+├── candidate-detail.html
+├── candidate-report.html
+├── candidates.html
+├── dashboard.html
+├── login.html
+├── resume-analysis.html
+├── skill-passport.html
+└── other Orbix application files
+```
+
+## Transcripts
+
+The submission includes the available session transcripts used during development.
+
+They are stored under:
+
+```text
+transcripts/
+```
+
+The directory contains the available Claude Code and Codex `.jsonl` session files, including the relevant main sessions and available subagent sessions.
+
+## Traces
+
+No separate `traces/` directory was available in the local project at submission time, so no trace files are claimed or fabricated in this submission.
+
+## Product Feedback
+
+Feedback was identified around the Codex harness usage-limit experience.
+
+In particular, the usage-limit error during the build process did not clearly distinguish whether the user should retry later or reconfigure billing/usage settings. This created uncertainty and cost development time during the submission workflow.
+