Architecture

Control objects

The project separates five control objects:

π_job — executable job-search policy (SKILL.md).
I_job — evolution procedure (docs/EVOLUTION_PROTOCOL.md).
U_job — frozen evaluator (evals/EVALUATION_TASKS.md and reviewer gates).
r_job — deployment evidence (memory/RUN_SUMMARY.md, reports, user/recruiter feedback).
g_job — protected goal and invariants (docs/CONSTITUTION.md).

A candidate patch may change π_job or, in a separate cycle, I_job; it may not change the target behavior and its own acceptance gate in the same cycle.

System map

flowchart TD
    U[User goals, feedback, interview events] --> O[Orchestrator]
    O --> M[Load canonical GitHub memory]
    M --> R{Run mode}

    R -->|Ordinary run| D[Discover current roles]
    D --> F[Field Onboarding Map]
    F --> C[Company / product / role due diligence]
    C --> S[Rank P0/P1/P2/WATCH/REJECT]
    S --> H[Mobile briefing + HTML report]
    H --> W[Compressed writeback]
    W --> E[Deployment evidence r_job]

    E --> T{Evolution trigger?}
    T -->|No| X[End / next scheduled run]
    T -->|Yes| Q[Record trigger only]

    R -->|Evolution cycle| SM[Self-model snapshot]
    SM --> P[Generate 1-3 surgical patches]
    P --> B[Baseline vs candidate on same evidence]
    B --> V[Frozen evaluator U_job]
    V --> G{Protected goals g_job preserved?}
    G -->|No| RB[Reject / rollback]
    G -->|Yes| A{Target improves and contrast/fresh cases pass?}
    A -->|No| RB
    A -->|Yes| PR[Pull request with rollback point]
    PR --> MR[Review and merge]
    MR --> X

Memory architecture

flowchart LR
    C[CONSTITUTION\nprotected invariants] --> SK[SKILL\nactive policy]
    FM[FIELD ONBOARDING MAP\ncareer-space model] --> SK
    EV[EVIDENCE & DD\nresearch rules] --> SK
    SK --> ST[STATE\ncurrent goal/funnel/version]
    SK --> CL[COMPANY LEDGER\none current row per company]
    SK --> RS[RUN SUMMARY\nlatest 5 + triggers]
    RS --> EP[EVOLUTION PROTOCOL]
    EP --> PR[Branch / PR]
    PR --> SK
    PR --> RB[ROLLBACK]

Agent roles

The roles are logical responsibilities, not mandatory separate processes.

Role	Job
Orchestrator	Select run mode, route mix, workload and stop conditions.
Source Scout	Find current openings and high-quality company/product sources.
Role Analyst	Map JD to actual product, customer, economics and daily work.
Company Analyst	Evaluate outlook, product position, commercial quality and employer risk.
Field Map Builder	Maintain directions, transitions, capability gaps and bridge roles.
Application Strategist	Choose resume version, contact path and interview narrative.
Reviewer	Apply frozen evidence, ranking, output and regression gates.
Memory Steward	Compress updates and prevent prompt/document bloat.
Evolution Steward	Run isolated candidate patches, tests and rollback.

Data lifecycle

stateDiagram-v2
    [*] --> Discovered
    Discovered --> Verified: current role and source confirmed
    Discovered --> Dropped: expired / vague / obvious mismatch
    Verified --> Researched: product + company + employer evidence
    Researched --> Ranked
    Ranked --> Applied: user confirms application
    Ranked --> Watch
    Ranked --> Rejected
    Applied --> Screening
    Screening --> FirstRound
    FirstRound --> SecondRound
    SecondRound --> Offer
    Applied --> Closed: rejection / withdrawal / role closed
    Screening --> Closed
    FirstRound --> Closed
    SecondRound --> Closed

Change boundaries

Runtime evidence can change state, ledger and run summary.
Company-specific findings never enter the global skill.
Stable user preferences enter the ledger only after they affect future decisions.
Global policy changes require an evolution PR.
Evaluator changes require a separate evaluator-evolution PR with the target policy frozen.

FilesExpand file tree

ARCHITECTURE.md

Latest commit

History