School-Core Wayfinder Map

Durable topology map for the Agent School stack. This document answers who owns what, which artifact crosses each boundary, and where evidence returns. It does not replace a CE plan, a bd issue, or the runtime implementation.

Canonical location: school-core/docs/wayfinder-map.md Maintained with: docs/route-envelope.md, docs/school-core-architecture.md, docs/pipeline-explainer.md, and the active CE plan for the current change.

The route-envelope schema is the companion contract for the identifiers and state transitions shown in this map.

The stack in one view

                         CONTROL PLANE

 KnowledgeCore ──► deterministic intent route ──► Wayfinder map
        │                                             │
        └──────────────────────┬──────────────────────┘
                               ▼
                         CE plan + U-IDs
                               │
                               ▼
                    bd beads + dependencies
                               │
                               ▼
                         CE work execution
                               │
                               ▼
                         EXECUTION PLANE

                  FirstMate ──► Orca ──► Hermes
                  dispatch       runtime    agent/tools/model
                               │
                               ▼
                         git artifact + evidence
                               │
                               ▼
                         ASSURANCE PLANE

       project verify gate ──┬── Entire sensor ──┬── CTO/COO review
                             │                  │
                             └──────┬───────────┘
                                    ▼
                         acceptance + bookbag verdict
                                    │
                 ┌──────────────────┴──────────────────┐
                 ▼                                     ▼
          bd lifecycle update                     AgentMail human gate
                 │
                 ▼
                         LEARNING PLANE

      trajectory + consolidation + score + RouterExperience
                                    │
                                    ▼
                         next bounded route decision

                         OBSERVATION SURFACES

          Kanban / last_run / activity / reports = read models

Ownership boundaries

Concern	Authority	Evidence it emits	Must not become
Durable task lifecycle	`bd`	bead status, dependencies, claim, close	A second Kanban/bookbag tracker
User intent and durable context	KnowledgeCore	anchor/context references	A task executor
System topology	This Wayfinder map	ownership and dependency edges	A duplicate implementation plan
Implementation decisions	CE plan	requirements, decisions, U-IDs, tests	Mutable execution state
Worktree/agent dispatch	FirstMate	crew status, report, fallback	A quality judge
Runtime isolation	Orca	worktree, terminal, automation identity	Curriculum or scoring policy
Agent behavior	Hermes	session/tool/model evidence	Task lifecycle authority
Mechanical verification	`verify_gate.py` / project checks	test/typecheck/lint result	A subjective grader
Intent-aware diff review	Entire	checkpoint/diff findings	The sole acceptance gate
Quality judgment	CTO + COO	verdicts, findings, score	A task queue
Verdict record	`bookbag.py`	accepted, findings, score	Full task tracking
Human decision rail	AgentMail	approve/reject/fix	Automated merge authority
Learning state	trajectories, consolidation, scores, RouterExperience	bounded outcome/history	Unverified policy mutation
Operational visibility	Kanban, `last_run`, activity	read-only status projection	Routing authority

Canonical task identity

A task should be traceable through this chain:

KC anchor
  ↔ wayfinder_map
  ↔ ce_plan_id + plan_unit (U-ID)
  ↔ bd_id
  ↔ cycle_session_id
  ↔ firstmate_crew_id
  ↔ orca_worktree_id + terminal_id
  ↔ hermes_session_id
  ↔ git base/branch/commit + Entire checkpoint
  ↔ verify result
  ↔ bookbag verdict
  ↔ bd close/update
  ↔ learning record

Not every execution path exposes every identifier yet. Missing identifiers are observability gaps, not permission to invent a value. The route-envelope work should add the missing joins incrementally and preserve redaction boundaries.

Lifecycle paths

Normal implementation path

request
  → deterministic intent classification
  → KnowledgeCore pull
  → Wayfinder lookup/update
  → CE plan or existing plan unit
  → bd bead claim
  → CE work
  → FirstMate/Orca/Hermes when admitted
  → artifact and evidence
  → verify gate + Entire + CTO/COO
  → bookbag verdict
  → bd update/close
  → trajectory and compound record

Low-risk solo fast path

existing bd bead
  → deterministic low-risk admission (`fast_lane.py`)
  → CE work in current/disposable worktree
  → local verify + Entire
  → bd close with evidence

The fast path retains bd, verification, Entire, and close-out evidence. It only removes unnecessary crew/school overhead for work that is already bounded. Admission fails closed without a bd identity, one-file/low-risk scope, or all required checks. summarize_lane_metrics() must show lower overhead with no higher rework before the policy is promoted.

Failure and repair path

failure evidence
  → classify edge and mode
  → choose one bounded next action
      ├─ retry same route
      ├─ change capability/role
      ├─ repair task contract
      ├─ repair runtime/tool
      ├─ escalate to teacher/faculty
      └─ request human decision
  → record reason and evidence
  → stop after no progress

A failed task must not silently become a successful task merely because a fallback produced a response. Lifecycle success, task quality, and pipeline cost remain separate signals.

Current implementation truth

Area	Current state	Boundary to preserve
Capability bundle	Implemented and recorded for role/profile/skills/tools/gate	The resolver is additive while older role maps remain compatible
FirstMate → Orca → Hermes crew path	Wired behind a flag with bounded fallback and artifact checks	Do not claim every model/delegation route is proven by worktree success
Verify gate	Wired with explicit production preflight and reusable soft-skip/strict behavior	A skipped gate is not a pass
Entire	Wired as a non-blocking intent-aware sensor	It informs review; it does not replace CTO/COO acceptance
CTO/COO review	Independent correctness/security and completeness lenses	Teachers judge; they do not dispatch or edit
`bd`	Repository task authority	Bookbag and Kanban must not duplicate lifecycle state
Kanban	Read model of `last_run` and issue state	It displays state; it does not choose the next route
RouterExperience	Adaptive combo selection with static cold-start fallback	It may optimize model choice inside a route, not the outer CE workflow
CE route decision	Recorded in route/bookbag evidence	Future routing effect remains an open feedback-loop task
SkillOpt	Installable, but no reproducible corpus/held-out benchmark contract yet	Do not claim lift until the measurement spine exists

Design invariants

One authority per concern. bd owns lifecycle; CE owns plan decisions; Orca owns runtime; school-core owns judgment and growth.
Compiler before critic. Mechanical verification runs before subjective review whenever the project can execute the artifact.
Entire is evidence, not authority. Intent-aware findings enrich the review record without silently changing the acceptance rule.
A report is not proof by itself. The report, branch, commit, base, and inspected artifact must refer to the same work.
Adaptive routing stays inside deterministic routing. Exploration may choose among approved model combos, but it must not change the task's workflow without a recorded policy decision.
One policy change per learning cycle. Observe, choose one correction, act, verify independently, record, and stop on no progress.
Read models do not become authorities. Kanban, activity, and dashboards project state from authoritative records; they do not mutate task policy.
No fabricated evidence. Missing credentials, missing artifacts, skipped verification, and unavailable runtimes are explicit terminal states.

Follow-up map

The implementation work is tracked under Beads epic brandonbennett-een:

.3  durable Wayfinder map                         ← this artifact
.1  canonical cross-layer route envelope
.2  normalized failure-edge/outcome taxonomy
.4  CE U-ID ↔ bd/runtime evidence join
.9  deterministic route → verifier resolution
.5  Entire findings → learning record
.6  post-bead compound learning loop
.7  solo-dev fast lane
.8  bounded SkillOpt experiment contract

The dependency order intentionally delays SkillOpt until the pipeline can separate skill quality from model, routing, runtime, verifier, and task-contract effects.