Mandatory AGGG2.0 Context Reading

The following is the imported AGGG2.0 project contract. It applies only when those files exist in the selected project root; it does not establish paths or loaders for the Oper workspace.
Read the canon in tiers, not all at once: Tier A — always and in full;
Tier B — per task type, before acting in that domain. Reading everything at
once overloads the agent: it stops reading to the end and compliance drops
(instruction compliance ceiling). Do not answer from memory — files are
edited, content changes.

Tiered reading (which file when)

Tier	Files	When to read
A — always	AGENTS.md — pointer: where things are, file mirrors, QA/linters; CYCLE.md — work cycle (re-contextualization → database → research → plan → action → verification → findings → report); CLAUDE.md — persona, web-research-first, coding rules ("truth is in primary sources")	every task in the workspace, in full, even a small one
B — by type	CAMOUFOX.md — web research (patterns, gotchas); DB-FIRST.md — database before answer; AGENT-LSP.md — code: types, references, edits; CODE-GRAPH.md — diff/review; SETUP.md — install/update	the ones the task touches, BEFORE acting in that domain

Rules: Tier A — always. Tier B — all files the task touches (code + diff →
AGENT-LSP.md AND CODE-GRAPH.md). Do not read all of Tier B "just in
case" — progressive disclosure: exactly your set.

Skills by Task Type (read BEFORE the task — mandatory)

The seven files above are the foundation; next, read the skill matching the task type,
BEFORE starting work (progressive disclosure: not all at once, but the one for your task):

Skill	When to read
skills/nodumb/	Before a non-trivial feature, expensive branching decision, redesign, second edit without new facts
skills/ask-nodumb/	Product/UX task: feature, screen, flow, concept — before designing the solution
skills/changelog-discipline/	Any code change in a project with CHANGELOG; release preparation
skills/system-feedback/	After actions: does the user understand what happened; loading/empty/error states
skills/fable-judge/	After claiming "done" on a non-trivial task: adversarial verification (report = claims, re-run checks, verdict) — CYCLE.md phase 5
skills/fable-loop/	Non-trivial multi-step task with orchestration (parallel sub-agents, unified plan, intent gate) — CYCLE.md phase 4
skills/fable-domain/	Creating a domain skill (domain research → bundle: workflow + flowchart + adapter + trap + eval)

These are mandatory skills (nodumbmode, assembled from real failures):
ignoring them on a matching task type = a violation.

Paths (two equivalent options — use whichever is available):

When an AGGG2.0 project root is selected and verified (root discovery is project-local; see that project's `SETUP.md` if present):
  $AGGG2_ROOT/AGENTS.md
  $AGGG2_ROOT/CYCLE.md
  $AGGG2_ROOT/CLAUDE.md
  $AGGG2_ROOT/CAMOUFOX.md
  $AGGG2_ROOT/DB-FIRST.md
  $AGGG2_ROOT/AGENT-LSP.md
  $AGGG2_ROOT/CODE-GRAPH.md

Relative from this skill's project-local root, only when the seven files are actually present:
  ../../AGENTS.md
  ../../CYCLE.md
  ../../CLAUDE.md
  ../../CAMOUFOX.md
  ../../DB-FIRST.md
  ../../AGENT-LSP.md
  ../../CODE-GRAPH.md

Workflow

1. Read Tier A in full (Read tool): AGENTS.md, CYCLE.md, CLAUDE.md — even if you "remember what's in them."
2. Read Tier B per task type (table above) — BEFORE acting in that domain.
3. Read the task-type skill (table below) — mandatory.
4. Produce the canon-check — a mandatory artifact: a table "file → read → 1 main rule from it". Without it, compliance is not verifiable: the user must see the canon was actually read, not declared.
5. Then work strictly by the rules from the files:
   • web research is the first action for tasks with unknowns or choices, not the last;
   • question about folder content or sherpa-voice → search the database first
     (db/aggg2.db, db/sherpa-voice.db, db/research.db), then answer;
   • research conclusions and findings → record in db/research.db
     (python3 db-tools/findings.py add ...).
6. If a file is unavailable or the path has changed — stop and report, don't continue from memory.

Gotchas

• The canon-check is mandatory (step 4): the table "file → read → 1 main rule". Skipping it = the skill was not fulfilled — compliance is checked, not declared.
• Tiers don't cancel Tier B — a code task without AGENT-LSP.md/CODE-GRAPH.md is a violation; still read exactly your set, not the whole canon "just in case".
• Don't cache file contents between sessions — they're alive, edited, growing. Re-read each time the skill triggers.
• The Oper copy lives at .agents/skills/aggg2-mandatory-reads/ and is discovered by the installed runtime's skill directory loader. Project-local mirrors under projects/AGGG-2.7/ are separate artifacts; do not infer that their ../../ paths resolve from the Oper copy or from an external harness.
• ⚠️ No hardcoded absolute paths in code/docs — root is determined automatically (VERSION marker) or via $AGGG2_ROOT. Don't hardcode mount points (/run/media/..., C:\...) — they don't exist on other machines.
• If the selected project root or any required canonical file cannot be verified, stop and report the missing evidence; do not invent contents or substitute Oper paths silently.

References

• ../../AGENTS.md
• ../../CYCLE.md
• ../../CLAUDE.md
• ../../CAMOUFOX.md
• ../../DB-FIRST.md
• ../../AGENT-LSP.md
• ../../CODE-GRAPH.md