Contextualizing

Walk the Spec Tree from product root to a target node, deterministically collecting and reading all context: ancestor specs along the path, lower-index siblings' specs at each directory level, and all ADRs/PDRs. Emit <SPEC_TREE_CONTEXT target="..."> marker with a structured context manifest.

This is full injection — every collected document is read into the conversation. No heuristic selection.

PREREQUISITE: Check for <SPEC_TREE_FOUNDATION> marker. If absent, invoke /understanding first.

Invoke with the full path from spx/ to the target node:

/contextualizing 21-infra.enabler/32-parser.outcome

Never use bare node names — indices are sibling-unique, not globally unique:

# Wrong: ambiguous
/contextualizing 32-parser.outcome

# Right: unambiguous
/contextualizing 21-infra.enabler/32-parser.outcome

COMPLETE CONTEXT OR ABORT. NO EXCEPTIONS.

• Every node along the path must have its spec file ({slug}.md)
• Missing spec file = ABORT with remediation guidance
• Read order: product root → ancestors → target (top-down)
• All ADRs and PDRs at all levels must be read — no skipping based on title relevance
• Lower-index siblings' specs must be read at each directory level — they constrain the target
• Test files are excluded — context is about specs and decisions, not evidence

BOOTSTRAP MODE: When the target path doesn't exist yet and the operation is authoring, return an empty manifest with bootstrap=true instead of aborting. This allows creating the first node in an empty tree.

Phase 0: Locate target node

# Find the product file
Glob: "spx/*.product.md"

# Verify target path exists
Glob: "spx/{target-path}/*.md"

If the product file is missing, ABORT: "No product file found in spx/. Create one with /authoring first."

If the target path doesn't exist:

• If operation is author → return empty manifest with bootstrap=true
• Otherwise → ABORT: "Target path not found: {path}. Check the path or create it with /authoring."

Extract the path segments from product root to target. Each segment is a directory to walk.

Phase 1: Load product-level context

# Read product spec
Read: spx/{product-name}.product.md

# Read product guide if present
Read: spx/CLAUDE.md  (if exists)

# Read ALL product-level ADRs and PDRs
Glob: "spx/*-*.adr.md"
Glob: "spx/*-*.pdr.md"

Read EVERY file returned by the globs. Do not filter by title. Decision records contain cross-cutting constraints that may not be obvious from the title.

Verification: Count files returned by globs. Count files actually read. These must match.

Phase 2: Walk the tree from root to target

For each directory along the path from product root to the target node:

2a. Read the directory's spec file

# The spec file is {slug}.md (no type suffix, no numeric prefix)
Read: spx/{path-to-dir}/{slug}.md

ABORT if the spec file is missing.

2b. Read all ADRs and PDRs in this directory

Glob: "spx/{path-to-dir}/*-*.adr.md"
Glob: "spx/{path-to-dir}/*-*.pdr.md"

Read EVERY file returned. Verification: glob count must equal read count.

2c. Read all lower-index siblings' specs

The target node has an index (e.g., 43 in 43-feature.outcome). All sibling nodes with a lower index constrain the target and must be read.

# List all sibling directories (same parent, different from target)
Glob: "spx/{parent-path}/*-*.{enabler,outcome}/"

# For each sibling with a lower index than the target:
Read: spx/{parent-path}/{sibling-dir}/{sibling-slug}.md

Lower-index siblings' ADRs/PDRs are NOT read — only the sibling's spec itself. The dependency encoding means the sibling's existence constrains the target, but the sibling's internal decisions are its own concern.

2d. Note same-index siblings (independent)

Siblings with the same index as the target are independent — they neither constrain nor are constrained by the target. List them but do not read.

Phase 3: Load target node context

# Read target spec
Read: spx/{target-path}/{slug}.md

# Read target ADRs and PDRs
Glob: "spx/{target-path}/*-*.adr.md"
Glob: "spx/{target-path}/*-*.pdr.md"

# Enumerate children (if any)
Glob: "spx/{target-path}/*-*.{enabler,outcome}/"

# Check for tests directory
Glob: "spx/{target-path}/tests/*"

# Check for escape hatches
Glob: "spx/{target-path}/PLAN.md"
Glob: "spx/{target-path}/ISSUES.md"

If PLAN.md or ISSUES.md exist, read them. These are non-durable escape hatches left by previous agents via /handoff. They contain deferred plans or known issues that the next agent must be aware of.

Phase 4: Emit context marker and summary

Emit the <SPEC_TREE_CONTEXT> marker with all collected information:

<SPEC_TREE_CONTEXT target="{full-target-path}">

Product: {product-name}
Target: {target-path} ({enabler|outcome})

Documents loaded:
  Product spec: {product-file}
  Ancestor specs: {count} read
  Lower-index sibling specs: {count} read
  ADRs: {count} found, {count} read
  PDRs: {count} found, {count} read

Hierarchy:
  {product-name}
  └── {ancestor-1} ({enabler|outcome})
      └── {ancestor-2} ({enabler|outcome})
          └── {target} ({enabler|outcome}) ← TARGET

Children: {count} ({list if any})
Tests: {exists|missing}
Escape hatches: {PLAN.md | ISSUES.md | none}
Lower-index siblings read: {list}
Same-index siblings (independent): {list}
Higher-index siblings (depend on target): {list}

</SPEC_TREE_CONTEXT>

When a required document is missing, ABORT immediately with:

1. What's missing — exact file path expected
2. Why it's needed — what context it provides
3. How to fix — specific remediation action

Do NOT proceed with partial context. The whole point of deterministic context is completeness.

Failure 1: Skipped ADRs/PDRs based on title relevance

Agent globbed 12 decision records but only read 3 whose titles seemed relevant. The answer to the user's question was in one of the 9 skipped documents. The verification gate ("glob count must equal read count") prevents this.

Failure 2: Missed lower-index siblings

Agent walked the ancestor chain but didn't read lower-index siblings' specs. A lower-index enabler contained infrastructure the target depended on. The dependency encoding means lower-index = constraining, so they must be read.

Failure 3: Read higher-index siblings

Agent read ALL siblings including higher-index ones. Higher-index siblings may depend on the target but don't constrain it. Reading them wastes context window and may introduce irrelevant information.

Context loading is complete when:

• Product spec located and read
• All product-level ADRs/PDRs read (glob count = read count)
• Every ancestor along the path: spec read, ADRs/PDRs read
• Lower-index siblings' specs read at each directory level
• Target spec read
• Target ADRs/PDRs read
• Children enumerated
• Test directory status checked
• Escape hatches (PLAN.md, ISSUES.md) checked and read if present
• <SPEC_TREE_CONTEXT target="..."> marker emitted with full manifest
• No ABORT conditions triggered (or appropriate error shown with remediation)