skillscript-runtime v0.17.3

Closes Perry's 1ea3d625 Finding 2 (execute_skill leaks full child
final_vars into the parent — 252KB MCP-response blowup on a 2-line
emit). Substrate-author's 6fb6ac1c empirical pre-flight (0/16 skills
reach ${R.final_vars.X}) confirmed the migration-free cut. Three-way
convergence: Perry's design ack (8ffa16d6) + substrate-author's
pre-flight (6fb6ac1c) + Perry's discipline-corrected catch
(5a7e4041) shaped the final scope.

Skills are functions — declare what you return

New # Returns: X, Y, Z frontmatter header declares the export surface
for execute_skill composition. The runtime filter at the composition
boundary (src/composition.ts:filterFinalVarsByReturns) drops every
final_vars entry whose name isn't in the declared list. Internal
scratch (large JSON, intermediate computations, debug values) stays
local to the child execution — never serialized into the caller's R
or the top-level MCP execute_skill response.

Concrete shape:

# Skill: get-weather
# Vars: LOCATION=Valdese
# Returns: SUMMARY, TEMP_F, CONDITIONS

fetch:
    shell(command="curl -s 'wttr.in/${LOCATION|url}?format=j1'") -> RAW
    $ json_parse ${RAW} -> PARSED

shape: fetch
    $set TEMP_F = ${PARSED.current_condition.0.temp_F}
    $set CONDITIONS = ${PARSED.current_condition.0.weatherDesc.0.value}
    $set SUMMARY = "${LOCATION}: ${TEMP_F}°F and ${CONDITIONS}"
    emit(text="${SUMMARY}")
default: shape

Caller binding -> R gets R.final_vars = {SUMMARY, TEMP_F, CONDITIONS} —
NOT RAW or PARSED. Always-exported (no declaration needed): outputs,
transcript, errors, target_order, fallbacks, agent_delivery_receipts.

Default: no Returns → no final_vars exported

Skills without # Returns: export nothing from final_vars. Caller still
sees outputs + transcript + execution metadata — the natural pattern
substrate-author's audit found in 100% of existing composers.
Pre-adoption rule (per project_pre_adoption_rule) makes the breaking
cut cheap; the migration is free for the canonical ${R.outputs.text}
pattern.

New lint: unknown-returns-ref (tier-1)

Fires when a name in # Returns: isn't bound anywhere in the skill body
(not in # Vars:/# Requires:, not output-bound, not a foreach iterator).
The skill claims to return something it never produces — structurally
broken, hard-blocked at compile.

Watch-items (not built, documented)

• transcript nests across composition depth. When A calls B calls C,
  A's R.transcript holds B's emissions; B's R.transcript (visible via
  declared returns or transcript-of-transcript chains) can hold C's.
  Bounded by intended emits per child; additive across siblings. Per
  Perry's 5a7e4041 discipline-corrected catch — fix-ahead-of-need not
  warranted today; revisit if deep-composition-with-large-emits bites.
• unexported-final-var-access (tier-2 advisory) — deferred to
  v0.17.4. Would catch ${R.X} where X isn't in the called skill's
  # Returns:. Requires cross-skill resolution + accessor-pattern
  parsing; warrants its own ring.

Test coverage

16 new tests (parser: 6, lint: 5, runtime: 5). Three-test discipline per
feedback_three_test_discipline_per_dispatch_shape. Full suite: 1613
tests pass.

Migration

Bundled examples + scaffold skills: zero changes needed. None reach
${R.final_vars.X} today (verified by substrate-author's audit). When
adopters want structured access to a composed child's results, they
add # Returns: X to the child skill.

---