Synthesize a quality-vision proposal for a brownfield codebase by merging the already-inferred nfr-spec, quality-profile, physical-architecture, and logical-architecture proposals, then attaching scan-index signals (ADRs, CI/lint/test config files, commit cadence) to the relevant ISO 25010 characteristics. Produces a Tier 1 proposal at architecture/quality-vision.yaml. Used exclusively by tech-architect during /codify.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Instruções da origem · Visualização somente leitura
name
infer-quality-vision-from-code
description
Synthesize a quality-vision proposal for a brownfield codebase by merging the already-inferred nfr-spec, quality-profile, physical-architecture, and logical-architecture proposals, then attaching scan-index signals (ADRs, CI/lint/test config files, commit cadence) to the relevant ISO 25010 characteristics. Produces a Tier 1 proposal at architecture/quality-vision.yaml. Used exclusively by tech-architect during /codify.
version
0.1.0
user-invocable
false
model
sonnet
allowed-tools
Read, Write, Grep, Glob
deprecated
true
deprecated_note
#434 ProductOS realignment — superseded by the command model; retained for Phase E reference, not installed
infer-quality-vision-from-code
Called by tech-architect during the /codify play, AFTER infer-nfr-spec-from-code,
infer-quality-profile-from-code (test-engineer output), infer-physical-architecture-from-code,
and infer-logical-architecture-from-code have already written their proposals. Produces
architecture/quality-vision.yaml at
{stm_base}/{issue}/evidence/codify/proposals/architecture/quality-vision.yaml, plus a
companion decision manifest and resolution trace.
Purpose
By the time this skill runs, four upstream proposals already describe what the codebase
does for quality: quality-profile.yaml (ISO 25010 relevance + target levels),
nfr-spec.yaml (per-NFR mechanisms + verification methods), physical-architecture.yaml
(specific named products), logical-architecture.yaml (component IDs + responsibilities).
This skill does synthesis, not fresh inference. The primary source for every vision
field is the upstream proposals; scan-index is used only to attach ADR references and
CI/lint/test config files to the right characteristic, and to read commit cadence for the
reliability narrative. It does NOT re-pick tools, re-set thresholds, or re-target levels.
Re-inference across boundaries is a discipline violation (see Boundaries).
Output is a Tier 1 proposal — learning_category: quality, sub_category: architecture.
Per DSD, high-confidence tier-1 entries flow through the batch-confirm surfacing stream at
the /codify checkpoint; knowledge gaps flow through the 1-by-1 stream.
Input
Receive from the /codify play orchestrator via JSON contract. All paths absolute.
Field
Required
Description
scan_index_path
yes
Path to scan-index.json produced by scan.py. Secondary input — used only for ADRs, config_files, and git.commits_analyzed.
stm_base
yes
STM root, resolved from stm.base-path in .garura/core/config.yaml.
Map with absolute paths to the four upstream proposals: nfr_spec_yaml, quality_profile_yaml, physical_architecture_yaml, logical_architecture_yaml. All four are required — this skill cannot run without them.
Process
1. Validate inputs
Confirm scan_index_path exists and is valid JSON. Missing → scan_index_missing.
Confirm every path in related_proposal_paths exists and parses as YAML. Any missing
or malformed → missing_related_proposal with offending_path set. This skill is
synthesis-only; without the four upstream proposals it cannot produce a valid vision.
Confirm output_path parent directory exists or is creatable. Failure →
output_parent_missing.
2. Resolution Protocol walk (write resolution-trace first)
Per core/components/memory/standards/rules/resolution.md. Record every probe.
R1 — STM. Skipped for /codify (reason: codify-bootstrap).
R2 — Product LTM. Probe {product_base}architecture/quality-vision.yaml. If
present and LOCKED, surface overlaps as alignment_confirmed entries. Absent → proceed.
R3 — KB. Probe {kb_base}/knowledge/quality/ (fallback
core/components/memory/knowledge/quality/) for vision-narrative templates per ISO
25010 characteristic. KB narratives seed phrasing; synthesis adapts to this product.
R4 — Web. Not invoked. Closed-universe over upstream proposals + scan-index + KB.
3. Parse the four upstream proposals into a synthesis table
Build one row per ISO 25010 characteristic whose quality-profile.yaml relevance is
not not_applicable. For each row, pull columns directly from the proposals — never
from fresh inference:
target_level ← quality-profile.yaml:characteristics[].target (with any adjustment
already recorded in nfr-spec.yaml).
design_linkage.components ← logical-architecture.yaml:components[].id for every
component whose responsibilities address this characteristic (read the responsibilities
text; no new mapping beyond what logical-architecture already states).
design_linkage.nfrs ← nfr-spec.yaml:nfrs[].id filtered to
characteristic == <row>.
tooling ← named products from physical-architecture.yaml (library_pins, observability,
auth_infra, ci stack) that are already identified as serving this characteristic. If
physical-architecture named a tool, use it verbatim with its version. Do NOT re-pick.
thresholds ← quantified values from nfr-spec.yaml:nfrs[].target and from
quality-profile.yaml:characteristics[].target. These are echoed, not synthesized.
vision_narrative ← a 2-5 sentence synthesis that frames why this characteristic
matters for this product. Source material: (a) quality-profile rationale, (b) epic
descriptions already available in .garura/product/scope/epics/ (from earlier /codify
stages or locked LTM), (c) KB narrative templates from R3. This is the ONE field
genuinely authored here — everything else is echo + attach.
4. Attach scan-index signals — bounded to three uses only
Scan-index is a secondary input; it contributes attachments, not decisions.
docs.adrs → ADR references. Match each ADR title to a characteristic by keyword
(e.g., "ADR-0007: connection pooling" → performance_efficiency). Attach as
design_linkage.adrs[]. Ambiguous titles attach to every plausible characteristic
and record as a manifest decision.
config_files → lifecycle gates. Lint (eslint.config.*, ruff.toml, .flake8),
type (tsconfig.json, mypy.ini), test (jest.config.*, vitest.config.*,
pytest.ini), and CI (.github/workflows/*.yml, .gitlab-ci.yml) map to lifecycle
gates under maintainability and reliability. Use the scan-index excerpt (do NOT re-read
source) to determine stage: pre-commit, PR blocking, nightly, pre-deploy.
git.commits_analyzed → reliability cadence signal. Commit frequency + revert
rate feeds the reliability narrative as a cadence signal. Narrative only — do NOT
derive thresholds from commit counts.
No other scan-index fields feed this skill. trees, entry_points, patterns,
manifests are out of scope — they served upstream proposals already.
5. Compose the top-level vision statement
Synthesize a 3-5 sentence vision_statement from (a) product name (via ltm_context →
project-profile.yaml), (b) the two or three highest-relevance characteristics from
quality-profile.yaml, and (c) primary use cases from the MVP proposal already written
earlier in /codify. The statement names which characteristics are load-bearing and why,
in the product's own terms.
6. Emit per-characteristic entries
One entry per synthesis row from step 3 under quality_vision.iso_25010 with fields
narrative, target_level, design_linkage, tooling, thresholds, lifecycle_gates.
Field shapes mirror core/components/skills/derive-quality-vision/SKILL.md — defer there.
Knowledge gap handling. If quality-profile.yaml marks a characteristic relevant but
nfr-spec.yaml has no matching NFR AND physical-architecture.yaml names no tool, emit
knowledge_gap: true with a note identifying the silent upstream. NEVER fabricate a
narrative or invent tooling to close the gap.
7. Write decision manifest
Inferred decisions here are narrow — most fields are echoed from upstream:
decision_id prefix
decision_type
Decided
D-iqvc-001
vision-narrative-framing
Which driver (business risk, compliance, UX) leads each narrative — synthesized from quality-profile rationale + epic text + KB template.
D-iqvc-002
adr-to-characteristic-attach
Which characteristic(s) each docs.adrs entry attaches to.
D-iqvc-003
config-file-to-lifecycle-gate
For each config file, which characteristic + gate stage.
D-iqvc-004
knowledge-gap-flag
Per characteristic, whether upstream coverage is complete or a gap is recorded.
Standard manifest shape (schema_version, skill, generated_at, decisions[]). Each
decision carries decision_id, decision_type, tier (per DSD), grounding_source
(kb_path or upstream-proposal pointer), recommendation, alternatives_considered,
agent_reasoning_summary. Write the manifest BEFORE the primary artifact.
8. Write primary artifact and return
Write quality-vision.yaml to output_path. Return the output contract to
tech-architect.
Output
Primary artifact — {output_path}
Shape mirrors core/components/skills/derive-quality-vision/SKILL.md (the canonical
schema) with meta.source_type: inferred_from_code and tier: 1. Top-level skeleton:
meta:source_type:"inferred_from_code"evidence:# upstream proposal paths + three scan-index pointers-"<nfr_spec_yaml>"-"<quality_profile_yaml>"-"<physical_architecture_yaml>"-"<logical_architecture_yaml>"-"scan-index.json#/docs/adrs"-"scan-index.json#/config_files"-"scan-index.json#/git/commits_analyzed"confidence:"high"|"medium"|"low"learning_category:"quality"sub_category:"architecture"tier:1quality_vision:vision_statement:"<3-5 sentences>"iso_25010:<characteristic-key>:narrative:"<2-5 sentences>"target_level:"<echoed from quality-profile / nfr-spec>"design_linkage: { components: [...], nfrs: [...], adrs: [...] }
tooling: [ { name, version, purpose } ] # every name from physical-architecturethresholds: [ "<quantified value>" ] # every value echoed, not synthesizedlifecycle_gates: [ { gate, tool, threshold } ]
# knowledge_gap: true + note — when upstream is silentknowledge_gaps: [ { characteristic, silent_upstream, note } ]
excluded_characteristics: [ { characteristic, reason } ]
Decision manifest — {decision_manifest_path}
Standard Garura decision-manifest shape. See step 7 for the inferred-decision catalog.
Resolution trace — {resolution_trace_path}
R1..R3 probes with source, path, outcome (hit | miss | skipped), extracted
payload where applicable. R4 omitted.
status:failurewhat_failed:"<code>"detail:"<specific error>"evidence:offending_path:"<file path if applicable>"
Codes:
missing_related_proposal — one or more of nfr_spec_yaml, quality_profile_yaml,
physical_architecture_yaml, logical_architecture_yaml is absent or unreadable.
This skill cannot synthesize without all four.
scan_index_missing — scan_index_path absent or not valid JSON.
insufficient_signal — quality-profile.yaml declares every characteristic
not_applicable (nothing to write), OR every relevant characteristic is a knowledge
gap because upstream proposals are silent. Return with overall_confidence: low and
the knowledge-gap list populated; do NOT fabricate entries.
ltm_resolution_failed — R2 probe errored (existing LTM quality-vision not parseable).
output_parent_missing — output_path parent cannot be created.
Boundaries
Synthesis, not fresh inference. Tool picks come from physical-architecture.yaml.
Target levels come from quality-profile.yaml / nfr-spec.yaml. Component IDs come
from logical-architecture.yaml. If this skill finds itself choosing between Jest and
Vitest, setting a coverage threshold, or picking an observability stack — STOP. That
is an upstream concern; re-picking is a boundary violation.
Scan-index is secondary. Only docs.adrs, config_files, and
git.commits_analyzed feed this skill. Other fields are out of scope.
Read-only against the codebase. Source files are not opened.
Writes stay under {stm_base}/{issue}/evidence/codify/proposals/architecture/.
Never writes to .garura/product/. Promotion is /garura:enrich's job.
Knowledge gaps are gaps, not placeholders. When upstream is silent, emit
knowledge_gap: true with a pointer — never fabricate to close the gap.
Does not re-run scan.py. Stale scan-index is the orchestrator's concern.
Always Tier 1.learning_category: quality, sub_category: architecture — /codify
routes confirmations through the batch-confirm flow per DSD.
