docs(specs): fold adversarial review findings into spec-derived scenario cards design

This commit is contained in:
Jesse Vincent
2026-07-04 20:10:22 -07:00
parent 2f3258c1fa
commit 21244c6595

View File

@@ -1,7 +1,9 @@
# Spec-Derived Scenario Cards — Design
Date: 2026-07-04
Status: approved (design review with Jesse, 2026-07-04)
Status: draft — adversarially reviewed (2x opus, findings folded in); one
decision open for Jesse at spec review: the card-author role boundary
(flag-only vs flag-then-fix-via-TDD). The spec is written flag-only.
Builds on: `2026-07-04-agentic-end-to-end-testing-design.md` (the skill this
extends; same branch)
@@ -18,7 +20,10 @@ new `agentic-end-to-end-testing` skill's card format, but no skill documents
how cards derive from a spec, no spec template asks for the table, and the SDD
pipeline has no hook to run any of it.
### Evidence (2026-07-04 experiment, 4 live runs)
### Evidence (2026-07-04 card-authoring experiment, 4 live runs; write-up at
`~/Documents/agentic-e2e-testing-corpus/live-runs-2026-07-04/CARDS-EXPERIMENT.md`,
raw artifacts alongside it — distinct from the same directory's RESULTS.md,
which records the earlier scenario-execution runs)
- With only a spec pointer (no table), card authors did NOT drift in the
current environment (n=2) — but the environment was contaminated (a
@@ -50,7 +55,9 @@ pipeline has no hook to run any of it.
- No quorum/eval-lab scenarios (per Jesse; the checker script and in-skill
discipline carry repeatability).
- No new plugin dependencies. Scripts use bash + POSIX tools only.
- No retroactive backfill of tables into existing specs.
- No bulk backfill campaign adding tables across existing specs. (Per-spec
backport during card authoring is allowed and specified in §2 — it is the
bootstrap path, not a campaign.)
## Design
@@ -69,7 +76,12 @@ section — a table with one row per scenario:
from the *requested* behavior at spec time, before implementation. This
line is a contract: cards must later carry it verbatim.
That is the entire brainstorming edit — no new checklist steps, no changes to
Two touchpoints, both small: the conditional above, plus one line added to
brainstorming's existing **Spec Self-Review** checklist — "user-facing
surface but no E2E scenario cards table? Add it." — so an omitted table is
*detected*, not merely discouraged (an unenforced prose conditional would
not deliver the "institutionalize" goal; downstream, SDD keys off the
table's presence, so silence would silently mean "no e2e"). No changes to
the question flow. Placement and exact wording are settled during
implementation under writing-skills discipline (RED baseline first:
brainstorm runs on a user-facing feature today do not produce such tables;
@@ -77,28 +89,41 @@ micro-test the wording; GREEN re-run).
### 2. agentic-end-to-end-testing: `authoring-cards-from-a-spec.md`
New supporting file, routed from SKILL.md §3 (one line: cards derive from the
spec when one exists) and reflected in §9's pipeline sentence. Contents:
New supporting file, routed from SKILL.md's "The scenario card" section (one
line: cards derive from the spec when one exists) and reflected in the
"Integration" section's pipeline sentence. (SKILL.md has no numbered
sections; reference headers by name.) Contents:
- **With a scenario table:** one card per row. The row's Falsification line
lands in the card's Expected section **verbatim**. The spec is
authoritative wherever the app's behavior disagrees — flag the
disagreement in the report; never adapt the card to observed behavior.
- **Without a table:** mine the spec's user-visible requirements into
behaviors; write the falsification lines; backport them to the spec (so
the contract exists for the next cycle) and say so in the report.
- **Without a table (bootstrap path):** mine the spec's user-visible
requirements into behaviors; write the falsification lines; add an "E2E
scenario cards" table to the spec carrying them (this is the sanctioned
per-spec backport), and flag the spec edit prominently in the report for
human review — the author must not present a self-written table as a
pre-locked contract. On this path the checker verifies transcription
consistency, not pre-implementation locking; the file says so plainly.
The locked-contract guarantee only exists when the table predates
implementation.
- **Coverage check:** every user-facing claim in the spec maps to a card or
a stated exclusion with a reason.
- **Role boundary:** the card author never modifies product code, test code,
or existing cards' assertions. A failing card plus root cause is the
deliverable, not a fix. (Provisional decision — flag-only; Jesse may widen
it to "flag, then fix via TDD" at spec review.)
deliverable, not a fix. (Open for Jesse at spec review: widen to "flag,
then fix via TDD." The 2026-07-04 experiment shows why ambient norms do
not decide this consistently: given the same failing card, one author
fixed and one declined.)
- **Dispatch snippet:** a short template for dispatching a fresh card-author
subagent (seeded from the historical card-authoring dispatch in the
corpus), naming: the spec path (authoritative), the card format, the
verbatim rule, the role boundary, and the report shape.
- **Mechanical check:** after authoring, run the checker script (below); a
clean pass is part of the author's report.
verbatim rule, the role boundary, the checker-run requirement, and the
report shape.
- **Mechanical check:** after authoring, the author runs the checker script
(below) and includes its output in the report; the dispatching agent
re-runs the checker independently before accepting the report —
self-attestation is not the gate.
### 3. subagent-driven-development: optional final step (core-skill edit)
@@ -107,14 +132,23 @@ final whole-branch review, plus one line in Integration:
- **Trigger (observable predicate):** the spec contains an "E2E scenario
cards" section, or the human asked for e2e verification. Otherwise the
step does not exist.
step does not exist. **Wiring:** SDD's entry step reads the plan, not the
spec — so the subsection instructs the controller, at skill start when it
reads the plan, to also open the spec the plan names and check for the
section; if present, record the pending e2e step in the todo list and
progress ledger so compaction cannot lose it.
- **Flow:** after the final review passes, the controller uses
superpowers:agentic-end-to-end-testing — dispatch a card-author subagent
(per `authoring-cards-from-a-spec.md`), then a runner subagent (per
(per `authoring-cards-from-a-spec.md`), run the checker independently on
the author's output, then dispatch a runner subagent (per
`runner-prompt.md`) against the built branch.
- **Failure handling mirrors the final-review contract:** card FAILs are
findings — ONE fix subagent with the complete list, then re-run the failed
cards. The card author never fixes; the fix wave does.
cards. The card author never fixes; the fix wave does. Fix-wave commits
land after the final whole-branch review, so they get their own focused
review (the task-review gate over the fix diff) before finishing —
unreviewed product changes must not ship on the strength of a green
re-run alone.
- **Placement:** before superpowers:finishing-a-development-branch, so
"ready to merge" includes live-scenario evidence.
@@ -131,24 +165,50 @@ Bash + POSIX tools (awk/grep/sed), no other dependencies. Usage:
check-cards-against-spec <spec.md> <cards-dir>
```
Matching semantics (normative — two implementers must not be able to build
different checkers):
- **Table location:** find the heading whose text case-insensitively equals
"E2E scenario cards" (any heading level); use the first markdown table
after it. No such heading or table → checks 2-3 are skipped and the
script exits non-zero with a "no scenario table" diagnostic (callers on
the bootstrap path run it only after the backport).
- **Columns** are identified by header name, case-insensitive (`Card`,
`Covers`, `Falsification`), not by position.
- **Cell unescaping:** `\|` in a table cell is unescaped to `|` before any
comparison.
- **Normalization:** collapse every run of whitespace (spaces, tabs,
newlines) to a single space and trim the ends; no other transformation;
comparisons are case-sensitive after normalization.
- **Matching is fixed-string** on the normalized text (no regex — the
falsification lines contain metacharacters and backticks by design).
- **Consequence, stated in the authoring file:** falsification lines are
prose contracts, not literal aligned output. Column-alignment assertions
(`TOTAL 20.85` with meaningful spacing) belong in the card's Expected
body, not in the table line, because normalization collapses runs of
spaces.
Checks, each reported individually, exit 0 only if all pass:
1. The spec's "E2E scenario cards" table parses (>= 1 row; every row has a
non-empty Card and Falsification cell).
2. Every table row has a corresponding `<cards-dir>/<card>.md`.
3. Every card contains its row's Falsification line verbatim,
**whitespace-normalized** (markdown re-wrapping must not fail the check —
the naive-grep false negative is a proven failure mode).
4. Every card has the skill's required sections (What this covers /
Pre-state / Steps / Expected / Cleanup). Sharp edges is not required —
it accretes during runs, and demanding it pre-run forces padding.
3. Every card contains its row's Falsification line verbatim under the
semantics above.
4. Every card has the skill's required parts, matched per the card format's
actual syntax: `**What this covers**` as bold inline text; `Pre-state`,
`Steps`, `Expected`, `Cleanup` as `##` headings. Sharp edges is not
required — it accretes during runs, and demanding it pre-run forces
padding.
5. Extra cards (in dir, not in table) are reported as a warning, not a
failure — authors may add cards beyond the spec's minimum.
Good `--help` and per-failure diagnostics (file, expected line, what was
found). Developed TDD: the script's failing tests come first, exercised
against fixture spec/card pairs; whether those fixtures are committed
follows house precedent for skill scripts, settled in the plan.
against fixture spec/card pairs that include a falsification line containing
`|` (escaped in the table) and regex metacharacters; whether those fixtures
are committed follows house precedent for skill scripts, settled in the
plan.
## Decisions
@@ -164,20 +224,36 @@ follows house precedent for skill scripts, settled in the plan.
## Testing plan (writing-skills Iron Law)
1. **Checker script:** ordinary TDD; red tests first.
1. **Checker script:** ordinary TDD; red tests first (including the
pipe/metacharacter fixture case).
2. **Brainstorming edit:** RED — baseline brainstorm run(s) on a small
user-facing feature; confirm no scenario table is produced today. GREEN —
with the edit, the spec contains a well-formed table (the checker's table
parser is the objective judge). Micro-test the conditional's wording.
3. **Card-authoring file:** RED exists already (the 2026-07-04 experiment is
the baseline; its artifacts are archived in the corpus live-runs
directory). GREEN — re-run the experiment's Arm-A prompt with the new
file available; cards must pass the checker and the report must flag the
spec disagreement.
4. **SDD edit:** RED — a scaled-down SDD run (tiny plan, spec-with-table)
without the hook: controller does not author/run cards. GREEN — with the
hook: controller reaches for the e2e skill after final review, and card
FAILs produce a fix wave, not a weakened card.
parser judges structure) AND a negative gate check: a brainstorm of a
non-user-facing change must NOT emit a table (the conditional's gate is
the failure-prone half). Table *quality* (falsification lines written
from requested behavior, actually falsifiable) is judged by human review
of the GREEN specs, not by the parser. Micro-test the conditional's
wording.
3. **Card-authoring file:** the honest framing of the 2026-07-04 experiment:
drift did not occur in the baseline (contaminated environment), so drift
prevention is sourced from project history, not claimed as
experimentally validated. What the experiment DID document as failures:
(a) the role-boundary split — one of two authors modified product code
without authorization; (b) verbatim compliance required an explicit
instruction. So: RED = the archived Arm-B1 run (unauthorized fix) and
Arm-A runs (no verbatim traceability without instruction). GREEN — rerun
both arm prompts with only the new file available (no special
instructions in the dispatch): authors must lift lines verbatim, pass
the checker, flag the spec disagreement, and NOT touch product code.
4. **SDD edit:** RED — a scaled-down SDD run (tiny plan, spec-with-table,
and a seeded assembly-level defect that unit tests pass but a card's
falsification line catches) without the hook: controller does not
author/run cards. GREEN — with the hook: controller reaches for the e2e
skill after final review, the seeded defect produces a card FAIL, and
the FAIL produces a fix wave plus focused re-review — not a weakened
card. (Without the seeded defect the discriminating half of this test
never fires.)
## Out of scope / future