msd-core/docs/TESTING-SUITES.md

# Testing Suites

This project's `tests/` directory uses **filename suffix markers** to group tests into named suites. The harness `scripts/run-tests.cjs` filters by suite when given `--suite <name>`. Without a flag it runs every `*.test.cjs` file (the historical default — unchanged).

> Tracked by issue [#3597](https://github.com/open-gsd/gsd-core/issues/3597).

## Suites

| Suite | Filename pattern | What goes here |
|---|---|---|
| `unit` | `*.test.cjs` (no other marker) | Default fast lane. Pure logic, no network, no external processes beyond `gsd-tools`. Most tests live here. |
| `integration` | `*.integration.test.cjs` | Cross-module flows: full installer end-to-end, multi-tool orchestration, anything that crosses two or more bin entry points. |
| `install` | `*.install.test.cjs` | Tests that perform a real install/uninstall against a sandbox project. Slower; PR CI skips these on PRs and runs them on `main` push only. |
| `security` | `*.security.test.cjs` | Adversarial input, prompt-injection guards, fixture-driven hostile-payload sweeps. |
| `slow` | `*.slow.test.cjs` | Anything that routinely takes >5s wall-clock or holds significant memory. |
| `qa` | `*.qa.test.cjs` | End-to-end walks that drive the real `gsd-tools` binary across multiple loop steps against one accumulating temp project, with invariant oracles after every step. Slower than `unit`; excluded from the fast lane. |
| `all` | (any) | Explicit alias for "no filter". Equivalent to running with no `--suite` flag. |

## How to place a new test

1. Pick the most specific bucket above.
2. Name the file with the matching suffix: `tests/<feature>.<suite>.test.cjs`.
3. If unsure, leave the suffix off — the file lands in `unit`, the default fast lane.

Examples:
- `tests/agent-frontmatter.test.cjs` — `unit`
- `tests/prompt-injection-guards.security.test.cjs` — `security`
- `tests/installer-end-to-end.install.test.cjs` — `install`
- `tests/sdk-mutation-stress.slow.test.cjs` — `slow`
- `tests/loop-walk.qa.test.cjs` — `qa`

The suite-suffix convention was chosen over a directory layout (`tests/security/`) so the 545+ existing test files don't need to move. Existing files all classify as `unit` until someone explicitly retags them.

## Regression tests

**Do not create new top-level `tests/bug-NNNN-*.test.cjs`,
`tests/fix-NNNN-*.test.cjs`, or `tests/issue-NNNN-*.test.cjs` files.** Add the
regression case to the owning module's main test file instead (e.g. a
`describe('regressions')` block in `tests/<module>.test.cjs`).

`node --test` spawns one child process per FILE, so file count — not test
count — is the unit of CI overhead, and it is worst on Windows lanes where
every spawn is Defender-scanned. The 2026-06 CI audit found 244 one-off
`bug-*` files (~38% of the suite). That population is grandfathered in
`scripts/lint-regression-test-names.allowlist.json` and enforced by an
identity ratchet (`npm run lint:regression-names`, part of `npm run lint:ci`),
which also bans `fix-*` and `issue-*` NNNN-prefixed files the same way:

- A **new** `bug-*`, `fix-*`, or `issue-*` NNNN-prefixed file fails CI — fold
  it into the owning module's file.
- **Deleting/consolidating** a grandfathered file requires pruning its
  allowlist entry, so the baseline only ever shrinks.
- **Inherited drift** (the failure names files your PR didn't add — e.g. the
  base branch merged `bug-*`/`fix-*`/`issue-*` files without feeding the
  allowlist, or you rebased and carried a pre-rebase allowlist): run
  `node scripts/lint-regression-test-names.cjs --update` and commit the
  regenerated allowlist. Snapshot artifacts like this allowlist (and
  `docs/INVENTORY.md`) must be regenerated **after** rebasing, never carried
  through a rebase.

### A folded suite may appear only once per host

When a standalone file is folded into its owning module's test file, the moved
suite is wrapped in a self-contained block carrying a marker:

```javascript
// ────────────────────────────────────────────────────────────────────────
// Folded from tests/bug-376-claude-js-hook-gsd-rewriter.test.cjs — …
// ────────────────────────────────────────────────────────────────────────
{
  const { describe: __foldDescribe } = require('node:test');
  __foldDescribe("folded:bug-376-claude-js-hook-gsd-rewriter (…)", () => { … });
}
```

Because the block is self-contained, a **second verbatim copy in the same host
parses, registers, and passes — twice.** Nothing in a green suite reports it.
[#3271](https://github.com/open-gsd/gsd-core/issues/3271) found 25 such copies
(~5,800 lines) across three install suites, all from a single stale-base
re-application during the consolidation epic. The cost is not only wasted CI on
every lane: it is a `DEFECT.GENERATIVE-FIX` trap, because a contributor fixing
one of those regressions edits the copy they found and leaves the other
asserting the old behavior, with the suite still green.

`local/no-duplicate-fold-marker` (`eslint-rules/no-duplicate-fold-marker.cjs`,
error under `tests/**/*.cjs`) reports the second and every later occurrence of a
`folded:<marker>` title in one file, naming the line the first occurrence sits
on. **When it fires, delete the copy it points at** — the two blocks are the
same suite, so the fix is removal, never an `eslint-disable`.

It keys on the whitespace-delimited token after `folded:`, which matters in both
directions. A narrower key that stops at `.` would collide
`feat-443-effort-fast-mode.integration` with `feat-443-effort-fast-mode` — two
genuinely distinct suites that coexist in `tests/model-resolver.test.cjs`. Keying
on the *whole* title instead would let a re-fold under a different batch label
slip through, which is exactly the shape #3271 took. Titles without a `folded:`
prefix, non-literal titles, and the same marker appearing in two *different* host
files are all left alone.

The ratchet deliberately covers only `bug-*`. Files named `feat-NNNN-*` /
`enh-NNNN-*` are *feature* test files — one (or one per suite) per feature is
the sanctioned layout (see the #443 strategy below), not a one-off regression
pattern. If `issue-*`/`perf-*` one-offs start accumulating the same way
`bug-*` did, extend the ratchet's regex and regenerate the allowlist.

## Workflow & agent size budget

> Tracked by issue [#1074](https://github.com/open-gsd/gsd-core/issues/1074).
> Bytes (not lines) per [#717](https://github.com/open-gsd/gsd-core/issues/717);
> LF-normalized per [#683](https://github.com/open-gsd/gsd-core/issues/683).

Workflow files (`gsd-core/workflows/*.md`) and agent files (`agents/gsd-*.md`)
both ship in the installed runtime and are loaded into context — workflows on
every command, agents on every subagent dispatch — so their byte size is a real
cost. Two sibling guards (`tests/workflow-size-budget.test.cjs` and
`tests/agent-size-budget.test.cjs`) keep that cost from creeping up invisibly,
sharing one byte-counter (`measureMdFiles`). Growth is caught by two independent
layers:

| Layer | What it does | Where |
|---|---|---|
| **Differential attribution size ratchet** (primary, #2724 / ADR-2719 §4) | The same computed-attribution check that replaced the golden-install-parity fixtures also reports growth in any `gsd-core/workflows/*.md` or `agents/gsd-*.md` file, with the exact byte delta, comparing PR HEAD against `next`. Unacknowledged growth is a hard failure; shrinkage needs no acknowledgment. No committed snapshot — nothing to regenerate by hand. | `tests/emitted-attribution.test.cjs` (real-tree test) via `tests/helpers/emitted-diff.cjs` |
| **Loose tier hard caps** (backstop) | Absolute outer red lines per tier — workflows: `XL ≤ 98304`, `LARGE ≤ 61440`, `DEFAULT ≤ 40960` bytes; agents: `XL ≤ 57344`, `LARGE ≤ 49152`, `DEFAULT ≤ 24576` bytes. A cap is **never raised** when a file approaches it: crossing it means *extract*, not bump. Independent of the ratchet above — unaffected by #2724. | `XL/LARGE/DEFAULT_CAP` in each guard file |

`discuss-phase.md` additionally has a thin-dispatcher target of `< 32000` bytes
(the discuss-phase progressive-disclosure split, #717). A net-new agent is
DEFAULT-tier and already bounded by the DEFAULT cap — no separate new-agent cap
is needed. (This tier-cap machinery is distinct from the separate 45 KB-*char*
extraction-evidence threshold on `gsd-planner` enforced by
`tests/planner-decomposition.test.cjs` — that one proves mode sections were
extracted; this one bounds total agent bytes.)

### How-to: a workflow or agent grew and CI is red

The differential attribution check reports the file and the byte delta. To resolve:

1. **Justify the growth in your PR** (a sentence in the description is enough) —
   the acknowledgment entry (below) is the review record that the larger size
   was a deliberate, seen decision, not silent drift.
2. **Add an acknowledgment entry** in `tests/emitted-drift-ack.json` naming the
   file and the reason, per `CONTEXT.md`'s `### Emitted Artifact Provenance`
   entry. This is deliberately a committed file, not a flag — the entry appears
   in your PR diff, so touching it *is* the visible signal.
3. **Or shrink it instead of acknowledging.** Prefer extraction when the growth
   is incidental: for a workflow, move per-mode bodies to
   `workflows/<name>/modes/`, templates to `workflows/<name>/templates/`, and
   shared prose to `gsd-core/references/`; for an agent, lift shared boilerplate
   into `gsd-core/references/` and `@`-reference it — then load it **LAZILY**. Do
   *not* convert them to eager `@-required_reading` includes: that shrinks the
   file's bytes without shrinking loaded context, so it games the guard while
   making the real cost worse. See `workflows/discuss-phase/` for the
   progressive-disclosure pattern.

If a hard cap (not the ratchet) is what failed, an acknowledgment will **not**
help — that is the signal to extract, per step 3.

### Reference

| Artifact | Role |
|---|---|
| `scripts/workflow-size.cjs` | Single source of truth — LF-normalized byte counter (`lfByteCount`) + generic `measureMdFiles(dir, predicate)` (backs both workflows and agents) + workflow enumeration (`listWorkflowStems`, `measureWorkflows`). Imported by both guards and by `tests/helpers/emitted-runtime.cjs`'s `currentSizes()` so they can never measure differently. |
| `tests/emitted-attribution.test.cjs` + `tests/helpers/emitted-diff.cjs` | The differential attribution check and its size ratchet (ADR-2719). The sole mechanism for both emitted-content propagation AND per-file size growth as of #2724. |
| `tests/emitted-drift-ack.json` | Committed acknowledgment file for unattributable emitted-content ripples and for size growth. Absent = no acks; its presence is the alarm. |
| `npm run regen:derived` | Runs every remaining generator in dependency order (build → registry → ADR index → capability matrix → inventory manifest → manifest versions → `tests/fixtures/install-tree/*.json`). |
| `tests/workflow-size-budget.test.cjs` | The workflow tier hard-cap guards, plus the `discuss-phase` progressive-disclosure checks. |
| `tests/agent-size-budget.test.cjs` | The agent tier hard-cap guards (the agent analog). |

`tests/workflow-size-baseline.json`, `tests/agent-size-baseline.json`,
`tests/fixtures/golden-install-parity/*.json`, `scripts/update-size-baseline.cjs`
(`npm run size:baseline`), and `scripts/git-merge-regen-driver.cjs`
(`npm run setup:merge-driver`) are all removed by
[#2724](https://github.com/open-gsd/gsd-core/issues/2724): they were pure
functions of the source tree, conflicted on every merge that touched them, and
their functions are now served by the differential attribution check above.
`tests/fixtures/install-tree/*.json` is the one artifact family that stays
committed and normally-merged (ADR-2719 §7) — it conflicts on 0 of 7, its diffs
are readable, and it preserves "the installer stopped shipping X" as a hard
absolute failure with no attribution reasoning involved. Regenerate it with
`npm run gen:install-tree` (folded into `npm run regen:derived`).

## The QA smell ratchet

`tests/loop-walk.qa.test.cjs` (the `qa` suite) is the QA-walk harness's own
self-test. Separately, `scripts/qa-smell-ratchet.cjs` drives that same harness
end to end against the real `gsd-tools` binary and turns its findings into a
CI gate — run it with `npm run lint:qa-smells`.

The gate has three independent inputs. The harness's oracles
(`tests/qa/oracles.cjs`) supply the first two:

- A **violation** is the engine breaking a documented contract. It always
  fails the build — baseline or no baseline, acknowledged or not.
- A **smell** is legal-but-questionable behavior. A smell **never fails a
  build on its own merits**. What fails is the absence of a decision about
  it: an **unacknowledged NEW smell**, or a **STALE** entry in
  `tests/qa/smell-baseline.json` (one that stopped firing — the baseline is
  shrink-only, so a fixed or changed scenario must be pruned, not left
  behind).

The third comes from the scenarios themselves, not from an oracle:

- A **scenario expectation failure** is a step's declared `expect` not
  holding — the scenario asserted `percent: 100` and the engine returned
  something else. Like a violation, it is **never acknowledgeable**: it
  carries no fingerprint, so there is no `key` to put in a baseline entry or
  an ack fragment. Fix the engine, or correct the expectation.

Expectation failures were invisible to the gate until
[#3597](https://github.com/open-gsd/gsd-core/issues/3597): `buildReport`
counted them in `totals.violations` while `collectFindings` read only oracle
violations, so a failing scenario printed `0 violations` and exited 0. The
`multi-workstream` scenario failed on every CI run for three weeks without
reddening a build. The invariant that keeps the two honest — asserted in
`tests/loop-walk.qa.test.cjs` — is:

```
collectFindings().violations.length
  + collectFindings().expectationFailures.length
  === report.totals.violations
```

Note also that `scripts/qa-smell-ratchet.cjs` only invokes its own `main()`
under `require.main === module`. That guard is what lets the QA suite
`require()` the script to test `collectFindings` without kicking off a real
20-scenario walk as an import side effect — the reason the gate's own logic
had no test before #3597.

Every smell must terminate in exactly one of TWO states — there is no third
"accepted with a good explanation" state:

1. **REAL** — an assigned defect. File it, then acknowledge the smell with an
   entry (baseline entry or `tests/qa/smell-acks/` fragment) carrying that
   `issue` number.
2. **FALSE POSITIVE** — the oracle itself is wrong. Fix the oracle
   (`tests/qa/oracles.cjs`) so it stops firing. It is NEVER baselined.

When the ratchet reports a NEW smell, there are exactly two legitimate
responses — fix the detector, or file a defect and cite its issue number:

1. **Fix the underlying behavior (or the oracle, if it's a false positive)**
   so the smell stops firing.
2. **File a defect and acknowledge it** by adding a fragment under
   `tests/qa/smell-acks/` — the ratchet's failure output prints a paste-ready
   skeleton naming the required `key`, `id`, `scenario`, and `issue` fields.
   `issue` MUST be a positive integer naming the tracking issue; a free-text
   `reason` may accompany it as an optional human note but can NEVER
   substitute for `issue` — "write an explanation" is not a way to acknowledge
   a smell. See `tests/qa/smell-acks/README.md` for the full shape and
   lifecycle.

Run `node scripts/qa-smell-ratchet.cjs --update` to regenerate
`tests/qa/smell-baseline.json` from the current run, folding in any acked
fragments and pruning stale entries. `--update` never invents an issue
number: a genuinely new smell is written with `issue: null` and a TODO
`reason`, and the very next plain (non-`--update`) run REJECTS that entry —
forcing a human to triage it before it can ship. The baseline only ever
shrinks: growth happens by adding an acknowledgment carrying a real issue
number (a reviewable diff), never by widening the generator's tolerance and
never by prose alone.

## Running suites locally

```bash
npm test                    # everything (backcompat — same as before)
npm run test:unit           # only unit
npm run test:integration    # only integration
npm run test:install        # only install
npm run test:security       # only security
npm run test:slow           # only slow

npm run test:coverage       # backcompat — coverage over EVERY test
npm run test:coverage:unit  # fast coverage signal — only unit suite
npm run test:coverage:all   # alias for test:coverage
```

Direct harness invocation also works:

```bash
node scripts/run-tests.cjs --suite security
node scripts/run-tests.cjs --suite=security
node scripts/run-tests.cjs --files "tests/command-contract.test.cjs tests/core.test.cjs"
node scripts/run-tests.cjs --files-from .ci-selected-tests.txt
```

`npm run test:affected` (scripts/run-affected-tests.cjs) is a **local-only**
convenience that selects tests via the `require()` dependency graph of your
working-tree diff. CI does not use it — CI selection is the rule table in
`scripts/ci-test-scope.cjs`, which is the authoritative mapping. If the two
disagree, trust (and fix) the rule table.

Unknown suites exit non-zero with the list of valid suites. Empty suites (e.g. `--suite security` before any security-tagged file exists) exit `0` with a `no tests in suite "..."` notice on stderr so CI lanes don't go red while a suite is being populated.

## The live-config hermeticity guard

Every `run-tests.cjs` invocation snapshots GSD's own install footprint in each
live runtime config directory before the suite and re-checks it afterwards. It
exists because the failure it catches is silent by construction: a test that
resolves a config directory from the ambient environment instead of a sandbox
writes into *your real* `~/.claude` (or `$GSD_HOME/.gsd`, or a Kimi
`config.toml`), and nothing reports it. CI cannot catch this class at all —
CI never has `CLAUDE_CONFIG_DIR` and friends set.

The guard watches only what GSD unambiguously owns — its top-level install
footprint plus `gsd-`-prefixed children of directories shared with the host
agent — never whole config roots, because a host agent legitimately writing
`history.jsonl` mid-run would make the guard cry wolf, and a guard that cries
wolf gets switched off.

Two environment variables control it:

| Variable | Effect |
|---|---|
| `GSD_STRICT_LIVE_CONFIG_GUARD=1` | A detected write **fails the run**. Set on the Linux/macOS lanes of every CI job that runs the suite. |
| `GSD_SKIP_LIVE_CONFIG_GUARD=1` | Skips the check entirely. |

Unset, the guard **reports and does not fail** — deliberately, not timidly. On
its first CI run it surfaced pre-existing leaks on the Windows lane, where
`os.homedir()` reads `USERPROFILE` and ~190 test sites sandbox `HOME` alone.
Those are real and worth fixing, but they are a different defect class, and a
brand-new gate that instantly reds an unrelated lane gets reverted rather than
obeyed. Windows lanes therefore stay report-only until that sweep lands; this
repo has the pattern already, in the `local/no-source-grep` ESLint rule that
shipped at `warn` and was promoted to `error` after its cleanup (ADR 452).

`GSD_SKIP_LIVE_CONFIG_GUARD` is a bypass on a safety check, so it is documented
here rather than left to be discovered in the source: an undocumented bypass is
one people eventually set without knowing what they turned off. If you need it
routinely, that is a bug report, not a workflow.

Reported paths are labelled `CREATED`, `MODIFIED`, `DELETED`, or `UNVERIFIED`.
`UNVERIFIED` means a scan bound was hit and the path could not be attested
either way — it is never the same as clean.

## CI matrix

The `Tests` workflow runs every PR through a scoped gate generated by
`scripts/ci-test-scope.cjs`.

All lanes run on **Node 24** — the `engines.node` floor (`>=24.0.0`) and the
only supported runtime.

| Lane | Scope |
|---|---|
| `ubuntu-latest` (scoped) | scoped tests — fast PR signal |
| `ubuntu-latest` (full, sharded) | unit + integration + security |
| `windows-latest` | scoped Windows/path/shell tests |
| `macos-latest` | full parity when required |

- **Scoped tests** are selected from the changed paths, plus a small CLI/package
  smoke set. They are for confidence on the affected surface, not for counting
  tests.

The default PR gate runs the broad `unit` (under the c8 coverage gate),
`integration`, and `security` suites once on Ubuntu / Node 24, scoped tests on
a second Ubuntu / Node 24 lane, and scoped tests on Windows / Node 24.
"Scoped" means the diff-selected list from the rule table — not the full suite
and not a fixed smoke set (the fixed smoke list is only the empty-selection
fallback). The Windows lane's list is the Windows-sensitive subset of the
selection, plus **every changed test file, unconditionally** (the #494
invariant, narrowed): a modified test is exercised on the divergent OS before
merge at per-file cost, without paying for the three full parity lanes.

PRs touching workflow, package, test-runner, install, release, or
Windows-sensitive surfaces also run the full parity matrix on macOS and the
older Windows runtime, plus `install` and `slow` on the primary Ubuntu lane.
Everything (including the full parity matrix) runs on every push to `next`,
which covers the residual macOS / Windows cross-product for scoped PRs.

Coverage runs inside the Ubuntu / Node 24 full lane (not a separate job — that
duplicated the entire unit run) and stays single-lane because multiplying
coverage across OS/runtime lanes adds cost without improving the threshold
signal. Note the gate's deliberate blind spot: it measures
`gsd-core/bin/lib/*.cjs` only — `scripts/`, `hooks/`, and `bin/` are
unenforced, and `stryker.config.mjs` additionally excludes ~48% of lib lines
from mutation testing (see the UNMUTATED list there). Widening either gate is
tracked work, not an accident to "fix" silently by raising thresholds.

To inspect the scope locally:

```bash
npm run ci:test-scope -- --files "commands/gsd/plan-phase.md"
node scripts/ci-test-scope.cjs --base origin/next --head HEAD
```

## Chunk packing and the test timing table

`scripts/run-tests.cjs` does not hand the whole selected file list to one
`node --test` process. It packs the files into **chunks**, each spawned
separately, because Windows caps a command line at 32,767 characters and because
each chunk gets its own 600s timeout (`RUN_TESTS_CHUNK_TIMEOUT_MS`) and a fresh
process, which bounds memory pressure.

How files are distributed across those chunks decides whether the slowest chunk
sits near that timeout while the others idle. The packer weights each file by its
**measured duration**, read from `tests/test-timings.json`, and places files with
LPT (longest-processing-time-first: heaviest file first, each into the currently
lightest chunk). Before #2456 the weight was guessed from the filename, which
mis-ranked files badly enough that the slowest chunk ran ~3.9x the lightest.

### Reference

| Knob | Default | Meaning |
|---|---|---|
| `RUN_TESTS_MAX_FILES_PER_CHUNK` | `60` | Per-chunk weight budget. Weights are normalized so an **average-cost** file weighs 1, so this still reads as "about 60 average files". |
| `RUN_TESTS_MAX_CMDLINE_CHARS` | `28000` | argv ceiling per chunk, with headroom under the Windows 32,767 limit. |
| `RUN_TESTS_TIMINGS_FILE` | `tests/test-timings.json` | Path to the timing table. Tests override it to inject a synthetic cost profile. |
| `RUN_TESTS_CHUNK_TIMEOUT_MS` | `600000` | Per-chunk timeout. |

The timing table is **advisory and deliberately un-gated**. There is no `--check`
mode and no CI lint that fails on staleness, because timing data legitimately
varies run to run. A file missing from the table falls back to the table's median
weight, and a missing or unparseable table falls back to uniform weight — so
drift costs chunk *balance*, never a red build. A count-based floor additionally
guarantees the packer never produces fewer chunks than plain count-based packing
would, so a badly stale table cannot collapse the suite into a few fat chunks.

### How-to: regenerate the timing table

Regenerate when the suite's cost profile has visibly drifted — after adding or
removing expensive tests, not on a schedule. The input is a `node:test` reporter
event stream from a `gsd-test` run:

```bash
node scripts/gen-test-timings.cjs \
  ~/.local/state/gsd-test/runs/<run-id>/test-events-linux-node24.jsonl \
  ~/.local/state/gsd-test/runs/<run-id>/test-events-macos-node24.jsonl
```

Pass every lane you have. A file's recorded time is the **max** across the
supplied streams, not the mean: the packer exists to keep the *slowest* lane's
slowest chunk away from the timeout, so the conservative bound is the right one.
Keys are sorted so a regeneration diff shows only the files whose cost moved.

## Best practices for forward-compat (Node 24/26)

- Use `process.execPath` when spawning Node in tests so each matrix lane exercises the lane's Node version.
- Avoid stack-trace or error-message prose assertions. Assert `err.code`, structured JSON fields, or enums — Node minor releases routinely tweak error wording.
- Prefer `node:test`, `node:assert/strict`, and `node:test` mocks. No external test frameworks.
- Coverage uses `c8` and propagates `NODE_V8_COVERAGE` through the harness's child process.

---

## Test strategy: #443 effort + fast_mode engine

> Feature: unified cross-provider effort and fast_mode knobs (issue #443).
> Test files: `tests/model-resolver.test.cjs` (unit),
> `tests/model-resolver.test.cjs` (integration).

### Testing pyramid

| Layer | File | What it covers |
|---|---|---|
| **Unit** | `feat-443-effort-fast-mode.test.cjs` | Pure logic: cascade rules, clamping, escalation math, malformed config handling, schema key validation. No CLI subprocess. |
| **Integration** | `feat-443-effort-fast-mode.integration.test.cjs` | Architecture-level invariants: cross-provider validity, totality across the 33-agent registry, CLI JSON contract, config round-trip, fast-mode honesty. Real subprocesses via `runGsdTools`. |
| **E2E** *(pending)* | *(not yet wired)* | Propagation layer: effort frontmatter / `CLAUDE_CODE_EFFORT_LEVEL` env actually reaching a spawned Claude Code subagent. See "Gaps" below. |

### Architectural invariants

Each invariant exists to prevent a specific class of production failure.

#### (a) Cross-provider validity

**What:** `renderEffortForRuntime(runtime, universalEffort).value` must always
be a member of the runtime's real provider enum. Ground-truth enums are defined
as local constants in the test — not sourced from the implementation.

```
PROVIDER_EFFORT_ENUMS = {
  claude: Set { 'low', 'medium', 'high', 'xhigh', 'max' }   // Anthropic output_config.effort
  codex:  Set { 'minimal', 'low', 'medium', 'high', 'xhigh' } // OpenAI model_reasoning_effort
}
```

**Why:** Passing a value outside these sets results in a 400 from the real API.
The clamping logic (`max -> xhigh` for codex; `minimal -> low` for claude) must
hold for every cell of the VALID_EFFORTS × runtimes matrix.

#### (b) Param/channel contract

**What:** Each runtime exposes a stable `param` string (the native API field
name) and `channel` (how the value is propagated). Unknown runtimes return
`param: null, channel: null` and pass the effort value through unchanged.

**Why:** Callers read `.param` to construct the dispatch payload. A regression
here would silently drop effort from subagent invocations.

#### (c) Resolve-execution JSON contract

**What:** The `gsd-tools resolve-execution <agent>` command emits a JSON object
with all eight keys present and typed correctly: `model` (string), `profile`
(string), `effort` (VALID_EFFORTS member), `effort_rendered` (string),
`effort_param` (string|null), `effort_propagation` (string|null), `fast_mode`
(boolean), `fast_mode_supported` (boolean).

**Why:** Orchestrators and workflow dispatchers parse this JSON. A missing or
mistyped field silently breaks downstream consumers.

#### (d) Totality across the real registry

**What:** For every agent in the 33-agent registry, `resolveEffortInternal`
returns a VALID_EFFORTS member (never undefined/null), `resolveFastModeInternal`
returns a strict boolean, and `renderEffortForRuntime('claude', effort)` stays
within the claude provider enum.

**Why:** A catalog addition that introduces a missing `routingTier` mapping
would otherwise produce `undefined` and propagate silently.

#### (e) Fast-mode honesty invariant

**What:** When the runtime is `claude`, `fast_mode_supported` in
resolve-execution output is always `false`, regardless of the fast_mode config.
`RUNTIMES_WITH_FAST_MODE` contains only `'api'`.

**Why:** Claude Code's `/fast` toggle is session-level only. Emitting
`fast_mode: true` as frontmatter on a Claude subagent is a silent no-op.
Advertising `fast_mode_supported: true` for claude would cause orchestrators to
believe the knob was wired when it is not.

#### (f) Precedence first-valid-wins

**What:** Both effort and fast_mode use a layered cascade. The test table covers
all four effort layers (invocation override → agent_overrides →
routing_tier_defaults → default) and all five fast_mode layers, including the
case where an invalid value at a higher layer correctly falls through.

**Why:** Silent precedence bugs (e.g., a numeric value in agent_overrides not
being rejected) would override intentional user config.

#### (g) Dynamic-routing composition

**What:** `resolveEffortForTier` escalates effort by attempt number
independently of the model tier mapping. The test verifies the effort ladder
(`low -> medium -> high -> xhigh -> max`), the `max` clamp, the
`max_escalations` cap, and that `escalate_on_failure: false` suppresses
escalation entirely.

**Why:** Effort escalation and model escalation share configuration
(`dynamic_routing`) but must operate independently; coupling them would cause
over-escalation or under-escalation.

#### (h) Config-tooling round-trip

**What:** `gsd-tools config-set` accepts all new key namespaces
(`effort.default`, `effort.routing_tier_defaults.<tier>`,
`effort.agent_overrides.<agent>`, `fast_mode.enabled`,
`fast_mode.routing_tier_defaults.<tier>`, `fast_mode.agent_overrides.<agent>`)
without an "Unknown config key" error, and values set via `config-set` are
reflected in `resolve-execution` output.

**Why:** The schema validation gate (`VALID_CONFIG_KEYS` + `DYNAMIC_KEY_PATTERNS`)
is separate from the resolver logic. A key missing from the schema would produce
a silent write failure and appear as a bug only at runtime.

### Coverage targets

| Suite | Target |
|---|---|
| Unit | Every cascade rule, every fallthrough, every clamp. All function branches in `resolveEffortInternal`, `resolveFastModeInternal`, `resolveEffortForTier`, `renderEffortForRuntime`. |
| Integration | All 8 architectural invariants. All 33 registered agents. All 6 provider × effort combinations for the valid-enum check. Full config-set key namespace. |

### Gaps / not yet covered

**E2E orchestrator-spawn-propagation layer (pending follow-up wiring):**
The integration tests verify that GSD resolves and renders effort values
correctly. They do NOT verify that the rendered values actually reach a spawned
Claude Code or Codex subagent at runtime. Specifically uncovered:

- `CLAUDE_CODE_EFFORT_LEVEL` env var being set and read by a spawned claude subprocess
- `output_config.effort` frontmatter key surviving the AGENTS.md template substitution
- `model_reasoning_effort` field surviving serialization into a Codex API request body
- Fast-mode `speed: "fast"` field reaching an `api`-runtime request when `fast_mode_supported: true`

These require spawning real subagents (or stubs thereof) and asserting on the
process environment / request payload — a scope that belongs in a future E2E
suite under `*.slow.test.cjs` or dedicated fixture-driven integration work.