@playform/hook-dsh-test-suggester
v0.2.1 - UNPUBLISHED - plan:
SEARCH-SUGGESTER-PLUGIN
The role
The test-suggester hook is a DeepSeek Harness plugin that answers the test task intent with a
target set.
When the model searches for proof (a grep for PASS/FAIL/error TS, a read of a Test.ts, a
bash error sweep), the hook classifies the call deterministically and names the likely targets
before the model finishes searching.
It is the test consumer of the suggester family
(SEARCH-SUGGESTER-PLUGIN.md, sections 3, 4,
8): a hook with no public API, an embedded pattern pack (the two-consumer extraction rule, with no
shared core service until the family proves two consumers), and one delivery surface.
How it works
-
The post-execute listener.
The tools/post-execute listener (waterfall, plain registration, running after the CCR's
{ prepend: true } listener) calls next() first.
Then the deterministic classifier runs over the call probe (tool name + arguments JSON), the
gate applies, and one additionalContexts notice is prepended to the downstream decision.
Content is never replaced, so a ccr-compressed result is unaffected.
-
The pre-step listener.
The agent/pre-step listener resets the note counter each turn when a user message enters.
Every listener is fail-open: on any error a warning is logged and the downstream decision is
returned unchanged.
A note never fails a tool call.
The pack
| Intent | Threshold | Trigger shape | Targets |
|---|
Suite | 0.75 | PASS|FAIL|error TS|assert|suite, Source/Test/ paths | Source/Test/<Suite>/Test.ts, the suite README, Maintain/Run-Smokes.mjs, the README badge |
Counts | 0.80 | N PASS, expected/PASS count, badge | README badge, SCHEME.md, the site content counts |
Confidence is sigmoid(Σ weights − 0.6) (the plan §3.5).
Every candidate carries a one-phrase reason, and a candidate without a reason is never emitted.
The intents are pack-driven: the built-in pack (the default of this monorepo) lives in
@playform/ets-lib-dsh-pack and is projected here as the test/ domain.
A repository that ships .dsh/pack.yml (or a user-layer pack at
~/.dsh/playform/workspaces/<slug>.yml) gets its OWN test/ intents, and a repository without a
pack stays silent.
The reconciliation
On a judged call in a package, the smoke claims of the package are reconciled with the tree it
actually has (the wave 2c re-spec of the completion spec,
03-SUGGESTERS.md).
The evidence is each ## Expected count of Source/Test/<Folder>/README.md, the folders that
exist, each Test.ts of a folder, the roster of the aggregate runner
(Target/Test/<Folder>/Test.js), and, when the workspace graph service is present, its Test nodes
for the package.
A drift (a folder without a README count, a README count whose Test.ts is gone, a suite the runner
never runs, a graph node whose file is gone) emits ONE note naming the drifted files and the
expected/actual counts.
It is read-only: node:fs reads, never a write.
A consistent tree logs one clear line per package root and stays silent, and the note is gated
exactly like the intent notes (workspace pack, enabled, cooldown, dedupe, per-turn cap).
The noise controls
- Confidence gate.
k ≥ max(threshold, intent threshold) (default 0.75).
- Cooldown. No second note for the same intent within
cooldownCalls (6) judged calls, per
agent.
- Per-turn cap. At most
maxNotesPerTurn (3) notes between user turns.
- Exclusion. Candidates under
Target/, node_modules, .git, dist are never suggested,
because the build mirror is generated and never edited.
- Fail-open and abstention. An empty surviving candidate set logs
<name>: abstained ambiguous <Intent> <k> and emits nothing.
- Workspace gate. A call that belongs to another workspace (no pack whose signature matches,
with the built-in default
Documentation/ARCHIVE-STYLE.md) logs abstained foreign-workspace and
emits nothing, and a target absent from the workspace is dropped before the gate
(abstained no-companions when the whole target set of the intent is gone).
Every activation, refusal, and suggestion is a ledger line in the own ledger of the hook (one log
per module), asserted byte-exactly by the end-phase smokes.
The ledger
test-suggester: activated (bundle 0.2.1, intents 2, threshold 0.75)
test-suggester: disabled (enabled=false)
test-suggester: suggest Suite 0.84 4 candidates call grep
test-suggester: abstained ambiguous Suite 0.62
test-suggester: abstained no-companions Suite 0.62
test-suggester: abstained foreign-workspace Suite 0.62
test-suggester: suppressed cooldown Suite
test-suggester: suppressed cap Suite
test-suggester: reconciliation clear <package> (<e> expected, <a> actual)
test-suggester: reconciliation drift <n> drifts <package> (<e> expected, <a> actual) call <tool>
test-suggester: reconciliation skipped deduped <package>
test-suggester: reconciliation skipped cooldown <package>
test-suggester: reconciliation skipped cap <package>
The configuration
See SCHEME.md for the full table.
Every knob is volatile (hot-editable without remounting).
The hook ships disabled (enabled: false in cordis.patch.yml), because the acceptance rate
must be measured (the §10.3 kill criterion of the plan) before it is trusted.
The plan
The package, the Source, the smoke suites (five folders), and the build pass are complete, and the
reconciliation landed (the wave 2c re-spec of the completion spec).
The ETS twin and the live activation (the profile row flip and the restart) are the remaining
end-phase steps (Documentation/Plans/POC-BRIEF.md section 8).
License
CC0-1.0.
