@playform/hook-dsh-edit-suggester
v0.2.1 - UNPUBLISHED - plan:
SEARCH-SUGGESTER-PLUGIN
The role
The modification-intent companion suggester for the
DeepSeek Harness plugin family (the Classic
release).
A DeepSeek Harness plugin that hooks the tools/post-execute waterfall for the modification
tools (write, edit, raw-write).
When the edited path matches one of the convention families, it prepends ONE labeled
additionalContexts note that names the standard companions the model has not touched yet: "this
file rarely stands alone."
It is the second consumer of the pattern algebra of the search-suggester and of its two-consumer
extraction trigger
(SEARCH-SUGGESTER-PLUGIN.md sections 1.4 and
8).
Why
The evidence of the search-suggester shows that the edit is where the discovery happens: the model
edited its way to the target sets that it never thought to open.
The convention of the repo is that a file rarely stands alone.
The version lives in 36 manifests, 6 Apply.ts files, the CHANGELOG and the prose, and a
Config.ts edit must also land in the Config section of the README, SCHEME.md, the
cordis.patch.yml row, and the byte-exact test.
The search consumer deliberately refuses to own the edit transform, and this consumer owns the
NOTIFY side of it: it names the companions, never the change.
The families
Seven, in priority order: Version (0.85, because a wrong version companion set is costly),
Config, Symbol, Test, Doc, Prose, Workflow.
Each family is a pack row: its path rules (regexes on the edited path with weights) and its standard
companion set, where each companion is a path or glob with a prior and one reason phrase ("the
Config section", "the byte-exact assertion", "the long tail").
A companion without a reason is not emitted.
The families 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 edit/ domain.
A repository that ships .dsh/pack.yml (or a user-layer pack at
~/.dsh/playform/workspaces/<slug>.yml) gets its OWN edit/ families, and a repository without a
pack stays silent.
The gate
The same algebra as the search consumer: k = sigmoid(sum(weights) - 0.6), the candidate scoring
over the pickable heuristics (see below), and the noise control (a cooldown, a dedupe, a
per-turn cap, and the already-touched filter).
Companions that the model edited earlier this turn are dropped, and when nothing remains the verdict
is an already-touched abstention.
Every abstention and suppression is a ledger line.
The learning layer (D8)
Every emitted note, and every gated silence worth learning from (low-confidence,
already-touched), records a decision: the family, the companion rows with their six pickable
heuristic values (scope, graph, prior, literal, history, accept), the effective weight
table, the enabled list, and a pending outcome placeholder.
The record is logged (decided ...) and stored through the optional pattern service
(ctx.get("pattern"), probed per use, never an inject).
Without that service it lives in a bounded in-memory fallback (storeLimit, default 512), which is
per instance and documented as such.
The note opens a follow-through window counted in judged calls (never a timer).
A later call that names one of the companions of the note is an accept for that companion and
its family (accepted <Family> <path>), and a window that reaches learnWindow judged calls
untouched closes as a soft reject (rejected <Family> <n> companions).
Both fill the outcome of the decision record and move the learned accept rates.
The score is sum over the enabled heuristics of weight * value:
| Heuristic | Value | Default weight |
|---|
scope | 1.00 - the call matched the family's path rules | 0.35 |
graph | 0.50 - the neutral graph-distance placeholder | 0.00 |
prior | the pack's prior for the companion | 0.25 |
literal | 0.50 - the neutral literal (no generated index) | 0.25 |
history | the learned per-companion accept rate (neutral 1/3) | 0.15 |
accept | the learned per-family accept rate (neutral 1/3) | 0.00 |
The default row is exactly the PoC formula, and a missing store answers the neutral 1/3, so the
default behavior is unchanged byte-for-byte.
A user can disable any heuristic (heuristics) or re-weight any of them (weights), and the
decision record plus the decided line show the choice.
The delivery
One note, never a content replacement and never a claim of the Transform: the post-execute listener
runs after the ccr and the normalize family and only adds additionalContexts.
The note carries the plugin's own source label
({kind: "edit-suggester", form: "notice", summary}).
The grep/read/bash calls are not judged, because this is the edit side of the family and it is
disjoint from the trigger set of the search suggester.
The waterfall discipline
next() is always awaited first, and pass-through returns the downstream decision unchanged.
The whole transform is fail-open: an internal error logs a warning and the decision passes through,
so a note never fails a tool call.
The configuration
See SCHEME.md for the full schema.
Every knob is volatile (hot-editable, no remount).
The learning layer adds learn, learnWindow, heuristics, weights, and storeLimit.
Ships disabled - enable by explicit ACTIVATE in the profile.
The plan
License
CC0-1.0.
