Reassess preparation incrementally from verified dependencies

.claude-plugin/marketplace.json

@@ -8,7 +8,7 @@
       "name": "dk-workflows",
       "source": "./",
       "description": "dk-workflows",
-      "version": "1.1.23"
+      "version": "1.2.0"
     }
   ]
 }

.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "dk-workflows",
-  "version": "1.1.23",
+  "version": "1.2.0",
   "description": "Portable repository workflows, configurable reviewers, and guarded drafting.",
   "author": {
     "name": "Dmitry Kropachev"

.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "dk-workflows",
-  "version": "1.1.23",
+  "version": "1.2.0",
   "description": "Portable repository workflows, configurable reviewers, and guarded drafting.",
   "author": {
     "name": "Dmitry Kropachev"

PROVENANCE.json

@@ -61,5 +61,12 @@
       "source": "local-skills/shared",
       "commit": "57382c9a8cf574e221891b6777d4ea8892956266"
     }
+  ],
+  "owned_changes": [
+    {
+      "source": "dk-code-factory/dk-workflows",
+      "version": "1.2.0",
+      "description": "Incremental preparation, conclusion graph, immutable evidence lineage, refresh validation and regression coverage."
+    }
   ]
 }

README.md

@@ -1,6 +1,6 @@
 # dk-workflows
 
-Independent MIT plugin for Claude Code and Codex. Version 1.1.23.
+Independent MIT plugin for Claude Code and Codex. Version 1.2.0.
 
 ## Install
 
@@ -226,3 +226,12 @@ schedule an update. All later manifest reads use that retained copy, including
 when the host removes the original installation during lookup. Entry failures
 without a usable runtime receipt stop work; offline release checks still return
 a retained runtime and token and remain nonblocking.
+
+Preparation v2 retains stable conclusion dependencies and immutable evidence snapshots.
+`prepare --previous-preparation FILE` compares complete Git entries and governing
+sources, reuses proven unchanged conclusions, and batches affected conclusions into
+one reassessment. Identical and supported operational refreshes use zero model calls;
+`--full-analysis` requests full reassessment. The refresh report exposes changed inputs,
+affected/reused IDs and actual call counts. Review coverage, session identities, CI,
+Decisions synchronization and iteration budgets retain their existing gates. See
+[the preparation artifact contract](runtime/dk-pr-decisions/references/preparation.md).

claude/skills/dk-pr-decisions/SKILL.md

@@ -8,3 +8,5 @@ Use Claude Code native agents and tools. `${CLAUDE_PLUGIN_ROOT}` is this plugin
 Before work, read [the lifecycle procedure](../../../common/lifecycle.md), then run its entry command with `--host claude`. Retain its returned token and follow its structured action. Always run its completion command with that token in a finalization step, including on failure or interruption.
 
 Read [the shared procedure](../../../runtime/dk-pr-decisions/procedure.md) and perform the requested workflow. Resolve paths beginning with `PLUGIN_ROOT/` against this plugin's root. Preserve the shared procedure's validation and publication guards.
+
+On preparation refresh, select the active artifact with `prepare --previous-preparation FILE`; follow [the incremental preparation contract](../../../runtime/dk-pr-decisions/references/preparation.md). Preserve review budgets and exact session/coverage gates.

claude/skills/dk-pr-iter/SKILL.md

@@ -8,3 +8,5 @@ Use Claude Code native agents and tools. `${CLAUDE_PLUGIN_ROOT}` is this plugin
 Before work, read [the lifecycle procedure](../../../common/lifecycle.md), then run its entry command with `--host claude`. Retain its returned token and follow its structured action. Always run its completion command with that token in a finalization step, including on failure or interruption.
 
 Read [the coordinator entry](../../../runtime/dk-pr-iter/procedure.md), then read each linked stage contract before acting in that stage. Give child agents only their assigned role contracts and pinned evidence paths. Resolve paths beginning with `PLUGIN_ROOT/` against this plugin's root. Preserve the shared procedure's validation and publication guards.
+
+On preparation refresh, select the active artifact with `prepare --previous-preparation FILE`; follow [the incremental preparation contract](../../../runtime/dk-pr-decisions/references/preparation.md). Preserve review budgets and exact session/coverage gates.

claude/skills/dk-pr-review/SKILL.md

@@ -8,3 +8,5 @@ Use Claude Code native agents and tools. `${CLAUDE_PLUGIN_ROOT}` is this plugin
 Before work, read [the lifecycle procedure](../../../common/lifecycle.md), then run its entry command with `--host claude`. Retain its returned token and follow its structured action. Always run its completion command with that token in a finalization step, including on failure or interruption.
 
 Read [the shared procedure](../../../runtime/dk-pr-review/procedure.md) and perform the requested workflow. Resolve paths beginning with `PLUGIN_ROOT/` against this plugin's root. Preserve the shared procedure's validation and publication guards.
+
+On preparation refresh, select the active artifact with `prepare --previous-preparation FILE`; follow [the incremental preparation contract](../../../runtime/dk-pr-decisions/references/preparation.md). Preserve review budgets and exact session/coverage gates.

codex/skills/dk-pr-decisions/SKILL.md

@@ -8,3 +8,5 @@ Use this Codex session and native subagents when available. Use the loaded skill
 Before work, read [the lifecycle procedure](../../../common/lifecycle.md), then run its entry command with `--host codex`. Retain its returned token and follow its structured action. Always run its completion command with that token in a finalization step, including on failure or interruption.
 
 Read [the shared procedure](../../../runtime/dk-pr-decisions/procedure.md) and perform the requested workflow. Resolve paths beginning with `PLUGIN_ROOT/` against this plugin's root. Preserve the shared procedure's validation and publication guards.
+
+On preparation refresh, select the active artifact with `prepare --previous-preparation FILE`; follow [the incremental preparation contract](../../../runtime/dk-pr-decisions/references/preparation.md). Preserve review budgets and exact session/coverage gates.

codex/skills/dk-pr-iter/SKILL.md

@@ -8,3 +8,5 @@ Use this Codex session and native subagents when available. Use the loaded skill
 Before work, read [the lifecycle procedure](../../../common/lifecycle.md), then run its entry command with `--host codex`. Retain its returned token and follow its structured action. Always run its completion command with that token in a finalization step, including on failure or interruption.
 
 Read [the coordinator entry](../../../runtime/dk-pr-iter/procedure.md), then read each linked stage contract before acting in that stage. Give child agents only their assigned role contracts and pinned evidence paths. Resolve paths beginning with `PLUGIN_ROOT/` against this plugin's root. Preserve the shared procedure's validation and publication guards.
+
+On preparation refresh, select the active artifact with `prepare --previous-preparation FILE`; follow [the incremental preparation contract](../../../runtime/dk-pr-decisions/references/preparation.md). Preserve review budgets and exact session/coverage gates.

codex/skills/dk-pr-review/SKILL.md

@@ -8,3 +8,5 @@ Use this Codex session and native subagents when available. Use the loaded skill
 Before work, read [the lifecycle procedure](../../../common/lifecycle.md), then run its entry command with `--host codex`. Retain its returned token and follow its structured action. Always run its completion command with that token in a finalization step, including on failure or interruption.
 
 Read [the shared procedure](../../../runtime/dk-pr-review/procedure.md) and perform the requested workflow. Resolve paths beginning with `PLUGIN_ROOT/` against this plugin's root. Preserve the shared procedure's validation and publication guards.
+
+On preparation refresh, select the active artifact with `prepare --previous-preparation FILE`; follow [the incremental preparation contract](../../../runtime/dk-pr-decisions/references/preparation.md). Preserve review budgets and exact session/coverage gates.

runtime/dk-pr-decisions/references/api.md

@@ -53,10 +53,10 @@ All former top-level command results now live under `data`; publication results
 
 ## Preparation and combined publication
 
-`prepare --input-manifest FILE --output-dir PRIVATE_DIR` delegates to
+`prepare --input-manifest FILE --output-dir PRIVATE_DIR [--previous-preparation FILE] [--full-analysis]` delegates to
 `scripts/preparation.py` without a discussion scan or any publication. It preserves
-this API v2 envelope; `data` contains `artifact`, `preparation_id` and `readiness`.
-The artifact has its own `preparation_version: 1`; read
+this API v2 envelope; `data` contains `artifact`, `preparation_id`, `readiness` and a `refresh` report.
+The artifact has its own `preparation_version: 2`; read
 [the preparation contract](preparation.md) for complete evidence, semantic analysis,
 question retention, placement, proportionate tests, and downstream gates.
 `verified: true` retains its original provenance/freshness meaning.

runtime/dk-pr-decisions/references/preparation-contract.md

@@ -0,0 +1,135 @@
+# Preparation data contract
+
+Use with [the preparation procedure](preparation.md). Return JSON only. Evidence
+references name indexed original sources, optionally with locators (`production/0#L42`).
+All strings below are nonempty unless stated otherwise; evidence arrays are nonempty.
+
+## Initial response
+
+Return `{"analysis": ANALYSIS, "graph": NODES}`. Analysis contains:
+
+| Field | Required content |
+| --- | --- |
+| `boundaries` | Nonempty conclusion arrays for `correctness`, `remediation_depth`, `performance`, `compatibility`, `complexity`. Cover supported cases/exclusions, allowed depth, relevant workloads/limits, supported interfaces/environments and justified design costs. |
+| `placement` | Conclusions for `production`, `tests`, `reuse`, `entrypoints`, `dependencies`, each with a nonempty `locations` string array. |
+| `test_strategy` | Nonempty array: string fields `behavior`, `existing`, `missing`, `level`, `location`, `command`, `rationale`, `limitation`, plus `evidence`. Use `none` when applicable. Level: `inspection`, `unit`, `integration`, `system`, `performance`. |
+| `questions` | Open prior questions and resolved contradictions, using the shape below. Empty is valid. |
+| `reviewer_addenda` | Each configured reviewer name mapped to tailored instructions. |
+| `worker_instructions` | Boundaries, selected remedy, principal code/test locations, helpers, scenarios, commands and stopping criteria. |
+| `completion_criteria` | Required behavior, placement, sufficient validation, compatibility and cumulative complexity. |
+| `sync_required`, `sync_reason` | Boolean and explanation; empty reason allowed only when false. |
+
+A conclusion has `text`, `agreement`, boolean `material`, `rationale`, `evidence`.
+Agreement: `established` (explicitly agreed), `interpretation` (grounded routine choice),
+or `proposal` (unsettled). Material proposals and open material questions produce
+`needs_decision`; required unpublished synchronization produces `sync_pending`.
+
+Questions require `id`, `status` (`open` or `resolved`), boolean `material`,
+`statements`, `affected_work`, `recommendation`, `proposed_changes`, `evidence`,
+and at least two `options`, each with `choice`, `benefits`, `costs`. A resolved
+question also needs `resolution` citing the exact answer or applicable precedence/
+supersession and reasoning. Preserve open IDs; graph updates cannot settle choices.
+
+`NODES` has exactly one node per analysis conclusion:
+
+```json
+{"location":"/boundaries/correctness/0","depends_on":[],
+ "paths":[{"path":"src","kind":"subtree"}],"sources":["production/0"],"decisions":[]}
+```
+
+Locations are JSON pointers: each boundary, placement, test, question, reviewer
+addendum, worker/completion string, and combined `/sync_required` value
+`{"sync_required":false,"sync_reason":""}`. Escape reviewer names as JSON pointers.
+Initial edges use locations; runtime assigns stable `c-…` IDs and stores nodes by ID.
+Each node needs supporting paths, sources or dependencies. Sources are input keys;
+Decisions are existing authoritative IDs. Paths are normalized repository-relative
+whole `file` or `subtree` watches; `.` permits a root subtree. Cover absent-file
+assumptions; no globs, symbol/line watches, call graphs or similarity matching.
+Runtime adds governing source watches and reviewer/worker/completion dependencies
+on governing conclusions; AI supplies other relevant edges.
+
+## Reassessment response
+
+Return exactly `dispositions`, `additions`, `impacts` arrays:
+
+```json
+{"dispositions":[{"id":"c-ID","action":"retain","reason":"Inspected evidence"}],
+ "additions":[],"impacts":[{"change":"path:new.py","affected":[],"rationale":"Inspected impact"}]}
+```
+
+- Each affected/additionally affected existing node and every dependent needs one
+  disposition: `id`, `action`, `reason`. `retain` preserves its value exactly;
+  `replace` requires `value`; `remove` retires the ID. Only replace supplies a value.
+- Optional `watches` replaces all fields: `depends_on`, `paths`, `sources`, `decisions`.
+  Existing edges use stable IDs. Dependents must explicitly rewire removed targets.
+- Additions require `location`, `value`, `watches`, `reason`. List locations append
+  with `/boundaries/DIMENSION/-`, `/test_strategy/-`, `/questions/-`; singleton slots
+  must be absent. Added governing conclusions require guidance dispositions.
+- New or updated nodes may depend on `addition:N`, the response's zero-based addition
+  index in canonical decimal without leading zeros. Forward edges and cycles work;
+  runtime preallocates stable IDs and resolves references before validation. Stored
+  IDs survive replacement; retired IDs cannot return. Runtime rebuilds list indexes.
+- Each unmapped change requires `change`, `affected` existing IDs (possibly empty),
+  `rationale`. Repeated `affected` IDs are coalesced. Missing mappings never prove
+  irrelevance. Invalid or dangling references, repeated `impacts[].change` or
+  `dispositions[].id` entries, ID collisions, missing dispositions or lost open
+  questions block.
+
+If impact cannot be bounded, return only
+`{"full_reassessment":true,"reason":"Why impact cannot be bounded"}`.
+The one full fallback uses the same disposition protocol for all existing nodes;
+a full phase cannot escalate again. Explanations belong in the refresh report.
+
+## Evidence normalization
+
+Repository sources match relative paths; known GitHub records match stable IDs.
+Opaque inputs keep exact byte hashes and path identities. A `complete` field in
+application data declares no fetch completeness. Supported collectors enforce theirs.
+Unknown fields, source additions/removals, authorship, text, nested order and precedence
+remain significant. Exact supported exceptions:
+
+- Context transport paths, head and local-checkout status are operational; metadata
+  ignores `headRefOid`, Decisions ignore only record `head_sha` (`data.head_sha` in
+  an envelope), patches derive from Git. All other Decisions envelope fields govern.
+  Manifest locations are transport, with each source compared separately. Reviewer
+  descriptors compare structurally; verified prior preparation supplies lineage.
+- Complete inline feedback uses integer `schemaVersion:1`, `threads`/`items` and
+  validated `threadsSha256`/`itemsSha256`. Conversation has no version marker,
+  `reviews`/`comments`/`items` and validated `sha256`. Only known collection order,
+  JSON formatting and supported thread `isResolved` are operational. Unknown fields
+  retain content/order, including other schemas' collection/hash names. All original
+  text records remain required; new, edited, removed, bot or self-authored text governs.
+- Integer `preparation_history_version:1` requires `governing` and `operational`
+  objects. Integer `evidence_version:1` requires `governing`, permits `operational`
+  and supports stable `source_id` across relocation. Other marker types stay opaque.
+  Keep choices, questions, constraints, unanswered requests and supersession under
+  `governing`; unknown fields govern. Excluded operational fields are nonnegative
+  integer `cycles`/`cycle_retries`/`repairs`/`reruns`, string `head_sha`/`run_id`/
+  `checked_at`, known CI `status`, and `publication_receipts` arrays with string
+  `url`/`head_sha`/`sha256`/`outcome`/`timestamp`/`artifact`/`kind` fields only.
+  Unknown receipt fields keep the collection governing. Known statuses: `pending`,
+  `queued`, `in_progress`, `success`, `failure`, `neutral`, `cancelled`, `skipped`,
+  `timed_out`. Operational state never changes Decisions or review-session identity.
+
+## Stored artifact and verification
+
+Immutable `preparation_version:2` artifacts retain `analysis`, graph/retired IDs,
+input fingerprints, reviewer descriptors, readiness, PR/base/head/body/Decisions
+identity, parent path/ID/hash and migration provenance. Snapshots are fingerprinted
+and read-only outside the repo. Historical validation checks retained integrity and
+graph structure; current gates also check live inputs and current mandatory rules.
+Migration preserves original v1 storage and grants no coverage or cycle budget.
+
+Artifact and API `refresh` report mode, changed paths/sources, unmapped changes,
+binding/identity changes, affected/reused conclusions, explanations and actual
+`model_call_count`, including corrections. Dependency metadata and refresh explanations
+remain outside binding analysis. Tree-equivalent review reuse preserves original
+provenance only when existing scope/rules/Decisions/instruction/base checks pass.
+
+Completion verification has `verification_version:1`, current `preparation_id`,
+`pr_url`, `head_sha`, and `checks` for `behavior`, `placement`, `test_sufficiency`,
+`compatibility`, `complexity`. Each check contains boolean `satisfied`, `rationale`,
+and nonempty `evidence`: `{"path":"ABSOLUTE_PATH","sha256":"FILE_HASH"}`.
+Inspection reports, code and test results qualify; missing, false or stale evidence
+blocks completion. Both stages use the same current preparation, verification and
+mutable run record; an immutable history snapshot cannot replace live choices.