sungjunlee · sungjunlee · Aug 8, 2026 · Aug 8, 2026 · Aug 8, 2026 · Aug 8, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -8,6 +8,8 @@ Each entry links the GitHub issue (the canonical spec) and the merge PR (the shi
 
 ### Removed
 
+- **backlog-triage judgment heuristics** — replaced ~265 lines of semantic judgment (`scanPhraseEdges`/`scanBlocks`/`scanDependsOn`, title-Jaccard duplicates in relate/stale, and rule-based `buildPriorityActions`/`buildMilestoneActions`) with prompt-level rubric guidance. Scripts now emit deterministic signals only (mentions, merged-PR links, dates, labels); the model's judged blocks/depends-on/duplicate edges and priority/milestone proposals enter the report through the new `triage-report.js --model-actions PATH`, keeping anchors, dedupe, and the Apply Checklist deterministic. Net −491 script/test lines. Closes [#358](https://github.com/sungjunlee/dev-backlog/issues/358).
+
 - **Monthly Progress sync and sprint issue publication** — removed 3,758 lines of scripts and dedicated tests after their intended adoption windows finished. Only one monthly issue was ever created (`Progress: April 2026`, #46; none in May–July), and the four sprint issues (#230, #234, #237, #239) were all created on 2026-07-03/04 with none since. No core lifecycle script or dev-relay integration invokes either feature, and `task-progress-reporting` accumulated no Learnings. Closes [#340](https://github.com/sungjunlee/dev-backlog/issues/340).
 - **Required task mirrors** — the GitHub-native core now resolves task specification, AC, and lifecycle directly from live Issues. Fresh setup and complete sprint execution require no `backlog/tasks/` or `backlog/completed/`; `sync-pull --legacy-export` remains an explicit one-way diagnostic/rollback boundary. Closes [#347](https://github.com/sungjunlee/dev-backlog/issues/347).
 - **Zero-adopter local tracker and generic compatibility machinery** — measured adoption found 0 of 17 consumers selecting a non-default tracker and GitHub remotes in all 18 known consumers. The local JSON store, local lifecycle tests, and generic/local design surface are removed; GitHub failure is fail-loud with no fallback, while Backlog.md remains manual import/explicit export compatibility only. Closes [#348](https://github.com/sungjunlee/dev-backlog/issues/348).

diff --git a/README.md b/README.md
@@ -310,7 +310,8 @@ SNAP=backlog/triage/.cache/<ts>.json
 node $SKILL/triage-collect.js
 node $SKILL/triage-relate.js --snapshot $SNAP --json > /tmp/relate.json
 node $SKILL/triage-stale.js  --snapshot $SNAP --json > /tmp/stale.json
-node $SKILL/triage-report.js --snapshot $SNAP --relate /tmp/relate.json --stale /tmp/stale.json
+# model-authored: blocks/depends-on/duplicate edges + priority/milestone actions
+node $SKILL/triage-report.js --snapshot $SNAP --relate /tmp/relate.json --stale /tmp/stale.json --model-actions /tmp/model-actions.json
 
 # Apply phase (opt-in): review the report, check accepted proposals, then
 node $SKILL/triage-apply.js backlog/triage/<date>-report.md --apply

diff --git a/backlog/triage-config.yml b/backlog/triage-config.yml
@@ -5,4 +5,3 @@ activity_days:
   warm: 14
   cold: 60
 stale_days: 60
-duplicate_threshold: 0.75
diff --git a/skills/backlog-triage/SKILL.md b/skills/backlog-triage/SKILL.md
@@ -25,8 +25,8 @@ Phase 1 — Report (default, read-only)     Phase 2 — Apply (explicit mutation
 | Phase | Step | Completion boundary |
 | --- | --- | --- |
 | Report | Collect | One `gh` fetch writes a snapshot JSON; downstream steps use `--snapshot PATH` and do not re-fetch. |
-| Report | Analyze | Classification, relationships, stale/obsolete signals, Alignment, and Decision Review are computed from the same snapshot/spec evidence. |
-| Report | Render | One markdown report is written with anchored proposals and a consolidated Apply Checklist. |
+| Report | Analyze | Classification and deterministic signals come from scripts; the model judges blocks/depends-on/duplicates, priority, and milestone actions from the snapshot and writes them to a `--model-actions` JSON file. Alignment and Decision Review are prompt-driven from the same evidence. |
+| Report | Render | `triage-report.js` validates the model actions, merges them with deterministic signals, and writes one markdown report with anchored proposals and a consolidated Apply Checklist. |
 | Apply | Review | A human accepts proposals by flipping paired checkboxes from `[ ]` to `[x]`; unchecked anchors remain inert. |
 | Apply | Dry-run | `triage-apply.js <report.md>` prints intended `gh` mutations without writing. |
 | Apply | Mutate | `triage-apply.js <report.md> --apply` executes only accepted actions; `--yes` is required for non-interactive apply. |
@@ -68,14 +68,16 @@ The report is a derived artifact under `backlog/triage/`. GitHub Issues remain t
 Required sections:
 
 - `## Classification` — issue buckets by theme, label, age, activity, and milestone state.
-- `## Relationships` — mentions, blocks, depends-on, duplicate candidates, and merged closing PR links.
+- `## Relationships` — deterministic mentions/comment-mentions/merged-PR links from scripts, plus model-judged blocks, depends-on, and duplicate-candidate edges passed via `--model-actions`.
 - `## Obsolete Candidates` — anchored close/revisit proposals with evidence.
-- `## Priority Proposals` — anchored priority proposals with rationale.
-- `## Milestone Suggestions` — anchored milestone proposals grouped into candidate sprint clusters.
+- `## Priority Proposals` — model-judged anchored priority proposals with rationale (delivered via `--model-actions`).
+- `## Milestone Suggestions` — model-judged anchored milestone proposals grouped into candidate sprint clusters (delivered via `--model-actions`).
 - `## Alignment` — objective coverage, orphan work, neglected objectives, contradictions, and proposed charter changes; when no charter exists, record that alignment was skipped.
 - `## Decision Review` — `Do Now`, `Shape First`, `Defer`, and `Drop / Close`.
 - `## Apply Checklist` — consolidated review surface for every anchored action.
 
+Judgment distribution: scripts own deterministic signals (issue refs, merged-PR links, dates, labels); the model owns semantic judgment (blocks, depends-on, duplicates, priority and milestone proposals). The model's judgment is rendered by `triage-report.js` from a `--model-actions` JSON file, so anchor formatting, dedupe, and the Apply Checklist stay deterministic.
+
 Full section examples and rubric details live in `references/classification.md`, `references/relationships.md`, `references/stale.md`, `references/decision-review.md`, and `references/apply.md`.
 
 ## Relationship To dev-backlog
@@ -105,19 +107,55 @@ node "$skill_dir/scripts/triage-collect.js" --dry-run --json
 node "$skill_dir/scripts/triage-apply.js" backlog/triage/YYYY-MM-DD-report.md
 ```
 
+Model-judged actions (`--model-actions`) are a JSON array of action objects. Sections `priority` / `milestone` / `obsolete` carry a positive `issueNumber`, `verb`, `summary`, and `args` with the mutation payload; section `relationship` carries `args.from` / `args.to` / `args.kind` for a model-judged edge. `triage-report.js` validates every entry before rendering (see `references/apply.md` for the anchor grammar each verb maps to):
+
+```json
+[
+  {
+    "section": "priority",
+    "verb": "set-priority",
+    "issueNumber": 42,
+    "args": { "value": "high", "reason": "customer-reported outage blocks the auth theme" },
+    "summary": "Set priority:high on #42 — customer-reported outage blocks the auth theme"
+  },
+  {
+    "section": "milestone",
+    "verb": "assign-milestone",
+    "issueNumber": 43,
+    "args": { "name": "Sprint W34" },
+    "cluster": "auth",
+    "sprintName": "Sprint W34",
+    "summary": "Assign Sprint W34 to #43 — auth cluster"
+  },
+  {
+    "section": "obsolete",
+    "verb": "close-duplicate",
+    "issueNumber": 44,
+    "args": { "target": "#12", "reason": "open issue duplicates closed #12" },
+    "summary": "Close duplicate #44 into #12 — open issue duplicates closed #12"
+  },
+  {
+    "section": "relationship",
+    "verb": "edge",
+    "args": { "from": 45, "to": 46, "kind": "blocks", "evidence": { "phrase": "Blocks #46" } },
+    "summary": "Blocks edge 45 -> 46"
+  }
+]
+```
+
 Useful scripts:
 
-- `scripts/triage-collect.js [--repo OWNER/REPO] [--limit N] [--json] [--dry-run] [--with-comments] [--with-closed-issues]` — fetch open issues and write `backlog/triage/.cache/<ISO-timestamp>.json`; `--with-comments` enables comment-mention edges and `--with-closed-issues` enables duplicate-of-closed signals, see `references/classification.md`.
-- `scripts/triage-relate.js --snapshot PATH [--json]` — detect mentions, blocks, depends-on, duplicates, and merged PR links.
-- `scripts/triage-stale.js --snapshot PATH [--since N] [--json]` — flag stale/obsolete candidates with evidence.
-- `scripts/triage-report.js --snapshot PATH [--relate PATH] [--stale PATH] [--active-sprint PATH] [--out PATH] [--json]` — render report; creates `.bak` on overwrite.
+- `scripts/triage-collect.js [--repo OWNER/REPO] [--limit N] [--json] [--dry-run] [--with-comments] [--with-closed-issues]` — fetch open issues and write `backlog/triage/.cache/<ISO-timestamp>.json`; `--with-comments` hydrates comment bodies and `--with-closed-issues` enriches the snapshot with recent closed issues for the model's duplicate judgment, see `references/classification.md`.
+- `scripts/triage-relate.js --snapshot PATH [--json]` — deterministic edges: mentions, comment-mentions, and merged closing PR links.
+- `scripts/triage-stale.js --snapshot PATH [--since N] [--json]` — deterministic stale/obsolete candidates from dates and labels.
+- `scripts/triage-report.js --snapshot PATH [--relate PATH] [--stale PATH] [--active-sprint PATH] [--model-actions PATH] [--out PATH] [--json]` — render report; model-judged actions (blocks/depends-on/duplicate edges, priority and milestone proposals) come from `--model-actions` JSON; creates `.bak` on overwrite.
 - `scripts/triage-apply.js <report.md> [--apply] [--yes] [--json]` — parse accepted anchors and execute/dry-run GitHub mutations.
 - `scripts/triage-apply.integration.test.js` — opt-in live integration test against the disposable sandbox repo; requires `TRIAGE_APPLY_INTEGRATION=1` and `GH_TOKEN`.
 
 ## References
 
 - `references/classification.md` — bucketing rules and YAML config schema.
-- `references/relationships.md` — relationship heuristics and evidence format.
+- `references/relationships.md` — deterministic edge rules and the model-judged blocks/depends-on/duplicate rubric.
 - `references/stale.md` — obsolescence signals, thresholds, and suggested-action grammar.
 - `references/apply.md` — anchor grammar, parse rules, idempotency contract, and apply-log schema.
 - `references/decision-review.md` — prompt-driven Do Now / Shape First / Defer / Drop rubric.

diff --git a/skills/backlog-triage/references/classification.md b/skills/backlog-triage/references/classification.md
@@ -15,7 +15,6 @@ comment_fetch_concurrency: 5
 closed_issue_days: 180
 closed_issue_limit: 200
 stale_days: 60
-duplicate_threshold: 0.75
 ```
 
 - `theme_keywords` maps a theme name to title-keyword substrings. The first matching theme wins.
@@ -24,7 +23,7 @@ duplicate_threshold: 0.75
 - `comment_fetch_concurrency` bounds `--with-comments` fan-out when comment hydration is enabled.
 - `closed_issue_days` bounds the lookback window for `--with-closed-issues`.
 - `closed_issue_limit` caps how many recent closed issues are collected for snapshot v2 enrichment.
-- `stale_days` and `duplicate_threshold` are collected as config-as-data for downstream scripts (`triage-stale`, `triage-relate`); `triage-collect` does not apply them yet.
+- `stale_days` is collected as config-as-data for `triage-stale`; `triage-collect` does not apply it yet.
 
 ## Snapshot shape
 

diff --git a/skills/backlog-triage/references/decision-review.md b/skills/backlog-triage/references/decision-review.md
@@ -95,3 +95,28 @@ Evidence absent: `spec/system-map.md`.
 ```
 
 Do not include anchors for non-mutating recommendations. Only include close/relabel/milestone anchors when the normal apply contract supports the action and a human can accept it with a checkbox.
+
+## Priority and Milestone Proposals
+
+Priority Proposals and Milestone Suggestions are also prompt-driven and delivered to `triage-report.js --model-actions`. The old script heuristics become judgment guidance:
+
+### Priority Proposals (`section: "priority"`, `verb: "set-priority"`)
+
+Suggest `priority:high` only for issues that are all of:
+
+- not already `priority:high` or `priority:critical`
+- not cold (recent activity or warm bucket)
+- not already proposed for close
+- high-leverage by at least one of: sits in a theme with several recent/warm issues, participates in relationship edges, or is called out by an active sprint or Decision Review
+
+Rationale must name the trigger (theme heat, relationship edges, sprint focus), not just restate the issue title.
+
+### Milestone Suggestions (`section: "milestone"`, `verb: "assign-milestone"`)
+
+Suggest a milestone only for issues that:
+
+- have no milestone yet
+- are not already proposed for close
+- are active or related (relationship edges, active theme, or a priority proposal above)
+
+Set `args.name` to a concrete candidate sprint (`Sprint W<next>`), `args.cluster` to the dominant theme, and put the same values in top-level `sprintName` / `cluster` for grouping. Only suggest a milestone when the issue is genuinely ready for sprint planning; otherwise leave it unassigned.
diff --git a/skills/backlog-triage/references/relationships.md b/skills/backlog-triage/references/relationships.md
@@ -1,18 +1,17 @@
 # Relationships
 
-**Purpose.** `triage-relate.js` reads a previously collected issue snapshot and emits read-only relationship edges for snapshot-resident signals:
+**Purpose.** The `## Relationships` section of the triage report maps how open issues connect to each other and to merged closing PRs. Deterministic signals come from `triage-relate.js`; semantic judgment (blocking, dependency, duplication) is the model's job and is rendered from the model's own edge JSON.
+
+## Script-generated Edges (deterministic)
+
+`triage-relate.js` reads a previously collected issue snapshot and emits read-only edges for snapshot-resident signals that need no interpretation:
 
 - `mentions` from plain `#123` references in issue bodies
 - `comment-mentions` from plain `#123` references in optional issue comments
-- `blocks` from explicit blocking / closing phrases in issue bodies
-- `depends-on` from explicit dependency phrases in issue bodies
 - `merged-pr-link` from per-issue merged closing PR metadata
-- `duplicate-candidate` from title-token Jaccard overlap
 
 Every emitted edge carries evidence taken directly from the snapshot so downstream report rendering can show why the relationship was inferred without re-fetching from GitHub.
 
-## Implemented Heuristics
-
 ### `mentions`
 
 - Source: `issue.body`
@@ -42,29 +41,6 @@ Every emitted edge carries evidence taken directly from the snapshot so downstre
   - `match`: matched issue reference
   - `snippet`: normalized sentence/line fragment containing the match
 
-### `blocks`
-
-- Source: `issue.body`
-- Keywords used by `scanBlocks`:
-  - `blocks #123`
-  - `closes #123`
-- Confidence: `1`
-- Evidence:
-  - `phrase`: normalized matched phrase, for example `Blocks #123`
-  - `snippet`: normalized sentence/line fragment containing the phrase
-
-### `depends-on`
-
-- Source: `issue.body`
-- Keywords used by `scanDependsOn`:
-  - `blocked by #123`
-  - `depends on #123`
-  - `depends-on #123`
-- Confidence: `1`
-- Evidence:
-  - `phrase`: normalized matched phrase, for example `depends on #123`
-  - `snippet`: normalized sentence/line fragment containing the phrase
-
 ### `merged-pr-link`
 
 - Source: `issue.closing_prs`
@@ -79,31 +55,55 @@ Every emitted edge carries evidence taken directly from the snapshot so downstre
   - `pr.mergedAt`: merge timestamp
   - `pr.url`: closing PR URL when present
 
-### `duplicate-candidate`
+## Model-judged Edges (semantic)
+
+Blocking, dependency, and duplication require reading issue intent, so they are judged by the model reading the snapshot, not by phrase matching or title token overlap. The model emits them as `section: "relationship"` entries in a `--model-actions` JSON file; `triage-report.js` validates each entry and merges it into the Relationships path. Wire shape:
 
-- Source: `issue.title`
-- Threshold: `backlog/triage-config.yml -> duplicate_threshold`
-- Confidence: Jaccard similarity score
-- Canonicalization:
-  - compare each issue pair once
-  - emit a single edge with the smaller issue number as `from`
+```json
+{
+  "section": "relationship",
+  "verb": "edge",
+  "args": {
+    "from": 100,
+    "to": 101,
+    "kind": "blocks",
+    "confidence": 1,
+    "evidence": { "phrase": "Blocks #101" }
+  },
+  "summary": "Blocks edge 100 -> 101"
+}
+```
+
+`args.kind` must be one of the script's deterministic kinds plus the semantic kinds below; `triage-report.js` rejects unknown kinds.
+
+### `blocks`
+
+- Read `issue.body` and comment bodies for explicit statements that issue X blocks issue Y, or closes it once completed.
+- Only emit for issues that exist in the snapshot; never for `#999`-style dangling references.
+- Confidence: `1` when the phrasing is explicit, lower when inferred.
 - Evidence:
-  - `score`: rounded Jaccard score (`4` decimal places)
-  - `overlap`: sorted shared title tokens
-  - `titles.from`: lower-numbered issue title
-  - `titles.to`: higher-numbered issue title
+  - `phrase`: normalized matched phrase, for example `Blocks #123`
+  - `snippet`: normalized sentence/line fragment containing the phrase
 
-## Jaccard Tokenization Rules
+### `depends-on`
+
+- Read `issue.body` and comment bodies for explicit dependency statements: `blocked by #123`, `depends on #123`, `depends-on #123`, or equivalent intent.
+- Only emit for issues that exist in the snapshot.
+- Confidence: `1` when explicit, lower when inferred.
+- Evidence:
+  - `phrase`: normalized matched phrase, for example `depends on #123`
+  - `snippet`: normalized sentence/line fragment containing the phrase
 
-Title similarity uses the following normalization before scoring:
+### `duplicate-candidate`
 
-- lowercase the title
-- extract tokens with regex `[a-z0-9]+`
-- drop one-character tokens
-- deduplicate tokens per title by converting to a set
-- compute `overlap / union`
+- Compare open issues against each other and against closed issues, judging semantic duplication from titles, bodies, labels, and comments — not title-token overlap alone.
+- Emit one canonical edge with the smaller issue number as `from`.
+- `args.evidence`:
+  - `reason`: short human-readable why this is a duplicate candidate
+  - `titles.from`: lower-numbered issue title
+  - `titles.to`: higher-numbered issue title
 
-If the union is empty, the score is `0` and no edge is emitted.
+A duplicate of a closed issue should be proposed as an Obsolete Candidate (`section: "obsolete"`, `verb: "close-duplicate"`), not only as a relationship edge — see `references/stale.md`.
 
 ## Evidence Schema
 
@@ -169,13 +169,12 @@ Evidence payloads vary by kind:
 
 ```json
 {
-  "score": 0.8,
-  "overlap": ["flow", "oauth", "refresh", "token"],
+  "reason": "same OAuth refresh flow as #200 with no substantive delta",
   "titles": {
     "from": "OAuth token refresh flow",
     "to": "OAuth token refresh flow redesign"
   }
 }
 ```
 
-`triage-relate.js` is intentionally still read-only. Close or duplicate proposals belong to `triage-stale.js` and still require report review plus an accepted apply checkbox before any GitHub mutation.
+`triage-relate.js` is intentionally read-only. Close or duplicate proposals still require report review plus an accepted apply checkbox before any GitHub mutation.