Arguments. <fetch|diff|status|apply> [--ci] [vA..vB|vX|text]. Full form: [--ci] [vA..vB|vX|text]. Actions: fetch (default on passive mention), diff, status, apply (explicit only). --ci applies to apply only
Pre-computed context
- Read marker (no fetch): !
bash "${CLAUDE_PLUGIN_ROOT}/skills/changelog/scripts/changelog-status.sh" --no-fetch 2>/dev/null | head -4 || echo "(status script unavailable)"
Variables
Arguments: $ARGUMENTS
Scope
Ingests Claude Code changelog entries and hands their decisions to the repo's stage skills. Covers the arc: read upstream changes → orient on repo impact → research new features → scope with the user → run what fits the session through the marketplace stage skills, one PR per owner plugin → file the rest as work items → native drift. diff reports decisions about components, not a list of items.
Distinct from:
/harness-ops:known-issues. Tracks CC bugs/workarounds. This skill integrates CC feature changes into repo config/docs
- Any release-triage automation the consumer runs (issue filing per release). This skill decides, holistically across a release, what changes in the repo, and hands each decision to the skill that plans, implements or verifies it; it re-implements none of them
Input modes and range
Three ways to provide changelog content (priority order):
- User pastes text. Skill parses inline changelog from conversation context; the releases the pasted text names are the range, so the cap and the version check apply to them and not to the repository's default feed
- Explicit range or version.
/harness-ops:changelog diff v2.1.257..v2.1.263 covers both ends inclusive; apply v2.1.263 covers that one release
- Default range.
/harness-ops:changelog diff (no range) runs from the read marker to the newest published release
The read marker, the default range, and the replay cap are defined in
context/read-actions.md and computed by
scripts/changelog-status.sh; every action starts by running it.
Version awareness
On every apply or diff invocation, compare the newest release in the range against the active terminal's CC version. The status script reports both as latest and installed and emits a warn line when the installed version is older:
- If the newest release in the range > installed version: warn user. "You're applying v2.1.263 changes but running v2.1.260. Update CC first (
claude update) or changes may reference features not yet available in your session."
- If the newest release in the range = installed version: proceed normally
- If the newest release in the range < installed version: fine. Catching up on older releases
Read marker and replay cap
One line in the repository's upstream ledger for Claude Code releases (default
docs/upstream/claude-code.md, override HARNESS_OPS_CHANGELOG_LEDGER) records the newest release
the repository has been read against. status reads that line; with no ledger it falls back to the
highest version named in a Conventional Commits SUBJECT of the form
chore(<scope>): address Claude Code v<A>..<B> changelog, and it never reads commit bodies, because
a body's "verified against Claude Code v" is a doc's recency stamp, not an apply.
A range wider than ten releases or 300 core items exceeds the replay cap. diff and apply then
stop and recommend a docs-conformance recheck of the components against the current docs, followed
by a marker set at the newest published release: the docs carry the cumulative state, and replaying
items past the cap costs more than it returns.
Action Router
Parse $ARGUMENTS to extract the action (first token) and remaining arguments.
| Action |
Description |
Detail |
apply |
Hand-off: ingest → explore → research → scope gate → stage skills per owner-plugin PR, larger decisions filed as issues → native drift. Consumes the diff working set |
See "Action: apply" below |
fetch |
Fetch + display changelog for a version or range. Read-only |
See "Action: fetch" below |
diff |
Resolve the range, apply the cap, emit decision rows by owner surface plus a docs-lag section. Read-only |
See "Action: diff" below |
status |
The read marker and its source, installed vs newest release, the default range, the cap verdict |
See "Action: status" below |
help |
Show action table |
(inline) |
Routing (model-invocable):
- Empty args or passive CC version mention →
fetch or diff (read-only). Never apply.
- Version-only token (
v2.1.152) or range-only token (v2.1.150..v2.1.152) without explicit apply intent → fetch for that version or range.
apply only when user explicitly requests integration (apply, apply changelog, /harness-ops:changelog apply, or unambiguous implement-this-release intent).
If action is unknown, show action table.
Action: apply. User intent gate
apply mutates the repo. Run only on explicit user intent per routing above. When the model
detects a new CC release in conversation, default to fetch or diff and offer apply. Do not
auto-start the hand-off.
apply plans, edits and verifies nothing itself. It reads the release, asks the user which decisions
are in scope, and hands each one to the marketplace skill that owns that stage, as Phases 3 and 4 set out.
Unattended: apply --ci
Only the literal --ci token selects this path, and it needs an explicit release or range. It
replaces every question in Phases 3 to 5 with the evidence rule in
context/ci-mode.md: only a verified, cited correct row confined to one owner
plugin and not security-relevant is committed, on a local branch per owner plugin; every other row
becomes a handoff. It asks nothing, pushes nothing, opens no PR, files nothing and never moves the
read marker. It writes ci-plan.json for the calling workflow, rewriting it as each step finishes so a
stopped run still publishes what it finished, and ends with ci-plan: <path> and
status: <stopped|planned>. Where the phases below say to ask, confirm or file, --ci follows that
file instead.
Phase 0. Ingest
Resolve the range, check the cap, and check version alignment:
- Resolve the range (first match wins):
- Changelog text already in conversation → the releases its release markers or version headings name ARE the range; pass them as
--range <lowest>..<highest> so the cap is judged on the pasted releases and never on the repository's default feed. Pasted text with no version at all skips the cap
- A range or version was given →
--range as given
- Otherwise → no
--range; the default range from the read marker applies
- Run the status script with that
--range. If cap reads exceeded, stop and relay the recommend line; the hand-off does not run past the cap. Relay any warn line per "Version awareness" above
- Reuse the working set when
diff saved one for this range and it passes the checks under "Persistence" in context/decisions.md: apply consumes it, resolves no content, and re-fetches only what a recheck trigger names. Otherwise continue
- Resolve content: pasted text is parsed as-is; otherwise slice the releases the
releases line names out of a local copy of the changelog per the fetch route in context/read-actions.md
- Parse into structured items. Each item gets: a stable id (
2.1.257-001), summary, category (feature / fix / UI / internal), affected surface (if identifiable). Ids and the working-set directory are defined in context/decisions.md
Phase 1. Explore
Orient on repo impact for EACH changelog item. Run bash "${CLAUDE_SKILL_DIR}/scripts/discover-surfaces.sh" first: its output is the surface list, and context/repo-surfaces.md gives per-class examples and scoped grep patterns:
- Grep/Glob each feature name, setting name, hook event, CLI flag across ALL surface classes the script printed
- Classify each item per
context/classification-rubric.md into one action lens per owner surface it touches: correct, replace, adopt, note, or skip
- Group by owner surface and write each correct, replace and adopt row with its required sentence. A skip item leaves no row
Output: decision rows grouped by owner surface, per context/decisions.md. The
work fans out by release cluster with no Workflow script; see "Fan-out shape" there.
Phase 2. Research
For rows needing enrichment (correct rows with behavioral changes, adopt rows with unclear scope):
Spawn parallel research subagents. One per feature cluster, each receiving the explorers' questions (use a Claude Code documentation-focused agent type when available)
Instruct each subagent to ground every claim in the page itself (a docs page through bash "${CLAUDE_PLUGIN_ROOT}/scripts/fetch-docs.sh" --cache --out <dir> <slug>, a changelog entry or GitHub issue by curl) and to return citations with each claim. Treat any uncited subagent claim as unverified and re-verify it against official docs before acting on it. A local verifier then checks the repo-side facts. Before dispatching it, run bash "${CLAUDE_PLUGIN_ROOT}/lib/read-verification.sh" --plugin harness-ops --skill changelog and copy its line into the output (levels). Under light, and off, which acts as light here, dispatch no local verifier, here or in apply's re-check of saved rows, and never check those facts yourself in its place: mark each row it would have checked unverified and write verification: skipped (<level>)
Research targets per item type:
- New frontmatter field → exact syntax, interaction with existing fields, docs gap
- New hook event → schema, sync/async, input/output shape
- New CLI flag → syntax, settings.json equivalent (or lack thereof), valid values
- Behavioral change → before/after, migration path, breaking implications
- Bug fix → what was broken, what surfaces affected, historical data impact
Synthesize research into enriched rows
Phase 3. Scope gate
Present the decision rows to the user via AskUserQuestion or structured markdown:
| # | Owner surface | Lens | Items | Required sentence | Action needed |
|---|---------------|------|-------|-------------------|---------------|
| 1 | <component> | correct | <ids> | <false vs true> | <specific update> |
| 2 | <component> | adopt | <ids> | <problem solved> | <adopt, decline, or defer pending probe> |
The user picks which rows are in scope: "all rows", "just correct", or specific rows by number. A
row left out is reported at the end and recorded only when the user declines it, with its reopen
condition in the ledger's Declined table.
Then sort each in-scope row by whether it fits this session. A row fits when its edits stay in
its owner plugin and one session can plan, make and verify them. A row is too large when it
redesigns a component, changes behavior across several plugins, or needs work this session cannot
finish and verify. Show the sort with the rows and let the user move a row across.
Phase 4. Hand off
- Rows that fit: work them in this session's branch, one branch and PR per owner plugin. Use
/planning:plan for the edit plan (if the planning plugin is installed), /implementation:implement
for the edits (if the implementation plugin is installed) and /verification:confirm for the
check (if the verification plugin is installed). A stage whose plugin is missing is not done by
hand here: report the rows that needed it as not executed, naming the missing plugin. A
replace row is never edited: nominate it into /harness-ops:audit-native-overlap
- Rows too large: file one issue per row through
/work-items:track (if the work-items plugin
is installed), carrying the owner surface, lens, required sentence, item ids and range. That
issue is worked later in its own PR, behind an interview-style human gate before any edit. When
/work-items:track is not installed, report the row and file nothing. List the rows to file and
file them only after the user confirms that batch; a row the user does not confirm is reported, not
filed
- Docs lag: hand the docs-lag pairs to
/harness-ops:known-issues. They are never a decision row
- Commit and PR shape: one PR per owner plugin, each carrying that plugin's
CHANGELOG.md
entry. The upstream ledger update is the last PR and references the others. Every commit subject
reads chore(<plugin>): address Claude Code v<A>..<B> changelog, the ledger's taking the
ledger owner's scope. The ledger PR moves the read marker only once every row in the range is
applied in a merged PR, nominated, recorded as declined or deferred, or filed, and stops below the first
release that still has a row outside those states, so no unfinished row drops out of the next
default range. status reports it from the ledger and, until the ledger exists, from that subject
Phase 5. Native-surface drift
Re-read what Claude Code ships and file what moved. Run the inventory self-check, a full
--binary-only --docs extraction, and overlap detect and self-check, then
scripts/native_drift.py to diff against the previous run's summary and evaluate the overlap
store's recheck triggers. Report surface changes (added, removed, renamed, reclassified),
invocability and marker changes, docs cross-check changes, and new overlap candidates. File one
work item per new candidate, fired trigger, and degraded or broken self-check through
/work-items:track, deduped by a native-drift:<kind>:<surface>:<component> key; a self-check
degraded only by a CLI version past the validated build proposes revalidation instead. A repository
with no overlap store is report-only: report, file nothing, keep the baseline. Commands,
the summary's location, items, dedupe and approval: context/native-drift.md.
Its commands write this skill's directory as <skill-dir>, which is ${CLAUDE_SKILL_DIR}; put that
path in place of the placeholder before running one.
When the harness-config plugin is installed, also run its effort-pin drift check directly, with
that plugin's install directory in place of the placeholder:
bash <harness-config plugin root>/skills/audit/scripts/check-effort-pins.sh
Report its lines with the native-drift report. Exit 1 means model-config's effort tables or
per-model defaults moved, or a pin names a level the page does not list: each flagged pin waits on
a person to re-decide it and rebaseline, as the /harness-config:audit effort-pins scope
documents. Exit 3 means the page was unread or reshaped, so the check made no claim; say so.
Without the harness-config plugin, report that the effort-pin check was not run.
Then list the behavior probes the range touches, against the same local copy of the changelog:
python3 "${CLAUDE_PLUGIN_ROOT}/skills/behavior-probes/scripts/probe.py" recheck --changelog <changelog.md> --range <A..B>
Each line names a case, the changelog item id and the tag that matched. Report them with the drift
report. Rerunning them live costs model runs, so offer /harness-ops:behavior-probes live for those
cases and run nothing until the user agrees.
End the run with a report that leads with what waits on the user (rows left out of scope, rows
filed or reported unfiled, stages skipped for a missing plugin), then the PRs opened and what each
changed, what /verification:confirm showed, the docs-lag pairs handed off, and the Phase 5 drift
report with the items filed or skipped.
Actions: fetch, diff, status (read-only)
The three read-only actions stop short of any edit. Full steps in context/read-actions.md.
Its fetch command writes this skill's directory as <skill-dir>, which is ${CLAUDE_SKILL_DIR}; put
that path in place of the placeholder before running it.
fetch. Read the raw changelog through the plugin's fetch-changelog.sh, which picks the source and writes the .md to a file; slice releases locally with the plugin's changelog-releases.sh --blocks) and display a version, a range, or the newest release. No edits
diff. Run the status script; stop at an exceeded cap with its recommendation; otherwise Phase 0 (ingest) + Phase 1 (explore) + Phase 2 (research) over the releases in range, stopping before the scope gate. Emits decision rows grouped by owner surface, each with its lens and required sentence, plus a docs-lag section, and saves its working set for apply. Answers "is this range worth an apply?"
status. Run the status script and relay: the read marker and its source (ledger line or commit subject, never a commit body), installed vs newest release, the default range, and the cap verdict with its recommendation
Next
- Candidates and fired triggers filed by Phase 5:
/harness-ops:audit-native-overlap.
- Items filed as raw intake, and rows filed by Phase 4:
/work-items:triage.
- A revalidation item:
/harness-ops:inventory.
- Behavior probes the range touches:
/harness-ops:behavior-probes live.
Reference index. Load on demand
| File |
Load when |
context/read-actions.md |
Running fetch, diff, or status; the read marker, range, cap, and fetch route are defined there. |
scripts/changelog-status.sh |
Every action's first step; --help lists its output lines and flags. Covered by scripts/changelog-status.test.sh. |
context/decisions.md |
Writing or reading decision rows, choosing where a decision is recorded (plugin CHANGELOG, audit-native-overlap nomination, ledger, filed issue), fanning out, or saving and reusing the working set. |
context/repo-surfaces.md |
Phase 1 explore, after running scripts/discover-surfaces.sh: per-class examples of what an item changes, and scoped grep patterns. |
context/classification-rubric.md |
Assigning a lens (correct, replace, adopt, note, skip) to an item, and defending a skip. |
context/ci-mode.md |
Running apply --ci: the branch rule, handoff kinds, branches, and the ci-plan.json format. |
context/native-drift.md |
Running apply Phase 5: the extraction commands, the previous-run summary, the drift report, and filing its items. |
scripts/native_drift.py |
Phase 5's summary, diff and trigger evaluation. Covered by scripts/test_native_drift.py. |
1---2name: changelog-23description: "Claude Code changelog ingestion into the current repo: fetch (read-only display), diff (impact analysis over a release range, no edits), status (read marker, default range, replay cap), and apply (one PR per owner plugin, larger decisions handed off as work items, explicit user intent only). Use when: 'new cc version', 'what changed in claude code', a new CC release is mentioned, or the user pastes changelog text."4argument-hint: "<fetch|diff|status|apply> [--ci] [vA..vB|vX|text]"5user-invocable: true6disable-model-invocation: false7shell: bash8metadata:9 source: https://skillmd.com/skills/melodic-software/changelog-210 workflow-stage: anytime11 summary: Ingest a Claude Code release changelog and integrate its changes into the repo12---1314**Arguments.** `<fetch|diff|status|apply> [--ci] [vA..vB|vX|text]`. Full form: <action> [--ci] [vA..vB|vX|text]. Actions: fetch (default on passive mention), diff, status, apply (explicit only). `--ci` applies to `apply` only1516## Pre-computed context1718- Read marker (no fetch): !`bash "${CLAUDE_PLUGIN_ROOT}/skills/changelog/scripts/changelog-status.sh" --no-fetch 2>/dev/null | head -4 || echo "(status script unavailable)"`1920## Variables2122Arguments: `$ARGUMENTS`2324## Scope2526Ingests Claude Code changelog entries and hands their decisions to the repo's stage skills. Covers the arc: read upstream changes → orient on repo impact → research new features → scope with the user → run what fits the session through the marketplace stage skills, one PR per owner plugin → file the rest as work items → native drift. `diff` reports decisions about components, not a list of items.2728Distinct from:2930- `/harness-ops:known-issues`. Tracks CC bugs/workarounds. This skill integrates CC feature changes into repo config/docs31- Any release-triage automation the consumer runs (issue filing per release). This skill decides, holistically across a release, what changes in the repo, and hands each decision to the skill that plans, implements or verifies it; it re-implements none of them3233## Input modes and range3435Three ways to provide changelog content (priority order):36371. **User pastes text**. Skill parses inline changelog from conversation context; the releases the pasted text names are the range, so the cap and the version check apply to them and not to the repository's default feed382. **Explicit range or version**. `/harness-ops:changelog diff v2.1.257..v2.1.263` covers both ends inclusive; `apply v2.1.263` covers that one release393. **Default range**. `/harness-ops:changelog diff` (no range) runs from the read marker to the newest published release4041The read marker, the default range, and the replay cap are defined in42[context/read-actions.md](context/read-actions.md) and computed by43`scripts/changelog-status.sh`; every action starts by running it.4445## Version awareness4647On every `apply` or `diff` invocation, compare the newest release in the range against the active terminal's CC version. The status script reports both as `latest` and `installed` and emits a `warn` line when the installed version is older:4849- If the newest release in the range > installed version: **warn user**. "You're applying v2.1.263 changes but running v2.1.260. Update CC first (`claude update`) or changes may reference features not yet available in your session."50- If the newest release in the range = installed version: proceed normally51- If the newest release in the range < installed version: fine. Catching up on older releases5253## Read marker and replay cap5455One line in the repository's upstream ledger for Claude Code releases (default56`docs/upstream/claude-code.md`, override `HARNESS_OPS_CHANGELOG_LEDGER`) records the newest release57the repository has been read against. `status` reads that line; with no ledger it falls back to the58highest version named in a Conventional Commits SUBJECT of the form59`chore(<scope>): address Claude Code v<A>..<B> changelog`, and it never reads commit bodies, because60a body's "verified against Claude Code v<X>" is a doc's recency stamp, not an apply.6162A range wider than ten releases or 300 core items exceeds the replay cap. `diff` and `apply` then63stop and recommend a docs-conformance recheck of the components against the current docs, followed64by a marker set at the newest published release: the docs carry the cumulative state, and replaying65items past the cap costs more than it returns.6667## Action Router6869Parse `$ARGUMENTS` to extract the action (first token) and remaining arguments.7071| Action | Description | Detail |72|--------|-------------|--------|73| `apply` | Hand-off: ingest → explore → research → scope gate → stage skills per owner-plugin PR, larger decisions filed as issues → native drift. Consumes the `diff` working set | See "Action: apply" below |74| `fetch` | Fetch + display changelog for a version or range. Read-only | See "Action: fetch" below |75| `diff` | Resolve the range, apply the cap, emit decision rows by owner surface plus a docs-lag section. Read-only | See "Action: diff" below |76| `status` | The read marker and its source, installed vs newest release, the default range, the cap verdict | See "Action: status" below |77| `help` | Show action table | *(inline)* |7879**Routing (model-invocable):**8081- Empty args or passive CC version mention → **`fetch`** or **`diff`** (read-only). Never **`apply`**.82- Version-only token (`v2.1.152`) or range-only token (`v2.1.150..v2.1.152`) without explicit apply intent → **`fetch`** for that version or range.83- **`apply`** only when user explicitly requests integration (`apply`, `apply changelog`, `/harness-ops:changelog apply`, or unambiguous implement-this-release intent).8485If action is unknown, show action table.8687## Action: apply. User intent gate8889**`apply` mutates the repo.** Run only on explicit user intent per routing above. When the model90detects a new CC release in conversation, default to `fetch` or `diff` and offer `apply`. Do not91auto-start the hand-off.9293`apply` plans, edits and verifies nothing itself. It reads the release, asks the user which decisions94are in scope, and hands each one to the marketplace skill that owns that stage, as Phases 3 and 4 set out.9596### Unattended: `apply --ci`9798Only the literal `--ci` token selects this path, and it needs an explicit release or range. It99replaces every question in Phases 3 to 5 with the evidence rule in100[context/ci-mode.md](context/ci-mode.md): only a verified, cited `correct` row confined to one owner101plugin and not security-relevant is committed, on a local branch per owner plugin; every other row102becomes a handoff. It asks nothing, pushes nothing, opens no PR, files nothing and never moves the103read marker. It writes `ci-plan.json` for the calling workflow, rewriting it as each step finishes so a104stopped run still publishes what it finished, and ends with `ci-plan: <path>` and105`status: <stopped|planned>`. Where the phases below say to ask, confirm or file, `--ci` follows that106file instead.107108### Phase 0. Ingest109110Resolve the range, check the cap, and check version alignment:1111121. **Resolve the range** (first match wins):113 - Changelog text already in conversation → the releases its release markers or version headings name ARE the range; pass them as `--range <lowest>..<highest>` so the cap is judged on the pasted releases and never on the repository's default feed. Pasted text with no version at all skips the cap114 - A range or version was given → `--range` as given115 - Otherwise → no `--range`; the default range from the read marker applies1162. **Run the status script** with that `--range`. If `cap` reads `exceeded`, stop and relay the `recommend` line; the hand-off does not run past the cap. Relay any `warn` line per "Version awareness" above1173. **Reuse the working set** when `diff` saved one for this range and it passes the checks under "Persistence" in [context/decisions.md](context/decisions.md): `apply` consumes it, resolves no content, and re-fetches only what a recheck trigger names. Otherwise continue1184. **Resolve content**: pasted text is parsed as-is; otherwise slice the releases the `releases` line names out of a local copy of the changelog per the fetch route in [context/read-actions.md](context/read-actions.md)1195. **Parse** into structured items. Each item gets: a stable id (`2.1.257-001`), summary, category (feature / fix / UI / internal), affected surface (if identifiable). Ids and the working-set directory are defined in [context/decisions.md](context/decisions.md)120121### Phase 1. Explore122123Orient on repo impact for EACH changelog item. Run `bash "${CLAUDE_SKILL_DIR}/scripts/discover-surfaces.sh"` first: its output is the surface list, and `context/repo-surfaces.md` gives per-class examples and scoped grep patterns:1241251. Grep/Glob each feature name, setting name, hook event, CLI flag across ALL surface classes the script printed1262. Classify each item per `context/classification-rubric.md` into one action lens per owner surface it touches: **correct**, **replace**, **adopt**, **note**, or **skip**1273. Group by owner surface and write each correct, replace and adopt row with its required sentence. A skip item leaves no row128129Output: decision rows grouped by owner surface, per [context/decisions.md](context/decisions.md). The130work fans out by release cluster with no Workflow script; see "Fan-out shape" there.131132### Phase 2. Research133134For rows needing enrichment (correct rows with behavioral changes, adopt rows with unclear scope):1351361. Spawn **parallel research subagents**. One per feature cluster, each receiving the explorers' questions (use a Claude Code documentation-focused agent type when available)1372. Instruct each subagent to ground every claim in the page itself (a docs page through `bash "${CLAUDE_PLUGIN_ROOT}/scripts/fetch-docs.sh" --cache --out <dir> <slug>`, a changelog entry or GitHub issue by `curl`) and to return citations with each claim. Treat any uncited subagent claim as unverified and re-verify it against official docs before acting on it. A local verifier then checks the repo-side facts. Before dispatching it, run `bash "${CLAUDE_PLUGIN_ROOT}/lib/read-verification.sh" --plugin harness-ops --skill changelog` and copy its line into the output ([levels](https://github.com/melodic-software/claude-code-plugins/blob/main/docs/contracts/verification-level/README.md)). Under `light`, and `off`, which acts as `light` here, dispatch no local verifier, here or in `apply`'s re-check of saved rows, and never check those facts yourself in its place: mark each row it would have checked `unverified` and write `verification: skipped (<level>)`1381393. Research targets per item type:140 - New frontmatter field → exact syntax, interaction with existing fields, docs gap141 - New hook event → schema, sync/async, input/output shape142 - New CLI flag → syntax, settings.json equivalent (or lack thereof), valid values143 - Behavioral change → before/after, migration path, breaking implications144 - Bug fix → what was broken, what surfaces affected, historical data impact1451464. Synthesize research into enriched rows147148### Phase 3. Scope gate149150Present the decision rows to the user via `AskUserQuestion` or structured markdown:151152```markdown153| # | Owner surface | Lens | Items | Required sentence | Action needed |154|---|---------------|------|-------|-------------------|---------------|155| 1 | <component> | correct | <ids> | <false vs true> | <specific update> |156| 2 | <component> | adopt | <ids> | <problem solved> | <adopt, decline, or defer pending probe> |157```158159The user picks which rows are in scope: "all rows", "just correct", or specific rows by number. A160row left out is reported at the end and recorded only when the user declines it, with its reopen161condition in the ledger's Declined table.162163Then sort each in-scope row by whether it fits this session. A row **fits** when its edits stay in164its owner plugin and one session can plan, make and verify them. A row is **too large** when it165redesigns a component, changes behavior across several plugins, or needs work this session cannot166finish and verify. Show the sort with the rows and let the user move a row across.167168### Phase 4. Hand off1691701. **Rows that fit**: work them in this session's branch, one branch and PR per owner plugin. Use171 `/planning:plan` for the edit plan (if the planning plugin is installed), `/implementation:implement`172 for the edits (if the implementation plugin is installed) and `/verification:confirm` for the173 check (if the verification plugin is installed). A stage whose plugin is missing is not done by174 hand here: report the rows that needed it as not executed, naming the missing plugin. A175 `replace` row is never edited: nominate it into `/harness-ops:audit-native-overlap`1762. **Rows too large**: file one issue per row through `/work-items:track` (if the work-items plugin177 is installed), carrying the owner surface, lens, required sentence, item ids and range. That178 issue is worked later in its own PR, behind an interview-style human gate before any edit. When179 `/work-items:track` is not installed, report the row and file nothing. List the rows to file and180 file them only after the user confirms that batch; a row the user does not confirm is reported, not181 filed1823. **Docs lag**: hand the docs-lag pairs to `/harness-ops:known-issues`. They are never a decision row1834. **Commit and PR shape**: one PR per owner plugin, each carrying that plugin's `CHANGELOG.md`184 entry. The upstream ledger update is the last PR and references the others. Every commit subject185 reads `chore(<plugin>): address Claude Code v<A>..<B> changelog`, the ledger's taking the186 ledger owner's scope. The ledger PR moves the read marker only once every row in the range is187 applied in a merged PR, nominated, recorded as declined or deferred, or filed, and stops below the first188 release that still has a row outside those states, so no unfinished row drops out of the next189 default range. `status` reports it from the ledger and, until the ledger exists, from that subject190191### Phase 5. Native-surface drift192193Re-read what Claude Code ships and file what moved. Run the inventory self-check, a full194`--binary-only --docs` extraction, and overlap `detect` and `self-check`, then195`scripts/native_drift.py` to diff against the previous run's summary and evaluate the overlap196store's recheck triggers. Report surface changes (added, removed, renamed, reclassified),197invocability and marker changes, docs cross-check changes, and new overlap candidates. File one198work item per new candidate, fired trigger, and degraded or broken self-check through199`/work-items:track`, deduped by a `native-drift:<kind>:<surface>:<component>` key; a self-check200degraded only by a CLI version past the validated build proposes revalidation instead. A repository201with no overlap store is report-only: report, file nothing, keep the baseline. Commands,202the summary's location, items, dedupe and approval: [context/native-drift.md](context/native-drift.md).203Its commands write this skill's directory as `<skill-dir>`, which is `${CLAUDE_SKILL_DIR}`; put that204path in place of the placeholder before running one.205206When the harness-config plugin is installed, also run its effort-pin drift check directly, with207that plugin's install directory in place of the placeholder:208209```bash210bash <harness-config plugin root>/skills/audit/scripts/check-effort-pins.sh211```212213Report its lines with the native-drift report. Exit 1 means model-config's effort tables or214per-model defaults moved, or a pin names a level the page does not list: each flagged pin waits on215a person to re-decide it and rebaseline, as the `/harness-config:audit` `effort-pins` scope216documents. Exit 3 means the page was unread or reshaped, so the check made no claim; say so.217Without the harness-config plugin, report that the effort-pin check was not run.218219Then list the behavior probes the range touches, against the same local copy of the changelog:220221```bash222python3 "${CLAUDE_PLUGIN_ROOT}/skills/behavior-probes/scripts/probe.py" recheck --changelog <changelog.md> --range <A..B>223```224225Each line names a case, the changelog item id and the tag that matched. Report them with the drift226report. Rerunning them live costs model runs, so offer `/harness-ops:behavior-probes live` for those227cases and run nothing until the user agrees.228229End the run with a report that leads with what waits on the user (rows left out of scope, rows230filed or reported unfiled, stages skipped for a missing plugin), then the PRs opened and what each231changed, what `/verification:confirm` showed, the docs-lag pairs handed off, and the Phase 5 drift232report with the items filed or skipped.233234---235236## Actions: fetch, diff, status (read-only)237238The three read-only actions stop short of any edit. **Full steps in [context/read-actions.md](context/read-actions.md)**.239Its fetch command writes this skill's directory as `<skill-dir>`, which is `${CLAUDE_SKILL_DIR}`; put240that path in place of the placeholder before running it.241242- **`fetch`**. Read the raw changelog through the plugin's `fetch-changelog.sh`, which picks the source and writes the `.md` to a file; slice releases locally with the plugin's `changelog-releases.sh --blocks`) and display a version, a range, or the newest release. No edits243- **`diff`**. Run the status script; stop at an exceeded cap with its recommendation; otherwise Phase 0 (ingest) + Phase 1 (explore) + Phase 2 (research) over the releases in range, stopping before the scope gate. Emits decision rows grouped by owner surface, each with its lens and required sentence, plus a docs-lag section, and saves its working set for `apply`. Answers "is this range worth an `apply`?"244- **`status`**. Run the status script and relay: the read marker and its source (ledger line or commit subject, never a commit body), installed vs newest release, the default range, and the cap verdict with its recommendation245246---247248## Next249250- Candidates and fired triggers filed by Phase 5: `/harness-ops:audit-native-overlap`.251- Items filed as raw intake, and rows filed by Phase 4: `/work-items:triage`.252- A revalidation item: `/harness-ops:inventory`.253- Behavior probes the range touches: `/harness-ops:behavior-probes live`.254255## Reference index. Load on demand256257| File | Load when |258|---|---|259| `context/read-actions.md` | Running `fetch`, `diff`, or `status`; the read marker, range, cap, and fetch route are defined there. |260| `scripts/changelog-status.sh` | Every action's first step; `--help` lists its output lines and flags. Covered by `scripts/changelog-status.test.sh`. |261| `context/decisions.md` | Writing or reading decision rows, choosing where a decision is recorded (plugin CHANGELOG, audit-native-overlap nomination, ledger, filed issue), fanning out, or saving and reusing the working set. |262| `context/repo-surfaces.md` | Phase 1 explore, after running `scripts/discover-surfaces.sh`: per-class examples of what an item changes, and scoped grep patterns. |263| `context/classification-rubric.md` | Assigning a lens (correct, replace, adopt, note, skip) to an item, and defending a skip. |264| `context/ci-mode.md` | Running `apply --ci`: the branch rule, handoff kinds, branches, and the `ci-plan.json` format. |265| `context/native-drift.md` | Running `apply` Phase 5: the extraction commands, the previous-run summary, the drift report, and filing its items. |266| `scripts/native_drift.py` | Phase 5's summary, diff and trigger evaluation. Covered by `scripts/test_native_drift.py`. |