dsh-plugin-upgrade-rc1

dsh plugin --profile <name> add dsh-plugin-upgrade-015

Install & uninstall

dsh plugin --profile web add dsh-plugin-upgrade-015            # from npm
dsh plugin --profile web add "github:PerryLink/dsh-plugin-upgrade-015#main"   # from source
dsh plugin --profile web remove dsh-plugin-upgrade-015         # uninstall (reversible)

Installing the bundle only registers a skill; removing the row removes the skill. The CLI is a normal npx target and needs no profile at all.

Compatibility

What you get

Two halves, one seam catalog:

• A bundled agent skill (plugin-upgrade-015) — the merged corridor card and a fix-and-verify loop. The body first routes the caller to the leg that matches its peer band; the model loads it only when a task actually needs it, and the package contributes no system-prompt paragraph and no tool.
• A zero-dependency CLI (dsh-plugin-upgrade-015-scan) — reports file:line facts for the merged twenty seams (S3, S8, S9, M1, S4, S5, S6, S7, S2, S1, S10, C1, C2, P1, C4, C5, H1, H2, H4, H3) re-read from the harness tag ranges on 2026-09-09 (leg A) and 2026-09-10 (leg B). Exit 1 on any error-severity hit, so it drops straight into CI.

The point is the failure mode this corridor exists to kill: this span's breakage is mostly silent, from both ends. The type line can be stale, so the repo compiles against the old catalog (seam M1), and the bare conversation client slot was deleted with no alias while ctx.slots.inject() only runs its callback when the declaration exists — so a client half that still targets it stops mounting with no error, no log line and no failed build (seam C1). Three classes of breakage survive typecheck + test:

1. the local gate compiles a stale type line — old paths alias, or dev/test types pinned at 0.1.5-alpha.* (seam M1);
2. a log writer omits the V3 required stream field, so the session imports and then refuses to resume (seam S3);
3. the tests are mocked against the old shape, so they pass while the host drops the contribution (seam C1).

Honest sizing: leg B's workspace sweep found that the family's client halves use only 8 slot keys, all of which survive in rc.1 — for them the rc.1 breakage is latent, not actual. Third-party client plugins that targeted the bare conversation key are the ones that break, and they break quietly. Leg A's sweep found the opposite texture: 40 repos, 11 of them hit M1, and fixing the stale path exposed real TypeScript errors in 3 repos that were previously "green".

Quick start

# 1. install the bundle into your profile
dsh plugin --profile web add dsh-plugin-upgrade-015

# 2. verify the row mounted
dsh --profile web --dump-config | grep -A3 'id: dsh-plugin-upgrade-015'

# 3. scan the plugin you are upgrading
npx dsh-plugin-upgrade-015-scan --repo ../my-plugin

Then ask the agent to use the plugin-upgrade-015 skill, or drive the loop yourself with the card at skills/plugin-upgrade-015/references/v0.1.3-alpha.1-to-v0.1.5-rc.1.md (leg A is §1, leg B is §2, the merged seam index is §3).

Configuration

Every key is optional and lives in the profile patch:

- insert:
    - id: dsh-plugin-upgrade-015
      name: dsh-plugin-upgrade-015
      config:
        skillName: plugin-upgrade-015

The plugin mounts loud: a missing SKILL.md, an empty body, or a frontmatter without name fails the mount instead of registering an empty skill.

Surfaces

Skill — plugin-upgrade-015 (model- and user-invocable by default). Body: leg routing table, the 8 hard rules and the 6-step loop. References: the merged corridor card. Scripts: the detector, shipped inside the skill directory so relative paths resolve.

CLI — dsh-plugin-upgrade-015-scan:

dsh-plugin-upgrade-015-scan [--repo <path>] [--json <out.json>] [--seams S3,C1,P1] [--quiet]

Exit codes: 0 no error-severity hit · 1 at least one error-severity hit · 2 usage or scan failure. A clean scan is necessary but not sufficient — the exit criterion is a real-host smoke, plus a resume round-trip for log writers (leg A) and a real browser assertion for the client half (leg B).

The twenty seams

Order follows the catalog in lib/scan.mjs (leg A first, then leg B), which is also the order test/card.test.mjs pins the card to.

S7, S2, S1, S10, C4, C5, H1, H2 and H4 are deliberately advisory: they have legitimate matches (a repo that already uses the new API, a documentation snapshot, a plugin's own model-id table, a Node ChildProcess.pid), so the scanner reports them as leads for manual review rather than failures. M1 and P1 are structured checks — they resolve package.json and tsconfig*.json instead of matching text — and H3 is card-only: documented, id-parity checked, and deliberately without a detector (CARD_ONLY = ['H3']).

What this does not cover

• A hop after 0.1.5-rc.1. The merged span ends at rc.1 by construction: the harness hop 0.1.5-rc.1 → 0.1.5-rc.2 added no plugin-facing seam (this package's own dev/test pin now runs on the 0.1.7-rc.1 line so the catalog is verified against the newest published types; the compat workflow's probe still anchors 0.1.6-alpha.2). Anything that adds a seam later is a new package — a card that drifts is worse than no card.
• The 0.1.1 → 0.1.2 hop. Use the community convergence skill.
• Restating across legs. Leg A owns the session-format seams (assistant/message.stream, SessionHandleReadResult, EpochHeader.system, ctx.agent, Inbox, SystemPrompt.persona, the V3 log generation) and leg B does not restate them — the whole packages/core/session/src diff in leg B's range is two added event-type literals and one comment line. Each leg's card section keeps its own scope statement.
• The DSH user-facing upgrade path. This package upgrades plugin source code, not a user's harness installation.
• Theme tokens. docs/web-styling.md has zero changes in leg B's range.
• Proof. A clean scan is a hypothesis. The exit criterion is a real-host smoke (temp DSH_HOME, target CLI, plugin add <tarball>, --dump-config) plus a resume round-trip for session-log writers (leg A) and a real browser assertion for every client-side hit (leg B).

Security boundaries

• Read-only scan. The CLI never writes inside the scanned repository; --json writes only to the path you pass.
• No network, no shell. The scanner imports nothing beyond Node's standard library and never spawns a process.
• No secrets. Nothing in the package reads credentials, environment tokens, or session data.
• Sandboxed smoke recipe. The card's real-host check uses a mkdtemp DSH_HOME; it never touches your real ~/.dsh.

Development

npm install                        # or: pnpm install (the repo ships a pnpm-lock.yaml)
npm test                           # node --test: scanner, card<->catalog parity, real Cordis + SkillRegistry
npm run verify:self-contained      # every import resolves inside the package
npm run verify:artifacts           # the packed tarball carries the skill, CLI and patch, and excludes tests
npm run check:readmes              # five-language README consistency
npm pack

The scanner has a synthetic fixture pair per leg: fixtures/leg-a-bad-repo (leg A's session/config seams, every error seam present on purpose) with fixtures/leg-a-good-repo (adapted), and fixtures/bad-repo (leg B's client-slot seams) with fixtures/good-repo (adapted) — plus a live negative on a family repository already pinned to 0.1.5-rc.1, so a regression in the catalog fails the suite rather than a downstream user. test/card.test.mjs asserts that the merged card index and lib/scan.mjs name exactly the same twenty seam ids with the same severities, and that CARD_ONLY is exactly ['H3'] — the evidence-binding rule as a machine gate.

Topics

dsh, dsh-plugin, deepseek-harness, deepseek, cordis, plugin-upgrade, migration, skill, version-card, scanner, client-slots (mirror package.json keywords; dsh-plugin is the ecosystem's visibility channel).

PerryLink DSH Plugin Family

This project is one of the 87 actively maintained DeepSeek Harness plugins from PerryLink. The family roster is 95 plugins in total — 5 frozen, 3 retired — and stays collapsed below so every README keeps a short first screen. Expand it to see every row, with the reason in the Status column.