fix: read Git's absolute git dir so a Windows worktree is not the base tree

[hkwuks]

Summary

Adds DeepSeek Harness (dsh) as a fourth host. The repository root becomes a dsh plugin package, so a dsh user installs the skill and the SessionStart hook with one command and no hand-written profile YAML:

dsh plugin --profile web add dsh-taskflow

package.json declares dsh.bundle.patch — the marker dsh reads — and dsh/index.js mounts dsh's own @deepseek-ai/dsh-skill-filesystem over this package's skills/ and @deepseek-ai/dsh-hooks-claude-code over the new hooks/hooks-dsh.json. Both are dsh's packages rather than copies: duplicating the skill provider would duplicate frontmatter parsing, watching and catalog caching; duplicating the bridge would duplicate the matcher and the output codec.

Three measured facts shape the wiring. Each was verified against an installed dsh 0.1.5-rc.2 rather than taken from documentation, and each is the reason for a choice that would otherwise look arbitrary:

1. A patch's baseUrl is the profile directory, not the plugin's. No bundle patch can name the files that ship beside it — so dsh/cordis.patch.yml is a single insert with no config, and every path lives in dsh/index.js, derived from import.meta.url.
2. The bridge substitutes ${CLAUDE_PLUGIN_ROOT} inside the command string but never exports it. hooks/session-start picks its output shape from the environment, so without the CLAUDE_PLUGIN_ROOT= prefix the hook takes its fallback branch, emits a top-level additionalContext, and the bridge's codec discards it silently — exit 0, valid JSON, no context. That prefix is the entire reason hooks-dsh.json is a separate file instead of reusing hooks/hooks.json.
3. dsh's SessionStartSource is startup | resume | clear | compact. There is no fork, so the dsh matcher omits it.

No hook script changes. hooks/session-start is byte-identical; the dsh wiring adapts to the scripts, not the reverse. The other three hosts' wiring files are untouched.

TaskFlow traceability

• Task: TaskFlowDocs/2026-09-19-dsh-host/
• Scope: root package.json; dsh/index.js; dsh/cordis.patch.yml; hooks/hooks-dsh.json; host lists in README.md, README.zh-CN.md, hooks/README.md, skills/taskflow/references/runtime.md, ROADMAP.md; the release-literal set in RELEASE.md, hooks/release-check, CHANGELOG.md, hooks/smoke-test. Approved Plan: TaskFlowDocs/2026-09-19-dsh-host/plan.md (v1, approved 2026-09-19 19:29 +08:00).
• Base branch: main
• Target repository: hkwuks/TaskFlow

Verification

• bash hooks/smoke-test → ALL SMOKE PASSED
• git diff --check → clean
• Relevant Skill/plugin validation → see below
• Results recorded in the TaskFlow Plan

Composition layer. dsh --profile <p> --dump-config → exit 0, with - id: taskflow / name: dsh-taskflow composited over the base and web-app layers. Verified through the literal documented command dsh plugin --profile <p> add <repo-root>, not a hand-edited profile.

Mount layer. Booting that profile with an independent observer plugin read ctx.skills.list() and got exactly ["taskflow|taskflow|custom"] — one entry, from the provider this plugin registers.

Seam layer. Driving dsh's own libraries over the hook this repository actually ships:

• matchesMatcher accepts startup, resume, clear, compact and rejects fork, end, and the empty string.
• parseHookOutput reads the hook's real stdout back as 397 characters of SessionStart context.
• Two negative controls: the same stdout read as PreToolUse yields nothing (proving the hookEventName guard admitted it), and dropping the CLAUDE_PLUGIN_ROOT prefix loses the context (proving that prefix is load-bearing, not decorative).

Repository checks. bash hooks/release-check . → STATUS: pass with the fourth manifest in the compared set. bash hooks/repository-check . → needs-user-input, naming this task's uncommitted directory at the time it ran — expected mid-task, and its exit code is unchanged by that report.

Mutation check on the new coverage. The smoke-test section added here was verified to fail on both mutations it exists to catch: removing the CLAUDE_PLUGIN_ROOT prefix, and pointing the command at a root with no hooks.

No API call was made; the approved scope excluded a live model run.

Review boundaries

• No secrets or opaque remote payloads included
• No unrelated task or user files changed
• Remote/base assumptions are stated
• Known limitations and follow-ups are documented

Base and remote. origin is https://github.com/hkwuks/TaskFlow.git, branch feature/dsh-host, cut from main at 7dcff62 with no divergence at cut time. main was not merged into this branch during the work.

Two files in this PR are not part of the dsh change, and both are worth a look rather than a rubber stamp:

• TaskFlowDocs/repository-docs/index.md — the SessionStart hook regenerating its own metadata dates. The identical diff exists independently in the base checkout, byte-for-byte; it is not a manual edit and not divergent.
• TaskFlowDocs/todo.md — this task's own intake entry, TF-20260919-55110d. Another session added a different entry (TF-20260919-d6005c, on personal-rules layout) to todo.md in the base checkout, which is not in this branch. Both are appends at the same anchor, so the merge driver handles the merge — but if that other entry is still uncommitted at merge time, it needs landing separately.

Known limitations

• No live model turn. A failure that only appears at the model seam — a malformed additionalContext, a hook that times out — is out of reach of the verification above; the seam test drives dsh's real parser to narrow the gap, not to close it.
• Windows is untested for this host. dsh's own Windows story is unverified here, and hooks/hooks-dsh.json names no commandWindows equivalent.
• ctx.loader.import is not a documented plugin-authoring surface. It is how dsh's own loadProfileDirectory mounts nested entries, so it is not private to the launcher, but a future dsh release could change it.
• The npm name dsh-taskflow is unclaimed and the package could be published later, but nothing in this PR publishes, tags, or pins anything.