dsh-requirements-alignment
DeepSeek Harness 的运行时需求漂移防护——在长期运行的代理工作期间，使其始终符合用户意图。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-requirements-alignment@0.4.2Requirements Alignment 以 dsh-requirements-alignment 发布，当前版本为 0.4.2。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Runtime requirement drift guard for DeepSeek Harness.
Keep long-running agents aligned with user intent while they work.
dsh-requirements-alignment turns the user's request into a durable requirement baseline — goal, protected constraints, must-preserve behavior, allowed scope, and settled user decisions — and guards it while the agent executes. The agent works silently until a step would materially change the task direction; only then does the plugin surface a drift candidate to you, records your decision, and updates the baseline.
Canonical alignment state lives in the durable AlignmentStateStore sidecar — an official storage-domain domain over the storage-json backend — never in session events. The DSH Session log holds only official DSH-recognizable events, so a bare DSH build (without this plugin) reads any new session, and resume, fork, and compaction recover the same state. The alignment/* event vocabulary is kept solely for legacy compatibility, migration, and test/fold fallbacks; production never appends it.
You decide the direction. The agent decides the engineering.
Plan Mode asks: "Is this the right implementation plan?"
Requirements Alignment asks: "Are we still solving the right problem?"
Plan Mode prevents a bad plan from starting. Requirements Alignment prevents a good plan from drifting.
Plan Mode is the official review-approve step before implementation. Requirements Alignment is intent continuity during execution. They compose: plan, approve, execute — and this plugin keeps the execution on the approved direction. It never modifies plan mode, exit_plan_mode, or any @deepseek-ai/* core package.
| Mechanism | What it does |
|---|---|
| System-prompt policy section | In auto mode (default) the drift-guard policy is contributed to every agent's prompt at order 60. It teaches the agent to hold a requirement baseline, monitor silently, and detect direction-level drift — scope expansion, constraint conflict, user-visible behavior change, architecture shift, invalidated assumptions, user direction change. |
establish_baseline tool | Records the baseline (goal, explicitConstraints, mustPreserve, allowedScope, userDecisions, openDirectionDecisions). Silent — it never asks the user. Recording again bumps the baseline revision. |
report_drift tool | Records a drift candidate (reason, description, requiredChange), asks you one question through the native user-questions channel, and records your decision. The default approve / stay-within-scope options are always offered; the two defaults map to approve / reject, a model-supplied alternative direction you pick — or your own free-text answer — maps to revise with your exact words as the note, never to a silent rejection. The tool result returns your exact choice (the note) and the required baseline change to the agent, so it never re-asks what you picked. Only alignment state managed by this plugin contributes to the requirement baseline — unrelated ask_user_question calls (plan mode, other plugins) never pollute it. |
/align command | Manual entry: reports the current alignment status (baseline revision, goal, protected constraints, drift count, last drift, last decision, current status, and whether the mode is the profile default or a runtime override) and steers a fresh alignment inspection into the agent. It inspects; it never blocks execution. |
/align-mode command | Always-on mode switch. No argument prints the three-layer snapshot (effective / profile default / runtime override). /align-mode auto|manual|off persists a runtime override; /align-mode reset drops it. Stays registered in Off so a live switch to Off is reversible without editing settings.yaml. |
| Durable state | Canonical alignment state is written to the durable AlignmentStateStore sidecar (official storage-domain → storage-json backend), keyed by session lifecycle identity, so it survives resume, fork, and compaction — and a bare DSH build without this plugin still reads new sessions. The session log itself only ever receives official DSH events; alignment/* remains a legacy/migration/fold fallback only. |
# from anywhere; path is anchored to your invoking directory
dsh plugin --profile web add <path-to-this-checkout>
# or from the registry once published
dsh plugin --profile web add dsh-requirements-alignment
The plugin is a profile bundle (dsh.bundle.patch + cordis.patch.yml), so it installs through the standard plugin mechanism and adds two rows:
requirements-alignment — the controller (policy section, /align, /align-mode, both tools);requirements-alignment-ask-user — the model-facing question tool.Install the bundle and start a normal DSH task. Auto mode is enabled by default; clear tasks run with zero interruption, and you are only asked when the execution is about to change direction.
dsh plugin --profile web add dsh-requirements-alignment
Use /align any time you want to inspect whether the current execution still matches the requirement baseline.
Alignment mode is a four-layer model that can change at runtime without a profile restart:
valid session override -> valid persisted runtime override -> valid profile default -> auto
| Layer | Source | Persisted where |
|---|---|---|
Session Override (sessionMode) | Your per-session switching (/align-mode session). Only the calling session is affected. | The durable requirements_alignment_modes sidecar (official storage-domain → storage-json), keyed by session lifecycle identity |
Profile Default (defaultMode) | The composition/profile config — `mode: auto | manual |
Runtime Override (overrideMode) | Your runtime switching (setMode / the settings document). | settings.yaml via the DSH Settings service (@deepseek-ai/dsh-settings) |
| Effective Mode | effectiveMode = valid session override ?? valid runtime override ?? valid profile default ?? auto; Effective Source reports which layer produced it (session / override / profile). | derived |
- id: requirements-alignment
name: dsh-requirements-alignment
config:
mode: auto # profile default; a runtime override wins over this
/align-mode session. Only the calling session's effective mode changes; other live sessions and the shared runtime override never move. Persisted per session (keyed by id + createdAt + cwd), so a resumed session restores its override, and a fork inherits the effective session override at its seed boundary (then becomes independently changeable). An invalid persisted value fails open to the next valid layer.effective = your last runtime override. An invalid persisted value (for example a hand-edited mode: banana) never fails startup: the plugin falls back to the profile default and repairs the document once.resetMode call or replacing the settings section with {}) drops the runtime override. effective = defaultMode, source = profile. It never re-writes the current effective mode as a new override./align-mode always agree.Hot switching — mode transitions are per-agent register/dispose operations on the agent's OWN scope (agent.ctx), not a "change the config and restart" step. When a session's effective mode changes, that session's agent is re-synced: the outgoing capability set is disposed and the incoming one registered in the agent's scope. Two live sessions hold disjoint capability sets with zero leakage. Auto → Manual → Off → Auto can be cycled live with no duplicates, no listener leaks, and no profile restart. /align-mode is the always-on control command: an Off session has no alignment capabilities at all (no policy, no tools, no /align) but keeps /align-mode so it can switch itself back.
Runtime Mode backend: implemented. Web floating capsule: implemented (v0.4.1). The capsule — a bottom-right collapsible float in the DSH Web
shell.overlayslot — switches the session and shared layers via the plugin's loopback management API, so it can never disagree with/align-mode. The runtime override is also persisted through the official DSH Settings service (@deepseek-ai/dsh-settings), and an externalsettings.yamlhot edit is picked up live. The profile default remainsmode:in the profile bundle.
Auto is the recommended default. Clear tasks run with zero interruption; you are only asked when the execution is about to change direction.
| Mode | Policy section | Alignment tools (establish_baseline, report_drift) | /align (+ /align-migrate) | /align-mode |
|---|---|---|---|---|
| Auto (recommended) | yes | yes | yes | yes |
| Manual | no | yes | yes | yes |
| Off | no | no | no | yes |
Every cell above describes ONE session's effective mode. Since v0.4.0 the capabilities are registered in that session's agent scope (agent.ctx), so two live sessions can hold different effective modes at the same time — the matrix applies per session, never globally.
report_drift only for a real direction change./align, which reports status and steers a fresh inspection./align. /align-mode (registered at plugin scope) stays so you can switch this session back to Auto or Manual without editing the profile or settings.yaml./align-mode operates on the shared layers (runtime override + profile default); /align-mode session operates on only the current session:
/align-mode # this session's four-layer snapshot
/align-mode session # same four-layer snapshot
/align-mode session auto|manual|off # set ONLY this session's override
/align-mode session reset # drop this session's override -> shared layers
/align-mode auto|manual|off # shared runtime override (all sessions fall back to it)
/align-mode reset # drop the shared runtime override -> profile default
/align-mode session is explicit so you cannot accidentally change every session while intending to change only the current one. Two concurrent sessions can hold different effective modes with zero leakage; changing Session A never changes Session B or the shared runtime override. A resumed session restores its override, and a fork inherits the effective session override at its seed boundary (then becomes independently changeable). The four-layer snapshot identifies the exact source of the current session's effective mode (session override / runtime override / profile default).
Canonical alignment state (baselines, drifts, decisions, manual checks) lives in the independent AlignmentStateStore sidecar, and session-mode overrides live in the separate requirements_alignment_modes sidecar. Switching Auto → Manual → Off → Auto (shared or per session) only changes which capabilities are registered in that session's agent scope; it never deletes a baseline, never deletes either sidecar, never clears state, and never rewrites session events. A baseline established in Auto is still there after Off and back, and a session override is never touched by a shared reset.
mode: off (as profile default, a runtime override, or a session override) leaves the bundle in the profile. The row is still loaded; an Off session has no alignment capabilities, and /align-mode session auto|manual|off can switch it back live. That is not the same as uninstalling. (A session that predates the persistence-compatibility fix may still carry legacy alignment/* events in its log; current production never appends them.)
# disable the controller only (leaves the ask-user tool mounted)
# in the profile's cordis.patch.yml:
# - id: requirements-alignment
# disabled: true
# full uninstall — DSH returns to its previous behavior
dsh plugin --profile web rm dsh-requirements-alignment
Every registration is a Cordis effect disposer: unloading removes the plugin-scope /align-mode and explicitly unwinds every per-agent capability set (policy section, /align + /align-migrate, both tools). Canonical alignment state remains in the durable sidecars. Only sessions written by older versions keep legacy alignment/* events in their log — the current plugin never appends them to live sessions.
The policy section is present in every agent's system prompt. Behavior at task start:
establish_baseline (silent) before the first substantive edit, pinning the constraints, then works. No user question is involved.ask_user_question, records the baseline, and works.During execution the agent is fully silent unless an action would materially change the baseline (drift). There are no periodic checks, no tool-call counting, no per-file questions. When a drift candidate appears, the agent calls report_drift before acting; the tool result names your exact choice back to the agent (the note and any required baseline change), it records the outcome and, if you approved or revised the direction, the baseline advances to the next revision. The same choice is projected in the per-session baseline summary, so an interrupted or crashed run that resumes knows exactly what you picked without asking again.
A delegated instruction such as "pick whatever makes sense" does not waive the one start question for a greenfield idea.
/align# profile cordis.patch.yml (or a --patch overlay):
- id: requirements-alignment
config:
mode: manual
Manual mode contributes no policy section — the agent works normally until you invoke the command:
/align
/align records the inspection, reports the current alignment status, and steers a fresh alignment check into the agent (which may then run the drift protocol if it finds a candidate). It never takes over the workflow and never blocks execution. The steered check uses the durable sidecar baseline, not the session event log.
/align-mode # show effective / profile default / runtime override
/align-mode manual # persist a runtime override and hot-switch now
/align-mode reset # drop the override; return to the profile default
User: Fix the submit bug. Don't change the UI or the public API.
Agent: [records the baseline silently, fixes the bug — no questions]
User: The result-page filter is the only thing to improve. Do not refactor backend logic.
Agent: [working… discovers the backend filter itself is broken and a correct fix would
need backend changes]
Agent: [report_drift → you are asked]
User: Stay within the current scope.
Agent: [improves the UI only, leaves the backend untouched]
User: The app is single-user and local-only. Now make it work across devices.
Agent: [detects an architecture shift]
Agent: [report_drift → you are asked]
User: Approve the direction change — multi-user with accounts and cloud sync.
Agent: [records the updated baseline (revision advances) and implements]
When a step needs you, the agent asks and waits instead of guessing. The session below is a real run of a long publish: the log shows the agent asking you to finish browser authorization, then continuing after you did.
Waiting for browser authorization, then authorization completed and publish resumed
What the session log can prove: the agent asked the user to complete browser authorization for publish and waited for an answer; after the user completed authorization, the publish job finished with exit 0 and the session continued. The log does not record a later registry listing or any outcome beyond that job's exit code.
The plugin records one of these reasons on every drift candidate:
| Reason | Meaning |
|---|---|
scope-expansion | Doing materially more than asked (e.g. "optimize the page" → "refactor all state management"). |
constraint-conflict | An explicit constraint blocks the way ("keep the API" — but the API must change to continue). |
behavior-change | A decision changes product behavior, UX, defaults, or compatibility without prior authorization. |
architecture-shift | Local→cloud, backend, auth, multi-user, sync, persistence model, public API, schema, migration. |
data-model-change | The data model must change in a way the user did not authorize. |
compatibility-change | Existing callers, formats, or APIs would break. |
assumption-invalidated | The implementation rested on a key assumption the code now disproves, and continuing needs a new direction. |
user-direction-change | The user introduced a new direction mid-task. |
The agent decides autonomously: filenames, helper placement, variable naming, map vs loop, routine refactors, formatter, lint, test placement, ordinary library use, the repository's established stack, small internal designs that do not change observable behavior, in-scope bug fixes, and necessary test additions.
DSH child agents cannot ask the user (ask_user_question and report_drift reject with DELEGATED_CALLER for owned children). A child that would need to change the baseline does not decide: it includes a Requirement drift candidate block — reason, current baseline, required change, decision needed — in its final report (or the report tool when available). The parent owns the user interaction and runs the drift protocol.
| Key | Default | Meaning |
|---|---|---|
mode | auto | Profile default layer of the runtime mode (auto — policy section + tools + commands; manual — tools + commands only; off — inert, nothing registered). A valid persisted runtime override wins over it at startup and while running; reset drops the override and returns to this value. |
section | shipped policy | Deployment-owned policy text replacing the shipped one (auto mode). Must be non-empty when provided. |
Unknown config keys fail at load (same stance as dsh-plan-mode).
@deepseek-ai/* packages; the only host-side file touched is the profile's bundle list / patch, which is exactly the mechanism DSH provides for installing plugins.storage-domain sidecar, persists a runtime mode override through the DSH Settings service when one is mounted, and steers one user message on /align. Production never appends alignment/* session events.mode: guard can build on the same events (the fold already derives drift-pending and baseline-update-pending) without changing the architecture.report_drift in a fraction of runs (measured honestly in the acceptance report: 3/4 in the RC benchmark); agent-detected constraint conflicts are more reliable. The policy is written to maximize the natural rate; the mechanism itself is deterministic once invoked./align needs a command adapter. UI-less spines (the headless profile, ACP automation) do not dispatch slash commands; the command is exercised by the Web client and by the unit tests / dogfood driver./align runs accumulate checkpoints (reads stay O(1) at the head, storage grows with the mutation count).settings.yaml hot edit may take up to one poll cycle (~2s) to appear. Full baseline history is still inspected through /align text; the capsule shows the latest snapshot only.shell.overlay slot and the webServer service. If the web profile is not running, or the shell.overlay slot is unavailable, the capsule simply does not mount and /align-mode remains the control surface.The release gate runs type checking, linting, a production build, and the Node test suite:
pnpm run check
Real DSH dogfooding boots real dsh profiles with an isolated DSH_HOME. Three run modes keep development fast and honest:
powershell -File scripts/dogfood.ps1 -Smoke # development: 02-typo, 03-bugfix, 04-scope-drift, 09-drift-choice
powershell -File scripts/dogfood.ps1 -Scenario 12-interrupt-revise # one scenario
powershell -File scripts/dogfood.ps1 # FULL correctness suite (RC gate): 01..13 minus the 05 benchmark
powershell -File scripts/dogfood.ps1 -Benchmark05 # natural benchmark: 3 runs, reports NATURAL DRIFT TRIGGER N/M
-FailFast aborts at the first failed check; -TimeoutSec <n> (default 600) is a hard per-scenario timeout that kills the process tree. Scenario tasks for natural-behavior cases (03, 04, 05) contain NO protocol instructions; protocol-forced mechanism cases (01, 06, 07, 08, 09, 10, 11, 12, 13) are reported separately — the natural drift trigger rate is its own metric, never presented as a mechanism verification. Scenario 13 (session-scoped mode) exercises the two-session isolation probe (switchTopLevelOnSubagent). The full suite must run under danger-full-access (see docs/PROJECT-MEMORY.md).
The packed-artifact smoke packs the current tarball, installs it into a disposable profile, boots Auto → Manual → Off (/align, establish_baseline, and the policy section are asserted from the assembled system prompt and live registries — not a loose word match), removes it, and verifies the profile restores cleanly:
powershell -File scripts/packed-smoke.ps1
The current v0.4.2 gate uses DeepSeek Harness 0.1.1-rc.2 throughout. The
package-level Web client graph injects only the two packages that provide
actual client entries (dsh-client-runtime and dsh-client-locale); the
browser module separately injects the Cordis services slots and locale.
dsh-client-ui-slots is therefore neither bundled nor requested as a client
graph node, while the shell.overlay behavior is unchanged.
The v0.2.1 release gate verified:
The v0.2.2 persistence-compatibility gate verified:
statusCache session-identity + 3 align-driver lazy-resolution regressions)apply() before the controller exists → later reads resolve the sidecar (revision 1), never the legacy foldbaseline recorded + revision >= 1)npm pack --dry-run passes (exports targets all present; no v0.3.0 runtime-mode / hot-switch files)KNOWN_SESSION_EVENT_TYPES = 44 official known event types; alignment/* = 0 official known event types. The invariant is semantic: the registry intersects with the upstream known set and never contains alignment/* types — the fixed count is informational, not contractual.The v0.3.0 runtime-mode / hot-switching gate verified:
/align-mode, sidecar fold)effectiveSource = profile); an invalid persisted override falls back and repairsalignment/* events/align, /align-mode, direct establish_baseline, uninstall, and manifest restoration verified. External model completion was unavailable (QUOTA: Insufficient Balance) and is reported separately rather than claimed as an E2E pass.The v0.4.0 session-scoped-mode gate verified:
session-mode and session-mode-store suites: four-layer resolution, two-session isolation, /align-mode session, fork inheritance, durable sidecar, identity binding, per-agent registration failure rollback)/align + /align-migrate in the agent's own scope; Manual keeps tools + commands; Off registers nothing (only the plugin-scope /align-mode survives). Two live sessions hold disjoint capability sets with zero leakage.setOverride/clearOverride are durable-first (failed writes never commit); identity binding prevents id-reuse leakage; fork children inherit the parent override once and become independent.QUOTA: Insufficient Balance) — reported separately, never claimed as a mechanism pass.QUOTA.The v0.4.1 Web capsule + DSH rc.1 compatibility gate verified:
@deepseek-ai/* change; no DSH Core patch)./align-mode and the management API never claim the target is
active. The only advertised non-converged states are the explicit
capability-degraded and pending source-compensation ones (both exposed on the
status payload with the ACTUAL active capability mode).0.1.1-rc.1 exactly; KNOWN_SESSION_EVENT_TYPES
is the official rc.1 semantic set (alignment/* types are never runtime-
registered — the invariant is intersection with the official known set, not a
hand-maintained count).parentSession/seedLength) migrate and reload through the real rc.1 reader
with seq continuity, header, packed chunk rows, resume end-seed, and fork
lineage preserved. Only whitelisted legacy events become ignorable;
everything else is byte-preserved.0.1.1-rc.1 DSH installation: the
tarball installs via the real dsh plugin command (bundle reconciled),
--dump-config composes the plugin's two rows, a real headless boot mounts
the plugin service and the full capability matrix is verified live — policy
section in the assembled prompt, establish_baseline + report_drift
tools, /align + /align-mode from the real registries. /align-mode
persistence fails loud on the entry-only port when storage-domain is absent
(verifies the no-silent-live-only contract); removing the bundle leaves no
leftover rows. External model completion was unavailable (QUOTA) and is
reported separately, never claimed as a model E2E pass.shell.overlay registration no longer
uses a function-valued label; the render test asserts exactly one accepted
registration.The v0.4.2 DSH rc.2 + client-graph gate verified:
0.4.2.dsh-client-runtime +
dsh-client-locale; test/client-manifest.test.ts prevents the pure/core
dsh-client-ui-slots package from returning and ties source + built bundle
to one shell.overlay occupant using the slots + locale services.0.1.1-rc.2 packages. The packed smoke checks the physical dsh
launcher, headless/base core, commands, session, storage, storage-domain,
and settings versions before boot and fails on any drift.alignment/* events./align//align-mode/hot switching/remove,
with no leftover package or config row.QUOTA: Insufficient Balance, while the credential-free release
recheck returned MISSING_CREDENTIAL. Model-dependent E2E is NOT RUN and
is not included in the green deterministic result.Detailed evidence and the bounded-run caveat are recorded in the repository acceptance report.
pnpm install # dependencies (Node >=22.18.0; Node 24 also verified)
pnpm run typecheck # tsc (src + test)
pnpm run lint # eslint (src + test)
pnpm run build # tsc → lib/
pnpm test # node:test
pnpm run check # all of the above
Real dogfooding (boots real dsh profiles with an isolated DSH_HOME; smoke mode for development, full suite + natural benchmark + packed add/rm smoke for the RC gate):
powershell -File scripts/dogfood.ps1 -Smoke
powershell -File scripts/dogfood.ps1 # full correctness suite (RC gate)
powershell -File scripts/dogfood.ps1 -Benchmark05 # natural drift benchmark
powershell -File scripts/packed-smoke.ps1 # packed add/rm smoke (RC gate)
# run a single scenario:
powershell -File scripts/dogfood.ps1 -Scenario 05-arch-shift
See docs/ARCHITECTURE.md for the design decisions and the exact capability seams used.
0.1.1-rc.2 (verified against the npm registry releases and
a real rc.2 DSH installation through the packed-artifact add/boot/remove
smoke; migration parity is byte-for-byte with the rc.2 writer/reader).@deepseek-ai/cordis 4.x, @deepseek-ai/dsh-* 0.1.1-rc.2 (exact pins, no
prerelease range drift).>=22.18.0. Development/test gates pass on Node 22.23 and 24.18; CI
pins the lower supported line at Node 22.18 and also runs Node 24.alignment/status events still count as manual checks, and a session without the new events simply reports revision 0 / "unknown" instead of crashing.MIT. See LICENSE.
?sessionId=undefined request can ever be built. Verified in unit tests E–H
against the production bundle with a real hooks/effects renderer, and the
rebuilt bundle + loopback management API guard contract were verified live
against the running DSH Web 0.1.1-rc.1.