v0.8.0-alpha.8 — Windows Canonical-Path Hotfix

v0.8.0-alpha.8 — Windows Canonical-Path Hotfix

Cross-platform repository checks: PASS

Exact release artifact lifecycle: PASS

Goal Governor live E1: NOT RUN

Outcome value / multi-client portability: NOT PROVEN

This append-only hotfix supersedes alpha.7 without rewriting its public tag or evidence. The alpha.7 GitHub Actions matrix passed on Ubuntu but failed one assertion on both Windows Node versions because the runner supplied an 8.3 short temp path while Quickstart persisted the native canonical long path.

• The test now derives its expected identity with fs.realpathSync.native, exactly like production.
• Project-root confinement remains bound to canonical path, filesystem device and inode.
• Safe installer lifecycle, review-first Quickstart and the real-process offline Governor demo are otherwise unchanged from alpha.7.
• The final alpha.8 tarball is built from a clean commit, hash-bound through package-manifest.json and SHA256SUMS, and accepted through isolated dry-run, install, backup, force upgrade, rollback, uninstall, uninstall rollback and packaged Quickstart flows.
• No model, remote API or Live E1 execution was used for this hotfix.

The retained failure and artifact evidence are documented in release lifecycle and Windows CI finding. Alpha.7 remains available for audit but is superseded by this release.

v0.8.0-alpha.7 — Safe Trial and Guided Governance

Warning

Superseded by v0.8.0-alpha.8. Alpha.7 is retained for audit because its first Windows CI run exposed a canonical-path test expectation fixed in alpha.8.

v0.8.0-alpha.7 — Safe Trial and Guided Governance

Safe install lifecycle: PASS

Guided review scaffold: PASS

Real-process offline Governor demo: PASS

Local Researcher output probe: FAIL (honestly retained)

Goal Governor live E1: NOT RUN

Outcome value / multi-client portability: NOT PROVEN

This release makes the alpha safer and easier to evaluate without broadening
its product claims.

• dsh-researcher install --dry-run performs compatibility and target
  preflight without writing installer-owned paths. Automatic detection may
  still start external dsh --version; use an explicit package binding when
  that external process must be avoided.
• Exact DSH 0.1.0-rc.7 is required by default. Empty rc.7 CLI version output
  can be verified through constrained @deepseek-ai/dsh package metadata;
  missing, wrong or untrusted metadata fails closed.
• backup, uninstall and rollback preserve both managed preset states in
  content-hashed snapshots. Force upgrades and rollbacks create their own
  recovery points.
• project-cognition quickstart creates an external review workspace and
  removes manual hash transfer. It cannot approve goals, install verifiers or
  promote Project Cognition.
• npm run demo starts two real verifier child processes in an isolated
  fixture. Assistant prose and exit code 1 do not end the goal; only the
  matching captured exit code 0 reaches DONE. DSH envelopes remain an
  explicitly simulated offline boundary.
• The README now separates the currently trialable Project Research layer from
  the experimental Goal Governor value claim.

Two local Project Research probes were also run through DSH Web. Qwen3 14B
reached SAFE but lost the task and misused git_read; local DeepSeek R1 14B
produced an unsupported report and was rejected by the terminal doctor gate.
No successful report is fabricated from these runs. They support runtime-gate
behavior and leave Researcher outcome value unproven.

The Live E1 audit found that the final run still requires a pinned readable
rc.7 module root, capture-only visible-tool evidence, an immutable run lock,
dedicated cost controls and owner input in a real interactive TTY. No Live E1
execution occurred; the readiness audit invoked no model or remote API, and no
synthetic bundle is presented as live evidence. The only model calls in this release cycle were the two local
Ollama Researcher probes above; no remote API or model fee was used.

v0.8.0-alpha.6 — DSH Web installer truth alignment

v0.8.0-alpha.6 — Truthful DSH Web Onboarding

Published Researcher DSH Web smoke: PASS

Installer truth alignment: PASS

Goal Governor live E1: NOT RUN

Outcome value / multi-client portability: NOT PROVEN

The exact alpha.5 GitHub release tarball was installed into a fresh DSH_HOME
and launched from PowerShell through dsh web. It obtained a full SAFE
Researcher Runtime Certificate and rejected a later switch to Workspace Write
before another model response. No model-created file change occurred.

That post-release test found stale installer guidance. Current DSH Web exposes
Read Only with approval=ask; Project Research safely tightens approval to
never and the UI becomes Custom. Alpha.6 corrects Node, PowerShell and bash
installers and adds a consistency test. Runtime code is byte-identical to
alpha.5, so this release is the selected final acceptance version.

npx -y github:TLNing260310/dsh-researcher#v0.8.0-alpha.6 --force

No remote API or paid model was used. This remains a Researcher runtime smoke,
not protocol-defined Goal Governor E1 or outcome validation.

v0.8.0-alpha.5 — DSH Web Recompose and Terminal Safety

v0.8.0-alpha.5 — DSH Web Recompose and Terminal Safety

Researcher DSH Web smoke: PASS

Goal Governor live E1: NOT RUN

Outcome value / multi-client portability: NOT PROVEN

This release repairs three boundaries found by real local-model DSH Web
testing: post-creation preset attachment, prose-only completion without a
doctor result, and stale SAFE capability after permission drift.

Use Read Only before selecting Project Research. The preset tightens DSH
Web's approval policy from ask to never, attaches refusing write/edit stubs,
and reports Custom in the UI. Writable sandboxes are rejected.

npx -y github:TLNing260310/dsh-researcher#v0.8.0-alpha.5 --force

The local smoke is intentionally not labeled E1. One local Qwen run produced a
full SAFE runtime certificate but failed the requested project review; local
DeepSeek-R1 failed to call doctor. This release improves runtime reliability,
not demonstrated research outcome value.

The smoke report remains provisional evidence. Canonical Project Cognition is
not automatically promoted by this release; promotion still requires explicit
owner review, seal, install, and deterministic projection regeneration.

v0.8.0-alpha.4 — Fail-Closed Model Cost Admission

v0.8.0-alpha.4 — Fail-Closed Model Cost Admission

E1 infrastructure: READY

Live E1 / DSH re-scan: NOT RUN

Outcome value / multi-client portability: NOT PROVEN

This prerelease makes the user's DeepSeek test-cost rule part of the frozen E1 evidence chain. It performs no DSH installation, live model session or paid API call.

What changed

• The active Goal Governor protocol is frozen as v1.1. v1 was superseded before any live run (0 runs) and remains retrievable by exact alpha.3 tag, commit, blob path, byte length and SHA-256 in the v1 archive record.
• Manifest and run-lock schemas freeze base_url. Remote is exactly provider=deepseek-official, model=deepseek-v4-flash, base_url=https://api.deepseek.com; local also fixes provider=deepseek-official so the DSH DeepSeek-compatible adapter resolves a no-trailing-slash, explicit-port literal loopback base URL. Unknown routes and route/provider/model/base-URL drift fail closed.
• Beijing weekday windows [09:00,12:00) and [14:00,18:00) reject DeepSeek API. Outside those windows and on weekends, remote E1 remains official-Flash-only at the exact official base URL. Weekends waive only the time blackout, not run lock, budget, cost acknowledgement or route constraints.
• Before every child launch, the outer runner creates a frozen settings file with watch=false and injects the locked DEEPSEEK_BASE_URL. The child revalidates the effective URL through DSH's public DeepSeek resolver before create/resume and before and after every model followup. Admission is also recalculated before output, before DSH spawn and in each resumed process; the reserved interval and absolute deadline cap the child runtime.
• Evidence bundles preserve admission receipts—including a pre-spawn denial receipt when launch is refused—and bind process start, end and timeout to the absolute deadline. The offline scorer reconstructs successful-run decisions from frozen policy, route, timestamps and budget rather than trusting an ALLOW label.
• Historical Phase A protocols, locks, bundles and evaluation/runtime/eval-headless.mjs are explicitly audit-only and cannot serve as a new model runner.

Upgrade from alpha.3

npx -y github:TLNing260310/dsh-researcher#v0.8.0-alpha.4 --force

--force replaces the two installed preset directories under DSH_HOME; hand edits inside those directories are not preserved. It does not modify repository-local Project Cognition, approved goals or the verifier registry.

Operational boundary

The policy is mechanically enforced by the official repository E1 runner and checked again by the offline scorer. A loopback base URL proves only the adapter's first hop is local; it does not prove the local service avoids proxying to a remote API. The policy also does not establish host-clock or scheduler integrity, OS-wide egress isolation, provider billing identity, cryptographic human identity, honest bundle production or the absence of out-of-band API calls. A real E1 operator should additionally configure trusted time, a dedicated key, provider-side limits and alerts, and process/host egress controls appropriate to the desired assurance.

Reproduce the offline release gate

npm run check
git rev-parse HEAD
npm run release:artifacts -- --out <empty-directory-outside-repo> --expected-version 0.8.0-alpha.4 --expected-revision <full-git-object-id-from-above> --require-clean

Run the artifact command only after the final commit. The offline checks do not install or start DSH and do not call a model or network API.

Evidence boundary

Repository tests, doctor and E1 preflight establish mechanical behavior only. This release did not run DSH, live E1, a model or an API. The candidate still needs the DSH-dependent Gate 0 checks—including successful rc.7 resolution of the frozen remote and local routes—followed by all live E1 trajectories. In particular, the local route is implemented but not yet demonstrated runnable. Outcome value remains gated on the pilot and E2; a second adapter and model × client attribution remain gated beyond that.

v0.8.0-alpha.3 — Auditable Promotion and Stop Discipline

v0.8.0-alpha.3 — Auditable Promotion and Stop Discipline

E1 infrastructure: READY

Live E1 / DSH re-scan: NOT RUN

Outcome value / multi-client portability: NOT PROVEN

This prerelease closes correctness gaps found during the alpha.2 evidence audit. It adds no second adapter and performs no paid or live model run.

What changed

• Canonical state must be sealed. doctor no longer accepts a .project-cognition/state.json without a valid state_hash.
• The CLI now mechanically supports and guards the artifact transition draft → diff → seal --out → install; owner review remains an external repository-governance step, not an authenticated CLI event. Install accepts only the sealed artifact, binds the reviewed current hash, requires the exact next revision, and uses best-effort rollback for in-process state/projection write failures. doctor detects active/stale governance locks and a missing or mismatched canonical state/projection pair; it does not enumerate every temp/backup file, perform crash recovery, or claim cross-file power-loss atomicity.
• Goal decisions are recomputed from the preceding evidence. A logged terminal label cannot override the reducer; baseline and observation revision labels must stay aligned.
• goal status --format markdown gives a shared stopping card: MUST/gate state, attempts remaining, no-progress streak, time/tokens, and a single next action. Terminal cards explicitly say to stop polishing.
• Goal approval rejects unknown/superseded invariant references and orphaned local revisions.
• The E1 scorer now distinguishes valid conformance failures from passes. A complete failing bundle is FAIL_UNDER_TRUSTED_HOST, never a causal PASS.
• Raw E1 bundle bytes can receive a cryptographic commitment and an optional external Ed25519 signature. Verification proves that the private key corresponding to the supplied public key signed those bytes; it does not identify the key holder, prove DSH ran, show that the signer is honest, authenticate a TTY user, or establish outcome value.
• The scorer accepts one self-contained bundle directory only; split --run manifest and --artifacts roots are rejected. Existing path components for bundle, key, attestation, and output paths must not traverse symlinks or junctions, and evidence files must not be hard-linked.
• Release artifacts now include the package tarball, SHA256SUMS, and a machine-readable package manifest. A --require-clean build computes Git blob object IDs directly from working bytes (allowing only an equivalent CRLF checkout representation of an LF blob), disables replace-object aliases, and reconstructs its source from the verified HEAD blobs in an isolated directory; ignored files, arbitrary clean filters, replace refs, and local package-control files therefore cannot silently alter npm pack. CI runs once per main commit/PR across Windows and Linux at Node 22.12.0 and current LTS.

Upgrade from alpha.2

npx -y github:TLNing260310/dsh-researcher#v0.8.0-alpha.3 --force

--force replaces the two installed preset directories under DSH_HOME; hand edits inside those preset directories are not preserved. It does not run project-cognition init and does not modify repository-local .project-cognition/state.json, approved goals, or verifier registry. Existing canonical state therefore remains in place. An old editable cognition draft is not installable directly under alpha.3: copy the intended changes into a newly generated unique draft and review the new machine diff before sealing.

Canonical cognition migration

Installing an editable draft directly is intentionally no longer supported:

project-cognition cognition draft --root . --out <unique-review-dir-outside-repo>/state.r<N>.draft.json
project-cognition cognition diff <unique-review-dir-outside-repo>/state.r<N>.draft.json --root .
# owner reviews the exact diff and base.state_hash
project-cognition cognition seal <unique-review-dir-outside-repo>/state.r<N>.draft.json --out <unique-review-dir-outside-repo>/state.r<N>.sealed.json
project-cognition cognition install <unique-review-dir-outside-repo>/state.r<N>.sealed.json --root . --replace --expect-current-hash <reviewed-base-state-hash>
project-cognition doctor .

Each review path must be unique because the CLI refuses overwrite and install does not clean review artifacts. Keep draft/sealed files outside the repository; if policy requires a workspace-local review directory, make it explicitly gitignored. Never place them in .project-cognition/, commit them, or include them in a package. --expect-current-hash prevents stale or concurrent replacement; it is not proof of reviewer identity.

Optional E1 bundle signature

npm run eval:e1:attest -- create --run <bundle> --private-key <external-private.pem> --out <external-attestation.json>
npm run eval:e1:score -- --run <bundle> --attestation <external-attestation.json> --trusted-public-key <external-public.pem>

The key and attestation must stay outside the evidence bundle. The top-level scorer-generated score.json is the only file excluded from the signed commitment.

Reproduce the offline release gate

npm run check
git rev-parse HEAD
npm run release:artifacts -- --out <empty-directory-outside-repo> --expected-version 0.8.0-alpha.3 --expected-revision <full-git-object-id-from-above> --require-clean

Run this only after the final commit. The builder disables Git replace-object aliases, verifies HEAD/index consistency, independently computes each working file's Git blob object ID without invoking clean filters, and permits only exact bytes or the reversible CRLF checkout form of an LF blob. It then reconstructs an isolated source tree directly from the bound HEAD blobs, samples revision and cleanliness again after offline npm pack, validates required/forbidden package entries and npm integrity, and records the source revision plus clean status in the manifest. It writes the tarball, checksums, and manifest; it does not install or start DSH.

Goal boundary limitation

Goal Contract v1 freezes in_scope, out_of_scope, and do_not_touch as semantic strings. The generic Goal Core/DSH runtime does not compile them into a filesystem allowlist or independently detect changed paths. E1 alone has a frozen machine path contract through manifest allowed_changes, enforced by its dedicated runner/scorer. Generic hard path scope remains a schema v2 and adapter-conformance requirement; alpha.3 does not claim it.

Evidence boundary

Repository tests and preflight establish mechanical behavior only. This release did not run DSH or a real model, so the current candidate has not closed the DSH-dependent parts of Gate 0. The next evidence work is to complete those checks and then run live E1 as defined solely by the frozen Goal Governor Evaluation Protocol; pilot, E2, a second adapter, and E3 remain gated behind it.

v0.8.0-alpha.2 — Truth and Evidence Governance

v0.8.0-alpha.2 — Truth and Evidence Governance

This alpha does not add another memory system or client adapter. It makes the existing Project Cognition + Goal Governor prototype say exactly which state is authoritative, how research findings may enter that state, and what evidence must exist before broader product claims or integrations.

E1 infrastructure: READY

Live E1: NOT RUN

Outcome value / multi-client portability: NOT PROVEN

Protocol-defined E1 infrastructure

• A machine-readable manifest freezes the protocol-defined cases, deterministic fixture hashes, DSH 0.1.0-rc.7, expected terminals, allowed paths, equal budgets, invalidity rules and resume/replay semantics.
• npm run eval:e1:preflight materializes each frozen fixture twice in isolated temporary directories, validates the frozen inputs and runs only local fixture verifiers. It performs zero network calls and zero model calls.
• npm run eval:e1:score -- --run <external-bundle-dir> checks raw host sidecars, call/result uniqueness and order, exact verifier arguments, worktree evidence, external interactive-gate provenance, stage-one seal binding and live/offline replay checkpoints, then writes score.json as PASS, FAIL or INVALID. Assistant prose is not evidence. A conforming package is reported as PASS_UNDER_TRUSTED_HOST: the scorer trusts the experiment operator and model-inaccessible bundle root, and does not claim external attestation against a malicious host fabricating an entire self-consistent bundle.
• npm run eval:e1:live -- ... --ack-live-cost is reserved for the next stage. It fails closed unless the exact DSH version, clean committed inputs, external run lock, isolated paths, pinned installed candidate and explicit cost acknowledgement all match; an architecture gate still requires direct interactive TTY input. That establishes the input channel and command linkage, not cryptographic human identity.
• The historical locked Phase A runtime is unchanged. No DSH installation, live model session or paid call was performed for this release.

One canonical project truth

• .project-cognition/state.json is the sole canonical source for project purpose, architecture boundaries, value claims, invariants, decisions and proof order.
• PROJECT_COGNITION.md is a deterministic generated projection and must not be edited independently.
• The session-log-backed research-state mechanism is now named the Research Session Ledger. It remains useful for replayable provisional claims and hypotheses, but it is not owner-ratified project truth.
• A finding reaches canonical state only through an explicit draft revision, authority/proof/evidence review, owner review, CLI seal/install and regenerated projection.

The full procedure is documented in Project Cognition Governance.

Claims are now independently testable

• V3A — Project Cognition longitudinal value: does persisted, freshness-aware project cognition improve multi-stage maintenance compared with reconstructing context?
• V3B — Goal Governor incremental value: with equal Research handoff content, does host governance reduce false completion, scope drift and correction cost?

The E2 B-versus-C comparison can test V3B. It cannot by itself establish V3A. A-versus-C may be reported only as a whole-stack effect.

Experiment C+ remains causal-invalid for the historical superiority claim. A future valid experiment may create a new claim under a new experiment identity; it cannot rehabilitate or rewrite C+.

Proof order

The governed sequence is now:

Gate 0 → E1 → non-inferential pilot → E2 → second-adapter conformance → E3

The frozen Goal Governor Evaluation Protocol is the sole source for E1's trajectories, run definitions, estimands, thresholds and invalidity rules. The pilot validates measurement readiness but produces no confirmatory outcome claim. A second adapter is deferred until E2 passes; E3 is deferred until that adapter first proves governed capabilities and live conformance.

Doctor is not freshness

project-cognition doctor . verifies schema/hash integrity, deterministic Markdown projection, Goal Contracts and Verifier Registry. It does not recalculate repository evidence or prove that evidence remains fresh. Freshness claims require explicit observed fingerprints via project-cognition cognition freshness.

Compatibility and evidence boundary

• Node.js requirement is consistently documented as >=22.12.0.
• DSH 0.1.0-rc.7 remains the only implemented adapter.
• Repository mechanical tests and DSH preset discovery remain evidence for implementation behavior, not for long-term maintenance value, live-model terminal reliability or cross-client compatibility.

Verification

npm run check
npm pack --dry-run

v0.8.0-alpha.1 — Project Cognition + Goal Governor

v0.8.0-alpha.1 — Project Cognition + Goal Governor

This prerelease turns dsh-researcher from a read-only research preset into an alpha evidence-and-completion-governance layer for DeepSeek Harness.

What is new

• Portable Cognition, Goal and Verifier cores with strict JSON schemas, canonical SHA-256 binding, revisions and deterministic Markdown projection.
• A second DSH preset, Governed Coding, in which the model reports attempts and evidence but the host alone owns terminal goal state.
• /researcher <question> for one read-only turn, /researcher on|off for guarded mode, /researcher goal for a draft contract, and /researcher run for an approved contract.
• Frozen verifier registry invocations bound to earlier real DSH tool call/result IDs.
• Explicit ALREADY_SATISFIED, CONTINUE, NEEDS_HUMAN, DONE, BLOCKED, STOPPED and CANCELLED decisions.
• Hard attempt and no-progress limits so SHOULD-level polishing cannot run forever.
• A portable project-cognition CLI and a self-describing .project-cognition/state.json for this repository.

Trust and evaluation fixes

• Assistant prose no longer counts as completion telemetry.
• DON'T BUILD no longer double-counts as BUILD.
• Research state imports are strict, bounded and atomic.
• The runtime doctor compares live state with deterministic replay.
• Read tools are confined to the project root, including symlink resolution.
• Experiment C+ is now explicitly INVALID FOR CAUSAL CLAIM; snapshot leakage cannot be rendered as Researcher superiority.

Verification at release candidate

• 83/83 Node tests pass, including schema-valid public examples and an isolated tarball install/load smoke test.
• project-cognition doctor . passes all cognition, projection, contract and registry checks.
• Both presets scan as loadable after a temporary DSH 0.1.0-rc.7 installation.
• Package dry-run contains the portable core, schemas, presets, tests, selected evaluation summaries and documentation.

Not yet proven

This release does not establish long-term maintenance gains, Researcher superiority, live-model terminal reliability, or compatibility with Codex, Claude Code, Kiro, Zed or OpenClaw. The next evidence gate is the preregistered live DSH and three-arm maintenance protocol in goal-governor-evaluation-protocol.md.

Install

npx -y github:TLNing260310/dsh-researcher#v0.8.0-alpha.1

Node.js >=22.12 and DeepSeek Harness 0.1.0-rc.7 are the verified environment.