# DeepSeek Harness Shared Analysis and Frozen-Cwd Evidence Bundle ## Traceability - Spec ID: deepseek-harness-shared-analysis-evidence-bundle - Story: #104 - Status: Implemented - Approved scope: [Issue #104](https://github.com/QoderAI/better-harness/issues/104) ## Intent DeepSeek Harness (DSH) already exposes independent persisted Session evidence and configured filesystem Skills plus cwd-sensitive Instructions. The canonical `/better-harness` workflow cannot combine those sources because DSH is rejected at the shared Asset Practices, Harness Report, and Evidence Bundle gates. This Story qualifies those three shared capabilities together. Evidence Bundle v3 freezes canonical cwd as part of evidence identity; Asset Baseline v2 retains the minimum current configured-snapshot provenance needed to keep configuration separate from historical Session observation. Existing generic inventory, lint, integrity, Project Harness, task-loop, and neutral Harness analysis owners remain canonical. The implementation is one Story and one PR. Report rendering and Checkup remain unsupported. ## Privacy blocker resolution The privacy blocker is resolved for the scope of #104 by inheriting the current shared cross-host Session Evidence behavior: - bounded sanitized ordinary user-intent prose may remain in `session-core-facts.candidates[].request.summary`; - existing credential, recognized-secret, identifier, injected-context, transcript-tail, and selected private-path sanitization remains in force; - the independent lead may retain its existing bounded request sample only where the shared lead contract already permits it; - #104 adds no DSH-specific Session privacy projector and no new lead text path. This inherited boundary is not claimed to be privacy-complete. The earlier DSH Session Evidence specification documented exposure of this same shared field as a pre-existing unresolved privacy concern. That concern remains unresolved and outside #104. Any future tightening belongs in shared Session and lead privacy owners for every host, not in a DSH-only Bundle rule. ## Story boundary Before #104, DSH advertises: ```text SESSION_ANALYSIS AGENT_CUSTOMIZE ``` After #104, DSH advertises exactly: ```text SESSION_ANALYSIS AGENT_CUSTOMIZE ASSET_PRACTICES HARNESS_REPORT EVIDENCE_BUNDLE ``` It remains absent from: ```text REPORT_RENDERING CHECKUP ``` The three new gates land together only after their executable routes and tests are complete. `HARNESS_REPORT` means neutral shared Harness evidence, not HTML, Markdown, Canvas, Studio, or another durable output. ## Canonical pipeline ```text /better-harness -> harness evidence-bundle -> Evidence Bundle v3 frozen context |- Session population -> DSH facts -> sessionEvidence |- existing Project Harness -> projectHarness |- DSH current configured inventory at frozen cwd | -> shared lint + public inventory + integrity | -> Asset Baseline v2 -> agentCustomize `- same Session population -> generic task-loop source -> independent current practice recollection at frozen cwd -> neutral evidence + summaryFacts -> lead -> complete / partial / failed Bundle -> specialists and lead reconciliation ``` Session and Project collection consume workspace, topology, analysis scope, and the frozen window. They do not consume configured cwd. Only current configured and practice collection consumes cwd. ## Acceptance scenarios ### AC-1: Exact capability promotion DSH gains exactly `ASSET_PRACTICES`, `HARNESS_REPORT`, and `EVIDENCE_BUNDLE` in addition to its existing two capabilities. `REPORT_RENDERING` and `CHECKUP` remain absent. ### AC-2: Evidence Bundle v3 and frozen cwd `EVIDENCE_BUNDLE_SCHEMA_VERSION` becomes `3`. Every emitted v3 context contains: ```js { workspace: "", cwd: "", provider: "dsh", language: "en" | "zh-CN", depth: "quick" | "normal", window: { since: "", until: "" }, evidenceLimit: 1 | 2 | 3 | 4 | 5, authority: { includeUserHome: boolean, includeMemories: boolean }, topology: "", analysisScope: "" } ``` The complete frozen context, including cwd, identifies the evidence collection. No new identity hash, v2 migration reader, or unrelated schema field is added. ### AC-3: Canonical cwd contract ```text effectiveCwd = explicit cwd ?? workspace ``` Workspace and cwd must exist as directories and are canonicalized through the real filesystem path. Canonical cwd must equal or be contained by canonical workspace. A lexical in-workspace symlink resolving outside workspace is rejected. The stable failures are `INVALID_CONFIGURED_CWD` and `CONFIGURED_CWD_OUTSIDE_WORKSPACE`. The implementation uses native `node:path` semantics and supports Windows, macOS, Linux, spaces, Unicode, drives, and UNC paths. Cwd is never derived from Session evidence. ### AC-4: Cwd propagation One canonical frozen cwd reaches Evidence Bundle context, Agent Customize, Asset Baseline, public configured inventory, Agent Lint, Harness report-run, and task-loop configured-practice collection. It does not become a Session analyzer or Session selection option and does not replace Project Harness's existing Git-root/workspace cwd. ### AC-5: Nested DSH Instruction selection For workspace `repo/packages/api` and cwd `repo/packages/api/src`, shared configured/practice collection includes an applicable `src/AGENTS.local.md`. With cwd equal to `repo/packages/api`, that nested local Instruction is not applicable. Existing #101 DSH native selection semantics are reused unchanged. ### AC-6: Asset Baseline v2 `ASSET_BASELINE_SCHEMA_VERSION` becomes `2`. A successful DSH baseline contains: ```js { kind: "agent-asset-baseline", schemaVersion: 2, status: "complete" | "partial" | "failed", scope: { provider: "dsh", workspace: "", cwd: "", includeUserHome: boolean, includeMemories: false }, configuredSnapshot, envelopes: { lint, inventory, integrity }, diagnostics } ``` Lint, inventory, and integrity consume one raw configured inventory snapshot inside the Baseline owner. Empty unsupported DSH collections are valid observed emptiness, not failures. ### AC-7: Compact configuredSnapshot A successful DSH baseline projects exactly: ```js { collectedAt: "", evidenceKind: "configured-not-observed", configurationSource: "qualified-defaults" | "caller-overrides", userHomeCollection: "included" | "not-authorized", instructionCollection: "enabled" | "disabled-by-byte-limit", qualification: { provider: "dsh", version: "0.1.1-rc.2", sourceSha: "b150a551b8d465e31e418e1b2eaf5e79bbb7d28e" }, runtimeResolution: { cordis: false, profile: false, preset: false, runtimeSkills: false } } ``` It excludes diagnostics arrays, configured paths, bodies, digests, symlink targets, and duplicate cwd. Failed raw inventory does not fabricate the field. ### AC-8: Current configuration is not historical observation The analysis may state that an asset is currently configured, when it was collected, or that current configuration was not observed in bounded Session evidence. It must not infer that a current Skill or Instruction existed during, was used by, or influenced a historical Session; that a same-name current asset is the historical asset; or that current absence proves historical absence. ### AC-9: Generic Skills and Rules practice coverage `collectProviderInventory()` derives `summary.practiceCoverageRows` from shared configured surfaces. Non-empty DSH Skills and Rules produce their existing generic rows with identity-deduplicated counts and bounded locators. Empty collections produce no phantom row. Representative existing generic hosts and Qoder Memory behavior remain compatible. ### AC-10: Asset Practices qualification DSH uses the existing public inventory, `agent-assets-review` lint, integrity, and Asset Baseline owners. No DSH-specific lint or integrity engine is added. User-home collection remains default-closed. ### AC-11: Bundle-facing path privacy Direct DSH Agent Customize retains its qualified lexical evidence. Bundle-facing Baseline findings use only: | Source | Locator | | --- | --- | | workspace itself | `` | | inside workspace | `/` | | inside Git root above workspace | `/` | | authorized home | `~/` | | unsafe off-tree | omitted or existing `` redaction | The rule applies to lint/integrity finding paths, owner routes, coverage paths, and bounded free text. Symlink target realpaths never enter output. ### AC-12: Configured content privacy Raw Skill bodies, Instruction prose, configured digests, symlink target realpaths, provider diagnostic content, raw credentials, and raw recognized secrets remain absent from Asset Baseline, the configured Bundle lane, and lead configured evidence. ### AC-13: Inherited Session request-summary boundary `PRIVATE_ORDINARY_PROSE_X` may survive in the existing bounded `request.summary`. `sk-test-secret-credential` and a recognized private home path must not survive raw. A bounded ordinary request sample may survive only in the existing eligible active-long lead projection. The old DSH request-summary privacy concern remains unresolved and outside #104. ### AC-14: Generic Harness Report DSH is admitted to `harness analyze` and reuses the existing task-loop source, report-source validation, neutral evidence formatter, and `summaryFacts` projection. Configured-practice collection receives frozen cwd; Session and repository scopes remain unchanged. `--canvas-out` remains Qoder/Cursor-only. ### AC-15: Multi-owner current snapshots P0 retains `collectionMode: "frozen-context-multi-owner"`. The Baseline shares one raw snapshot among its stages; the lead independently recollects current practice evidence using the same frozen parameters. Object identity, timestamp equality, and atomic filesystem snapshotting are not promised. ### AC-16: Project Harness remains generic DSH Bundle collection invokes the existing Project Harness owner with topology, analysis scope, and Git-root/workspace scope. Configured cwd is not forwarded to Project Harness. No DSH-specific Project Harness module is added. ### AC-17: Bundle completeness For normal depth, complete topology, all three specialist lanes available, and an available lead produce `complete`; a partial/unavailable specialist, incomplete topology, or unavailable lead produces `failed`. For quick depth, an incomplete specialist/topology with an available lead produces `partial`; lead or population/binding failure produces `failed`. Valid observed emptiness may remain available. ### AC-18: `/better-harness` flow and isolation The Skill's single Step 1 command passes `--platform dsh`, `--workspace`, and `--cwd`. Explicit inline/no-files DSH analysis may proceed through specialists and reconciliation. Durable HTML, Markdown, Canvas, Studio, and report-file output are not fabricated. Checkup remains unsupported. ## Public CLI The only new general option is `--cwd `. It is accepted by Evidence Bundle and the direct configured/practice and Harness analysis surfaces that must reproduce the same DSH Instruction selection. Existing `--dsh-home` and authority flags retain their current meaning. No byte-budget, candidate, snapshot, path-mode, or atomic-snapshot control becomes public. ## Evidence Bundle completeness | Component | Available/complete | Partial | Failure/rejection | | --- | --- | --- | --- | | Session | Valid facts and compatible binding | `unobserved`/`partial` coverage | Invalid discovery/facts/binding | | Project Harness | Evidence pack `status: ok` | Valid non-`ok` pack | Invalid envelope/collector error | | Agent Customize | Baseline v2 complete | Baseline v2 partial | Failed/invalid Baseline | | Lead | Valid evidence, summaryFacts, binding | No partial lead state | Invalid/unavailable/conflicting lead | | Normal Bundle | All available and topology complete | Not emitted | Any incomplete required input | | Quick Bundle | All complete | Incomplete specialist/topology with lead available | Lead/population/binding failure | ## Failure boundaries - Invalid cwd fails before configured asset reads. - Session discovery failure makes Session and lead unavailable. - Missing configured assets may be valid observed emptiness. - User-home disabled means not authorized, not globally absent. - Unresolved DSH runtime composition remains explicit provenance, not failure. - Rendering and Checkup requests continue through their existing unsupported capability contracts. ## Non-goals - report rendering, HTML, Markdown, Canvas, Studio, or durable output routing - Checkup, lifecycle, Quickstart, or the complete default durable workflow - complete Cordis, Profile, Preset, MCP, Plugin, Hook, or Custom Agent support - runtime/in-process/scoped Skill enumeration - atomic configured-filesystem snapshots - causal configured-to-Session correlation - DSH Session Analysis or configured-assets native-semantic redesign - DSH-specific lint, integrity, report, Project Harness, or privacy owners - shared Session privacy tightening or arbitrary-prose removal - configured asset body/content collection - schema migration infrastructure - upstream DSH changes - package/release/version/changelog/roadmap work ## Plan and tasks 1. Add RED tests for the exact capability set, Bundle v3, Baseline v2, and cwd. 2. Add RED tests for DSH Asset Practices, shared coverage, provenance, paths, configured content, and temporal separation. 3. Add RED tests for generic Harness analysis, Bundle composition/status, Project Harness isolation, Skill flow, rendering, and Checkup isolation. 4. Implement one shared canonical cwd resolver and propagate it only through current configured/practice paths. 5. Implement generic inventory/Baseline qualifications and path compaction. 6. Implement generic report and Bundle admission, then promote capabilities. 7. Update only the canonical DSH capability/reference documentation owners. 8. Run focused tests, full tests, Node 22, cross-platform CI, doc-link checks, package verification, and change-traceability review. ## RED test plan and review evidence | Acceptance | RED proof | | --- | --- | | AC-1 | Exact host capability projection and negative gates | | AC-2–AC-4 | Bundle schema/context, cwd validation, identity, and propagation spies | | AC-5–AC-7 | Nested DSH fixture, Baseline v2, and configuredSnapshot assertions | | AC-8 | Historical Session/current configuration negative claims | | AC-9–AC-10 | Generic inventory, lint, integrity, and compatibility tests | | AC-11–AC-12 | Synthetic path and configured-content canaries | | AC-13 | Existing shared summary sanitizer plus DSH Bundle admission canary | | AC-14–AC-16 | Report-run, task-loop, multi-owner, and Project Harness tests | | AC-17 | Normal/quick complete, partial, and failed matrix | | AC-18 | Better Harness Skill and capability-isolation tests | RED is valid only when the pre-change relevant suite is green and every new failure reaches missing #104 production behavior rather than a syntax, fixture, mock, import, platform, or Node-version error. The implementation begins only after the RED contract is independently reviewable.