Deepr v2.37.0

Deepr v2.37.0 ships bounded, evidence-first expert investigations for persistent expert councils, with separate factual and perspective learning lanes and stricter no-surprise-bills controls.

Highlights

• Persistent experts can investigate a question with distinct evidence lenses, challenge and revise positions across rounds, and produce provenance-indexed artifacts without mutating expert state during the run.
• Facts and perspectives are kept separate. Hypotheses, theories, stances, concepts, null hypotheses, and original ideas can remain useful without being mislabeled as verified facts or human-reviewed work.
• Explicit bulk learning preflights every selected expert and lane, hash-verifies producer artifacts, locks all targets, and applies atomically. A budget is a ceiling, not consent.
• Remote MCP, REST, WebSocket, A2A, queue, and worker paths now preserve scoped identity and ownership while failing closed on unproven authority, metered consent, or accounting.
• Claude Code is the only currently executable plan-quota adapter, and only after a fresh provider observation proves paid extra usage is disabled. Other plan CLIs remain visible but execution-blocked until Deepr can prove safe confinement and billing posture.
• URL retrieval, outbound MCP, local approvals, graph commits, report absorption, deployment secrets, reservations, and the append-only cost ledger received broad security and lifecycle hardening.
• The council and capacity guides now give exact commands, cost semantics, learning behavior, and honest works-now versus planned boundaries.

Validation

• GitHub CI passed lint, strict type islands, frontend tests and packaging, dependency and secret audits, and the full unit suite on Python 3.12, 3.13, and 3.14.
• 9,203 unit tests passed locally with 85% branch coverage, above the blocking 80% threshold.
• CodeQL passed for Python, JavaScript/TypeScript, and Actions.
• Live validation used local Ollama and safely proven subscription quota only. Paid and metered-at-margin validation spend was exactly $0.00.
• The corrected three-expert pilot staged six candidate writes but applied none. Semantic superiority and held-out evidence quality remain explicit evaluation gates, not release claims.

See docs/CHANGELOG.md for the complete record.

Deepr v2.36.2

Deepr v2.36.2 hardens metered accounting fail-closed paths and soft cost admission without re-enabling live metered chat.

Highlights

• Dual confirmation for live metered expert chat: substrate flag plus DEEPR_ALLOW_METERED_EXPERT_CHAT=1. The substrate flag remains false.
• After durable settle, chat and research paths mirror session spend only (no second ledger write). Deep-research final usage ledger rows charge overrun only.
• Metered skill tools use durable fixed-cost admission when explicitly allowed; SkillExecutor defaults allow_metered_tools=False. Skill manifests clamp budgets and timeouts and treat unknown cost tiers as high.
• Embedding cache document and query embeds share the dual metered-chat confirmation gate.
• Soft cost preflights fail closed on bookkeeping failure across citation validation, gap discovery, map-reduce, multi-pass, curriculum, synthesis, and task planner.
• Documented the metered expert-chat re-enable review checklist.

Validation

• GitHub CI passed on Python 3.12, 3.13, and 3.14 including lint, frontend, core-install, security, and coverage gates.
• This release validation made no paid provider calls.

See docs/CHANGELOG.md for the complete record.

Deepr v2.36.1

Deepr v2.36.1 hardens plan-quota process ownership and lands the durable metered expert-chat admission substrate without re-enabling live metered chat.

Highlights

• Plan-quota CLI stdout/stderr are each hard-capped at 8 MiB raw bytes. Overflow kills and reaps the process tree, returns output_limit_exceeded, and pairs unknown quota usage with a $0 cost event.
• Windows Job Objects and Linux detached supervisors own CLI descendants before launch; unsupported POSIX platforms fail closed before dispatch.
• Metered expert-chat complete/stream, research, and embedding paths share reserve, dispatch-mark, and settlement helpers when execution is enabled.
• Sessions serialize turns and pass min(estimate, remaining session budget) as each call ceiling; missing max_tokens derive from half the dollar hold.
• Live metered chat remains fail-closed (METERED_EXPERT_CHAT_EXECUTION_ENABLED = False) until final deep-research usage settlement, optional metered skill tools, double-count audit, and explicit spend confirmation clear.

Validation

• GitHub CI passed on Python 3.12, 3.13, and 3.14 including lint, frontend, core-install, security, and coverage gates.
• This release validation made no paid provider calls.

See docs/CHANGELOG.md for the complete record.

Deepr v2.36.0

Deepr v2.36.0 closes the verification and cost-accounting loop before widening agent autonomy.

Highlights

• Direct research now uses one finite request envelope for preview, reservation, queued metadata, provider payload, and settlement.
• Expert consultations have durable lifecycle journals, bounded local or explicit plan synthesis, preserved dissent, and cancellation-safe accounting.
• deepr eval deliberation provides an offline 11-case structural evaluator at $0 without enabling live multi-round deliberation.
• Plan-quota CLI attempts now write paired idempotent quota and $0 cost events across cancellation, timeout, and runner failure.
• Scheduled local maintenance detects GPU contention and records bounded 30-minute, 2-hour, then 6-hour waits without paid fallback.
• The bundled Deepr research skill, README, roadmap, architecture, capacity guide, model guide, security policy, and deployment docs now describe works-now, visible, and gated capacity consistently.

Safety posture

• One fully priced bounded API research job remains supported under a hard approved ceiling.
• Standalone metered expert chat, unsafe lifecycle mutations, live provider benchmarks, hosted context, automatic paid fallback, and metered multi-call campaigns fail closed until they share the durable transaction contract.
• The AWS research deployment is explicitly an infrastructure preview and blocks paid submission at both API and worker boundaries.

Validation

• 8,426 unit tests passed, 8 skipped, and branch coverage reached 84.62 percent against the 80 percent gate.
• GitHub CI passed on Python 3.12, 3.13, and 3.14, including strict types, frontend, package, security, dependency audit, and SBOM jobs.
• CodeQL passed for Python, JavaScript/TypeScript, and Actions.
• Wheel and sdist both passed isolated install, version, doctor, and 11 of 11 $0 deliberation smoke tests.
• The v2.36 release validation made no paid provider calls.

See docs/CHANGELOG.md for the complete change and migration record.

Deepr v2.35.0

Deepr v2.35.0 hardens the complete expert-learning loop: evidence acquisition, bounded deliberation, persistent belief updates, capacity selection, cost settlement, browser chat, queues, and release packaging.

Highlights

• Expert councils now preserve bounded roster, capacity, dissent, truncation, and synthesis status in durable consult traces returned consistently by CLI and MCP.
• Scheduled local sync, sync-all, and gap work use best-effort GPU contention signals. Busy capacity records a durable wait with 30-minute, 2-hour, then 6-hour retry guidance and never falls through to a paid backend.
• Topic learning and learn-web require replayable fetched evidence, persist source packs and snapshots, reject unsupported candidates, and leave under-ready runs retryable without advancing freshness.
• API-backed absorption now reserves and settles every extraction, contradiction, deduplication, and adjudication call under one caller-supplied ceiling. Local and prepaid-plan paths remain $0.
• Browser expert chat now has explicit API-only admission, bounded session budgets, durable per-turn accounting, truthful cancellation, stable session restoration, and cost breakdown reconciliation.
• CLI version and root help now avoid heavy provider and expert imports. Windows p95 startup measured about 223 ms for version and 217 ms for root help.
• GitHub Releases is now the working binary channel. Installers and deepr upgrade resolve the exact repository-hosted wheel and fail safely on missing or invalid release metadata.
• A measured architecture decision keeps Deepr Python-first. Rust, Go, Mojo, and free-threaded Python remain evidence-gated options for specific proven seams rather than speculative rewrites.

Validation

• Hosted CI passed lint, strict typing, security and SBOM, core-only install, frontend build and packaging, and unit coverage on Python 3.12, 3.13, and 3.14: https://github.com/blisspixel/deepr/actions/runs/29175666752
• CodeQL passed for Python, JavaScript/TypeScript, and Actions: https://github.com/blisspixel/deepr/actions/runs/29175666584
• Local no-key validation: 7,997 passed, 8 skipped, 84.77 percent branch coverage across 8,005 collected tests.
• Frontend validation: lint, TypeScript, production build, and 26 tests passed.
• The committed frontend payload reproduced byte-for-byte with Node 22.23.1 on Windows and Linux.
• Dogfood spend remained $0.20, within the $2 ceiling, with no additional paid calls.

Install

python -m pip install https://github.com/blisspixel/deepr/releases/download/v2.35.0/deepr_research-2.35.0-py3-none-any.whl

PyPI publication is not part of this release.

SHA-256

• deepr_research-2.35.0-py3-none-any.whl: A2710DE2EEE61FFE1CDB3F2A6F78894DCDDDB71381B4CA75AB0E5E059AB6D029
• deepr_research-2.35.0.tar.gz: ADC9FE0E20A25EAFCADB8AC714F6F9310CBAE082FB63EE2C13E17BA1D7586B80

v2.34.4

Highlights

• Adds deepr expert next NAME, a read-only, $0 structural navigator over
  claims, freshness, gaps, contradictions, and verified learning-loop evidence.
• Returns canonical argv arrays, checks capacity before scheduled compiled
  sync, and never executes work or claims semantic maturity.
• Keeps legacy profiles readable without persisting migrations from the
  read-only command.
• Confines belief reads to the exact containment-validated file selected below
  the configured expert root, with regressions for directory, file, and
  basename-switch symlink escapes.
• Corrects expert freshness guidance, local vector-store blockers, storage-root
  defaults, and sequential multi-device safety documentation.
• Adds research-backed roadmap designs for ExpertEventV2, forgetting-aware
  improvement, historically grounded Leonardo-informed and Beethoven-informed
  perspectives, and modern agent-harness control patterns.

Validation

• 7,696 unit tests passed with 8 intentional skips.
• 84.56 percent branch coverage against the 80 percent gate.
• Python 3.12, 3.13, and 3.14 CI passed.
• Ruff, formatting, strict mypy islands, file-size, complexity, security,
  documentation consistency, frontend lint, frontend tests, TypeScript,
  production build, package build, dependency audit, Gitleaks, CodeQL, and
  Dependency Graph passed.
• Two CodeQL path alerts were dismissed as documented false positives only
  after exact containment validation and adversarial symlink regressions proved
  the sinks safe.

Upgrade

pip install --upgrade deepr-research==2.34.4

v2.34.3

Security and lifecycle hardening

• Make provider cleanup identifiers and cost reservations server-owned, reject them atomically across public submissions, and redact them from public job views.
• Bind status and cancellation to each job's recorded provider.
• Confirm cancellation only after queue, cost, and provider-resource cleanup closure.
• Make completion settlement and cleanup retry-safe before durable terminal status.
• Keep unknown provider statuses active and observable without exposing provider-controlled diagnostics.
• Return fixed public metadata errors across REST and web submission paths.
• Treat cancelled work as terminal in Research Live and preserve retryable cancellation failures in context.
• Correct current command, support, and lifecycle documentation.

Verification

• 7,675 unit tests passed, 8 skipped, with 84.45 percent branch coverage.
• 230 focused lifecycle regressions and 15 frontend tests passed.
• Ruff, formatting, strict mypy, documentation consistency, repository ratchets, lock validation, package build, dependency audit, secret scan, frontend lint, TypeScript, and production build passed.
• Both original no-network cleanup authorization PoCs stop before provider deletion target selection.
• Validation spend: $0.

Deepr v2.34.2

Security

• Prevent unexpected provider response content from entering benchmark validation console or dashboard-captured output.
• Treat unexpected provider responses as failed validation and return a nonzero command exit.
• Return a fixed benchmark-start validation response without exposing caught exception details.

Verification

• 7,577 unit tests passed with 8 intentional skips and 84.43 percent branch coverage.
• Ruff, strict mypy, frontend lint, 13 frontend tests, production build, documentation consistency, repository ratchets, lock verification, and package build passed.
• Regression tests cover secret-bearing provider output, secret-bearing validation exceptions, failure totals, and process exit status.
• Two independent reviews found no remaining medium or high issue.
• Validation spend was $0.

Deepr v2.34.1

Product trust and recovery

• Bound dashboard benchmark approval to the exact tier, quick-run, and judge settings that produced its estimate.
• Added scoped recovery states across expert, trace, benchmark, and settings surfaces without discarding valid cached data.
• Preserved failed benchmark output and added direct setup and retry paths.

Spend and reliability hardening

• Added durable per-call reservations and canonical ledger settlement for benchmark evaluation, judge, and validation calls.
• Enforced adapter output and bounded search maxima before provider submission.
• Failed closed on unpriced models and provider paths without deterministic request-level spend ceilings.
• Serialized destructive demo-data actions and restored benchmark execution from any working directory.

Validation

• Full Python branch-coverage suite and frontend lint, tests, and production build pass locally.
• Strict typing, Ruff, docs consistency, security and complexity ratchets, package lock, package build, and dependency audit pass locally.
• Validation used no provider calls and cost $0.

Deepr v2.34.0

Deepr v2.34.0

Product trust and research continuity release.

• Research Studio now fails closed when OpenAI readiness or cost estimation is unavailable and reports actionable recovery state.
• Queue, ledger, stale-cleanup, live-update, and runtime-version semantics now agree across dashboard and documentation surfaces.
• Research drafts are validated and restored within the current browser tab without persisting uploaded file data.
• Existing drafts are preserved until a different linked prompt is explicitly accepted.
• Overview, Research, Results, and the global status bar now provide clearer narrow-screen hierarchy and accessible control state.
• Frontend draft validation, privacy, arbitration, and failure tests are a blocking CI gate.
• Package metadata now uses current SPDX license fields and builds cleanly as sdist and wheel.

Validation: 7,559 Python unit tests at 84 percent branch coverage in local Cycle 1 validation, 8 frontend draft-state tests, Ruff, strict mypy, TypeScript, frontend production build, documentation consistency, structural ratchets, lock validation, package builds, dependency audit, secret scan, CodeQL, dependency graph, and exact-commit GitHub CI are green. Validation spend: $0.