Changelog

Version 0.6.6 - “沈黛君 (Shen DaiJun)”

In memory of a mother who always gave and never asked anything in return.

This release implements PSP-9’s governed agentic platform: every stochastic action is a proposal, every effect is mediated by a deterministic kernel, every accepted state is measured on the realized candidate, and every reliability claim is conditional on recorded evidence — following the Stability is All You Need papers I, II, and III. On top of it, PSP-10 makes the loop’s prompts, resident context, and search governed artifacts of the same record.

Governed Tool Loop & Admissibility Kernel (PSP-9):

  • SRBN Tool Loop - New perspt-agent::toolloop implements Paper II’s harness contract as an executable loop: model-issued tool calls become Paper III proposals, the five-clause admissibility kernel (Definition 3.2) decides each one, and the acceptance gate reads the re-measured candidate — never the model’s account of it. An adversarial “lying model” fixture proves a model claiming success while every effect is denied can never reach a hard pass.

  • Five-Clause Kernel - check_full_admissibility replaces hardcoded contract/barrier passes with registered evaluators; an absent clause is recorded absent, never true. Promotion is one transaction debiting exactly the barrier’s certified increment c_t — the allowance and the debit are the same number.

  • Operational Safety Barrier - perspt-coding registers six versioned policy channels (protected paths, sandbox escape, secret exposure, network reachability, dependency policy, resource limits) with exact deterministic increments; correctness evidence stays in V.

  • Recovery Lattice - One shared, non-replenishing cascade budget across all five levels (Definition 8.1): exhaustion forces strict escalation, containment is unconditional, and every cascade terminates within b + k steps in one of Theorem 6’s four terminal classes.

Model Portfolio & Transport:

  • Multi-Provider Portfolio - The [providers] table holds several credentialed routes in one process; perspt providers prints the capability matrix with declared-vs-probed honesty. A one-entry portfolio keeps single-provider setups unchanged.

  • ModelTransport Port - The SDK declares the object-safe transport port; perspt-core provides the genai tool-calling driver; perspt-agent::transport is the single adapter — the tool loop cannot name a vendor type or credential (Gate S).

  • Portfolio Routing - Fully qualified provider::model routes with capability objectives, decorrelated-family review preference, and cross-provider failover chains. Failover is recovery, not retry, and draws on the shared cascade pool.

  • Typed Tool Catalog - Declarative footprints (writes ⊆ reads by construction), capability-filtered tool specs, and an open LanguageId adapter registry. sed_replace/awk_filter leave the model-facing catalog; edit_file fails closed on ambiguity. External tool servers register under the kernel and can never exceed session authority.

Measured Certificates:

  • Realizability Interface - srbn upgraded 0.1.1 → 0.3.0 (stabilize_realized, restore-best policy, hash-chain ledger). Coding uses unmeasured realization with projection_mismatch telemetry; the analytic Theorem 12.4 path is admitted only when every hypothesis — including Theorem 12.1’s step-size bounds — is represented by evidence.

  • Independence Statistics - Matched-stratum pairwise misses, degenerate marginals using direct joint misses, conservative Hoeffding upper bounds, and certification gating; ρ_eff reports n beside it and the correlation-driven energy attenuation helper is removed.

  • Conformal Feasibility Floor - A declared budget ρ fixes n ≥ 1/ρ − 1 labeled samples before anything can be autonomously accepted; reject-all is reported as insufficient calibration, never as a satisfied budget. Calibration is keyed by full exchangeability strata with shadow/active/stale readiness epochs.

  • Audit Replay - perspt replay <session> folds the durable event stream deterministically with tamper detection and no provider credentials; context compaction preserves an exact ControlFrame and unresolved tool calls, and a stale checkpoint is rebuilt, never patched.

Recovery, Resume & Calibration:

  • Recovery Ladder - The runtime now holds Theorem 6’s higher rungs live: level 2 revises the work graph with the exhausted attempt’s evidence and re-runs at a new generation, level 3 hands the node to the configured architect route, and level 4 durably revokes the session’s authority epoch — also wired to perspt abort and the TUI quit path.

  • Exact Mid-Loop Resume - Every gate acceptance writes a durable candidate checkpoint (control frame plus content-addressed file artifacts). perspt resume rebuilds the accepted candidate from it, re-mints a fresh capability from the grant intersection at the current durable epoch, and re-enters the loop with exactly the remaining budgets; a bumped epoch refuses the resume.

  • Delayed Audit Labels & Conformal Activation - Every promotion persists an audit-selected, unlabeled calibration sample; perspt audit ingests delayed labels (single-assignment) and recomputes the stratum threshold over labeled samples only. At the finite sample floor a new immutable epoch activates, after which hard-pass candidates above theta commit without a prompt with the certified accept ledgered; any other state backs off to approval.

  • Grant Ceilings Enforced - GrantPolicy::mint intersects every live capability with the grant ceilings (effects, paths, commands, deny-by-default network, approval strictness) and fails closed; persisted grant intent is signed over canonical bytes and verified against a local trust anchor on resume, never the embedded key.

  • Behavioral Provider Probes - perspt providers --probe observes each configured route on a live two-tool round trip (tool calling, multi-tool selection, parallel batching, schema validity) with evidence labelled behavioral; all three configured routes passed live.

  • Exploration-Only Mode - perspt agent --exploration-only runs an interactive explorer tool loop under a strictly read-only capability: every call passes the kernel and mutation attempts are recorded denials.

  • Parity Gate Passed - The governed tool loop ran against the ungoverned whole-file ablation baseline live: 100% hard-pass with zero false stability, failures preserved. No measurement justifies a typed whole-file tool yet; PSP-9 status moves to Final. The parity_bench harness is retired by PSP-10; evaluation moves to the optional perspt-benchmark crate, run manually.

  • Governance Surfaces - A dashboard Governance page shows authority epochs, signed grant intent, calibration epochs, pending delayed audits, and validator verdicts; perspt status, perspt ledger, and perspt abort now read and act on the PSP-9 surfaces instead of printing placeholders.

PSP-9 Completion (all phases Done):

  • Open Execution Plane - Tool execution moved behind an exact-name handler registry with a conformance fixture proving a new family registers at the composition root (catalog entries + handlers + derived grants) without editing the loop, the candidate, or the node assembly. The legacy 15-arm executor dispatch and its ungoverned sed_replace/awk_filter/run_command paths are deleted; grants derive from the assembled catalog intersected with an explicit withheld set.

  • Shipped Tool Families - A read-only system explorer (sys_info, sys_processes, sys_disk, sys_env — variable names only, values withheld) and a local DB explorer (db_list, db_schema, db_query — SELECT-only over an in-memory read-only DuckDB view with statement allowlisting and row/byte caps) under two new deliberate read-only effect kinds, SystemProbe and DataRead.

  • Domain Registry Wired - The runtime selects its domain package from the open DomainRegistry (--domain or detection); the research domain is registered and reachable; the closed CodingLanguage enum is retired in favor of the LanguageId adapter registry; PluginRegistry accepts registrations.

  • Governed Dependency Mutation - mutate_dependencies resolves through each language plugin’s dependency commands (cargo add, uv add/uv lock``+``uv sync, npm install) behind an explicit --allow-dependency-mutation grant, with Manifest/Lockfile footprints, package-name injection guards, pre-image journaling, and write-ahead external-effect bracketing; manifest and lockfile promote like any governed mutation.

  • Ledger-Folded Conversation (Gate O) - Model context is a pure fold over digest-chained conversation deltas recorded before they are applied; the durable checkpoint’s control frame carries the rolling fold digest, and resume refolds the ledger and refuses a conversation that is not derivable from the record.

  • Commuting Tool-Call Batches (Gate P) - Same-turn calls are planned over declared footprints: commuting mutators run in the model’s own arrival order, precise-footprint reads execute concurrently, and a colliding pair is returned as a ToolBatchConflict observation, never given an invented order.

  • Bounded Multi-Node Dispatch (Gate P) - --max-parallel-nodes (default 1, requires --yes) enables a governed forced-choice update_graph planning turn and a dispatcher running per-node lifecycle futures on borrowed structured concurrency with cancellation tokens, one shared non-replenishing recovery pool, first live GraphWrite lease use, and single-flight promotion; --max-parallel keeps its verifier-fan-out meaning.

  • Proposal Ensembles Superseded (Gate M/T) - The live proposal ensemble ([ensemble] / --ensemble-width) is removed by PSP-10 in favor of the bounded search forest: each branch is still one ledgered gate decision in its own reversible overlay with selection strictly by measured energy, and a lingering [ensemble] block is now a fatal startup error pointing to [exploration].

  • Certified Pairwise Bounds (Gates R/T) - The deterministic verifier suite records a verdict row beside the adjudicator’s for the same candidate and stratum; perspt audit labels verdicts with the calibration sample in one pass; the independence estimator runs over labeled matched verdicts and surfaces per-pair statistics in perspt status and the governance dashboard — Certified with n or the literal insufficient evidence.

  • Latest MCP Client Wired Into Runtime and Chat (Gates K/L/U) - The custom compatibility transport was replaced by the official Rust SDK and a strict MCP 2026-07-28 server/discover lifecycle with no legacy fallback. Stdio and stateless Streamable HTTP now cover tools, resources/templates, prompts/completions, explicit roots, opt-in sampling/elicitation, typed notifications, subscriptions/listen, MRTR, and bounded tasks. Remote tools remain locally admitted, bracketed, and replayed; protocol context primitives use reserved read-only operations.

  • MCP TUI and Boundary Hardening - /mcp and /help expose status, policy rejection, and elicitation actions; chat accepts interactive form/URL requests while the non-interactive agent declines them. Sampling preserves the local model route and forwards generation controls, subscription changes re-admit catalogs, task and sampling budgets are bounded, stdio children inherit no ambient credentials beyond executable-resolution variables, and bracketed paste now inserts Unicode/multiline clipboard content into the chat input.

  • Diagnostics and Rollback (Gate R) - The dashboard renders the work-graph revision lineage (Topology), a Backlog page with the arriving-potential gauge Φ(W) labeled as conditional diagnostics, and the governance independence pair table; perspt status shows Φ(W), the withheld topology_gap, and certification state; perspt ledger --rollback <session> restores promoted pre-images through the hardened descriptor-relative path, extends the hash chain with the rollback event, and labels the candidate UNSAFE — the delayed-label source for calibration and independence.

  • Promotion TOCTOU Hardening - Promotion, resume recovery, and rollback share one engine. Unix uses openat/renameat/unlinkat with O_NOFOLLOW component walks and held directory descriptors. Native Windows uses write-through replacement and rejects observed reparse points in its explicitly weaker reduced-isolation mode. Rollback failures are reported, never swallowed. The scheduler now tracks refined generations exactly.

  • Recovery Continuity - Transient provider errors (429, 5xx, timeouts) are retried with bounded exponential backoff inside the transport, per Gate P’s rule that rate-limit waits are throughput delays; recovery-ladder rungs reseed from the previous attempt’s best accepted candidate state instead of re-paying the whole cost (ladder_reseeded); and containment caused by exhausted transport preserves the session’s authority epoch so perspt resume remains viable after an outage, while governance containment still revokes.

  • Legacy Retirement - ~19,000 lines of orphaned PSP-5 orchestrator code, dead TUI entry points, the legacy dashboard LLM/Sandbox pages, and the empty legacy status branch are removed; remaining pages read the PSP-9 ledger.

Typed Prompts, Paged Context & Bounded Search (PSP-10):

  • Typed Prompt Sections & Codegen - The new perspt-prompt-macros crate compiles typed prompt sections into prompt programs; every compiled program and every model call is ledgered with ordered section provenance, the tool-surface hash, and its digest. perspt prompts (list, render, lint, manifest, explain-session) inspects and maintains the section libraries, and perspt agent --allow-experimental-prompts substitutes validated [prompts] bundle sections live — experimental until a change record passes paired evaluation (Gate AE).

  • Resident-Context Paging - Model context is a paged resident set with reserves declared in [context]: evicted pages stay addressable through the context_recall tool, every compaction and refusal is a recorded context event, and perspt context explain-turn explains a session’s recorded events.

  • Bounded Search Forests - After a gate failure the runtime can open a bounded search forest ([exploration]; at most three branch identities, sequential eager branches): exact no-goods are learned from rejected candidates, and every branch’s workspace reservation is durable and pre-action with cost-free refusals.

  • Multi-Node Dispatch with Integration Gate - --max-parallel-nodes dispatches work-graph nodes into per-node staging overlays behind one global integration gate, so concurrent nodes promote through a single measured decision point.

  • Evolving Test Evidence - Production coding gates now default to [verification] test_policy = "evolving": the resulting implementation, tests, and project configuration are measured together, so an intentional contract change is not forced to satisfy a superseded test. backward-compatible explicitly adds the restored historical regression view, while external-oracle overlays separately protected acceptance material on a private candidate copy and adds its command as a blocking test-stage verdict. The selected policy is shared by node, search-preview, and integration measurements and recorded at session start; benchmark hidden oracles remain withheld and fixed.

  • Crash Resume - perspt resume rebuilds a crashed session from the durable record, judging the work graph by its latest recorded state and refolding the conversation by digest chain.

  • Optional Benchmark Crate - The new perspt-benchmark crate (perspt-cli Cargo feature benchmark, non-default) evaluates and compares Perspt with other coding agents over a 30-task hidden-oracle corpus: suites smoke (1 arm, 8 tasks), adaptive (2 arms), and full (7 arms). It is credentialed and run manually only — never part of cargo test or CI.

Engineering Discipline:

  • Reproducible Toolchain - Added rust-toolchain.toml pinned to Rust 1.97.1 while retaining edition 2021; workspace packages inherit the same MSRV.

  • DuckDB 1.5.5 Boundary - Exact-pinned duckdb and libduckdb-sys at 1.10505.0, added dynamic ABI rejection, transactional versioned migrations, catalog-checked nullability changes, and post-DDL checkpoints.

  • Recoverable WAL Repair - perspt db repair --db-path PATH --discard-wal records WAL size and SHA-256, makes durable timestamped database/WAL backups, quarantines the WAL, verifies recovery read-only, and restores the original WAL on failure.

  • Shared MCP Runtime Foundation - Introduced the governed external-runtime boundary, mode filtering, local effect/footprint admission, namespacing, bounded observations, environment-referenced secrets, replay-only observations, and explicit reconciliation. The current release supersedes its initial tool-only transport with the official latest SDK client described above.

  • PSP-5 API Retirement - Removed the legacy orchestrator and Merkle-ledger facade from the public agent crate and refused resume of PSP-5 sessions; their existing rows remain inert forensic data.

  • PSP Code Check - New xtask checker enforces file ≤ 1408 lines, function ≤ 70 code lines (measured with syn, not brace counting), and line ≤ 108 columns, with a shrink-only baseline ratchet wired into CI (./check-rules.sh).

  • Workspace Decomposition - The 5,889-line orchestrator, types.rs, plugin.rs, store.rs, tools.rs, ledger.rs, and verification.rs split into focused submodules; workspace version fields now inherit from [workspace.package].

  • Native Windows Capability Contract - TUI chat, simple-chat, MCP, and the coding workflow compile natively. Coding process execution stays fail-closed by default and requires --allow-unisolated or [verification] allow_unisolated = true until a restricted-token/ACL sandbox backend exists; CI tests strict refusal and the acknowledged native verifier path separately.

  • Audit-Driven Hardening - A full-code audit closed a candidate-restore gap (files created after a checkpoint now roll back exactly via a pre-image journal), separated read paths from the promotable mutated set, routed promotion itself through the five-clause kernel, made the policy engine consume canonical command lines, inverted network-scope delegation to deny-by-default, and made ledger appends O(1) via stage-then-commit.

Version 0.6.2 - “Hózhó”

Hózhó (Navajo) — A state of perfect balance, harmony, and continuous self-improvement.

This iteration focuses on bringing greater stability, balance, and refinement to the core architecture. It introduces the full platform SDK implementation alongside the mathematical constructs outlined in the Stability is All You Need paper series (Papers I, II, and III) and formalizes PSP-8.

Stability Is All You Need & Platform SDK (PSP-8):

  • SDK Platform Crates - Introduced perspt-sdk, perspt-coding, and perspt-research to separate domain-neutral orchestration logic from target execution domains.

  • Measured Acceptance Gate - Implemented the PSP-8 spectral acceptance gate using symmetric eigensolvers (via nalgebra) to compute energy-slope convergence guarantees.

  • Goal Verification Loop - Integrated target validation kernels with directed LLM corrections and structured GoalVerdict/TaskType contracts.

  • Symbol Extraction & Verification - Added runtime parsing, dependency tracing, and semantic verification to enforce correct output states.

Core Capabilities & Providers:

  • Vertex AI Support - Added seamless Google Cloud Vertex AI provider support with automatic OAuth2 Bearer token resolution via gcp_auth ADC.

  • 29 LLM Adapters - Expanded provider configurations to handle 29 unique genai adapters, modernizing model defaults (e.g., gemini-3.5-flash, gpt-5.5, claude-fable).

  • Pipenv Integration - Added support for pipenv environment isolation to the Python execution plugin.

Documentation Overhaul:

  • Stability Foundations - Reworked core concept and developer documentation to align with the theoretical foundations of the Stability paper series.

  • Sphinx Book Polish - Redesigned the Sphinx book layout featuring a retro scientific textbook LaTeX cover, custom plots, and enhanced tables.

  • Refined Clinical Tone - Rewrote the user manual, tutorials, and configuration guides in a formal, clinical documentation style.

Workspace Maintenance & Upgrades:

  • Ratatui Upgrade - Bumped ratatui dependency to 0.30.2 in perspt-tui to support modern terminal layouts and components.

  • TUI Textarea Fork Migration - Migrated from the unmaintained tui-textarea to tui-textarea-2 version 0.11.0 to resolve API compatibility conflicts.

  • Dependency Upgrades - Bumped rustyline to 18.0, toml to 1.1, and tower-http to 0.7 across the workspace.

  • GitHub Actions Security - Upgraded GitHub Action workflows to use actions/checkout@v7 for improved stability and security.

Version 0.6.1 - “AKU”

AKU - sharp fixes from sharp ears.

Config Coherency & Schema-Driven Settings:

  • TOML Config Schema & Resolution - Added a robust TOML configuration schema with smart config-driven provider resolution.

  • Unified Provider Binding - Bound the configured provider cohesively across all modes including chat (TUI), simple-chat (CLI), and the autonomous agent mode.

  • Refined Config Commands - Redesigned perspt config command to be fully structured, interactive, and improved key initialization/init prompts.

TUI & CLI Slash Commands:

  • CLI Simple-Chat Enhancements - Integrated rustyline file-based history and introduced interactive slash commands inside simple-chat.

  • TUI Input Navigation & Persistence - Added persistent history paths and implemented fully UTF-8 safe input navigation in TUI inputs.

  • Conversation Persistence - Added new conversation save commands to easily export dialogues to local files.

Advanced Terminal UI Rendering:

  • LaTeX Math Transpilation - Integrated real-time LaTeX mathematical transpilation into the markdown rendering pipeline.

  • Intelligent ASCII Table Wrapping - Implemented self-wrapping logic for ASCII tables in the chat UI to ensure readable presentation on narrow views.

Documentation:

  • Comprehensive Feature Guides - Fully documented TUI/CLI slash commands, config schemas, and keyboard shortcut matrices in the user guides.

  • PSP-8 Stability Documentation - Added a dedicated Perspt Book chapter explaining how the three Stability is All You Need papers shape agent mode, the SDK-first roadmap, and future domain plugins.

Workspace Maintenance:

  • Crate Version Alignment - Bumped version of all workspace crates to 0.6.1.

Version 0.6.0 - “kukuza”

Nurturing the foundation, empowering the core.

Workspace-Wide Dependency Upgrades:

  • duckdb Upgrade - Bumped duckdb requirement to =1.10503.1 in the workspace root.

  • askama Upgrade - Bumped askama requirement to 0.16 in perspt-dashboard.

  • diffy Upgrade - Bumped diffy requirement to 0.5 in perspt-agent.

  • genai Upgrade - Bumped genai requirement to 0.6.1 in perspt-core.

  • starlark Upgrade - Bumped starlark requirement to 0.14 in perspt-policy.

API Alignments & Adaptations:

  • genai API Alignment - Updated all_model_names inside llm_provider.rs to pass () for the new ProviderConfig parameter.

  • starlark API Alignment - Adapted policy loading in engine.rs to use Module::with_temp_heap since Module::new was deprecated and made private in 0.14.

Specification & Ecosystem Maintenance:

  • PSP-7 Finalization - Formally transitioned the PSP-7 specification to Final status under the PSP-000001 process.

Version 0.5.9 - “xinli guanghua”

Perfecting the essence until the work needs no words to shine.

PSP-7: Robust Correction Loop Contracts:

  • Structured Artifact Bundle Format - Switched correction prompt from free-form File: ... output to a strict JSON { artifacts: [], commands: [] } schema. Includes exact target paths from evidence so the LLM targets the correct files, reducing parse failures.

  • AgentTools Integration - Routed correction commands through execute_correction_command() to integrate with plugin policy, user approval gates, and tool failure tracking.

  • Typed Parse Pipeline - Replaced Option-based bundle extraction with a 5-layer fail-closed typed parse pipeline. Added RetryClassification (Retarget, MalformedRetry, SupportFileViolation, Replan) population in CorrectionAttemptRecord.

  • Manifest Policy Enforcement - Added semantic validation to prevent implicit mutation of root manifests (Cargo.toml, package.json) unless explicitly listed as output targets, while preserving legal support files.

  • Strict Budget Exhaustion - Widened budget exhaustion checks from cost-only to any_exhausted() to properly respect step and revision caps before attempting LLM calls.

LLM Provider Maintenance:

  • genai Upgrade - Bumped genai dependency from 0.5.1 to 0.5.3 (stable patch release with bug fixes).

  • Dead Code Cleanup - Removed generate_response_with_history() and generate_response_with_options() which had zero callers across the workspace and were the only methods leaking genai::ChatOptions into the public API surface.

  • Clippy Fixes - Fixed clippy::unnecessary-sort-by and applied clippy::collapsible-match auto-fixes for Rust 1.95 compatibility.

Version 0.5.8 - “Qualitaetsveredelung”

Orchestration State Overhaul Release

“Qualitaetsveredelung - the craft of refining what exists until its quality speaks for itself. Not new features, but the quiet discipline of making every state transition truthful, every metric honest, and every dead path removed.”

Orchestration Correctness (Refs: #112, #113, #114, #116):

  • SessionOutcome enum - New SessionOutcome type (Success, PartialSuccess, Failed) derived from actual completed/escalated node counts. The Complete event now carries truthful outcomes instead of unconditional success: true.

  • NodeOutcome enum - execute_node() returns Result<NodeOutcome> where NodeOutcome is Completed or Escalated, replacing the previous Result<()> that could not distinguish outcomes.

  • Correct session outcome derivation - run_orchestration() and run_resumed_inner() track completed/escalated counts per node and derive the final SessionOutcome accordingly.

  • Always-on LLM telemetry - call_llm_with_logging() now records token usage (in/out), latency, and estimated cost via record_llm_usage() after every LLM call, regardless of --log-llm. The flag now only controls verbose prompt/response text persistence.

  • Budget envelope persistence - upsert_budget_envelope() called after each BudgetUpdated event to persist cost/step tracking to the database.

  • Sandbox-aware context retrieval - ContextRetriever in step_speculate() uses effective_working_dir(idx) (the node’s sandbox directory) instead of the workspace root. Sandbox file tree listings included in actuator and correction prompts for better generation grounding.

Type-Safe State Management (Refs: #114):

  • NodeState::from_display_str() - Case-insensitive canonical parser with legacy aliases (“running” -> Coding, “stable” -> Completed, “retrying” -> Retry). Replaces all ad-hoc string parsing across the codebase.

  • NodeState helpers - is_success() (true only for Completed), is_active() (true for Coding, Verifying, Planning, Retry, SheafCheck, Committing), and Display impl producing lowercase labels.

  • CLI state cleanup - All string-based state comparisons in status.rs, agent.rs, and resume.rs replaced with NodeState::from_display_str() and type-safe helper methods.

Dead Code Elimination:

  • Removed 16 unused functions across perspt-core, perspt-store, perspt-agent, perspt-policy, perspt-tui (~234 lines)

  • Downgraded canonicalize to pub(crate) in perspt-policy

  • Removed orphaned sha2 dependency from perspt-store

Bug Fixes (Refs: #107, #111):

  • Session status stuck at RUNNING - Status now persisted in end_session() with COALESCE-based finalization guarantee (#111)

  • LLM token counts always zero - Real provider token usage (prompt + completion tokens) extracted from genai ChatResponse::usage and persisted per request (#107, #110)

Tests:

  • 7 new tests covering NodeState parsing round-trips, SessionOutcome equality, NodeOutcome discriminants, and session outcome derivation from completed/escalated counts

  • Total test count: 359 (up from 352)

Documentation:

  • Updated README: fixed crate count (8 -> 9), deduplicated dashboard command, added missing agent flags (--single-file, --verifier-strictness, --output-plan, all --*-fallback-model), aligned contributing commands with CI gates

  • Updated Perspt Book: SRBN architecture (Phase 7 -> Commit & Outcome), CLI reference (logs always shows token metrics), developer architecture guide (orchestrator lifecycle, NodeOutcome, SessionOutcome in type inventory and data flow), workspace crates (removed dead is_safe_for_auto_exec)

  • Updated PSP-5: execution flow steps 8-11 with Completed/Escalated paths and SessionOutcome derivation, headless output with OUTCOME line, added Orchestration State Overhaul implementation appendix

Version 0.5.7 - “navikaran”

Dashboard UX Polish Release

“Bridging the purpose of Ikigai with the momentum of Kaizen - renewal through continuous, intentional refinement.”

Dashboard UX Improvements (PSP-6 continued):

  • Custom DaisyUI 5 themes - perspt-light and perspt-dark themes with orange/pink oklch palette (WCAG AA compliant), powered by @plugin "daisyui/theme" blocks

  • Theme toggle - Navbar button with sun/moon icons, localStorage persistence, and migration from legacy theme names

  • Friendly session names - Deterministic human-readable names (e.g. “bold-hawk”) derived from session UUIDs via hash-indexed adjective+noun arrays

  • Breadcrumb friendly names - All six session sub-pages show friendly name with UUID-on-hover tooltip

  • Session card layout - Stacked vertical cards with btn-outline sub-page buttons replacing ghost buttons

  • Task text formatting - whitespace-pre-line rendering for readable multi-line task descriptions

  • Collapse arrow fix - pe-10 padding on DAG and LLM collapse summaries to prevent arrow overlap with text

  • Decisions page resilience - All six store queries use unwrap_or_default() instead of ? early-return, preventing 503 errors on partial data

  • Paginated overview - 20 sessions per page with DaisyUI join pagination controls, backed by list_sessions_paginated() and count_sessions() store methods

  • Login page theme - Updated to perspt-light default

CI & Build:

  • Node.js in CI - Added actions/setup-node@v4 (Node 22) to CI test matrix and release workflows so npx @tailwindcss/cli runs on all runners

Store:

  • list_sessions_paginated(limit, offset) - LIMIT/OFFSET SQL for paginated session listing

  • count_sessions() - Total session count for pagination controls

Version 0.5.6 - “ikigai”

SRBN Sandbox Revision Flow Release

“A reason for being - the happiness of always being busy with what you love.”

Real-Time Web Dashboard (PSP-6):

  • perspt-dashboard crate - Axum 0.8 + Askama 0.15 + HTMX 2 + Tailwind v4/DaisyUI 5 web interface for monitoring agent execution

  • Read-only store access - SessionStore::open_read_only() with DuckDB AccessMode::ReadOnly for safe concurrent reads alongside the agent

  • Six monitoring pages - Overview (sessions), DAG (task graph), Energy (Lyapunov components), LLM (request telemetry), Sandbox (provisional branches), Decisions (escalations, sheaf validations, rewrites, plan revisions, repairs, verifications)

  • SSE live updates - Server-Sent Events stream node statistics every 2 seconds

  • Password authentication - Random token, HttpOnly/SameSite cookie, Secure flag on non-localhost deployments

  • ``perspt dashboard`` CLI command - Launches the dashboard server on a configurable port

  • 12 integration tests - Route smoke tests, SSE content-type, auth flow

  • Store extensions - get_session_energy_history(), get_all_sheaf_validations(), get_all_repair_footprints()

SRBN Sandbox Revision Flow (PSP-5 Phases 3-12):

  • PlanningPolicy - Adaptive agent gating with 5 policies (LocalEdit, FeatureIncrement, LargeFeature, GreenfieldBuild, ArchitecturalRevision). needs_architect() and needs_speculator() gate agent tier activation

  • FeatureCharter auto-creation - Policy-derived file/module/revision limits created before architect planning so the plan gate has bounds to enforce

  • Speculator lookahead gating - Speculator tier only activates for LargeFeature, GreenfieldBuild, and ArchitecturalRevision policies

  • BudgetEnvelope session restore - Step/cost/revision caps restored from DB during resume so interrupted sessions honour the original limits

  • Bundle path normalization - filter_bundle_to_declared_paths uses normalize_artifact_path for correct comparison of path variants (e.g. ./src/main.rs vs src/main.rs)

  • NodeState::Superseded - New terminal state for plan amendment (Phase 14 preparation). Updated is_terminal(), parse_node_state(), and NodeStatus conversion

  • Orchestrator module extraction - orchestrator.rs split into 9 submodules: mod.rs, bundle.rs, commit.rs, convergence.rs, init.rs, planning.rs, repair.rs, solo.rs, verification.rs

  • Centralized prompts - All 15 agent prompts consolidated in prompts.rs with constants and render_* helpers; duplicates removed from agent.rs

  • RepairFootprint-backed correction - build_correction_prompt uses RepairFootprint for precise, grounded repair context

  • Greenfield bootstrap ordering - Plugin-driven project initialization with correct pre-sheafify plugin re-detection

  • Provisional branch lifecycle - Sandbox-first execution with branch creation, merge, and flush cascade

  • Escalation classification - 5 categories with 9 rewrite actions and graph surgery support

Documentation:

  • Updated architecture docs with PlanningPolicy, FeatureCharter, and NodeState::Superseded documentation

  • Added Planning Policy and Feature Charter sections to SRBN architecture guide

  • Updated workspace crates docs with orchestrator submodule structure

  • Added speculator lookahead and budget restore documentation to agent mode guide

  • Fixed energy weight default (gamma=2.0) in advanced features guide

  • Refreshed all version references to 0.5.6

Version 0.5.5

PSP-5 Cross-Platform and CI Stabilization Release

Cross-Platform Fixes:

  • Windows sandbox path normalization - list_sandbox_files now returns forward-slash-separated relative paths on all platforms

  • Windows workspace-bound validation - validate_workspace_bound correctly detects absolute paths with Windows drive prefixes (C:\...) and normalizes backslash path separators before POSIX shell tokenization

  • Clippy and fmt CI compliance - Resolved items_after_test_module and useless_vec warnings that were failing CI on all platforms

Build and CI Improvements:

  • Removed accidental eval workspace member - .perspt-eval/rust_cli removed from workspace members; .perspt-eval/ added to .gitignore

  • Stabilized cargo doc - Added doc = false to CLI bin target to prevent output collision with the perspt library crate

  • Removed deprecated atty dependency - Replaced with std::io::IsTerminal for TTY detection

  • Lockfile refresh - Cleared hard cargo audit vulnerability failures via dependency updates

Documentation:

  • Updated workspace coding instructions to match current multi-crate architecture

  • Refreshed all version references across the Perspt Book and Sphinx configuration

Version 0.5.4

PSP-5 Compliance Release

SRBN Agent Enhancements:

  • Per-node error recovery - Retry logic respects ErrorType classification (Compilation, ToolFailure, ReviewRejection) with separate counters

  • Multi-file extraction - Actuator reliably extracts artifact bundles with multiple files from LLM responses

  • Multi-artifact bundles - Bundle protocol correctly handles write, diff, and command artifacts in a single node

  • Plugin-driven project initialization - All plugins use uv init --lib (Python), cargo init (Rust), npm init (JS) for proper project scaffolding

  • Degraded verification mode - When tool binaries are missing, falls back to heuristic verification with clear warnings

  • Sheaf validation - 7 validator classes for cross-node contract verification

Bug Fixes:

  • Fixed uv init to uv init --lib in plugin.rs (2 locations) and test_runner.rs (1 location) for correct src/ layout with [build-system]

  • Removed dead test_check_workspace_requirement test from orchestrator.rs

Documentation:

  • Complete rewrite of the Perspt Book for PSP-5 accuracy

  • Updated developer guide with full type inventory and architecture diagrams

  • Added tutorials for headless mode, Python ETL, Rust CLI, and scientific computing

  • Updated all model names and provider defaults to current versions

Version 0.5.3

  • SRBN energy convergence improvements

  • Provisional branch lifecycle management

  • Interface seal and flush cascade support

  • DuckDB migration from SQLite

Version 0.5.2

  • Initial SRBN orchestrator implementation

  • Per-tier model selection (--architect-model, etc.)

  • Merkle ledger with DuckDB backend

  • Plugin system for Python, Rust, and JavaScript

  • Agent TUI with dashboard, task tree, and review modal

  • Starlark policy engine

  • Basic sandbox command execution

  • 10 CLI subcommands (chat, agent, simple-chat, init, config, ledger, status, abort, resume, logs)

Version 0.5.1

  • Multi-provider chat support

  • TUI markdown rendering with code blocks

  • Simple CLI mode for scripting

  • Streaming response protocol with EOT signal

Version 0.5.0

  • Initial release

  • Basic chat interface with OpenAI

  • Configuration auto-detection from environment