chore(release): v0.8.0 #37

Merged

peersky merged 1 commit from release/quorum-rs/v0.8.0 into dev

2026-08-17 18:46:42 +00:00

forgejo-actions commented

2026-08-17 18:45:36 +00:00

Release v0.8.0

This PR was generated by the release-prepare workflow.

Review the changelog below and edit CHANGELOG.md directly if needed before merging.

[quorum-rs-v0.8.0] - 2026-08-17

🚀 Features

(openrouter) Per-agent web-search plugin config
(llms) Preserve HTTP status on non-400 4xx via LlmError::Api
(heartbeat) Add model_down flag so a dead remote model benches the agent
(worker) Report model_down on a 404 task failure, self-heal after cooldown
(agent-health) Escalate model-down bench backoff, reset on success
(agents) Resolve claim citations to exact proposal spans, error/retry on miss
(agents) Ground claim cites in the exec-agent path too (non-destructive)
(dashboard) Add agent diagnostics API (metrics + latest errors)
(dashboard) Surface paused/flagged state in agent diagnostics
(providers) Model-availability probe over a provider model catalog
(agent-health) Proactive model-availability self-bench via provider catalog
(dashboard) Guard the agent control plane with a bearer token
(agents) Self-reported agent health (healthy/degraded/down) on heartbeat
(dashboard) Fleet-wide API errors per agent (last 24h) view
(dashboard) NATS-backed agent event log powering 24h errors, tasks, and tool-calls views
(agents) One grounding seam for claim citations, with spans clients can highlight by

🐛 Bug Fixes

(remote) SSE auth via Authorization header, not ?token= query param
(crypto-core) Compile on wasm32-unknown-unknown
(tui) Clamp thread/nested scroll at bottom — no wrap to top
(prompt-exposure) Redact leaks post-recovery instead of retry-looping to a hard fail
(prompt-exposure) Re-scan after redact; block+retry if leak persists (H5)
(claude-recovery) Skip unreadable-dir sweep test when running as root
(cite) Stop whitespace-tolerant resolve panicking on a multibyte tail
(cite) Ground evaluation claims against the shown thought_process window
(context-tools) Resolve current-round anonymized candidates in read_proposal
(context-tools) Resolve current-round anonymized candidates in search_deliberation
(mcp-tools) Resolve current-round anonymized candidates in nsed_read_proposal/nsed_search
(llm-repair) Stop silently mis-repairing truncated numbers + cap python literal depth
(llm-repair) Anchor markdown eval columns + don't misclassify a proposal as a read
(llm-repair) Parity-check the incomplete-unicode trim so an escaped backslash survives
(xml-regex) Bound the tool-regex cache to stop unbounded growth
(scoring) Guard non-finite endorsement weights so a NaN can't mis-decide the winner
(middleware) Fail closed when a middleware guard can't be built
(dashboard) Stream job feed over fetch under auth + same-origin token guard
(dashboard) Fail closed — refuse a non-loopback bind with no token
(agents) Spell out café so the spell checker stops reading it as a typo
(agents) Say why a structured-output parse found nothing to parse
(dashboard) Stop a finished task reading as a failure, and reporting another phase's result
(agents) Give the tool-call bucket name a single owner
(agents) Do not ask a question there is no time to answer
(agents) Offer only the tools that have something to serve, and say why one was refused
(agents) Report a tool call that was abandoned instead of leaving it pending
(agents) Treat a retired model as down, not as a transient failure
(deps) Vendor the Swagger UI assets so docs.rs can build

💼 Other

(claims) Require verbatim claim quotes so the UI highlighter can locate them
(claims) Force verbatim claim quotes — no paraphrase (client substring-match)

🚜 Refactor

(agent-health) Abstract model-down detection behind a trait
(llm-repair) Dedup read_proposal/read_critiques positional arg parsing

📚 Documentation

(portal) Add frontmatter, short titles, mermaid diagrams for the /docs portal
(explanation) Add claim-citation grounding rationale
(openapi) Register orchestrator-budgets endpoint in agent dashboard spec
(reference) Add the openrouter agent-config block reference
(how-to) Add self-serve device registration guide
(portal) Frontmatter + short titles for device + provider-config guides
Tighten docs + make README an index (no PR/issue refs, less verbosity)
Link the hosted docs portal and MCP endpoint from the README
Drop the framework-name line from the docs index
Fix stale propose/evaluate log sample in first-agent tutorial
Repair cross-references that only resolved inside the superproject
Why a threaded conversation loses its live question
A tutorial for threading a conversation
Thread on the endpoint that can pick its council
Show the server-rendered form of a threaded turn
What the thread key is, and the ways a thread loses continuity
Say what a plain-LLM council actually needs from a thread

🧪 Testing

(tui) Fix stale selection assertion after the Down-clamp change
(llms) Lock chat_completion 4xx→LlmError classification end-to-end
(exec) Cover claim-cite grounding through evaluate()
(cite) Lock dash-label, list-bullet, and smart-single-quote forms
(context) Cover render_proposal offset-overflow and windowing
(context) Cover keyword filter on the candidate search branch
(llm-repair) Cover bare and max-hex incomplete unicode truncation
(llm-repair) Cover the python-AST depth-cap stack-overflow guard
(nsed-agent) Cover the no-op-redact fail-closed re-scan gate
(dashboard) HTTP-level coverage for the agent diagnostics endpoint
(openapi) Guard annotated-but-unregistered endpoints in agent dashboard
(agent-manager) Cover persist_agents_to_yaml config persistence
(crypto-core) Cover from_seed malformed-seed rejection
(native) Assert web-search search_prompt is emitted
(dashboard) Guard the bearer-auth audit fixes against regression
(agents) Stamp the phase on the completion this fixture pairs

⚙️ Miscellaneous Tasks

Classify a feature release as a minor bump
(release) V0.8.0

[quorum-crypto-core-v0.7.2] - 2026-07-30

🚀 Features

(serve) Register role-based workspace policies at boot
(middleware) Dylib entries accept config, injected into context metadata
(agents) Add middleware field to AgentConfig (deserialize-only)
(worker) Invoke before_prompt / on_provider_response / on_completion pipelines
On_job_complete hook — fire once at job-final via orchestrator's job_complete event
(agents) Middleware-declared proposal schema → forced structured tool call
(claude) Forceful MissingTerminalCall retry feedback — directive 'MUST call the tool now, no prose' so the subprocess complies with terminal-tool submission
(mcp) Nest the middleware-declared schema into nsed_propose's inputSchema
(mcp) Log nsed_propose input_schema override (runtime proof the nested schema is advertised)
(sdk) Branching thread model, session resume, delta-task efficiency
(tui) Email-style branching thread TUI + ask_user HITL
(serve) Inherit orchestrator address/token from ~/.nsed when config blank
(tui) Auto-open deliberation detail on send, ^C stop, full-editor steering
(tui) Resolve a thread's running job from the orchestrator, not local disk
(tui) Enter opens the live detail while deliberating; louder "no deliberation"
(tui) Multiline steering input + shared text-editing primitives
(workers) Per-agent max_concurrent_jobs to serialize racing jobs
(workers) ReAct retry on a provider_response reviewer block instead of failing the task
(observability) Opt-in outgoing prompt/request dumps (NSED_DUMP_PROMPTS_DIR)
(project-registry) Map project id -> agents holding the epic (pure core)
(project-registry) ProjectAdvertisement::from_verdict — build advert from before_prompt content
(project-registry) Publish project_advanced on job_complete (pull trigger)
(project-sync) Client epic replica — clone once, git pull on project_advanced
(project-registry) Worker advertises ProjectAdvertisement on before_prompt
(workers) Thread conversation_id to patch-deliberation for stable-worktree keying
(tui) App-level ask_user overlay — question surfaces + answerable on any screen
(tui) Inbox 'needs answer' CTA + fix ask_user answer never sending on reopen
(mcp-tools) Read-only file/line history tools + block raw git in deliberation
(crypto-core) DeviceIdentity — NATS user-nkey wrapper for self-serve register (T015)

🐛 Bug Fixes

(mcp) Unify proposal content-resolution across handler + recovery
(clippy) Restore #[allow(too_many_arguments)] on generate_structured_output
(mcp) Enforce the forced proposal schema on nsed_propose + retry on violation
(mcp) Detect rate-limit when claude exits before reading the prompt
(tui) Surface a failed deliberation instead of silently dropping the spinner
(tui) Readable thread tree — front indent, timestamp column, visible cursor on entry
(tui) Mark branch-offs as their own thread + drop the redundant re: tag
(mcp) Expose user tools (ask_user) to the claude agent — HITL clarify was dead
(mcp) Make ask_user unambiguous + integration-test the claude tool list
(tui) Resolve a thread's running job from persisted pending_job
(tui) Don't pin a stale "Fetching…" footer for background fetches
(tui) ^C stops the deliberation even without a locally-known job
(tui) Detect Ctrl+C under tmux (drop the Press-only gate) + show "Stopping…"
(tui) Revert ^D-breaking is_ctrl change; add tmux-proof x stop alias
(tui) Roll back the optimistic turn when the orchestrator rejects the submit
(agents) Run claude subprocess with cwd = per-job worktree override
(agents) Restore too_many_arguments allow on execute_phase_attempt
(claude) Resume (keep cache) after a quota-kill instead of restarting fresh
(react-loop) Validate the forced-schema envelope, not solution_content
(project-sync) Clear inherited git env so ops target the replica, not GIT_DIR
(claude) Mint a fresh uuid on post-collision restart, not the collided one
(claude) Don't clamp rate-limit wait to the phase budget — wait for reset, then continue
(mcp) Fresh-restart reuses the deterministic session uuid — stop the every-turn re-send
(serve) Wire the user-tool handler factory in build_worker so ask_user is advertised

💼 Other

(quorum-rs) Add fmt-check, clippy, audit make targets
(claude) Surface session-recovery outcome + retry failure for the lock race

🚜 Refactor

(agents) Rename content key pd_worktree → agent_working_dir
(workers) Reuse the react loop's retry for reviewer blocks — drop the second budget

📚 Documentation

On_provider_response idempotency note (reruns on retry) + debug log for dropped before_prompt transform
Middleware-declared proposal schema (forced structured output) + on_job_complete hook
Thread TUI — how-to, reference, explanation
(explanation) Add "Understanding rooms and policies"
(explanation) Add policy-as-model + session/thread explainer; forward-note rooms
(explanation) Document the agent working-directory override

⚡ Performance

(prompts) Static system message + per-turn header — restore prompt-cache reuse

🧪 Testing

(worker) Hook-invocation integration tests via mock middleware
(user-tools) Prove ask_user threads build_request -> AgentContext -> claude MCP surface
(ask_user) Prove claude_cli gets ask_user as user_ask_user over MCP

⚙️ Miscellaneous Tasks

Gitignore stray .git-worktrees/ dirs
TODO comments for deferred gaps (disable_native_tools enforcement, rmcp version pin)
(supply-chain) Fast-track crossbeam-epoch 0.9.20 (RUSTSEC-2026-0204)
(perf) Disable debuginfo for external deps to cut build RAM

Merging this PR fires release-publish.yml: tag creation + cargo publish for the affected crate(s).

## Release v0.8.0 This PR was generated by the `release-prepare` workflow. **Review the changelog below** and edit `CHANGELOG.md` directly if needed before merging. --- ## [quorum-rs-v0.8.0] - 2026-08-17 ### 🚀 Features - *(openrouter)* Per-agent web-search plugin config - *(llms)* Preserve HTTP status on non-400 4xx via LlmError::Api - *(heartbeat)* Add model_down flag so a dead remote model benches the agent - *(worker)* Report model_down on a 404 task failure, self-heal after cooldown - *(agent-health)* Escalate model-down bench backoff, reset on success - *(agents)* Resolve claim citations to exact proposal spans, error/retry on miss - *(agents)* Ground claim cites in the exec-agent path too (non-destructive) - *(dashboard)* Add agent diagnostics API (metrics + latest errors) - *(dashboard)* Surface paused/flagged state in agent diagnostics - *(providers)* Model-availability probe over a provider model catalog - *(agent-health)* Proactive model-availability self-bench via provider catalog - *(dashboard)* Guard the agent control plane with a bearer token - *(agents)* Self-reported agent health (healthy/degraded/down) on heartbeat - *(dashboard)* Fleet-wide API errors per agent (last 24h) view - *(dashboard)* NATS-backed agent event log powering 24h errors, tasks, and tool-calls views - *(agents)* One grounding seam for claim citations, with spans clients can highlight by ### 🐛 Bug Fixes - *(remote)* SSE auth via Authorization header, not ?token= query param - *(crypto-core)* Compile on wasm32-unknown-unknown - *(tui)* Clamp thread/nested scroll at bottom — no wrap to top - *(prompt-exposure)* Redact leaks post-recovery instead of retry-looping to a hard fail - *(prompt-exposure)* Re-scan after redact; block+retry if leak persists (H5) - *(claude-recovery)* Skip unreadable-dir sweep test when running as root - *(cite)* Stop whitespace-tolerant resolve panicking on a multibyte tail - *(cite)* Ground evaluation claims against the shown thought_process window - *(context-tools)* Resolve current-round anonymized candidates in read_proposal - *(context-tools)* Resolve current-round anonymized candidates in search_deliberation - *(mcp-tools)* Resolve current-round anonymized candidates in nsed_read_proposal/nsed_search - *(llm-repair)* Stop silently mis-repairing truncated numbers + cap python literal depth - *(llm-repair)* Anchor markdown eval columns + don't misclassify a proposal as a read - *(llm-repair)* Parity-check the incomplete-unicode trim so an escaped backslash survives - *(xml-regex)* Bound the tool-regex cache to stop unbounded growth - *(scoring)* Guard non-finite endorsement weights so a NaN can't mis-decide the winner - *(middleware)* Fail closed when a middleware guard can't be built - *(dashboard)* Stream job feed over fetch under auth + same-origin token guard - *(dashboard)* Fail closed — refuse a non-loopback bind with no token - *(agents)* Spell out café so the spell checker stops reading it as a typo - *(agents)* Say why a structured-output parse found nothing to parse - *(dashboard)* Stop a finished task reading as a failure, and reporting another phase's result - *(agents)* Give the tool-call bucket name a single owner - *(agents)* Do not ask a question there is no time to answer - *(agents)* Offer only the tools that have something to serve, and say why one was refused - *(agents)* Report a tool call that was abandoned instead of leaving it pending - *(agents)* Treat a retired model as down, not as a transient failure - *(deps)* Vendor the Swagger UI assets so docs.rs can build ### 💼 Other - *(claims)* Require verbatim claim quotes so the UI highlighter can locate them - *(claims)* Force verbatim claim quotes — no paraphrase (client substring-match) ### 🚜 Refactor - *(agent-health)* Abstract model-down detection behind a trait - *(llm-repair)* Dedup read_proposal/read_critiques positional arg parsing ### 📚 Documentation - *(portal)* Add frontmatter, short titles, mermaid diagrams for the /docs portal - *(explanation)* Add claim-citation grounding rationale - *(openapi)* Register orchestrator-budgets endpoint in agent dashboard spec - *(reference)* Add the openrouter agent-config block reference - *(how-to)* Add self-serve device registration guide - *(portal)* Frontmatter + short titles for device + provider-config guides - Tighten docs + make README an index (no PR/issue refs, less verbosity) - Link the hosted docs portal and MCP endpoint from the README - Drop the framework-name line from the docs index - Fix stale propose/evaluate log sample in first-agent tutorial - Repair cross-references that only resolved inside the superproject - Why a threaded conversation loses its live question - A tutorial for threading a conversation - Thread on the endpoint that can pick its council - Show the server-rendered form of a threaded turn - What the thread key is, and the ways a thread loses continuity - Say what a plain-LLM council actually needs from a thread ### 🧪 Testing - *(tui)* Fix stale selection assertion after the Down-clamp change - *(llms)* Lock chat_completion 4xx→LlmError classification end-to-end - *(exec)* Cover claim-cite grounding through evaluate() - *(cite)* Lock dash-label, list-bullet, and smart-single-quote forms - *(context)* Cover render_proposal offset-overflow and windowing - *(context)* Cover keyword filter on the candidate search branch - *(llm-repair)* Cover bare and max-hex incomplete unicode truncation - *(llm-repair)* Cover the python-AST depth-cap stack-overflow guard - *(nsed-agent)* Cover the no-op-redact fail-closed re-scan gate - *(dashboard)* HTTP-level coverage for the agent diagnostics endpoint - *(openapi)* Guard annotated-but-unregistered endpoints in agent dashboard - *(agent-manager)* Cover persist_agents_to_yaml config persistence - *(crypto-core)* Cover from_seed malformed-seed rejection - *(native)* Assert web-search search_prompt is emitted - *(dashboard)* Guard the bearer-auth audit fixes against regression - *(agents)* Stamp the phase on the completion this fixture pairs ### ⚙️ Miscellaneous Tasks - Classify a feature release as a minor bump - *(release)* V0.8.0 ## [quorum-crypto-core-v0.7.2] - 2026-07-30 ### 🚀 Features - *(serve)* Register role-based workspace policies at boot - *(middleware)* Dylib entries accept `config`, injected into context metadata - *(agents)* Add `middleware` field to AgentConfig (deserialize-only) - *(worker)* Invoke before_prompt / on_provider_response / on_completion pipelines - On_job_complete hook — fire once at job-final via orchestrator's job_complete event - *(agents)* Middleware-declared proposal schema → forced structured tool call - *(claude)* Forceful MissingTerminalCall retry feedback — directive 'MUST call the tool now, no prose' so the subprocess complies with terminal-tool submission - *(mcp)* Nest the middleware-declared schema into nsed_propose's inputSchema - *(mcp)* Log nsed_propose input_schema override (runtime proof the nested schema is advertised) - *(sdk)* Branching thread model, session resume, delta-task efficiency - *(tui)* Email-style branching thread TUI + ask_user HITL - *(serve)* Inherit orchestrator address/token from ~/.nsed when config blank - *(tui)* Auto-open deliberation detail on send, ^C stop, full-editor steering - *(tui)* Resolve a thread's running job from the orchestrator, not local disk - *(tui)* Enter opens the live detail while deliberating; louder "no deliberation" - *(tui)* Multiline steering input + shared text-editing primitives - *(workers)* Per-agent max_concurrent_jobs to serialize racing jobs - *(workers)* ReAct retry on a provider_response reviewer block instead of failing the task - *(observability)* Opt-in outgoing prompt/request dumps (NSED_DUMP_PROMPTS_DIR) - *(project-registry)* Map project id -> agents holding the epic (pure core) - *(project-registry)* ProjectAdvertisement::from_verdict — build advert from before_prompt content - *(project-registry)* Publish project_advanced on job_complete (pull trigger) - *(project-sync)* Client epic replica — clone once, git pull on project_advanced - *(project-registry)* Worker advertises ProjectAdvertisement on before_prompt - *(workers)* Thread conversation_id to patch-deliberation for stable-worktree keying - *(tui)* App-level ask_user overlay — question surfaces + answerable on any screen - *(tui)* Inbox 'needs answer' CTA + fix ask_user answer never sending on reopen - *(mcp-tools)* Read-only file/line history tools + block raw git in deliberation - *(crypto-core)* DeviceIdentity — NATS user-nkey wrapper for self-serve register (T015) ### 🐛 Bug Fixes - *(mcp)* Unify proposal content-resolution across handler + recovery - *(clippy)* Restore #[allow(too_many_arguments)] on generate_structured_output - *(mcp)* Enforce the forced proposal schema on nsed_propose + retry on violation - *(mcp)* Detect rate-limit when claude exits before reading the prompt - *(tui)* Surface a failed deliberation instead of silently dropping the spinner - *(tui)* Readable thread tree — front indent, timestamp column, visible cursor on entry - *(tui)* Mark branch-offs as their own thread + drop the redundant re: tag - *(mcp)* Expose user tools (ask_user) to the claude agent — HITL clarify was dead - *(mcp)* Make ask_user unambiguous + integration-test the claude tool list - *(tui)* Resolve a thread's running job from persisted pending_job - *(tui)* Don't pin a stale "Fetching…" footer for background fetches - *(tui)* ^C stops the deliberation even without a locally-known job - *(tui)* Detect Ctrl+C under tmux (drop the Press-only gate) + show "Stopping…" - *(tui)* Revert ^D-breaking is_ctrl change; add tmux-proof `x` stop alias - *(tui)* Roll back the optimistic turn when the orchestrator rejects the submit - *(agents)* Run claude subprocess with cwd = per-job worktree override - *(agents)* Restore too_many_arguments allow on execute_phase_attempt - *(claude)* Resume (keep cache) after a quota-kill instead of restarting fresh - *(react-loop)* Validate the forced-schema envelope, not solution_content - *(project-sync)* Clear inherited git env so ops target the replica, not GIT_DIR - *(claude)* Mint a fresh uuid on post-collision restart, not the collided one - *(claude)* Don't clamp rate-limit wait to the phase budget — wait for reset, then continue - *(mcp)* Fresh-restart reuses the deterministic session uuid — stop the every-turn re-send - *(serve)* Wire the user-tool handler factory in build_worker so ask_user is advertised ### 💼 Other - *(quorum-rs)* Add fmt-check, clippy, audit make targets - *(claude)* Surface session-recovery outcome + retry failure for the lock race ### 🚜 Refactor - *(agents)* Rename content key pd_worktree → agent_working_dir - *(workers)* Reuse the react loop's retry for reviewer blocks — drop the second budget ### 📚 Documentation - On_provider_response idempotency note (reruns on retry) + debug log for dropped before_prompt transform - Middleware-declared proposal schema (forced structured output) + on_job_complete hook - Thread TUI — how-to, reference, explanation - *(explanation)* Add "Understanding rooms and policies" - *(explanation)* Add policy-as-model + session/thread explainer; forward-note rooms - *(explanation)* Document the agent working-directory override ### ⚡ Performance - *(prompts)* Static system message + per-turn header — restore prompt-cache reuse ### 🧪 Testing - *(worker)* Hook-invocation integration tests via mock middleware - *(user-tools)* Prove ask_user threads build_request -> AgentContext -> claude MCP surface - *(ask_user)* Prove claude_cli gets ask_user as user_ask_user over MCP ### ⚙️ Miscellaneous Tasks - Gitignore stray .git-worktrees/ dirs - TODO comments for deferred gaps (disable_native_tools enforcement, rmcp version pin) - *(supply-chain)* Fast-track crossbeam-epoch 0.9.20 (RUSTSEC-2026-0204) - *(perf)* Disable debuginfo for external deps to cut build RAM --- > Merging this PR fires `release-publish.yml`: tag creation + `cargo publish` for the affected crate(s).

forgejo-actions added 1 commit

2026-08-17 18:45:36 +00:00

chore(release): v0.8.0

Publish Release / Create Tag & Sync (pull_request) Successful in 15s

Details

Publish Release / Publish dry-run (no side effects) (pull_request) Has been skipped

Details

Publish Release / Publish to crates.io (pull_request) Successful in 1m51s

Details

Publish Release / Create GitHub Release (pull_request) Successful in 3s

Details

6a37926107