Changelog¶
All notable changes to Archon Horizon are documented here.
The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.
While the major version is 0, the API and on-disk formats may change between
minor releases. See the pipeline setup guide for current
service configuration and catalog updates.
Unreleased¶
- The public dashboard demo uses a synthetic administrator with populated execution hosts, provider harnesses, reviewers, skills, prompts and subagent descriptors. Forge and Zulip tabs show local read-only previews without live credentials or external service connections.
0.2.0-alpha.2 — 2026-10-09¶
-
Align Python, frontend and README version metadata with the second alpha snapshot. Preserve
v0.2.0-alpha.1and add the historicalv0.1.5release tag. -
Graph milestones use a star shape, including historical
type: milestonenodes. The Nodes directory combines milestone, type, progress, and text filters with accurate pagination; custom node types are discovered from project sources. -
References now store PDFs, original TeX, archives, and companion documents in durable project artifacts. The dashboard supports upload, preview, download, and archive; dispatched agents can list, upload, and retrieve files with checksum verification and durable transfer recovery. Apply migration
0031_reference_filesbefore restarting an existing control plane.
Pipeline redesign¶
horizon,horizon-pipeline, andpython -m archon_horizonnow enter the same PostgreSQL control plane. The former SQLite control plane, its CLI commands, and/api/v2endpoints are removed; the current API is/api/v3and the dashboard is served at/pipeline.- Existing 0.1.5 data is not converted by the new Alembic migrations. Prepare
a separate installation and PostgreSQL database, review its configuration,
then run the explicit
migratecommand. See setup. - Missions now form a revision-checked ownership tree, while mathematical dependencies remain in the roadmap graph. Runs use durable assignments, retained obligations and provider context, leased executions, and a root maintainer that closes each phase. Preprocessing, formalization, and postprocessing share the same scheduling and review machinery.
- Workers journal uncertain API requests and Git publications for recovery. The dashboard provides phase-aware projects, missions, activity, roadmap progress, administration, and Forgejo/Zulip integrations. Installation extras separate control-plane, worker, client, and search dependencies.
- Release wheels include the built dashboard, grouped skills, reviewer and specialist descriptions, and migration history. Source distributions retain the frontend build inputs and development checks.
Earlier development notes¶
The notes below record intermediate work since 0.1.5. Some describe the
superseded SQLite platform or /api/v2 behavior; the pipeline redesign above
and the current guides define the shipped interfaces.
-
Host storage admission measures 10 GiB / 10% headroom against the user or group quota when one is set, not against the larger tmpfs or backing filesystem. Quota-limited
/tmpwith remaining space no longer marks the host unavailable or blocks new workers. -
Maintainer review treats a
formally_provedparent whosechildrenare still informal as stale graph: cascade bounded prerequisite updates into the same PR rather than merging an inverted DAG. -
Workspace and graph Forgejo issues are worker-accessible (
GET/POST/PATCH /api/v2/forge/issues). SessionStart briefings list open tickets so collaborators and maintainers can raise a lease before a larger change and close completed hygiene without a periodic scan. Zulip remains live discussion; theorem gaps stay on the graph. -
Packaged
horizon-efficiencyskill: filter context withls/rg/git loginstead of dumping threads or trees; use LSP before Lake; skip re-checking already labeled or extra-library evidence; do not weaken hypotheses for short-term progress; treat Zulip as targeted@mentions (humans via@**Name**, workers via[@label](...horizon-mention=1)), not a session log. Collaborator and maintainer profiles load it withhorizon-start. -
Project-library PRs are legacy. Skills and prompts no longer ask agents to create or review
library-{project}contributions; Lean is published in the shared workspace. Startup no longer auto-provisions a library repository,POST /api/v2/projects/{project}/libraryrefuses new creation, and briefings/maintenance no longer scan or count library PRs. SessionStart hooks and the Activity dashboard rewrite leftover launch snapshots that still saidN Library PRs open/N awaiting reviewonto the live DAG counts. The SessionStart briefing is injected once at a cold session start; resume, compact, and recovered-thread relaunches skip it. Existing forge library repositories are left in place. -
Ready graph PRs automatically queue one maintainer child in their owning live run, sharing its session budget, allocation, and cancellation. A live maintenance queue and PR lifecycle notices keep newly ready work from being abandoned; the maintainer stop gate rechecks that queue. The public profile is now
horizon-maintainer(reviewerremains a compatibility alias), with a dedicated graph maintenance skill. Managed PR creation and periodic discovery now enforce repository and wait-state labels, request thehorizon-maintainerreviewer, reject merges while awaiting the author, and replace stale wait labels with a terminal label after merge or closure. -
Missions can pin nodes that exist only in graph pull requests through
metadata.node_attempts. Open, closed, rejected, superseded, and merged PR versions remain addressable without becoming canonical DAG state; worker startup includes the proposed node document, andhorizon node show PROJECT NODE --pull Nreads it directly. New graph PRs receive a concise purpose and changed-node/impact body instead of being created without a description. -
Prompt runs now use explicit
horizon-collaboratorandhorizon-maintainerMarkdown profiles, independent of specialty templates and harness selection. Runs default to collaborator credentials; collaborators can queue scoped maintenance batches, and maintainers can return implementation work to collaborators without changing the caller's credential. Each session starts with a live briefing of global running/queued work, eligible slots, open Library/DAG PRs, and review wait states. Managed prompt reports use an exact five-section checklist and a durable stop gate; an incomplete mission or unused capacity continues scheduling, and a pending live audit cannot race a terminal report outcome. -
A node is one Markdown file (
node.md) with Git-like history stored in SQLite. Object IDs are the SHA-1 Git would assign to a repository that contains only that file: blob of the UTF-8 bytes, a one-entry tree (100644 node.md), and commits whosetreeis that tree.HEADisrefs/heads/main; open PRs arerefs/pull/{n}/head. Collaborators propose node creation, node edits, and DAG wiring in atomic, multi-node graph pull requests. Maintainers can auditably amend a changeset before merge, and PR impact reports identify affected dependants, missing nodes, and cycles. Simultaneous progresslabelsreplace the exclusive nodestage; legacy stages migrate on write, and synthetic warning revisions are no longer created. The dashboard node page shows the accepted Markdown, child nodes, and PRs. Forgejo links are relative to the dashboard host and open the node'snode.md. Selecting a PR renders the node as if that change were accepted. Published Lean lives in the shared workspace; a historicallibrary-{project}Forge repository may still exist. -
The dashboard no longer polls compact state, missions, or Activity. Those views load on open or on Refresh. Compact
/api/v2/stateis view-scoped (projects,hosts,activity) so unused catalogs, Markdown bodies, and graph activity stay off the wire. Objective and mission lists omit document bodies until an item is opened. Node listings omit live-run attachments unlessinclude_activity=true. Markdown skips KaTeX unless the page contains math, and DAG math labels render for the selected node only. Node directory search still pages from SQLite. -
Node directory search and pagination read a SQLite current-claim index (title/label/id/tags plus dependency edges) instead of materializing the compact graph projection. Statement bodies are still excluded from search.
-
Control-plane Python modules live under semantic packages in
src/archon_horizon/platform/(graph,control,api,scheduler,hosts,integrations,workers). Public names onarchon_horizon.platformare unchanged. Import modules from the new packages (archon_horizon.platform.api.http, and so on). Reinstall managed Forge Lean-metrics hooks after upgrade so they invokepython -m archon_horizon.platform.integrations.forge_metrics. -
Remote Tailscale Serve access is no longer treated as the loopback maintainer console. Serve identity and forwarded headers require a Horizon account even when the TCP peer is
127.0.0.1. The live dashboard proxy forwards Host and Tailscale headers to the API and does not answer/api/v2/statefrom its local read replica. -
Single-machine initialization is non-interactive.
horizon initwrites API, local user, and worker profiles without prompting; flags override bind, port, workers, forge, and Tailscale. Loopback dashboard access is the builtin maintainer console. Remote access uses Horizon accounts (hashed passwords, viewer/operator/maintainer roles, optional hashed API keys). New signups start as viewers. Managed Forge and Zulip identities are created per account and never shown as API keys. Each Forge user is added to project organizations onhorizon-viewers,horizon-operators, orhorizon-maintainersaccording to the Horizon role. -
Lean search splits Mathlib, published workspace, and extra libraries. Informal Mathlib uses the LSP
lean_leansearchtool; type-shaped Mathlib useslean_loogle. Published project Lean is indexed from the forge default branch (GET /api/v2/projects/{project}/search). Agents can add search-only clones (POST /api/v2/search/librarieswith name, git URL, and commit) and query them without compiling. Maintainers can remove a clone withDELETE /api/v2/search/libraries/{name}. The dashboard Search tab exposes the same workspace and pool queries, plus library add/remove. Locallean_search/horizon searchstill cover the worker checkout, including module headers. Text and header ranking use bm25s (NumPy/SciPy) and persist the inverted index next to the declaration cache. Name search uses an exact/prefix index (no subsequence ranking). Type patterns bind?a/_to one identifier or bracket group. Cold index builds extract Lean files in parallel. Search API calls and LeanSearch / Loogle MCP queries appear on the session Activity timeline. Indexes are complementary to walking the actual trees: agents should stillls/rgworkspace, Library, graph, Mathlib, and extra-library checkouts, and may inspect public repositories and pull requests. An empty query is not evidence that a listed source is empty. -
Zulip topic titles expose a leading status emoji (
💬open,⚠️waiting,📣operator,✅resolved,💾memory,🔧infrastructure,🚫stuck).GET /api/v2/zulip/topicsacceptsstatusandemojifilters and returnsstatus,mark, andtitleon each topic.horizon zulip topicslists them;horizon zulip prefix-topicsadds the channel default to unmarked titles. Queue planners rankmission_candidatesby priority then depth and inspect waiting/open/stuck/operator titles before dispatch. -
Graph node guidance requires semantic kebab-case labels, short mathematical titles, informal descriptions in textbook prose, and a
## Referencessection with a fenced BibTeX entry plus the source theorem number or section and page. Workers must not prefixlem-/thm-/def-or encode Lean types and implementation adjectives in the label, title, or informal body. -
Harness configuration refreshes register profiles without forcing a full scheduler pass for each host. The background scheduler admits queued work using the updated profiles while existing worker processes continue.
-
Agent writing guidance favors concise Markdown bullets, descriptive links, node/proof references, and textbook-style mathematical statements. Zulip posts use linked attribution footers and readable session mentions, preserving legacy provenance and notification detection. Reports link recognized proof, session, and workspace commit IDs without rewriting their stored content.
-
Lean source metrics retain library LOC and lexical axiom counts when a file has an unterminated comment or literal. Commit checks and reports show the exact source warning and location; malformed files cannot produce a clean axiom check or suppress metrics for the rest of MorganTianLib.
-
Quota/balance and startup failures retry the same session behind shared cooldowns; unchanged queue planners back off. Worker startup errors retain redacted phase/traceback diagnostics. Full disks and transient database errors preserve completed worker results and command receipts without rerunning work. Cleanup protects preparing sessions, ignored sources and operator locks, and telemetry errors no longer terminate the harness. API retries respect Retry-After and preserve admission keys after malformed responses. Disconnected response clients close cleanly without repeated error writes and tracebacks. Background notification fetches back off during API or local inbox outages. Assignment prompts bound legacy mission excerpts and omit the full catalog index; worker guidance prioritizes Lean repair loops and concise linked evidence.
-
Coordination notices are batched after 20 tools and two minutes, with up to three titles and prompt delivery for direct mentions. Worker Zulip posts notify mentioned sessions; mission events notify their parent/coordinator. Run frontiers suggest deep unfinished missions and expose node overlaps. Workers use short missions, optional history lookups and sustained Lean edit/check/fix work; default prompts omit ancestor mission documents. Infrastructure retries also refresh the assignment. Active DAG nodes retain their proof-status fill with a pink outline, and queue planners do not mark their root mission's nodes active.
-
Node and DAG views show clickable tags for a node's running sessions and for the sessions that created or last updated the node or its proofs.
-
Activity occupancy uses running sessions against the run's available scheduler slots, queued counts, and clickable running-session chips. The agent tree can show only live sessions, mission tags include involved nodes with the same hover previews as objectives, and a successful coordination-hook delivery is recorded as a notification event.
-
The missions dashboard pages compact summaries instead of waiting for every mission body, and shows a loading state until the first page arrives.
-
Release archives include the frontend rebuild inputs, tests, documentation, and notices, with CI checks excluding local configuration and operator state. Development instructions use the tracked skill catalog;
CLAUDE.mdimportsAGENTS.md. Obsolete frontend screens and empty legacy directories have been removed, and pytest discovery is scoped to Horizon's own tests. -
Zulip uses one channel per project plus shared
guide,issues, andblockerschannels.guideis a short human reference for topic marks; agents may read it and cannot post there. Project channels are named from the project title. New projects receive a kebab-case id from that title. Managed setup removes Zulip's defaultgeneral/sandbox/zulipstreams. Worker posts get a server-added provenance header. Agents followhorizon-zulipfor concise, human discussion rather than a progress log. -
Managed Lean checks prefer LSP feedback during editing and coordinate necessary builds across hosts using expiring leases and streamed SHA-256 artifact storage. Compatible Lake versions reuse actual root and dependency artifacts, with separate worker and maintainer caches. Each host shares Lake artifacts and dependency Git objects across checkouts; reflinks reduce source duplication where supported. The current worker and build setup is in the pipeline setup guide.
-
The formalization graph holds statements, proof alternatives, source evidence, and failed attempts. Objectives hold roadmaps, missions hold executable assignments, and project workspaces preserve shared mathematical files. Packaged skills follow this control-plane workflow.
0.1.5 — 2026-09-04¶
Formalization-quality skills and reviewers, a heartbeat benchmark, complete roadmap CLI mutations, and workspace-local session scratch.
Added¶
- Formalization quality program (AJCR lessons) — blueprint-first complete
proofs with bibliography/
\dcref/\sourceand explicit adaptation notes; mathlib naming/docstrings/tree layout with leanprover-community guide links; writablejanitorfor Lean/blueprint layout moves;restart-moduleskill for backup-and-rewrite when layered debt or fix-loops dominate; lean-quality and API composition guidance treatset_optionheartbeats as blockers. Corpus guidance was recorded in the 0.1.5 development history. definition-qualityskill +definition-quality-reviewer— Mathlib-based criteria for good/bad definitions; detect suspects from how theorems/lemmas consume them (shared consumer pain → definition root cause).horizon benchmark(+ dashboard Benchmark view //api/benchmark) — rank Lean files by summedset_optionheartbeat budgets so agents can find costly modules.- Roadmap CLI completeness —
horizon roadmap show,rename, fullerset(--project,--depends-on,--inbox-ref,--task-ref), andremove --cascade(default remove un-nests children and drops depends-on edges). Skills (horizon,horizon-start,strategy-convergence) and the janitor descriptor now treat an empty or frozen roadmap as unfinished strategy: agents should add/move/delete/rename items as the long-term plan changes, not only flip status on pre-seeded rows. - Focused advisory skills for review method, progress integrity, source fidelity, semantic adversarial checks, blueprint integrity, proof load-bearing, consumer dependencies, API composition, Lean quality, Mathlib orientation, verification evidence, graph traceability, provenance isolation, strategy convergence, run health, external boundaries, transcription fidelity, release reproducibility, and review adjudication.
honestyskill and read-onlyhonesty-reviewer— certificate taxonomy (proved/conditional/imported/axiom-backed/empty/unverified), hypothesis-packaging and vacuity probes, stale-evidence checks, and a source-frontier test for repeated wrapper/re-expression churn.source-discovery— a channel and provenance map for local references, GitHub repository/review history, Mathlib source and PRs, Tau Ceti artifacts, and Zulip discussions, including credential and access-boundary guidance.- Separate read-only reviewer descriptors for those lanes. The roster remains optional: no automatic review stage or dispatch gate was added.
- Workspace-local, per-session scratch under
.archon-horizon/tmp/: agent subprocesses receiveTMPDIR/TMP/TEMPthere, successful sessions reclaim their scratch automatically, andhorizon tmp clean --applysafely removes stale leftovers while protecting live runs.
Changed¶
work-reviewerstays narrowly focused on honest progress, anti-loop checks, and task/report/diff consistency; focused reviewers own the other lanes.- Ground is an optional workspace-wide perspective the lead may choose, not a scheduled second orchestrator.
- Logs run/session task chip — the task link sits in the same meta-chip row as time/usage (not as a separate trailing control); long task ids ellipsize so they no longer shove the other tags around.
- Stale-skill fallback prompt is bounded and no longer mandates helper
dispatch; headless Horizon asks for
honesty-reviewerbefore new certificate/\leanokclaims or repeated frontier moves.
Fixed¶
- Blueprint textbook chapters are served from
content.texthe same way hgraph does (include only reached chapters, strip preamble definitions, harvest KaTeX macros from the whole tree). - Published DAG caches still using
kind/content_typeare filtered to countable nodes so stale remarks cannot affect graph status.
0.1.4 — 2026-08-28¶
Added¶
horizon ledger status/horizon ledger prune [--gc]— inspect and retro-clean agent ledgers that still track Horizon state or generated hgraph trees (index-only remove; working tree untouched; optionalgit gcto reclaim pack space).lean-isolatorsubagent descriptor for isolating large Lean declarations that time out or fail to produce an.olean.- Fixed English date/time helpers for the dashboard (
formatChipDateTime, …).
Changed¶
- Agent ledger is source-only. The out-of-tree workspace ledger
(
.archon-horizon/vcs/workspace.git) records Lean, blueprints, andconfig.yamlonly. The entire Horizon control-plane tree (.archon-horizon/) and all**/hgraph/paths are excluded; they stay on disk for the live dashboard. Users version what they want in their own root.git. Session integration commits no longer stage inbox/tasks/runs/roadmap state. - Session integration commits only
config.yaml+ scoped project sources. - Agent dirty-check / commit reminders ignore non-source ledger noise (state tree and hgraph).
- Static dashboard export prefers the user root
.gitwhen present; ledger is not the publish path for Pages snapshots. - CLI entry routes through
archon_horizon.__main__:mainsoagent-hook-fastcan bypass the full Typer import on every tool boundary. - Run/session chips show date and time (
Aug 28, 15:20), not time alone.
Fixed¶
- Session model chip no longer inherits the first subagent's model.
observed_modeland the dashboard skip nested-subagent events, so a parent that rangpt-5.6-solis not labeled with a child model (e.g. luna). horizon initmodel prompt is harness-family-aware. Choosing Codex clears a leftover Claude default (opus/sonnet/…) and shows Codex-only help; the reverse for Claude Code. Pre-existingCLAUDE.md/AGENTS.md/.envare respected and announced.- Dashboard dates stay English (
en-US) so a French OS locale no longer yields localized month names in the UI. - Inbox order uses real latest activity (item timestamps, comments, history). Collapsed cards show an activity date chip.
- hgraph Lean extraction:
/--doc comments only; empty-body guard so declarations are not written with empty node bodies. - Secret guard re-stages with
git add -fafter redaction so ignored paths still get the scrubbed blob. - Static export / dashboard: bound history export, stamp export time, keep expanded chapter graphs intra-chapter only (also carried from the v0.1.3 follow-ups on this branch).
0.1.3 — 2026-08-18¶
Added¶
- Live Codex rollout telemetry now discovers native and nested subagents from spawn provenance while the parent is running, records their model/role/depth and terminal status, and surfaces context compactions separately from current request and cumulative token counters.
horizon attempt savepreserves rejected drafts and diagnostics as explicit session artifacts;horizon checkserializes resource-heavy Lean checks, coalesces identical concurrent requests, enforces timeouts, and records results.
Changed¶
- Session change panels distinguish agent and integration commits, rejected attempts, and recorded checks. Long sessions get elapsed-time progress/commit checkpoints, including task/roadmap/inbox writes.
- The transcript keeps only the newest Codex context snapshot visible while the append-only log retains every raw telemetry event. Current-request, cached, cumulative, context-window, and compaction values remain in session details.
- Live dashboard state derives large changing transcripts from a bounded tail and
only rematerializes subagents when appended rows contain child activity; the
latest state-generation timing is available at
/api/performance.
Fixed¶
- Headless Codex runs now reconstruct native and nested subagents from rollout spawn provenance, stream their activity before the parent exits, suppress duplicate reconciliation events, and distinguish completed, failed, interrupted, cancelled, and orphaned children.
- Both Codex compaction record shapes are normalized without conflating current request input, cached input, cumulative usage, or the model context window.
[0.1.2] — 2026-07¶
The "lightweight harness, finished" release, extended with multi-team collaboration on a shared workspace. Full rationale and measurements in the 0.1.2 development history.
Added¶
- Enforced freeze CLI —
horizon freeze add/list/removemaintains config-backed agent/project/file/declaration/blueprint-node rules; semanticinbox protectconstraints remain clearly separate. -
Release version helper —
scripts/version.pyupdates and checks the Python package, README badge, dashboard metadata, and demo workspace stamp together. -
horizon usage— the agent-facing consumption gauge: this session's live token/cost burn (streamed by the harness into<session>/usage.json), run totals, budget headroom, recent rate-limit signals, and any pause marker. workspace.budget(opt-in) —session_tokens_outcancels a live session that crosses it;run_tokens_out/run_cost_usdstop the run cleanly between sessions. A limit/budget stop writesruns/<id>/paused.json(reason, advertised retry window, focus, resume command) andhorizon runexits with code 3 so relaunch loops can tell "paused, resumable" from failure.- Vendored semantic graph — the dependency graph is generated from
per-node files under
<project>/hgraph/; agents attach comments and Maths/Lean reviews, andhorizon graph frontierranks what to prove next. Horizon owns the implementation and has no external graph dependency or secondary DAG engine. horizon ps— per-run process registry (runs/<id>/process.json): list live runs, reap zombie markers (--clean), kill a stuck run (--kill <run>); killed/crashed runs resume with--resume.- Session identity env vars — every session now sees
ARCHON_HORIZON_RUN/SESSION/SESSION_DIR/ROUND/ROUNDS/TASK/TASK_TITLE/PROJECTS(documented in thehorizonskill), so any engine can read its own context without prompt-parsing. - Engine-native orientation —
initwritesCLAUDE.md/AGENTS.mdpointers to thehorizonskill, so any engine launched in the workspace self-orients without pushed prompt prose. - Dashboard: hgraph-style little-squares mini-map + segmented progress bars on
the Blueprint table of contents;
/api/blueprintssplit out of/api/state. - Teams collaborating on one workspace — each
horizon runis a team; parallel teams coordinate through shared state, not synchronous meetings. Inbox items are now either shared (default) or owned by one task (a private per-team inbox); read-state is per-team (horizon inbox read/unread,inbox list --mine/--unread/--task <id>); and teams can direct-message one another withinbox add --to task:<id>/run:<id>. - Roadmap as a project board — roadmap items carry
owner, amilestonelabel (grouping/filtering, no due dates), and pinned commits (horizon roadmap set/add --owner --milestone,set --pin-commit/--unpin-commit,list --milestone/--owner);horizon task set --roadmap-ref/--inbox-reflinks a task to its milestone/inbox. A milestone-grouped/boarddashboard view and clickable local-reference chips (roadmap/task/inbox/node/commit ids in any rendered text) surface it in the UI; newGET /api/commit?sha=resolves a SHA. horizon permissions+workspace.delegation(default deny) — the standing consent an agent reads before launching work for other teams (new tasks, or a whole newhorizon run); spawning subagent workers within a team needs no permission.- Pre-command synchronizer — inside a session, every
horizoncommand first prints a cached, stderr-only digest (unread inbox for the task, session runtime/tokens, other live runs) so an agent stays aware without polling;ARCHON_HORIZON_NO_SYNC=1disables it and it never touches--jsonstdout. - Provenance-defaulted flags —
--author,--project, inbox owner/reader, and task default from the session'sARCHON_HORIZON_*env, so an agent rarely passes them.
Changed¶
- Agent startup names the horizon skill by absolute path, agent-authored state and hgraph comments retain run/session/task provenance, inbox triage defaults to open items, and the session contract requires inbox/roadmap maintenance.
-
Formalization notes now belong on hgraph nodes rather than in mathematical blueprint
.texsources. -
The automated prompt is now a task directive + "load the
horizonskill". All pushed role prose/policy/workspace state is gone; the skill is the (user-editable) contract and state is pulled through the CLI. The old.archon-horizon/prompts/overrides are inert. - Dashboard hot path — measured on a 162-run workspace, a poll cost ~12 s
of server work every 5 s; it is now ~85 ms when idle (ETag short-circuit
before compute) and ~0.5 s when changed (session-state cache persisted under
.archon-horizon/cache/, subagent materialization moved to the write path, incremental events parse, subtree-stamp collection caches). - Failure classification reads the engine's typed error events before grepping stderr; an advertised "retry after Xs" stretches the retry backoff.
horizon runhas one launch path (--supervisor,horizon run horizon, and focused runs all reduce to a focus + RunRecord).- The ledger secret guard now redacts instead of blocking. A high-confidence,
key-prefixed, case-sensitive credential match is replaced with
XXXXin the staged content (and re-staged) with a warning to rotate it — the commit is never blocked, and the scan fails open so a guard bug can't wedge commits. Being case-sensitive, long camelCase identifiers no longer false-match.ARCHON_HORIZON_ALLOW_SECRETS=1skips it; the separate silent-clobber/deletion guard is unchanged. - hgraph Lean declaration extraction now covers
structure/inductive/class(not justtheorem/lemma/def/abbrev/instance), allows Unicode/subscript declaration names, and honours the_root_.escape (theorem _root_.Foo.barresolves toFoo.bar), so more\lean{}references resolve instead of going stale. [temporary]inbox items auto-archive at run finish once they have been visible for a full run (consumed one-shot notes); items created during the run are kept for the next one.- Concurrent
horizon inbox addcalls get distinct ids (flock-serialised id allocation, fail-open). The UI's per-session commit log is latest-first, and the automated session prompt leads with the absolute skill path so a non-interactive engine never guesses a failing relative.claude/skills/...path.
Removed¶
- Dead config from the two-agent era (
roles,start_with,end_with,ground_subagents,subagent_harness,references.transcription,scheduler.unknown_write_set_policy) — old keys are ignored, not errors. Theworkspace.ground_agentblock is now fully gone too:discuss, the post-init advisor, andhorizon subagentall use the horizon harness. - The aggregate per-session "Changes" view (
/api/run/changes,/api/run/working-changesand their UI panel): progress is read from the commit-granular git view (commit cards + per-commit file diffs), which is the durable record anyway. - The former embedded DAG package, parser fallback, and standalone DAG command;
horizon graphis the single graph surface. - The vestigial cross-process commit-lock stack,
project_checkpoint,ProjectGit, git-notes helpers; the old graph exporters/queries/reporter; the in-memory inbox provider; prompt composition and its templates; write-domain enforcement (core/permissions.py); the legacy Python-rendered dashboard fallback (a no-dist install now gets build instructions; the JSON API stays live).
0.1.0 — 2026-06¶
First public release. Archon Horizon is a workspace-first successor architecture to Archon: a workspace is the global collaboration root and each Lean codebase is a member project, coordinated by a Ground agent (blueprints, DAG, roadmap, reports, inboxes) and a long-Horizon agent (autonomous Lean formalization).
Added¶
- Workspace model.
config.yaml+ a tracked.archon-horizon/state dir. Projects are embedded directories with their own build and (optional) out-of-tree git under.archon-horizon/vcs/<project>.git— no submodules. - Two abstract agents. Engine-agnostic Ground and Horizon agents with a collaboration-round driver across deterministic sync boundaries.
- Harness seam. A
Harnessinterface with a genericCommandHarness(Claude Code, Codex, or any CLI, selected by config), acodexharness, and an in-processNullHarness. Claude Code backends:default,vscode,desktop, and the optionalclaude-pheadless backend. Provider routing for Kimi/Moonshot and DeepSeek (direct keys or OpenRouter fallback). - Inboxes. Local (filesystem), GitHub (shadow via
gh), in-memory, and sharded providers, gated byarchon:accept/archon:pending/archon:rejectedlabels. No live injection — hints are observed at sync boundaries only. - Roadmap, tasks, runs, reports, events, memory stores with YAML/JSON codecs; tasks are human-only, agents organize pending work via the roadmap.
- Blueprint → DAG. A LaTeX-subset blueprint parser building dependency graphs for Lean/blueprint queries and dashboard rendering.
- Permissions & freeze. Deterministic freeze enforcement before dispatch and write-set locks; protections stored as persistent inbox items.
- Native subagents. Read-only starter descriptors (blueprint-reviewer,
diff-auditor) compiled per-engine into workspace-local
.claude/agents/.codex/agents; skills installed under.claude/skills/. - Dashboards. A live editable web dashboard (separate frontend) and a static HTML snapshot generator.
- Workspace search (
horizon search) over mathlib, workspace projects, and configuredexternal_libraries. - CLI (
horizon):init,setup,update,run,inbox,roadmap,task,project,skills,blueprint, graph queries,search,sync,dashboard. Every command supports--json(pure JSON on stdout, human chrome on stderr). - Version stamping.
horizon initrecords the Horizon version in.archon-horizon/version; workspace commands warn on drift and point athorizon init --update(refresh a stale workspace) orhorizon update(upgrade a stale tool). - Install / update.
curl … | bashinstaller (install.sh) and ahorizon updateself-update command.