Repository-owned governance for AI-assisted work
Agentic Execution Runtime
Formerly Agentic Project Kit. Keep project context, rules and work results in your repository when developing with an AI assistant — including an ordinary browser chat.
For maintainers who want inspectable context and controlled lifecycle steps. You provide the coding assistant and own the requirements and approvals.
An everyday use case
Continue tomorrow without rebuilding the project context
You discuss a change in a browser chat. The Kit prepares repository context and a bounded next step. You run the local workflow; it records check results and the remaining work. When you change chats or assistants, a generated handoff carries the repository state forward.
You can also use a local coding agent or controller. An autonomous multi-agent framework, paid agent platform or LLM API is not a prerequisite. Git, a local Kit installation and the permissions required by your chosen workflow still matter.
The useful result is an inspectable work process: what was requested, what ran, what passed or failed, and what is next. Checks enforce defined rules; the maintainer still reviews whether the change solves the right problem.
See the four interaction pathsStart here
Choose How You Want To Work
Start from the interaction mode, then follow the workflow to the relevant commands. File Transfer, Copy and Paste, Agent Direct, and the experimental GUI surface are generated as one workflow projection instead of competing command lists.
Brownfield evidence
The historical B1 threshold window consolidates 5 maintenance cycles as B1_EVALUABLE. Later observations appear in the generated evidence below. One familiar private repository does not establish general Brownfield portability.
Who it serves
Maintainers of continuing projects
Solo maintainers and small teams working with AI on Git-backed software or documentation, particularly across long sessions, reviews and releases. Python is a reference profile; equal maturity across all toolchains is not claimed.
Why use it
Durable context and bounded actions
Repository state replaces dependence on chat memory. Manifest-backed workflows expose effects, checks and evidence. Generated handoffs make the next session inspectable.
What it costs
Explicit process takes work
Adoption requires Git knowledge, maintained contracts, useful tests and review. More state files and closeout steps are real overhead. Green gates do not prove correct requirements or defect-free software.
Why
Repository Memory
Long LLM projects lose time when context lives in chat history. The Kit makes the repository the durable project memory: current state, handoff packages, command manifests, gates, reports and release evidence are inspectable before new work starts.
Install
Start Locally
python -m pip install agentic-project-kit
The public PyPI package is available. Git is still a system prerequisite for agentic-kit init and Git-backed workflows; pip cannot install Git portably. After installation, run agentic-kit --version, then choose agentic-kit init for a new governed repository or agentic-kit workspace adopt for an existing one.
Current State
Generated Status
Latest substantive work: PR #2549 (`Harden release run, curate the release changelog and classify PRs (#2549)`).
Next safe step: continue from fresh main with the next planned governed slice.
Architecture
Runtime Structure
The Kit separates human decisions, replaceable executors, repository contracts, deterministic checks and external integrations. CLI, GUI and transport surfaces use governed core operations. A private local controller can transport and execute bounded work while the Kit owns its process contracts.
Architecture and control flow · Local controller integration
- Manifest identity
a6ee797254ff- Reproduced manifest
a6ee797254ff- Identity verified
true- Build commit
7f233db18afa2fc2d0b383318f0a195d893f4aac
Evidence from actual use
Self-hosting, Greenfield and Brownfield
Using the Kit on itself exercises its own workflow. External adoption and maintenance tests exposed additional boundary defects and informed later repairs. Historical retests, recorded maintenance and current release notes are different evidence classes.
Greenfield: installed release
agentic-project-kit==1.0.13 · 2026-09-16 · recorded_pass
Historical released-package first-cycle work/PR/CI/merge/handoff retest in a private external repository; not a retest of every later release.
Brownfield: recorded maintenance
7 merged observations · latest: B1-COMM-SCI-20260831-007 · legacy seam count: 55 -> 44
Recorded maintenance observations in one familiar private repository; not an independent benchmark or general portability proof.
Release notes, not a new retest
What changed in 1.1.0
Source status: recorded. Shipped changes recorded in the current release section.
- Upgrade note for 1.0.x: transfer and typed work orders that run BOUNDED Kit commands, BOUNDED cockpit actions or unlisted commands now need the local --allow-unlisted flag; DESTRUCTIVE Kit commands, raw git/gh mutations and interpreters are refused in every order; transfer orders need a full 40-character expected_current_head bound to the commit they were written on; work-order run --execute, the next-turn.py slot and the current.sh slot report BLOCKED_RETIRED_LANE. Orders that ran under 1.0.25 can therefore stop and must be rewritten.
- Report a real conflict from transfer rebase-on-upstream instead of CONFLICT: no, keep the in-progress rebase as evidence and block when the conflict state cannot be observed; read .git/config with repeated keys such as several remote.origin.fetch lines; let commands sync-entrypoints --execute refresh the DPA readiness command-manifest ACK, so a manifest change no longer needs a separate release-prepare step.
- Add agentic-kit context report, which measures the files a successor is told to read (about 210,000 estimated tokens, 135,000 of them the command manifest), and agentic-kit context render --profile slice|planning|release|handoff, which renders a deterministic task packet of about 3,600 estimated tokens from the same authorities: current ACK, every execution-contract rule, manifest excerpts of the profile's commands, gates, open plan items and a checksummed path list for everything left out. Add the part 2 context benchmark harness with five tasks, deterministic success and ledger method checks, a no-LLM self-test, and its preregistration (docs/reports/context_budget_20261016/); the LLM runs await an owner decision.
- Fix transfer remote-next and transfer continue crashing with a ScannerError on every transfer order that starts with the required COMMAND_MANIFEST_ACK line, so no order could complete remote-next; found by the new offline gate benchmark (docs/reports/gate_benchmark_20261016/), which detects 11 of 11 injected faults on main against 5 of 10 on 1.0.25.
- Add agentic-kit audit-drift-signals, a report-only measurement of release-projection drift, refresh chains after product merges and Kit/consumer finding status, each with its denominator; agentic-kit ledger report now lists repeated failures by fingerprint and --write-finding-drafts writes local drafts that are never overwritten; the test suite no longer writes to the invocation ledger.
- Add a local invocation ledger: every agentic-kit process start (and in-process calls an orchestrator reports) is recorded with command name, timing, return code, Kit version, branch and call-tree ids, never arguments, in the git-ignored .agentic/tmp/invocation-ledger/; on in the Kit repository, opt-in elsewhere, bounded retention, and a ledger failure never changes a command's result. agentic-kit ledger report summarises it read-only. agentic-kit doc-registry register gains --status, --review-policy and --summary, so new runbooks are registered without hand edits.
- Retire the unrestricted work-order lanes: work-order run --execute, the next-turn.py fixed slot (next_turn_runner, GUI work-order run) and the current.sh agent command slot no longer execute code; they report BLOCKED_RETIRED_LANE and point to typed work orders, transfer orders and the planner executor (WOTB-04, WOTB-09). The commands are removed after a deprecation window.
- Bind the freshness anchor of a transfer order to the commit it was written on: a full 40-character hash equal to HEAD for an uncommitted order or to the parent of the commit that last changed it, checked again after switch and pull on every transfer remote-next path including transfer continue, so the order that runs is the order that was validated (WOTB-07, WOTB-08). Orders with short anchors, or changed after anchoring, now stop as stale.
Formal assurance: planning boundary
PAA: planned · review: review_required · P0: proposed
PAA is shown from the canonical plan. A planning status is not an implemented verification capability or a proof of software correctness.
Methods and generalization limits · Comparison with open-source and commercial tools
CLI
Normal Lifecycle
agentic-kit initCreate a governed project skeleton with selected profiles and policy packs.agentic-kit workspace adoptAnalyze an existing repository without writing workspace files.agentic-kit workspace initPlan or create a bounded operating-layer workspace.agentic-kit workspace dpa-intakeRun the deterministic DPA repository-intake orchestration.agentic-kit workspace upgradePlan or run deterministic workspace manifest schema upgrades.agentic-kit work startStart a human patch/slice workflow with the safe standard startup sequence.agentic-kit work recoverRun safe recovery/status commands after interrupted work.agentic-kit work rescueRescue local main work before realigning main to the local base ref.
Diagnostics
Common Blockers
agentic-kit audit-status-current-stateAudit STATUS.md current-state claims against handoff, release, and origin/main state.agentic-kit check-docsCheck required documentation coverage, current-state anchors and public workflow documentation before treating a slice as ready.agentic-kit docs-auditRun the umbrella documentation-system audit.agentic-kit doctorRun a compact project health check.agentic-kit workflow statusPrint the current workflow state and bounded evidence pointers.agentic-kit workflow-guard checkInspect workflow and governance files for known unsafe, stale, or ambiguous execution patterns before continuing.
GUI
Manifest Driven
The GUI entry point is shown only through claim evidence, and command grouping follows the manifest surface projection instead of a separate hand-maintained command list.
Communication
Handoff First
Successor handoff packages, transfer closeout, remote work evidence and copy-and-paste bridges are part of the governed workflow. The current package is regenerated by Kit commands after merges.
Evidence Status
Claims
Verified now
b1-brownfield-evaluableIn the historical B1 threshold window, five real Brownfield maintenance cycles reached the predeclared B1_EVALUABLE threshold while preserving the generalization boundary.cli-entry-pointThe package provides the agentic-kit command-line interface.command-manifest-synchronizationThe public command catalog is generated from the current command manifest and checked for synchronization.generated-command-catalogThe website command catalog is generated from the Kit's current command manifest.package-versionThe website package version is projected from pyproject.toml.public-source-install-documentedThe GitHub source projection remains documented for unreleased fixes and Kit development.pypi-package-installableDirect PyPI installation with python -m pip install agentic-project-kit is available for arbitrary users.python-requirementThe website Python requirement is projected from pyproject.toml.
Available but evolving
artifact-gc-dry-runagentic-kit artifact-gc supports dry-run behavior.gui-entry-pointThe package provides the agentic-kit-gui graphical entry point.pypi-trusted-publishing-readyPyPI and TestPyPI publication are prepared through separate gated Trusted Publishing automation paths.successor-handoff-packageThe Kit can produce a structured successor handoff package for continuing work in another session or model.workspace-adopt-read-onlyagentic-kit workspace adopt inspects without writing.workspace-init-preview-executeagentic-kit workspace init previews changes before explicit execution.
Planned
- No planned public claims are currently declared.
Not claimed
- Global DPA conformance for arbitrary foreign repositories is not claimed.
- Remote target-CI validation is not claimed when the target repository reports no checks; remote adoption PR mechanics are separate from target-owned CI evidence.
- Post-fix PyPI validation is not claimed until a later package is published and installed again from PyPI.
- Full GUI parity with every CLI command is not claimed.
- General Brownfield portability is not claimed from observations in one familiar private repository.
- Live GitHub Pages publication depends on either GitHub Actions Pages deployment or the generated docs-source fallback being current.
- Concept DOI
10.5281/zenodo.20101359- Primitives
106- Command groups
40- Project direction
active- Required claims
8