research-document GV-START-001

Agent Startup Guide

Agent Startup Guide

Mission

Praxis, Echelon Foundry's repository operating system, makes research, engineering, decisions, and handoffs durable without relying on conversation history or tribal knowledge.

Start Here

Older installations may have ./ros, a compatibility alias of ./praxis; new instructions use Praxis (DF-ROS-2026-A050).

  1. Read the governance index.
  2. Identify the task's scope and operating mode.
  3. Locate the applicable canonical domain records; inspect the repository and user changes before editing.
  4. State or record material unknowns, constraints, assumptions, and risks.
  5. Use the smallest process that preserves correctness, traceability, and continuity.
  6. Execute, validate, update affected records, and leave a handoff.

Detailed rules are in the Agent Operating Manual. Research packages follow the REP Specification; engineering follows the Engineering Standards.

Authority

Apply, in descending order: explicit user instruction; applicable safety, legal, and platform constraints; canonical governance; accepted domain REPs and theory; accepted architecture and decision records; current implementation; local convention; agent preference. A higher authority cannot authorize a violation of an applicable safety or legal constraint. When same-level sources conflict, prefer the narrower and newer accepted record and document the resolution; escalate if the outcome materially changes the authorized goal.

Core Rules

  • Never fabricate evidence, file reads, approvals, commands, test results, or certainty.
  • Preserve user work. Inspect before modifying; do not destroy or irreversibly migrate without authorization.
  • Make reasonable, reversible, in-scope decisions. Escalate high-impact irreversible, security/privacy-sensitive, legally ambiguous, or materially out-of-scope decisions.
  • Research by testing hypotheses against confirming and falsifying evidence. Engineering by establishing a baseline, defining acceptance criteria, making the smallest robust change, and testing in proportion to risk.
  • Important claims cite EV-, HY-, and TH- records when those records exist. Material decisions use DF-, which canonically means Decision Record.
  • Do not silently change canonical policy. Propose or record the change, its evidence, consequences, version, and migration path.
  • Do not claim a test passed unless it ran and passed. Name skipped or unavailable checks and their implications.
  • Treat execution telemetry as evidence: discover capabilities, distinguish zero from unavailable, preserve normalized and sanitized raw provider data, prefer deterministic collection, and never invent a metric.
  • Keep upstream drift bounded: start with ./praxis sync check --start, repeat the fetch-only check whenever 30 minutes have elapsed since the last successful check, and integrate upstream changes at a safe work boundary before final validation.
  • Not every edit needs a REP. Use the artifact threshold in the Agent Operating Manual.

Handoff

For substantial work, record: objective; work completed; files changed; decisions and assumptions; tests run and results; evidence added; unresolved questions; risks; and next recommended action. A capable successor must be able to continue without the originating conversation.

Work Protocol

Before meaningful mutation, identify the external work item and run ./praxis work begin --id ID --occurred-at TIMESTAMP (see the F# CLI note below for the timestamp — it must be the real current time, not an arbitrary one). That transition starts an execution-telemetry record; inspect ./praxis work context ID, classify the work, and ingest runtime telemetry that the current environment can expose. For substantial execution, use ./praxis step begin --name "..." --occurred-at TIMESTAMP and ./praxis step complete --occurred-at TIMESTAMP around meaningful plan units; do not create command-level noise or retroactive steps. Preserve unknown provider fields through the sanitized raw layer and record unsupported/unavailable capability explicitly. Perform the bounded work, gather configured evidence, commit and push it, record a durable checkpoint (see "Durable checkpoints and continuity" below), request a legal transition with ./praxis work complete --id ID --occurred-at TIMESTAMP --evidence TYPE=PATH (repeatable; finalizes active telemetry), then run ./praxis registry build and ./praxis validate, and commit and push the resulting Praxis state. Attribute canonical records you create or change with ./praxis provenance record (see Agent Identity and Provenance below). Use ./praxis work block --id ID --occurred-at TIMESTAMP --reason TEXT and ./praxis work resume --id ID --occurred-at TIMESTAMP rather than hand-editing context. Use ./praxis status when resuming unfamiliar work. Meaningful committed changes require machine-readable attribution; see docs/work-protocol.md and docs/development-telemetry.md. If meaningful changes were committed while no work item was active, reconcile them after the fact with ./praxis work reconcile --id ID --reason TEXT --commit REV --occurred-at TIMESTAMP (Git-evidenced, recorded as post-hoc, never a substitute for beginning work). Never touch, rewrite, or recommit files to manufacture attribution, and never create a work item only to absorb changes.

No externally-assigned ID yet? Check ./praxis work ready for capturable, unblocked repository work before assuming none exists, and use ./praxis add "..." to record a newly discovered obligation instead of leaving it as an unfiled comment or dropped observation (add does not require --occurred-at; it defaults to the real current time). ./praxis work start --id ID --occurred-at TIMESTAMP (begin is also accepted) promotes a ready backlog item into the protocol above. This local backlog is repository-scoped triage, not a project-management system; see the "Local backlog" section of docs/work-protocol.md.

Upstream synchronization and bounded drift

Long-running or stalled agents must not assume their starting view of main is still current. After selecting the worktree and before meaningful mutation, run ./praxis sync check --start. Repeat ./praxis sync check at coherent work boundaries whenever 30 minutes of wall-clock time have elapsed since the last successful check. Waiting on tools, external systems, or the user counts toward that elapsed time. Also run it immediately before final validation and handoff.

The check performs a bounded git fetch of the configured upstream and writes only per-worktree Git metadata. It never pulls, merges, rebases, switches, stashes, resets, discards, commits, or edits working files. ./praxis sync status and the additive upstreamSync block in ./praxis status report the starting and current upstream commits, elapsed freshness, ahead/behind counts, incoming and local paths, exact overlap, and whether final validation can be claimed against the fetched upstream snapshot. A failed check does not reset the 30-minute clock.

When upstream advanced, stop at the next safe boundary, inspect overlap, and integrate according to the repository's branch policy while preserving user work. Re-run affected tests after integration. Do not begin final validation while the report is stale, unavailable, or behind. If integration is unsafe or blocked, record the exact condition and hand off instead of forcing it. A fresh report proves only the latest successfully fetched snapshot; no local tool can prove that a remote did not move immediately afterward. Configure the policy at workProtocol.upstreamSync in ros.json; repositories with no usable upstream must disable it explicitly and record that limitation.

Durable checkpoints and continuity

An executor session is disposable. Repository state and Praxis state are the continuity boundary. No meaningful completed work may exist only in an executor's local environment: a successor on another machine, with no access to your filesystem, process, or conversation, must be able to continue. See docs/work-protocol.md ("Durable checkpoints and continuity") and DF-ROS-2026-A042.

  • A commit is local. A pushed commit is on a remote. A verified durable checkpoint is Praxis's own record that your HEAD, the checkpoint commit, and the head of your upstream remote branch were the same commit, with no meaningful uncommitted work: ./praxis work checkpoint --id ID --occurred-at NOW --summary "what is done" --next-action "what is next" [--step STEP-ID]. Praxis never commits, pushes, or stashes for you.
  • A historical checkpoint is that record; it is never rewritten. A currently recoverable checkpoint is one the remote still carries now; ./praxis work context ID --text and ./praxis status report both, separately.
  • Checkpoint at coherent recovery boundaries, not on a timer and not per edit: after a meaningful implementation slice or material telemetry step; before a risky change; before switching work items or repositories; before an intentional handoff; when context exhaustion or termination looks possible; before blocking after new work; before completing Git-backed work. Never create a meaningless commit to satisfy Praxis.
  • For Git-backed work the order is: commit, push, work checkpoint, work complete, then commit and push the Praxis state (.ros/). Where workProtocol.continuity.requireDurableCheckpoint is set, completion refuses anything else, and blocking after un-checkpointed work needs a checkpoint or --unrecoverable-reason TEXT stated truthfully. Work that changed nothing completes as before.
  • New observability is effective-current. Praxis preserves truthful historical gaps rather than restarting work or fabricating telemetry. Adopt step telemetry (./praxis telemetry step start|complete|fail) at the next material slice; never restart an execution or work item to gain it, never invent earlier steps, and never split earlier usage among steps. Missing historical step data is unavailable, not zero and not invalid.
  • Taking over active work whose executor disappeared: fetch, switch to the checkpoint's branch in a clean checkout, then ./praxis work continue --id ID --occurred-at NOW under your own identity. You get a new execution whose parent is the predecessor's; the predecessor is recorded as interrupted, never as you and never as successful. blocked -> resume remains for intentionally blocked work.

If the native Praxis executable is genuinely unavailable, use the single approved runtime-free envelope in docs/fallback-reconciliation.md; never hand-edit .ros state. Preserve only identity and telemetry the runtime actually exposes. Reconciliation is retryable: leave a pending journal/input in place and rerun rather than recreating a transaction under a new ID.

Agent Identity and Provenance

Every agent working under this repository has an explicit, machine-readable identity, and records it on the work it creates or changes. This applies to every provider and runtime, and equally to humans and automation. See docs/agent-provenance.md.

  1. Establish identity once, at the start of the execution. ./praxis work begin records who you are in the execution record, and every later command inherits that identity.
    • A known runtime (Codex, Claude Code, Gemini CLI, Copilot, GitHub Actions) is detected automatically.
    • Otherwise declare yourself with PRAXIS_ACTOR_KIND (agent|human|automation), PRAXIS_ACTOR (your stable agent ID), PRAXIS_TELEMETRY_PROVIDER, PRAXIS_TELEMETRY_MODEL, and PRAXIS_TELEMETRY_RUNTIME, or pass the matching flags on work begin (the legacy ROS_* names still work).
    • Check the result with ./praxis provenance identity.
  2. Never impersonate another agent, human, or execution. Never record work under an execution you did not run. Praxis refuses a contribution whose actor contradicts its execution.
  3. Never fabricate a provider, model, version, session, or agent name. Leave an unknown value unset: Praxis records it as unknown, which is correct.
  4. Preserve existing provenance. Never edit, reorder, or delete another contributor's provenance entry.
  5. Add your contribution; do not replace anyone else's.
  6. Attribute every requirement you create: ./praxis provenance record --id RQ-... --operation created.
  7. Attribute every meaningful modification you make to a canonical record (requirement, decision, evidence, hypothesis, experiment, theory, journal, mission, research package): --operation modified. Use reviewed or approved only for review or approval you actually performed.
  8. Propagate lineage when you derive one artifact from another: --derived-from SOURCE-ID. Lineage names the source. It does not make the source's author an author of your artifact.
  9. Make generated evidence, findings, and results traceable to your execution. Record them inside the work execution, and name supporting records with --evidence.
  10. Run ./praxis validate before finishing. Missing or contradictory provenance on new work is an error.

Identity recorded this way is provenance, not authentication. It is self-reported and cross-checked, not cryptographically proven.

No local runtime? Use remote execution

If you can reach this repository on GitHub but cannot run Praxis locally, for example because there is no .NET, you are governed the same way through typed remote requests. Follow docs/remote-agent-contract.md: discover with a praxis.describe request, send typed operations with a stable requestId and the expectedSha you read, and read results from the repository.

  • Remote execution is available only when .github/workflows/praxis-remote.yml exists. Remote mutation is available only where ros.json lists it under remote.capabilities.
  • The rules above still apply: never impersonate, never fabricate identity, and never touch files to manufacture attribution.
  • Do not work around a missing runtime by hand-editing .ros/ state.
  • Checkpoint and take over remotely with work.checkpoint (naming your own execution.id) and work.continue (protocol 1.3).
  • Can commit but cannot dispatch Actions? Commit the request as .praxis-inbox/<requestId>.json on a praxis-inbox/... branch. The inbox relays it unchanged (DF-ROS-2026-A045).

Lifecycle commands

Installation, verification, diagnosis and upgrade go through the standard lifecycle interface, implemented in F#. Install the praxis command (with ros as its compatibility alias) one of two ways (DF-ROS-2026-A044; npm is no longer a distribution channel):

  • No runtime needed: the self-contained native bundle from GitHub Releases (scripts/install-native.sh, install-native.ps1 on Windows, or echelon install praxis); see docs/native-installation.md.
  • With .NET 10: dotnet tool install -g EchelonFoundry.Praxis.
praxis init
praxis status
praxis verify
praxis upgrade
praxis doctor

In this source checkout the same commands are available as ./praxis init, ./praxis verify and so on. init is idempotent, every command is non-interactive, --dry-run and --check change nothing, and --json puts a single document on stdout. Exit codes are a documented contract: 0 success, 2 invalid arguments, 3 verification failed, 4 incompatible installation, 5 migration blocked, 6 prerequisite failure. See docs/cli.md, docs/installation.md and docs/upgrading.md.

Installation state lives in .echelon/ros.json; it is tool bookkeeping, not repository work, and is never treated as a meaningful change for attribution. Before editing a file the tool installed, check its ownership there: a tool-owned file is replaced on upgrade, so a local edit belongs in a user-owned or shared file instead.

F# CLI

./praxis in this source checkout, and in every project scaffolded by praxis init (both profiles), runs the F# CLI (DF-ROS-2026-A030, DF-ROS-2026-A049). Praxis's own repository is F#/.NET only (RQ-ROS-2026-A024): it owns no JavaScript, TypeScript, npm or Node tooling, and ./praxis architecture check (also part of ./praxis validate here) fails on any such file. Do not add one; implement the behaviour in F#. In this checkout ./praxis is a shell launcher for the built CLI; if it reports it needs building, run dotnet build Praxis.slnx --configuration Release first (CI always builds before ./praxis runs). In an installed project ./praxis runs the Praxis version the project pins, installing that native release on first use.

The command syntax is worth knowing rather than guessing from memory:

  • Every mutating command shown above except add requires an explicit --id ID (repeatable) and --occurred-at TIMESTAMP, rather than a positional ID with an implicit clock read. Pass the real current time (e.g. `date -u +%Y-%m-%dT%H:%M:%S.000Z`), not an arbitrary or backdated one: a telemetry execution's own startedAt always reads the real wall clock (matching production), and a later transition whose supplied --occurred-at predates it fails ./praxis validate with a spurious "capability state recording order must be chronological" finding — a real trap this decision's own preparation hit and diagnosed, not a defect to work around.
  • work start and work begin are both accepted, as are work complete and work done.
  • work context ID and work show ID take a positional ID.
  • docs/migrations/fsharp/STATUS.md is the authoritative ledger of any remaining command-surface gaps (e.g. telemetry finalize --input, a deliberately unported adapter-ingestion-at-finalize path).

This section's command-syntax notes apply equally to ./praxis in this source checkout and to any project's own bootstrapped ./praxis, since both run the same F# CLI.

Visual Engineering UI research

Managed by npx @echelon-foundry/visual-engineering. Do not edit inside this block.

Before designing, implementing, or reviewing UI:

  1. Run npx @echelon-foundry/visual-engineering verify and stop if it reports a failure.
  2. Read .visual-engineering/AGENT-INSTRUCTIONS.md.
  3. Read .visual-engineering/UI-FOUNDATIONS.md.
  4. Read .visual-engineering/UI-DECISION-CHECKLIST.md.
  5. Read .visual-engineering/UI-ANTI-PATTERNS.md.
  6. Consult .visual-engineering/RESEARCH-INDEX.md for provenance and deeper evidence.
  7. Inspect the product and its existing design system.
  8. Apply the research as decision criteria, not as a visual style.
  9. Report the context version, source commit, principles applied, verification performed, and justified deviations.

Do not copy Visual Engineering research into this repository by hand.

Communication Engineering

Communication Engineering is installed as evidence-bounded operational guidance. Before producing consequential communication, read:

  • .communication-engineering/COMMUNICATION-FOUNDATIONS.md
  • .communication-engineering/COMMUNICATION-DECISION-CHECKLIST.md
  • .communication-engineering/PURPOSE-OUTCOME-MATRIX.md
  • .communication-engineering/COMMUNICATION-ANTI-PATTERNS.md
  • .communication-engineering/RESEARCH-STATUS.md

Treat research maturity as a constraint. Do not turn provisional findings into universal rules, optimize persuasion at the expense of user autonomy, or substitute style for proof obligations.