README
¶
SOP
SOP — a standard operating procedure for agentic software development. SOP is a local-first orchestrator that runs a software development lifecycle with AI agents while keeping workflow control deterministic.
Agents do the work. SOP controls the process.
Coding agents are good at understanding requirements, planning, writing code and tests, reviewing, and diagnosing failures — but they should not be the source of truth for workflow state. SOP separates those responsibilities: agents reason and implement, and the orchestrator decides what is allowed to happen next. Detailed behavior lives in the documentation; the authoritative ownership model is docs/architecture/SOP-BOUNDARY.md.
Core Principles
SOP owns orchestration and lifecycle state.
Agents perform bounded work.
Models are replaceable reasoning engines.
Validation provides deterministic evidence.
Review and OpenJEV provide additional quality signals.
Human approval remains explicit where policy requires it.
- Deterministic control, agentic execution. State transitions, scheduling, retry bounds, merge eligibility, and failure classification are deterministic and model-free; models only produce content.
- The LLM creates content; the application controls the process. A model never decides whether its own output is valid.
- Machine state is not Markdown. Executable state lives in structured data
(
SQLite,plan.json); the orchestrator never parses human docs for its state. - Bounded autonomy. Retries and fix loops have explicit limits and end at a human gate when they must.
- Operator-owned acceptance checks. Deterministic checks provide objective completion evidence; a task's PASS alone does not verify its objective. See Execution.
- Local first, provider neutral. Runs locally with local or cloud models and does not require a cloud service. The smallest class is local-first: it runs a local model and falls back to a cloud model only when the local runtime cannot serve it (see docs/specs/MODEL-ROUTING.md).
SOP is not an unrestricted autonomous coding agent: the orchestrator defines boundaries, the agent performs specific work, validation checks it, and only then does state change. See docs/PRD.md for principles and requirements.
Governance ownership is defined in SOP Boundary. The consolidation guide records advisory compatibility, evaluation, and enablement boundaries.
Basic Architecture
PRD / PLAN
|
v
Agentic SOP
|
+-- Scheduler
+-- Agent Harness
+-- Validation
+-- Review
+-- OpenJEV
+-- Quality Policy
+-- Recovery
|
v
Repository
SOP separates reasoning from control; every component above is implemented. Detailed structure: docs/architecture/OVERVIEW.md.
Requirements
- Go (see
go.mod) and Git. - An AI command or agent harness for planning and implementation (a local model via Ollama/llama.cpp/MLX, or an external agent command).
- SQLite is provided through the Go driver; no external database is required.
Quick Start
Install the CLI (clone the repository and cd into it first):
./install.sh # macOS / Linux
sop version
.\install.ps1 # Windows (PowerShell)
sop version
Then, from a project you want SOP to manage:
cd ~/workspace/projects/book-rag
sop run
sop run discovers the planning source, prepares state and tasks, and executes
the task graph — idempotently. Full setup: docs/guides/GETTING-STARTED.md
and docs/guides/PROJECT-SETUP.md. Complete command
reference: docs/reference/CLI.md. To run SOP's model
classes on Claude, see docs/guides/HOW-TO-SETUP-WITH-CLAUDE.md.
For ad-hoc work, sop prompt runs a request through the same governed machinery:
sop prompt --capability review "Review internal/provider for architectural issues"
Details: docs/specs/PROMPT-EXECUTION.md.
The unified executable also provides sop controller serve for the dashboard and
sop decision decide|shadow for advisory decisions. Setup, optional requirements,
and legacy wrappers are covered in the
consolidation guide.
Agent Skills
Install the /sop* agent commands with ./install.sh --skills zed,
./install.sh --skills claude, or ./install.sh --all. Start with
the command examples,
installation, or
the Claude Code plugin guide.
SOP owns routing, validation, and approval for every command. Three RAG-domain
skills (/rag-schema, /rag-test, /rag-ingest) add a pgvector/RAG checklist on top
of the read-only capabilities, and /sop-rag-integrate advises whether a project needs
retrieval augmentation at all and plans a scoped integration (or recommends NO-RAG);
see docs/guides/RAG-SKILLS.md.
The provider-neutral /mcp-live-validation skill prepares pinned MCP candidates,
produces a versioned cross-agent validation handoff, and preserves approval-gated
start, rollback, and promotion boundaries; see
the MCP live-validation guide.
Planner–Executor Coordination
The merged interactive planner baseline lets an executor publish progress,
blockers, test results, and completion evidence with sop result publish. A
planner connected through sop mcp --planner can retrieve task results and
questions, then record a bound continue, correct, revise, or escalate
decision. The executor applies that decision with
sop planner continue --apply and resumes through sop run.
Planner decisions are advice, never human approval, and published results are
evidence, never automatic task completion. Resumption is currently interactive;
no unattended ChatGPT invocation or notification channel is claimed. PE-1 through
PE-4 implementation and live acceptance are merged; the merged PE-4 evidence
records all four tasks as LOCAL_DONE. Completing the plan remains a separate
operator-local SOP lifecycle action and is not inferred from that repository
evidence. See the
planner–executor implementation plan,
PE-4 acceptance evidence,
and CLI reference.
Optional evaluation tooling is also available outside the default execution
path: sop compare runs bounded
baseline/candidate comparisons, and the
Promptfoo adapter performs report-only behavioral
security evaluations. Neither tool can promote a candidate or alter deterministic
execution governance.
Basic Workflow
PLAN
↓
SELECT TASK
↓
PRECHECK
↓
EXECUTE
↓
VALIDATE
↓
REVIEW
↓
JEV / QUALITY
↓
HUMAN APPROVAL
↓
COMPLETE
Detailed state transitions and lifecycle rules live in the specifications: docs/specs/WORKFLOW.md, docs/specs/TASK-LIFECYCLE.md, and docs/specs/EXECUTION.md.
Status
SOP V1 and the subsequent routing, provider, recovery, distribution, approval,
agentic reliability, and context-efficiency layers are implemented. Optional local decision memory is implemented but disabled by default pending a valid held-out benefit evaluation. See
Project Status for delivery evidence and known
limitations, Backlog for future work, and the
performance baseline for measured results
and their scope. Repository consolidation is complete; its
evidence and compatibility boundaries
do not establish current local machine or CI status. The legacy sop-controller
and sop-decision-adapters GitHub remotes are currently archived read-only,
but archival readiness is not verified; see the
legacy transition checklist.
Documentation
Start at docs/README.md for the complete categorized index.
| Read about | Start here |
|---|---|
| Product requirements | PRD |
| Architecture and ownership | Overview, SOP boundary |
| Execution and lifecycle | Execution, task lifecycle |
| Providers and harnesses | Agent provider, model routing |
| CLI and configuration | Commands, configuration |
| Usage and development | Getting started, development |
| Decision-model findings | Decision-model findings |
| Plans and historical evidence | Plans, history |
License
Apache License 2.0
Directories
¶
| Path | Synopsis |
|---|---|
|
cmd
|
|
|
deceleval-advisory
command
Command deceleval-advisory runs the advisory (Experimental B) evaluation over the frozen decision-model corpus and writes the paired A/B artifact.
|
Command deceleval-advisory runs the advisory (Experimental B) evaluation over the frozen decision-model corpus and writes the paired A/B artifact. |
|
deceleval-baseline
command
Command deceleval-baseline runs the deterministic-only (Baseline A) evaluation over the frozen decision-model corpus and writes a reproducible results artifact for paired comparison with the advisory run (DME-004).
|
Command deceleval-baseline runs the deterministic-only (Baseline A) evaluation over the frozen decision-model corpus and writes a reproducible results artifact for paired comparison with the advisory run (DME-004). |
|
deceleval-jev
command
Command deceleval-jev runs the JEV shadow-mode evaluation (DME follow-up).
|
Command deceleval-jev runs the JEV shadow-mode evaluation (DME follow-up). |
|
deceleval-recovery
command
Command deceleval-recovery runs the DME-005 failure-recovery advisory experiment: it asks the local advisory adapter a failure-RECOVERY question (AUTO_FIX / CONTINUE / HUMAN) instead of the production change-RISK question, and measures the answer against the frozen labels with the SAME add-attention semantics.
|
Command deceleval-recovery runs the DME-005 failure-recovery advisory experiment: it asks the local advisory adapter a failure-RECOVERY question (AUTO_FIX / CONTINUE / HUMAN) instead of the production change-RISK question, and measures the answer against the frozen labels with the SAME add-attention semantics. |
|
rdm-pilot-corpus
command
Command rdm-pilot-corpus builds the source-backed planning-only decision-memory corpus from recorded SOP run artifacts.
|
Command rdm-pilot-corpus builds the source-backed planning-only decision-memory corpus from recorded SOP run artifacts. |
|
rdm004-eval
command
Command rdm004-eval runs the frozen RDM-004 A/B comparison.
|
Command rdm004-eval runs the frozen RDM-004 A/B comparison. |
|
sop
command
|
|
|
sop-controller
command
Command sop-controller is a thin wrapper over the shared controller service.
|
Command sop-controller is a thin wrapper over the shared controller service. |
|
sop-decision-adapter
command
Command sop-decision-adapter is a thin wrapper over the shared decision-adapter CLI.
|
Command sop-decision-adapter is a thin wrapper over the shared decision-adapter CLI. |
|
sop-memory
command
Command sop-memory ingests completed runs, rebuilds the local advisory index, and exercises the production advisory path for a task or failure.
|
Command sop-memory ingests completed runs, rebuilds the local advisory index, and exercises the production advisory path for a task or failure. |
|
sop-ollama-agent
command
Command sop-ollama-agent is the bootstrap Ollama coding-agent harness.
|
Command sop-ollama-agent is the bootstrap Ollama coding-agent harness. |
|
internal
|
|
|
activity
Package activity is SOP's structured activity model: a small, observer-only stream of "what the run is doing right now" events.
|
Package activity is SOP's structured activity model: a small, observer-only stream of "what the run is doing right now" events. |
|
adaptiveroute
Package adaptiveroute is the CTX-011 Adaptive Routing policy: evidence-driven model class selection that optimizes toward the smallest capable model.
|
Package adaptiveroute is the CTX-011 Adaptive Routing policy: evidence-driven model class selection that optimizes toward the smallest capable model. |
|
agent
Package agent defines the boundary between the application and an LLM.
|
Package agent defines the boundary between the application and an LLM. |
|
agentbin
Package agentbin resolves the location of the installed known-good sop-ollama-agent binary.
|
Package agentbin resolves the location of the installed known-good sop-ollama-agent binary. |
|
approval
Package approval is SOP's human approval application boundary: the single, authoritative place a client reads an active human approval gate and records a decision on it.
|
Package approval is SOP's human approval application boundary: the single, authoritative place a client reads an active human approval gate and records a decision on it. |
|
archtest
Package archtest holds deterministic architecture and provider-neutrality tests for the Phase 8 subsystems (CTX-001 through CTX-012).
|
Package archtest holds deterministic architecture and provider-neutrality tests for the Phase 8 subsystems (CTX-001 through CTX-012). |
|
autonomy
Package autonomy is SOP's centralized, deterministic risk-based autonomy policy: the single place that decides whether a failure or a requested action may be handled automatically or must wait for a human.
|
Package autonomy is SOP's centralized, deterministic risk-based autonomy policy: the single place that decides whether a failure or a requested action may be handled automatically or must wait for a human. |
|
bootstrap
Package bootstrap wires an environment (bootstrap) stage into the task graph.
|
Package bootstrap wires an environment (bootstrap) stage into the task graph. |
|
budget
Package budget is the canonical, explicit model of the deterministic execution limits an agent harness invocation runs under.
|
Package budget is the canonical, explicit model of the deterministic execution limits an agent harness invocation runs under. |
|
ci
Package ci renders a project's GitHub Actions verification workflow from its configured verification commands (T008).
|
Package ci renders a project's GitHub Actions verification workflow from its configured verification commands (T008). |
|
ciremediation
Package ciremediation fixes ordinary CI failures within a bounded loop.
|
Package ciremediation fixes ordinary CI failures within a bounded loop. |
|
cli
Package cli implements the sop command-line interface.
|
Package cli implements the sop command-line interface. |
|
commandpolicy
Package commandpolicy classifies commands deterministically before they are executed: SAFE, REQUIRES_APPROVAL, or DENIED.
|
Package commandpolicy classifies commands deterministically before they are executed: SAFE, REQUIRES_APPROVAL, or DENIED. |
|
commitgate
Package commitgate decides whether a task commit is allowed and, if so, creates it through a committer port.
|
Package commitgate decides whether a task commit is allowed and, if so, creates it through a committer port. |
|
completion
Package completion drives a plan to the end without supervision: it selects the next runnable task, drives it to CI, and completes validated work (merge -> refresh main -> DONE), repeating until nothing remains.
|
Package completion drives a plan to the end without supervision: it selects the next runnable task, drives it to CI, and completes validated work (merge -> refresh main -> DONE), repeating until nothing remains. |
|
config
Package config defines the project configuration model loaded from .agent-sdlc/config.yaml.
|
Package config defines the project configuration model loaded from .agent-sdlc/config.yaml. |
|
context
Package context defines the canonical, deterministic representation of the evidence supplied to an agent for one execution — the Context Engine.
|
Package context defines the canonical, deterministic representation of the evidence supplied to an agent for one execution — the Context Engine. |
|
continuation
Package continuation is the deterministic, read-only reconcile-before-continue gate for an already-active SOP plan.
|
Package continuation is the deterministic, read-only reconcile-before-continue gate for an already-active SOP plan. |
|
controller
Package controller is the SOP Controller service entry point, shared by the unified `sop controller serve` command and the legacy `sop-controller` executable.
|
Package controller is the SOP Controller service entry point, shared by the unified `sop controller serve` command and the legacy `sop-controller` executable. |
|
controller/assets
Package assets embeds the SOP Controller templates and static files.
|
Package assets embeds the SOP Controller templates and static files. |
|
controller/config
Package config holds runtime configuration for the SOP Controller dashboard.
|
Package config holds runtime configuration for the SOP Controller dashboard. |
|
controller/sopclient
Package sopclient is the controller's only boundary to SOP.
|
Package sopclient is the controller's only boundary to SOP. |
|
deceleval
Package deceleval loads and validates the decision-model effectiveness corpus: independently labeled failure scenarios used to compare SOP's deterministic classification (baseline A) with the advisory decision layer (experimental B).
|
Package deceleval loads and validates the decision-model effectiveness corpus: independently labeled failure scenarios used to compare SOP's deterministic classification (baseline A) with the advisory decision layer (experimental B). |
|
decision
Package decision is SOP's bounded decision boundary: a provider produces a choice with a confidence, and policy (outside the provider) decides what that choice means.
|
Package decision is SOP's bounded decision boundary: a provider produces a choice with a confidence, and policy (outside the provider) decides what that choice means. |
|
decision/bridge
Package bridge is SOP's deterministic bridge between the canonical SEAM-002 decision seam (internal/decision) and a provider-neutral, batched decision adapter profile.
|
Package bridge is SOP's deterministic bridge between the canonical SEAM-002 decision seam (internal/decision) and a provider-neutral, batched decision adapter profile. |
|
decision/command
Package command adapts an external, provider-neutral decision process to the decision.Provider interface.
|
Package command adapts an external, provider-neutral decision process to the decision.Provider interface. |
|
decision/commandprofile
Package commandprofile is SOP's production connector between the canonical SEAM-002 decision seam and a provider-neutral, batched decision adapter executable, bridged through the ratified M3 profile bridge (internal/decision/bridge).
|
Package commandprofile is SOP's production connector between the canonical SEAM-002 decision seam and a provider-neutral, batched decision adapter executable, bridged through the ratified M3 profile bridge (internal/decision/bridge). |
|
decisionadapter/cli
Package cli is the shared decision-adapter command implementation used by both `sop decision ...` and the legacy `sop-decision-adapter` executable.
|
Package cli is the shared decision-adapter command implementation used by both `sop decision ...` and the legacy `sop-decision-adapter` executable. |
|
decisionadapter/providers/clef
Package clef adapts the Clef decision model, served through an Ollama-compatible backend, to the provider-neutral decision contract.
|
Package clef adapts the Clef decision model, served through an Ollama-compatible backend, to the provider-neutral decision contract. |
|
decisionadapter/providers/fallback
Package fallback composes two decision providers: the backup serves only when the primary fails.
|
Package fallback composes two decision providers: the backup serves only when the primary fails. |
|
decisionadapter/providers/julia
Package julia adapts the Julia decision model, executed as ONNX inference, to the provider-neutral decision contract.
|
Package julia adapts the Julia decision model, executed as ONNX inference, to the provider-neutral decision contract. |
|
decisionadapter/providers/nimble
Package nimble adapts the Nimble decision model, served through an Ollama-compatible backend, to the provider-neutral decision contract.
|
Package nimble adapts the Nimble decision model, served through an Ollama-compatible backend, to the provider-neutral decision contract. |
|
decisionadapter/shadow
Package shadow implements an adapter-side, observational shadow evaluation harness.
|
Package shadow implements an adapter-side, observational shadow evaluation harness. |
|
decisionmemory
Package decisionmemory is the CTX-010 Decision Memory: a bounded, durable store of engineering decisions that future execution can consult.
|
Package decisionmemory is the CTX-010 Decision Memory: a bounded, durable store of engineering decisions that future execution can consult. |
|
decisionquality
Package decisionquality holds the RDM-004 decision-quality evaluation contract: the frozen held-out and tuning corpora, the pre-decision evidence rules, the freeze hashes, and the scoring rules frozen before the final run.
|
Package decisionquality holds the RDM-004 decision-quality evaluation contract: the frozen held-out and tuning corpora, the pre-decision evidence rules, the freeze hashes, and the scoring rules frozen before the final run. |
|
dotenv
Package dotenv reads an optional, gitignored .env file and applies it to the process environment.
|
Package dotenv reads an optional, gitignored .env file and applies it to the process environment. |
|
dqharness
Package dqharness runs the RDM-004 A/B comparison.
|
Package dqharness runs the RDM-004 A/B comparison. |
|
durablefs
Package durablefs provides a crash-safe file write: content is written to a temp file in the destination directory, fsynced, atomically renamed over the destination, and the destination directory is fsynced.
|
Package durablefs provides a crash-safe file write: content is written to a temp file in the destination directory, fsynced, atomically renamed over the destination, and the destination directory is fsynced. |
|
e2e/harness
Package harness provides deterministic test infrastructure for the full agent harness regression suite (PREJEV012).
|
Package harness provides deterministic test infrastructure for the full agent harness regression suite (PREJEV012). |
|
eval
Package eval runs a corpus of cases through a runner and aggregates the metrics a harness benchmark needs: task success, fix cycles, review findings, and elapsed time.
|
Package eval runs a corpus of cases through a runner and aggregates the metrics a harness benchmark needs: task success, fix cycles, review findings, and elapsed time. |
|
failure
Package failure classifies why a task's lifecycle stopped short of a pass and what SOP should do about it.
|
Package failure classifies why a task's lifecycle stopped short of a pass and what SOP should do about it. |
|
git
Package git provides SOP's deterministic Git boundary: a small adapter over the installed git CLI.
|
Package git provides SOP's deterministic Git boundary: a small adapter over the installed git CLI. |
|
github
Package github provides SOP's remote-workflow boundary: pushing branches, managing pull requests, inspecting checks, and merging.
|
Package github provides SOP's remote-workflow boundary: pushing branches, managing pull requests, inspecting checks, and merging. |
|
handoff
Package handoff produces a small, deterministic handoff capsule after a task completes, so later work can reuse the durable meaning of that task without re-reading its full history.
|
Package handoff produces a small, deterministic handoff capsule after a task completes, so later work can reuse the durable meaning of that task without re-reading its full history. |
|
handoff/caveman
Package caveman is an optional handoff.Compressor adapter that delegates compression to a configured command.
|
Package caveman is an optional handoff.Compressor adapter that delegates compression to a configured command. |
|
hipcompare
Package hipcompare implements the HIP-003 opt-in comparison harness: a small, controlled A/B experiment that runs one task manifest twice (baseline and candidate instruction sets) from the same starting commit in separate disposable worktrees, then records the two arms as machine-readable results and a short Markdown comparison.
|
Package hipcompare implements the HIP-003 opt-in comparison harness: a small, controlled A/B experiment that runs one task manifest twice (baseline and candidate instruction sets) from the same starting commit in separate disposable worktrees, then records the two arms as machine-readable results and a short Markdown comparison. |
|
jev
Package jev defines the boundary between SOP and JEV (the optional engineering-analysis capability described in docs/history/plans/PLAN-JEV-Implementation.md and docs/requirements/PRD-JEV.md).
|
Package jev defines the boundary between SOP and JEV (the optional engineering-analysis capability described in docs/history/plans/PLAN-JEV-Implementation.md and docs/requirements/PRD-JEV.md). |
|
mcp
Package mcp implements a minimal Model Context Protocol server over stdio.
|
Package mcp implements a minimal Model Context Protocol server over stdio. |
|
memindex
Package memindex is the separate, rebuildable local SQLite memory index for RDM-002B. It stores structured metadata in a normal table and the searchable text in an FTS5 virtual table, in a database file that is distinct from the SOP execution-state database.
|
Package memindex is the separate, rebuildable local SQLite memory index for RDM-002B. It stores structured metadata in a normal table and the searchable text in an FTS5 virtual table, in a database file that is distinct from the SOP execution-state database. |
|
mergegate
Package mergegate decides whether validated work may merge and, if so, merges it through an explicit Merger port.
|
Package mergegate decides whether validated work may merge and, if so, merges it through an explicit Merger port. |
|
model
Package model resolves SOP's optional model-routing layer: which model (and provider) a run uses, selected by a small/medium/large class from a layered configuration.
|
Package model resolves SOP's optional model-routing layer: which model (and provider) a run uses, selected by a small/medium/large class from a layered configuration. |
|
normalize
Package normalize is the CTX-006 Response Normalizer: the deterministic boundary between provider output and harness interpretation.
|
Package normalize is the CTX-006 Response Normalizer: the deterministic boundary between provider output and harness interpretation. |
|
ollamaagent
Package ollamaagent is a minimal, temporary coding-agent harness that lets SOP's command provider use a local Ollama model (deepseek-v4.1-flash:cloud) as its implementation agent.
|
Package ollamaagent is a minimal, temporary coding-agent harness that lets SOP's command provider use a local Ollama model (deepseek-v4.1-flash:cloud) as its implementation agent. |
|
orchadopt
Package orchadopt is the ORCH-012 Phase 9 multi-agent adoption gate: the deterministic, model-free evidence gate comparing the single-agent baseline against the multi-agent candidate on genuinely parallelizable tasks.
|
Package orchadopt is the ORCH-012 Phase 9 multi-agent adoption gate: the deterministic, model-free evidence gate comparing the single-agent baseline against the multi-agent candidate on genuinely parallelizable tasks. |
|
orchestration
Package orchestration defines a provider/model-neutral domain model for orchestrating work across the SOP lifecycle:
|
Package orchestration defines a provider/model-neutral domain model for orchestrating work across the SOP lifecycle: |
|
outcome
Package outcome verifies an objective's explicit acceptance criteria against the workspace, independently of any model self-report.
|
Package outcome verifies an objective's explicit acceptance criteria against the workspace, independently of any model self-report. |
|
outcomefile
Package outcomefile writes and reads versioned outcome documents as authoritative local evidence for the memory index.
|
Package outcomefile writes and reads versioned outcome documents as authoritative local evidence for the memory index. |
|
parallel
Package parallel runs dependency-independent tasks concurrently, each in its own isolated working directory, bounded by a maximum.
|
Package parallel runs dependency-independent tasks concurrently, each in its own isolated working directory, bounded by a maximum. |
|
perf
Package perf is SOP's single performance model: where a task and a run spend their time, and how many agent calls, validations, reviews, and fix cycles they cost.
|
Package perf is SOP's single performance model: where a task and a run spend their time, and how many agent calls, validations, reviews, and fix cycles they cost. |
|
planflow
Package planflow prepares a project for execution: it discovers or resolves the human planning source, keeps the machine plan (.agent-sdlc/plan.json) in step with it, and creates persisted tasks when none exist.
|
Package planflow prepares a project for execution: it discovers or resolves the human planning source, keeps the machine plan (.agent-sdlc/plan.json) in step with it, and creates persisted tasks when none exist. |
|
planner
Package planner turns a project PRD into a structured Plan by asking an Agent and validating the result.
|
Package planner turns a project PRD into a structured Plan by asking an Agent and validating the result. |
|
projectdetect
Package projectdetect detects a repository's project types from its manifests, structure, and explicit Makefile targets, and proposes validation commands for the existing configuration schema.
|
Package projectdetect detects a repository's project types from its manifests, structure, and explicit Makefile targets, and proposes validation commands for the existing configuration schema. |
|
prompt
Package prompt is the CTX-005 Prompt Compiler: the deterministic owner of the structure of the model input SOP sends.
|
Package prompt is the CTX-005 Prompt Compiler: the deterministic owner of the structure of the model input SOP sends. |
|
promptcache
Package promptcache is the CTX-008 Prompt Result Cache: a conservative cache of read-only model responses.
|
Package promptcache is the CTX-008 Prompt Result Cache: a conservative cache of read-only model responses. |
|
prompttuning
Package prompttuning is the CTX-012 Automatic Prompt Tuning controller: an evaluation-gated promotion boundary for prompt candidates.
|
Package prompttuning is the CTX-012 Automatic Prompt Tuning controller: an evaluation-gated promotion boundary for prompt candidates. |
|
provider
Package provider defines SOP's provider/runtime abstraction: the layer that reports which model server is configured, whether it is reachable, which models it serves, and what those models can do.
|
Package provider defines SOP's provider/runtime abstraction: the layer that reports which model server is configured, whether it is reachable, which models it serves, and what those models can do. |
|
provider/command
Package command adapts the external command provider to the provider interface.
|
Package command adapts the external command provider to the provider interface. |
|
provider/httpx
Package httpx provides the small, bounded HTTP helpers the provider adapters share: a GET that returns a status and a size-capped body, and a GET that decodes a JSON body.
|
Package httpx provides the small, bounded HTTP helpers the provider adapters share: a GET that returns a status and a size-capped body, and a GET that decodes a JSON body. |
|
provider/llamacpp
Package llamacpp adapts a llama.cpp llama-server (an OpenAI-compatible endpoint) to the provider interface.
|
Package llamacpp adapts a llama.cpp llama-server (an OpenAI-compatible endpoint) to the provider interface. |
|
provider/mlx
Package mlx adapts an Apple-Silicon MLX / oMLX runtime to the provider interface.
|
Package mlx adapts an Apple-Silicon MLX / oMLX runtime to the provider interface. |
|
provider/ollama
Package ollama adapts a local or cloud Ollama server to the provider interface: it reports reachability (/api/version), the served models (/api/tags), and a model's capabilities (/api/show).
|
Package ollama adapts a local or cloud Ollama server to the provider interface: it reports reachability (/api/version), the served models (/api/tags), and a model's capabilities (/api/show). |
|
provider/openai
Package openai adapts an OpenAI-compatible HTTP endpoint to the provider interface.
|
Package openai adapts an OpenAI-compatible HTTP endpoint to the provider interface. |
|
quality
Package quality is the deterministic quality gate.
|
Package quality is the deterministic quality gate. |
|
question
Package question defines SOP's smallest durable question/decision contract: a task may raise a structured question, an answer is recorded independently of the question, and the question carries a closed state vocabulary that is deliberately separate from any run stage (in particular, none of these states is the run's WAITING_FOR_PLANNER stage).
|
Package question defines SOP's smallest durable question/decision contract: a task may raise a structured question, an answer is recorded independently of the question, and the question carries a closed state vocabulary that is deliberately separate from any run stage (in particular, none of these states is the run's WAITING_FOR_PLANNER stage). |
|
ragclient
Package ragclient implements the SOP-side client for the versioned, retrieval-only RAG protocol (schema "rag-retrieval/v1", see docs/specs/RAG-RETRIEVAL-CONTRACT.md).
|
Package ragclient implements the SOP-side client for the versioned, retrieval-only RAG protocol (schema "rag-retrieval/v1", see docs/specs/RAG-RETRIEVAL-CONTRACT.md). |
|
recovery
Package recovery is SOP's deterministic, bounded execution-recovery policy (Phase 5).
|
Package recovery is SOP's deterministic, bounded execution-recovery policy (Phase 5). |
|
repoindex
Package repoindex builds a deterministic Structural Repository Index: a stable, machine-readable representation of repository structure for the Context Engine to consume later.
|
Package repoindex builds a deterministic Structural Repository Index: a stable, machine-readable representation of repository structure for the Context Engine to consume later. |
|
resume
Package resume recovers safely from process interruption.
|
Package resume recovers safely from process interruption. |
|
retrieval
Package retrieval is a deterministic lexical ranker (BM25) over the Structural Repository Index's candidate corpus.
|
Package retrieval is a deterministic lexical ranker (BM25) over the Structural Repository Index's candidate corpus. |
|
retrievalboundary
Package retrievalboundary is the backend-neutral bounded retrieval boundary.
|
Package retrievalboundary is the backend-neutral bounded retrieval boundary. |
|
retrievalgate
Package retrievalgate is the CTX-004 Retrieval Evaluation Gate: a deterministic, model-free evaluation deciding whether structural + BM25 retrieval materially improves over the baseline discovery behavior, before more retrieval complexity is justified.
|
Package retrievalgate is the CTX-004 Retrieval Evaluation Gate: a deterministic, model-free evaluation deciding whether structural + BM25 retrieval materially improves over the baseline discovery behavior, before more retrieval complexity is justified. |
|
review
Package review provides SOP's structured review boundary.
|
Package review provides SOP's structured review boundary. |
|
router
Package router is SOP's deterministic model-class router (Phase 3.5).
|
Package router is SOP's deterministic model-class router (Phase 3.5). |
|
run
Package run persists a single harness run: its artifacts and stage under .agent-sdlc/runs/<id>/.
|
Package run persists a single harness run: its artifacts and stage under .agent-sdlc/runs/<id>/. |
|
runmetrics
Package runmetrics is a deterministic, model-free aggregator over recorded run artifacts.
|
Package runmetrics is a deterministic, model-free aggregator over recorded run artifacts. |
|
runtrace
Package runtrace defines the canonical, versioned structured run trace: the important execution trajectory of an SOP run, persisted beside the other run artifacts as trace.json.
|
Package runtrace defines the canonical, versioned structured run trace: the important execution trajectory of an SOP run, persisted beside the other run artifacts as trace.json. |
|
scheduler
Package scheduler performs deterministic, dependency-aware task scheduling.
|
Package scheduler performs deterministic, dependency-aware task scheduling. |
|
security
Package security implements SOP's optional security scanner add-on.
|
Package security implements SOP's optional security scanner add-on. |
|
skill
Package skill describes and validates the shipped SOP agent skill surface: the canonical skill tree under skills/, and the per-agent adapters that install it.
|
Package skill describes and validates the shipped SOP agent skill surface: the canonical skill tree under skills/, and the per-agent adapters that install it. |
|
taskbuilder
Package taskbuilder turns a validated planner.Plan into executable domain.Tasks, validates the resulting dependency graph, and coordinates atomic persistence.
|
Package taskbuilder turns a validated planner.Plan into executable domain.Tasks, validates the resulting dependency graph, and coordinates atomic persistence. |
|
taskfile
Package taskfile loads a single task from a Markdown file, so the workflow can start from a hand-written spec instead of only a generated plan.
|
Package taskfile loads a single task from a Markdown file, so the workflow can start from a hand-written spec instead of only a generated plan. |
|
taskresult
Package taskresult stores results an executor publishes to SOP outside the governed run loop.
|
Package taskresult stores results an executor publishes to SOP outside the governed run loop. |
|
testfast
Package testfast covers the repository's `make test-fast` developer tooling (scripts/checks/test-fast.sh).
|
Package testfast covers the repository's `make test-fast` developer tooling (scripts/checks/test-fast.sh). |
|
testrunner
Package testrunner executes project-defined verification commands and reports deterministic, structured results.
|
Package testrunner executes project-defined verification commands and reports deterministic, structured results. |
|
toolharness
Package toolharness is SOP's controlled tool harness: the single place where a model may read and modify repository files, run policy-checked commands, and inspect Git state.
|
Package toolharness is SOP's controlled tool harness: the single place where a model may read and modify repository files, run policy-checked commands, and inspect Git state. |
|
validate
Package validate is the harness's validation stage: it runs a project's configured verification commands in a deterministic order and returns structured results.
|
Package validate is the harness's validation stage: it runs a project's configured verification commands in a deterministic order and returns structured results. |
|
vectoreval
Package vectoreval is the CTX-009 Vector Retrieval Evaluation: the second evidence gate in Phase 8.
|
Package vectoreval is the CTX-009 Vector Retrieval Evaluation: the second evidence gate in Phase 8. |
|
verifcache
This file adds the VER-002 portable cross-worktree evidence store and the append-only reuse-decision record.
|
This file adds the VER-002 portable cross-worktree evidence store and the append-only reuse-decision record. |
|
verify
Package verify is SOP's deterministic, model-free verification risk tier.
|
Package verify is SOP's deterministic, model-free verification risk tier. |
|
workitem
Package workitem defines SOP's unified execution input: the small adapter boundary that lets the same governed execution machinery accept planned project tasks and ad-hoc operator prompts.
|
Package workitem defines SOP's unified execution input: the small adapter boundary that lets the same governed execution machinery accept planned project tasks and ad-hoc operator prompts. |
|
worktreecleanup
Package worktreecleanup implements SOP's deterministic, model-free classification model for the /sop-cleanup skill (skills/sop-cleanup).
|
Package worktreecleanup implements SOP's deterministic, model-free classification model for the /sop-cleanup skill (skills/sop-cleanup). |
|
writing
Package writing owns SOP's single shared writing rule.
|
Package writing owns SOP's single shared writing rule. |
Click to show internal directories.
Click to hide internal directories.