Recorded workflows

// Package notebook contains two bounded, informational workflow examples.
// They are learning/evaluation material, independent of release or deployment gates.
package notebook

import (
	"context"
	"fmt"
	"github.com/tylergannon/gimbal"
)

//go:generate go tool polytype --validate

// Route connects a reader's question to evidence in the supplied corpus.
type Route struct {
	Question string   `json:"question"`
	Sources  []string `json:"sources"`
	Answer   string   `json:"answer"`
}

// Index is a compact question-driven map, not a vector database.
type Index struct {
	Routes []Route `json:"routes"`
}

// Task is a bounded delivery outcome with dependencies and acceptance evidence.
type Task struct {
	ID         string   `json:"id"`
	Outcome    string   `json:"outcome"`
	DependsOn  []string `json:"depends_on"`
	Acceptance string   `json:"acceptance"`
}

// Sprint is an ordered plan for the supplied goal and constraints.
type Sprint struct {
	Goal  string `json:"goal"`
	Tasks []Task `json:"tasks"`
}

// Review is an independent assessment against the original fixture.
type Review struct {
	Findings []string `json:"findings"`
}

// SemanticIndex creates a retrieval map and gives it to a fresh reviewer.
func SemanticIndex(ctx context.Context, env gimbal.Env, corpus string) (Index, error) {
	gimbal.Set(ctx, "corpus", corpus)
	author := gimbal.NewSession(ctx, gimbal.RoleBulkClassification, env.WorkDir)
	index, err := author.Generate[Index](ctx, indexPrompt)
	if err != nil {
		return Index{}, err
	}
	if err := ValidateIndex(index); err != nil {
		return Index{}, err
	}
	gimbal.SetJSON(ctx, "candidate", index)
	reviewer := gimbal.NewSession(ctx, gimbal.RoleCodeReview, env.WorkDir)
	review, err := reviewer.Generate[Review](ctx, indexReview)
	if err != nil {
		return Index{}, err
	}
	gimbal.SetJSON(ctx, "review", review)
	if len(review.Findings) != 0 {
		return Index{}, fmt.Errorf("index review: %v", review.Findings)
	}
	return index, nil
}

// SprintPlan creates an actionable plan and gives it to a fresh reviewer.
func SprintPlan(ctx context.Context, env gimbal.Env, brief string) (Sprint, error) {
	gimbal.Set(ctx, "brief", brief)
	planner := gimbal.NewSession(ctx, gimbal.RoleSprintPlanning, env.WorkDir)
	plan, err := planner.Generate[Sprint](ctx, sprintPrompt)
	if err != nil {
		return Sprint{}, err
	}
	if err := ValidateSprint(plan); err != nil {
		return Sprint{}, err
	}
	gimbal.SetJSON(ctx, "candidate", plan)
	reviewer := gimbal.NewSession(ctx, gimbal.RoleCodeReview, env.WorkDir)
	review, err := reviewer.Generate[Review](ctx, sprintReview)
	if err != nil {
		return Sprint{}, err
	}
	gimbal.SetJSON(ctx, "review", review)
	if len(review.Findings) != 0 {
		return Sprint{}, fmt.Errorf("sprint review: %v", review.Findings)
	}
	return plan, nil
}

const indexPrompt = `Build a compact semantic index of the corpus below. Return four question-driven routes: starting a local run, isolation of concurrent app/database stacks, ownership of workflow semantics, and updating a backend while runs are active. Each answer must cite the relevant supplied source IDs, distinguish requirements from demonstrated support, and be useful to a reader choosing what to open next. Do not invent capabilities. Use only the supplied corpus. Make no file changes.`
const indexReview = `Independently inspect the candidate against the original corpus. Check all four requested retrieval questions, exact source attribution, requirements versus proven support, and unsupported promises. Return concrete findings, or an empty findings list when the index is faithful. Make no changes.`
const sprintPrompt = `Turn the supplied brief into a small sprint of three to five tasks. Name bounded outcomes, earlier task dependencies, and observable acceptance evidence. Cover the local baseline, concurrent stack isolation, cancellation and cleanup, and human-legible observation. Respect the budget and exclusions. This is a proposed plan, not completed work. Make no file changes.`
const sprintReview = `Independently inspect the candidate against the original brief. Check complete coverage, feasible ordering, concrete acceptance, budget, and exclusions. Reject claims of completed implementation, universal restartability, or mandatory Docker for basic local workflows. Return concrete findings, or an empty findings list when the plan is actionable. Make no changes.`

// Package notebook contains two bounded, informational workflow examples.
// They are learning/evaluation material, independent of release or deployment gates.
package notebook

import (
	"context"
	"fmt"
	"github.com/tylergannon/gimbal"
)

//go:generate go tool polytype --validate

// Route connects a reader's question to evidence in the supplied corpus.
type Route struct {
	Question string   `json:"question"`
	Sources  []string `json:"sources"`
	Answer   string   `json:"answer"`
}

// Index is a compact question-driven map, not a vector database.
type Index struct {
	Routes []Route `json:"routes"`
}

// Task is a bounded delivery outcome with dependencies and acceptance evidence.
type Task struct {
	ID         string   `json:"id"`
	Outcome    string   `json:"outcome"`
	DependsOn  []string `json:"depends_on"`
	Acceptance string   `json:"acceptance"`
}

// Sprint is an ordered plan for the supplied goal and constraints.
type Sprint struct {
	Goal  string `json:"goal"`
	Tasks []Task `json:"tasks"`
}

// Review is an independent assessment against the original fixture.
type Review struct {
	Findings []string `json:"findings"`
}

// SemanticIndex creates a retrieval map and gives it to a fresh reviewer.
func SemanticIndex(ctx context.Context, env gimbal.Env, corpus string) (Index, error) {
	gimbal.Set(ctx, "corpus", corpus)
	author := gimbal.NewSession(ctx, gimbal.RoleBulkClassification, env.WorkDir)
	index, err := author.Generate[Index](ctx, indexPrompt)
	if err != nil {
		return Index{}, err
	}
	if err := ValidateIndex(index); err != nil {
		return Index{}, err
	}
	gimbal.SetJSON(ctx, "candidate", index)
	reviewer := gimbal.NewSession(ctx, gimbal.RoleCodeReview, env.WorkDir)
	review, err := reviewer.Generate[Review](ctx, indexReview)
	if err != nil {
		return Index{}, err
	}
	gimbal.SetJSON(ctx, "review", review)
	if len(review.Findings) != 0 {
		return Index{}, fmt.Errorf("index review: %v", review.Findings)
	}
	return index, nil
}

// SprintPlan creates an actionable plan and gives it to a fresh reviewer.
func SprintPlan(ctx context.Context, env gimbal.Env, brief string) (Sprint, error) {
	gimbal.Set(ctx, "brief", brief)
	planner := gimbal.NewSession(ctx, gimbal.RoleSprintPlanning, env.WorkDir)
	plan, err := planner.Generate[Sprint](ctx, sprintPrompt)
	if err != nil {
		return Sprint{}, err
	}
	if err := ValidateSprint(plan); err != nil {
		return Sprint{}, err
	}
	gimbal.SetJSON(ctx, "candidate", plan)
	reviewer := gimbal.NewSession(ctx, gimbal.RoleCodeReview, env.WorkDir)
	review, err := reviewer.Generate[Review](ctx, sprintReview)
	if err != nil {
		return Sprint{}, err
	}
	gimbal.SetJSON(ctx, "review", review)
	if len(review.Findings) != 0 {
		plan, err = planner.Generate[Sprint](ctx, "Revise the plan against the original brief and the recorded review. Keep it bounded and proposed. Give explicit observable criteria; distinguish the simple local baseline from the multitier trial.")
		if err != nil {
			return plan, err
		}
		if err := ValidateSprint(plan); err != nil {
			return plan, err
		}
		gimbal.SetJSON(ctx, "revised", plan)
		finalReviewer := gimbal.NewSession(ctx, gimbal.RoleCodeReview, env.WorkDir)
		finalReview, err := finalReviewer.Generate[Review](ctx, "Review only the revised plan against the original brief. Inspect whether the earlier findings are resolved, without expanding the assignment. Return concrete material findings or an empty findings list. Make no changes.")
		if err != nil {
			return plan, err
		}
		gimbal.SetJSON(ctx, "final-review", finalReview)
		if len(finalReview.Findings) != 0 {
			return plan, fmt.Errorf("sprint review after one revision: %v", finalReview.Findings)
		}
	}
	return plan, nil
}

const indexPrompt = `Build a compact semantic index of the corpus below. Return four question-driven routes: starting a local run, isolation of concurrent app/database stacks, ownership of workflow semantics, and updating a backend while runs are active. Each answer must cite the relevant supplied source IDs, distinguish requirements from demonstrated support, and be useful to a reader choosing what to open next. Do not invent capabilities. Use only the supplied corpus. Make no file changes.`
const indexReview = `Independently inspect the candidate against the original corpus. Check all four requested retrieval questions, exact source attribution, requirements versus proven support, and unsupported promises. Return concrete findings, or an empty findings list when the index is faithful. Make no changes.`
const sprintPrompt = `Turn the supplied brief into a small sprint of three to five tasks. Name bounded outcomes, earlier task dependencies, and observable acceptance evidence. The first task must prove the ordinary single-process CLI starts, runs and finishes with container/network infrastructure blocked; this baseline is separate from the app/database trial. Prepare live observation before the trial. Cover concurrent stack isolation, startup failure after partial allocation, cancellation of an active run, and cleanup within a stated time measured from failure/cancellation. State the measurable pass condition for each task. Respect the budget and exclusions. This is a proposed plan, not completed work. Make no file changes.`
const sprintReview = `Independently inspect the candidate against the original brief. Check complete coverage, feasible ordering, concrete acceptance, budget, and exclusions. Reject claims of completed implementation, universal restartability, or mandatory Docker for basic local workflows. Return concrete findings, or an empty findings list when the plan is actionable. Make no changes.`