@probelabs/visor 0.1.106 → 0.1.107

package/defaults/task-refinement.yaml

@@ -0,0 +1,620 @@
+version: "1.0"
+
+# Simple agent: task-refinement
+# - Collects user input, refines it with AI (skip code context), and loops until refined.
+# - Returns final refined text via the `finish` step output.
+
+steps:
+  ask:
+    type: human-input
+    group: task-refinement
+    # Allow multiple turns in the refinement loop; do not guard with an `if`
+    # No explicit event trigger; run in CLI by default but guard via if
+    prompt: |
+      {% assign last_refine = outputs_history.refine | last %}
+      {% if last_refine and last_refine.refined == false %}
+      {{ last_refine.text }}
+      {% else %}
+      Provide the task you want to accomplish. Be specific about constraints
+      (inputs, outputs, environment, success criteria).
+      {% endif %}
+    multiline: false
+    allow_empty: false
+    # No on_success.goto required — refine depends_on ask
+
+  refine:
+    type: ai
+    group: task-refinement
+    # Run only after 'ask' — dependency drives ordering
+    depends_on: [ask]
+    # Allow re-execution in loop turns (no if guard)
+    ai:
+      skip_code_context: true
+      disableTools: true
+      system_prompt: |
+        You are a helpful, precise task refinement assistant (role: requirements-analyst).
+        Your goal is to get to an agreed, testable task definition and clear acceptance criteria.
+        - Refine the user's task into an unambiguous, executable description with minimal assumptions.
+        - Define how correctness will be validated (objective success criteria and measurables).
+        - If information is missing, set refined=false and ask_user=true and put a single, specific
+          clarification question in the "text" field (one question at a time).
+        - If everything is sufficient, set refined=true and put the final refined wording in "text".
+        - Be succinct and concrete. Prefer measurable outcomes over vague phrasing.
+        - You can't mark plan as refined until explicit user confirmation/agreement is implied.
+    # Schema ensures the agent either finalizes or asks to clarify
+    schema:
+      type: object
+      additionalProperties: false
+      properties:
+        refined: { type: boolean, description: "true if the task is fully specified and accepted" }
+        text: { type: string, description: "final refined task or question to user" }
+      required: [refined, text]
+    prompt: |
+      <history>
+        {% assign umsgs = outputs_history.ask | default: [] %}
+        {% assign amsgs = outputs_history.refine | default: [] %}
+        {% assign merged = umsgs | concat: amsgs | sort: 'ts' %}
+        {% for m in merged %}
+          {% if m.refined != nil %}
+            <assistant>{{ m.text }}</assistant>
+          {% else %}
+            <user>{{ m.text }}</user>
+          {% endif %}
+        {% endfor %}
+      </history>
+
+      <input>
+        {{ outputs['ask'].text }}
+      </input>
+
+    # Loop control using fail_if + on_fail (no goto_js)
+    fail_if: "output['refined'] !== true"
+    on_fail:
+      goto: ask
+
+  plan-commands:
+    type: ai
+    group: task-refinement
+    depends_on: [refine]
+    reuse_ai_session: refine
+    session_mode: append
+    ai:
+      # Allow tools so the model can inspect the repo context if needed
+      disableTools: false
+      skip_code_context: true
+      system_prompt: |
+        You are a Task Validation Planner.
+        Produce a deterministic, minimal list of shell commands that, when executed in order,
+        validate that the refined task is complete for THIS repository/local env.
+        Constraints:
+        - Commands must be safe and non-destructive. Do not delete files or push to remotes.
+        - Prefer read-only checks and standard project commands (build, lint, test) when present.
+        - Each item is a single shell line; you may use && or || inside a line if necessary.
+        - Favor idempotent commands. Avoid interactive prompts.
+        - Detect package manager/tooling pragmatically (npm/pnpm/yarn/bun, eslint/biome, jest/vitest, etc.).
+        Return JSON with an array of strings under "commands" and optional "notes".
+    schema:
+      type: object
+      additionalProperties: false
+      properties:
+        commands:
+          type: array
+          minItems: 1
+          items: { type: string, minLength: 1 }
+        notes: { type: string }
+      required: [commands]
+    prompt: |
+      Refined task:
+      {{ outputs['refine'].text }}
+
+      If the repository has build/lint/test, include them. Otherwise propose basic checks that still
+      demonstrate completion (e.g., typecheck, compile, format verification, smoke run).
+
+      {% assign prev_conf_hist = outputs_history['confirm-interpret'] | default: [] %}
+      {% assign prev_run_hist = outputs_history['run-commands'] | default: [] %}
+      {% assign last_failed = nil %}
+      {% for r in prev_run_hist %}
+        {% if r.failed and r.failed > 0 %}
+          {% assign last_failed = r %}
+        {% endif %}
+      {% endfor %}
+      {% assign last_conf = prev_conf_hist | last %}
+      {% assign prev_run_count = prev_run_hist | size %}
+      {% assign last_run = prev_run_hist | last %}
+      {% if last_failed or prev_run_count > 0 %}
+      Previous attempt failed. Here are the details to learn from:
+      - Previous commands: {{ last_conf.commands | to_json }}
+      - Run results: {{ last_failed | default: last_run | to_json }}
+      Please revise the commands to address failures. Keep the list minimal and deterministic.
+      {% endif %}
+
+      Output strictly JSON per schema. No prose around it.
+    # No on_success routing needed; dependents naturally follow in the DAG
+
+  ask-confirm:
+    type: human-input
+    group: task-refinement
+    depends_on: [plan-commands]
+    prompt: |
+      Here is the proposed validation command list (to run sequentially):
+      {% for c in outputs['plan-commands'].commands %}
+      {{ forloop.index }}. {{ c }}
+      {% endfor %}
+
+      Confirm running these? Reply "yes" to accept, or provide edits:
+      - JSON array of commands, e.g. ["npm ci", "npm test"], or
+      - Plain text with one command per line.
+    placeholder: "yes | or paste modified list..."
+    multiline: true
+    allow_empty: true
+    default: "yes"
+
+  confirm-interpret:
+    type: ai
+    group: task-refinement
+    depends_on: [plan-commands, ask-confirm]
+    ai:
+      skip_code_context: true
+      disableTools: true
+      system_prompt: |
+        You are a confirmation interpreter. Given the planned command list and the user's reply,
+        decide whether to proceed to execution with a normalized list of shell commands, or
+        return for replanning.
+        - If the user says yes/approve, set proceed=true and provide commands as-is.
+        - If the user supplied edits (JSON array or one per line), parse, trim, dedupe, and set proceed=true with commands.
+        - If the reply indicates high-level changes (not runnable commands), set proceed=false and include a short reason.
+        Output strictly JSON per schema.
+    schema:
+      type: object
+      additionalProperties: false
+      properties:
+        proceed: { type: boolean }
+        commands:
+          type: array
+          items: { type: string, minLength: 1 }
+        reason: { type: string }
+      required: [proceed]
+    prompt: |
+      Planned commands:
+      {{ outputs['plan-commands'].commands | to_json }}
+
+      User reply:
+      {{ outputs['ask-confirm'].text }}
+
+      Return JSON per schema. If proceed=true, commands must be a non-empty array of shell lines.
+    fail_if: "output && output.proceed !== true"
+    on_fail:
+      goto: plan-commands
+
+  run-commands:
+    type: command
+    criticality: internal
+    group: task-refinement
+    depends_on: [confirm-interpret]
+    assume:
+      - "outputs['confirm-interpret']?.proceed === true"
+      - "(outputs['confirm-interpret']?.commands?.length ?? 0) > 0"
+    exec: |
+      node <<'NODE'
+      const { spawn } = require('child_process');
+      const cmds = {{ outputs['confirm-interpret'] | to_json }}.commands || [];
+      const results = [];
+      const runOne = (cmd) => new Promise((resolve) => {
+        const child = spawn('bash', ['-lc', cmd], { stdio: ['ignore', 'pipe', 'pipe'] });
+        let out = '', err = '';
+        const started = Date.now();
+        child.stdout.on('data', d => (out += d.toString()));
+        child.stderr.on('data', d => (err += d.toString()));
+        child.on('close', code => {
+          results.push({ cmd, code, stdout: out, stderr: err, durationMs: Date.now() - started });
+          resolve();
+        });
+      });
+      (async () => {
+        for (const c of cmds) { await runOne(c); }
+        const failed = results.filter(r => Number(r.code||0) !== 0).length;
+        process.stdout.write(JSON.stringify({ failed, results }));
+      })().catch(e => { process.stdout.write(JSON.stringify({ failed: 1, error: String(e) })); process.exit(0); });
+      NODE
+    output_format: json
+    schema:
+      type: object
+      additionalProperties: true
+      properties:
+        failed: { type: number }
+        results:
+          type: array
+          items:
+            type: object
+            additionalProperties: true
+            properties:
+              cmd: { type: string }
+              code: {}
+              stdout: { type: string }
+              stderr: { type: string }
+              durationMs: { type: number }
+            required: [cmd, code]
+      required: [failed]
+    fail_if: "output && Number(output.failed||0) > 0"
+    on_fail:
+      goto: plan-commands
+
+  finish:
+    type: log
+    group: task-refinement
+    depends_on: [run-commands]
+    if: "(outputs && outputs['run-commands'] && Number((outputs['run-commands'].failed||0)) === 0) && !(outputs && outputs['finish'])"
+    message: |
+      ✅ Refined Task:
+      {{ outputs['refine'].text }}
+
+      ✅ Validation commands (final):
+      {% for c in outputs['confirm-interpret'].commands %}
+      - {{ c }}
+      {% endfor %}
+    level: info
+    include_pr_context: false
+    include_dependencies: false
+    include_metadata: false
+
+tests:
+  defaults:
+    strict: true
+    ai_provider: mock
+  cases:
+    - name: one-pass-refinement
+      description: Single turn; AI is happy and returns refined text.
+      event: manual
+      fixture: local.minimal
+      mocks:
+        ask: "Build a small Node CLI that prints \"hello\""
+        refine:
+          refined: true
+          text: "Create a Node.js CLI (using Node >=18) that prints 'hello' when run; include usage example and exit code 0."
+        plan-commands:
+          commands: ["echo hello-build", "echo hello-lint", "echo hello-test"]
+        ask-confirm: "yes"
+        confirm-interpret:
+          proceed: true
+          commands: ["echo hello-build","echo hello-lint","echo hello-test"]
+        run-commands:
+          stdout: '{"failed":0,"results":[{"cmd":"echo hello-build","code":0},{"cmd":"echo hello-lint","code":0},{"cmd":"echo hello-test","code":0}]}'
+      expect:
+        calls:
+          - step: plan-commands
+            exactly: 1
+          - step: ask-confirm
+            exactly: 1
+          - step: confirm-interpret
+            exactly: 1
+          - step: run-commands
+            exactly: 1
+          - step: ask
+            exactly: 1
+          - step: refine
+            exactly: 1
+          - step: finish
+            exactly: 1
+        outputs:
+          - step: run-commands
+            path: failed
+            equals: 0
+
+    - name: multi-turn-refinement-loop
+      description: Two clarifying turns followed by a final refinement; manual-only chat.
+      event: manual
+      fixture: local.minimal
+      mocks:
+        ask[]:
+          - "Create a CI job"
+          - "GitHub Actions; run on push to main"
+          - "Use Node 18 and npm ci + npm test"
+        refine[]:
+          - { refined: false, ask_user: true, text: "Which CI platform and trigger conditions?" }
+          - { refined: false, ask_user: true, text: "What Node version and commands should run?" }
+          - { refined: true, text: "Set up GitHub Actions workflow: on push to main, use Node 18.x, cache npm, run npm ci && npm test." }
+        plan-commands:
+          commands: ["echo build", "echo lint", "echo test"]
+        ask-confirm: "yes"
+        confirm-interpret:
+          proceed: true
+          commands: ["echo build","echo lint","echo test"]
+        run-commands:
+          stdout: '{"failed":0,"results":[{"cmd":"echo build","code":0},{"cmd":"echo lint","code":0},{"cmd":"echo test","code":0}]}'
+      expect:
+        calls:
+          - step: ask
+            exactly: 3
+          - step: refine
+            exactly: 3
+          - step: plan-commands
+            exactly: 1
+          - step: ask-confirm
+            exactly: 1
+          - step: confirm-interpret
+            exactly: 1
+          - step: run-commands
+            exactly: 1
+          - step: finish
+            exactly: 1
+        prompts:
+          # Ask prompt should surface the last refine clarification
+          - step: ask
+            index: 1
+            contains:
+              - "Which CI platform and trigger conditions?"
+          - step: ask
+            index: last
+            contains:
+              - "What Node version and commands should run?"
+          - step: refine
+            index: last
+            contains:
+              - "Which CI platform and trigger conditions?"
+              - "Use Node 18 and npm ci + npm test"
+              - "What Node version and commands should run?"
+          # Keep prompt assertions resilient to minor formatting changes by using 'contains'
+          # instead of a single large regex.
+        outputs:
+          # Ensure the final successful refinement carries a timestamp and expected text
+          - step: refine
+            where: { path: refined, equals: true }
+            path: ts
+            matches: "^\\d{10,}$"
+          - step: refine
+            where: { path: refined, equals: true }
+            path: text
+            matches: "(?is).*GitHub Actions.*Node 18.*npm ci.*npm test.*"
+          - step: run-commands
+            path: failed
+            equals: 0
+
+    - name: plan-and-run-success
+      description: Plans commands, user confirms, runs successfully, finishes.
+      event: manual
+      fixture: local.minimal
+      mocks:
+        ask: "Add CI to run build, lint, and tests"
+        refine:
+          refined: true
+          text: "Ensure project builds, lints, and tests pass via CI"
+        plan-commands:
+          commands: ["echo build", "echo lint", "echo test"]
+        ask-confirm: "yes"
+        confirm-interpret:
+          proceed: true
+          commands: ["echo build","echo lint","echo test"]
+        run-commands:
+          stdout: '{"failed":0,"results":[{"cmd":"echo build","code":0},{"cmd":"echo lint","code":0},{"cmd":"echo test","code":0}]}'
+      expect:
+        calls:
+          - step: ask
+            exactly: 1
+          - step: refine
+            exactly: 1
+          - step: plan-commands
+            exactly: 1
+          - step: ask-confirm
+            exactly: 1
+          - step: confirm-interpret
+            exactly: 1
+          - step: run-commands
+            exactly: 1
+          - step: finish
+            exactly: 1
+        outputs:
+          - step: run-commands
+            path: failed
+            equals: 0
+
+    - name: plan-run-fail-then-refine
+      description: First run fails, planner adjusts, second run passes, finish.
+      event: manual
+      fixture: local.minimal
+      mocks:
+        ask: "Verify app passes tests"
+        refine:
+          refined: true
+          text: "Run tests to verify app correctness"
+        plan-commands[]:
+          - { commands: ["node -e \"process.exit(1)\""] }
+          - { commands: ["node -e \"process.exit(0)\""] }
+        ask-confirm[]:
+          - "yes"
+          - "yes"
+        confirm-interpret:
+          proceed: true
+          commands: ["node -e \"process.exit(1)\""]
+        run-commands[]:
+          - { stdout: '{"failed":1,"results":[{"cmd":"node -e \\"process.exit(1)\\"","code":1,"stderr":"fail"}]}' , exit_code: 1 }
+          - { stdout: '{"failed":0,"results":[{"cmd":"node -e \\"process.exit(0)\\"","code":0}]}' }
+      expect:
+        calls:
+          - step: ask
+            exactly: 2
+          - step: refine
+            exactly: 2
+          - step: plan-commands
+            exactly: 2
+          - step: ask-confirm
+            exactly: 2
+          - step: confirm-interpret
+            exactly: 2
+          - step: run-commands
+            exactly: 2
+          - step: finish
+            exactly: 1
+        prompts:
+          - step: plan-commands
+            index: last
+            contains:
+              - "Previous attempt failed"
+              - "Run results"
+
+    - name: confirm-amend-json
+      description: User pastes JSON array of edited commands; interpreter parses and proceeds.
+      event: manual
+      fixture: local.minimal
+      mocks:
+        ask: "Make sure project installs deps and runs tests"
+        refine:
+          refined: true
+          text: "Install dependencies and run tests"
+        plan-commands:
+          commands: ["npm test"]
+        ask-confirm: '["npm ci","npm test"]'
+        confirm-interpret:
+          proceed: true
+          commands: ["npm ci","npm test"]
+        run-commands:
+          stdout: '{"failed":0,"results":[{"cmd":"npm ci","code":0},{"cmd":"npm test","code":0}]}'
+      expect:
+        calls:
+          - step: ask
+            exactly: 1
+          - step: refine
+            exactly: 1
+          - step: plan-commands
+            exactly: 1
+          - step: ask-confirm
+            exactly: 1
+          - step: confirm-interpret
+            exactly: 1
+          - step: run-commands
+            exactly: 1
+          - step: finish
+            exactly: 1
+        outputs:
+          - step: confirm-interpret
+            path: proceed
+            equals: true
+          - step: confirm-interpret
+            path: commands.length
+            equals: 2
+
+    - name: confirm-amend-lines
+      description: User pastes multi-line commands with duplicates; interpreter dedupes and proceeds.
+      event: manual
+      fixture: local.minimal
+      mocks:
+        ask: "Run lint and tests"
+        refine:
+          refined: true
+          text: "Run lint and tests"
+        plan-commands:
+          commands: ["npm run lint","npm test"]
+        ask-confirm: |
+          npm test\nnpm test\n  npm run lint  
+        confirm-interpret:
+          proceed: true
+          commands: ["npm test","npm run lint"]
+        run-commands:
+          stdout: '{"failed":0,"results":[{"cmd":"npm test","code":0},{"cmd":"npm run lint","code":0}]}'
+      expect:
+        calls:
+          - step: ask
+            exactly: 1
+          - step: refine
+            exactly: 1
+          - step: plan-commands
+            exactly: 1
+          - step: ask-confirm
+            exactly: 1
+          - step: confirm-interpret
+            exactly: 1
+          - step: run-commands
+            exactly: 1
+          - step: finish
+            exactly: 1
+        outputs:
+          - step: confirm-interpret
+            path: proceed
+            equals: true
+          - step: confirm-interpret
+            path: commands.length
+            equals: 2
+
+    - name: user-declines-replan-then-accept
+      description: User declines; interpreter routes to replan; user then accepts; run succeeds.
+      event: manual
+      fixture: local.minimal
+      mocks:
+        ask: "Set up lint and test checks"
+        refine:
+          refined: true
+          text: "Ensure lint and tests run"
+        plan-commands[]:
+          - { commands: ["npm test"] }
+          - { commands: ["npm run lint","npm test"] }
+        ask-confirm[]:
+          - "No, add lint too"
+          - "yes"
+        confirm-interpret[]:
+          - { proceed: false, reason: "needs lint" }
+          - { proceed: true, commands: ["npm run lint","npm test"] }
+        run-commands:
+          stdout: '{"failed":0,"results":[{"cmd":"npm run lint","code":0},{"cmd":"npm test","code":0}]}'
+      expect:
+        calls:
+          - step: ask
+            at_least: 1
+          - step: refine
+            at_least: 1
+          - step: plan-commands
+            exactly: 2
+          - step: ask-confirm
+            exactly: 2
+          - step: confirm-interpret
+            exactly: 2
+          - step: run-commands
+            exactly: 1
+          - step: finish
+            exactly: 1
+        outputs:
+          - step: confirm-interpret
+            index: last
+            path: proceed
+            equals: true
+
+    - name: run-fail-then-success-loop
+      description: First run fails; planner sees history; second run succeeds; prompt shows failure context.
+      event: manual
+      strict: false
+      fixture: local.minimal
+      mocks:
+        ask: "Verify app tests pass"
+        refine:
+          refined: true
+          text: "Run tests to verify app correctness"
+        plan-commands[]:
+          - { commands: ["node -e \"process.exit(1)\""] }
+          - { commands: ["echo ok"] }
+        ask-confirm[]:
+          - "yes"
+          - "yes"
+        confirm-interpret[]:
+          - { proceed: true, commands: ["node -e \"process.exit(1)\""] }
+          - { proceed: true, commands: ["echo ok"] }
+        run-commands[]:
+          - { stdout: '{"failed":1,"results":[{"cmd":"node -e \\"process.exit(1)\\"","code":1,"stderr":"fail"}]}' , exit_code: 1 }
+          - { stdout: '{"failed":0,"results":[{"cmd":"echo ok","code":0}]}' }
+      expect:
+        calls:
+          - step: ask
+            at_least: 1
+          - step: refine
+            at_least: 1
+          - step: plan-commands
+            exactly: 2
+          - step: ask-confirm
+            exactly: 2
+          - step: confirm-interpret
+            exactly: 2
+          - step: run-commands
+            exactly: 2
+        prompts:
+          - step: plan-commands
+            index: last
+            contains:
+              - "Previous attempt failed"
+              - "Run results"