[Bug]: Speckit skills are not interactive

[jkrishnarao2003]

Bug Description

I am using spec-kit latest version 1.0.6 for 3 projects. These are the problems I faced.

1. Most of the skills don't ask questions from the user. It just start doing it's join as and when I submit the /speckit-*** command
2. In my recent case there were 60 tasks that were documented by /speckit-tasks. But when I tried /speckit-implement command it' straight away implemented 25 tasks until I realized that it's continuously implementing the tasks without asking me for review. My point is the command should be interactive like implement task by task and then ask user's review and user's confirmation before proceeding to implement next task.

Steps to Reproduce

I tried these commands in the order FYI

1. /speckit-constitution
2. /speckit-specify
3. /speckit-plan
4. /speckit-tasks
5. /speckit-implement

Expected Behavior

1. /speckit-*** skills should be interactive with the user and ask questions to user to get more context
2. /speckit-*** skills should implement task by task and then ask user's review and user's confirmation before proceeding to implement next task.

Actual Behavior

1. /speckit-*** skills are behaving mostly autonomous and not asking user's inputs for getting more context.
2. /speckit-*** skills are not asking for user's review and continually implementing on it's own.

Specify CLI Version

1.0.6

AI Agent

GitHub Copilot

Operating System

Windows 11

Python Version

Python 3.13.15

Error Logs

No error logs

Additional Context

No response

[github-actions]

Bug assessment — skills-not-interactive: Invalid · severity low

---

Bug Assessment: Speckit skills are not interactive

• Slug: skills-not-interactive
• Created: 2026-09-15T22:22:41Z
• Source: issue [Bug]: Speckit skills are not interactive #4582
• Verdict: invalid
• Severity: low

Report (summarized)

The reporter used Spec Kit 1.0.6 with GitHub Copilot on Windows 11 and ran /speckit-constitution, /speckit-specify, /speckit-plan, /speckit-tasks, and /speckit-implement. They expected every skill to ask for additional context and expected /speckit-implement to pause for review and confirmation after each task. They report that the skills proceeded autonomously and that implementation continued through approximately 25 of 60 tasks without an approval prompt. No error logs were provided.

No issue comments supplied additional reproduction details or context.

Symptom

The reported behavior is consistent with the documented command design: the implementation skill is instructed to process all tasks in tasks.md in order, while the quickstart asks the user to invoke each skill one at a time and review its result before continuing. The core skills are not specified to prompt for arbitrary clarification at every stage, and clarification is a separate /speckit-clarify workflow.

Reproduction

1. In a project initialized with Spec Kit, create a feature through /speckit-constitution, /speckit-specify, /speckit-plan, and /speckit-tasks.
2. Invoke /speckit-implement with a populated tasks.md.
3. Observe that the command instructs the agent to execute the implementation tasks in order rather than asking for confirmation between individual tasks.

[NEEDS CLARIFICATION: The report does not include the generated command file, Copilot transcript, project configuration, or the exact task file, so the claimed count of 25 completed tasks cannot be independently tied to a specific agent execution.]

Suspected Code Paths

• templates/commands/implement.md:2 — describes /speckit-implement as processing and executing all tasks defined in tasks.md.
• templates/commands/implement.md:149-171 — directs the agent to execute the implementation plan phase by phase, report progress, and mark completed tasks; the only explicit user confirmation gate is for unchecked checklists at approximately lines 81-82.
• presets/lean/commands/speckit.implement.md:17-21 — the lean preset likewise directs the agent to execute tasks in order and halt only on failure, with no per-task approval gate.
• README.md:61-63 — documents the intended interaction boundary as invoking each /speckit-* skill one at a time and reviewing the result before continuing.
• templates/commands/clarify.md — provides the dedicated workflow for identifying underspecified areas and asking targeted clarification questions; the other lifecycle commands are not required to conduct an open-ended interview.

Root Cause Hypothesis

High confidence: this is an expectation/documentation mismatch rather than an implementation defect. The requested per-task confirmation behavior is not part of the current /speckit-implement contract; its contract is batch execution of the task plan, subject to checklist and failure gates. Whether Copilot pauses in practice is additionally agent-dependent, but the supplied report does not show a Spec Kit error or a deviation from the template instructions.

Proposed Remediation

Preferred: Close as not a bug and clarify usage expectations in the issue response: use /speckit-clarify when requirements need targeted questions, invoke skills one at a time, and review the result between skills. For a manual checkpoint, split the work into smaller task groups or provide an explicit instruction to stop after a chosen task before invoking /speckit-implement.

If maintainers want a new behavior, track it as an enhancement: add an explicit opt-in interactive mode (for example, a command argument or integration option) that pauses after each task and requires confirmation. The mode would need to define how task completion is recorded, how tests and dependent tasks are handled, and how it behaves across agents that do not support interactive prompts consistently.

Files likely to change:

• templates/commands/implement.md — only if an opt-in checkpoint mode is approved.
• presets/lean/commands/speckit.implement.md — keep preset behavior aligned with the core template.
• README.md and relevant guides — document the supported interaction modes and checkpoint semantics.
• Relevant command-template or integration tests — lock down generated content and the opt-in behavior.

Tests to add or update:

• Verify the default implement template remains batch-oriented and retains its existing checklist gate.
• If an interactive mode is introduced, verify it emits a checkpoint after each completed task, waits for confirmation, preserves task markers, and resumes correctly after approval.
• Add coverage for malformed or unsupported mode arguments and for agents/integrations that cannot provide reliable interactive pauses.

Risks & Considerations

• Changing the default to pause after every task would substantially slow existing workflows and could break unattended or CI-oriented agent usage.
• Interactive behavior is partly controlled by the coding agent, so prompt text alone cannot guarantee a pause across all supported integrations.
• A new mode must avoid leaving tasks marked complete when implementation or validation has not actually finished.
• The reported version (1.0.6) is older than the current checked-out code, so any version-specific regression would require a transcript and exact generated command files to assess.

Open Questions

• [NEEDS CLARIFICATION: Which specific skill failed to ask a question when its input was materially underspecified?]
• [NEEDS CLARIFICATION: Did the Copilot transcript include an explicit request to pause after each task, or was the expectation inferred from the workflow description?]
• [NEEDS CLARIFICATION: Does the behavior reproduce with the current Spec Kit release and a freshly initialized project?]

Generated by 🐛 Assess Bug from Labeled Issue for #4582 · copilot · gpt52codex · 2.29 AIC · ⌖ 5.45 AIC · ⊞ 25.5K · ◷

[mnriem]

You can supply additional instructions with each skill invocation. By default, /speckit-implement works through the task plan rather than asking for approval after every task.

To request one task at a time, try this in Copilot chat:

/speckit-implement Implement only T001. Stop afterward for my review.
Do not start another task until I approve.

This gives Copilot an explicit instruction; it is not a mechanically enforced approval gate.

Could you share the exact invocation you used, including any text after the skill name? If you already requested clarification or a review stop and it was ignored, please include a short, redacted conversation excerpt showing that.

Drafted for @mnriem with assistance from GitHub Copilot (model: GPT-6 Astra; interactive comment drafting).

[jkrishnarao2003]

I did instructed the skill to stop next task until I approve. By the time I realized the prompt already implemented 25 tasks. So my whole point is - as a best practice, AI skills should adhere HIL (Human In the Loop) unless user instructed to become autonomous.