v1.10.2

Patch Changes

• c784af7: Kiro explorations no longer fail when Kiro throttles them. A throttled exploration (session context with AI, or the plan's requirements and design stages) now drafts from the repository summary instead, and says so. A failed kiro-cli or claude run now shows its real error line instead of the first lines of tool progress, and "throttled" counts as a rate limit, so single-shot calls retry it. Spec plan documents no longer start with the model's narration ("Now I have enough context…"): everything before the stage's first section is dropped.

v1.10.1

Patch Changes

• b19d284: recover now refuses the base branch (main, develop…) in the CLI, the dry run and the menu, and asks for the failed branch or its ticket. Recovering the base branch used to read a tree without anything the pipeline wrote, so the draft and the spec plan built on it planned around work that already existed. The pre-filled recovery goal no longer becomes the title of the context and of the plan built from it; anything you add to the goal is still kept.

v1.10.0

Minor Changes

• 703513e: Add sinscribe recover to take over a branch an automated pipeline could not finish. It fetches, finds the branch from a ticket or its name, and opens it in place or in its own worktree (--worktree). The AI then reads what the pipeline left (specs, plans, logs, its commits, and an optional --from diagnosis), read-only, and drafts the branch's session context as a takeover brief with Failure: and Remaining: lines. It never proposes weakening the tests. After you approve it, continue with a quick bugfix prompt or the spec plan. The menu gains Recover a failed branch for the current branch, and -p --save drafts and saves the context without review for scripts.

  Small terminals are handled better along the way. The header's brand line and subtitle no longer wrap past the rows reserved for them. The AI session-context screens drop their extra lines instead of growing taller than the terminal (which froze the CLI). An approved context is no longer dropped when it is approved before the menu finishes reading the repository.

v1.9.0

Minor Changes

• 85013b8: Agent-facing output now follows tracked standards. agents --target both writes AGENTS.md as the canonical file and a CLAUDE.md that imports it with @AGENTS.md, so they no longer drift apart. agent-setup stops pinning model: sonnet (definitions inherit the session's model) and gives read-only roles a restricted tools: list. Both commands check what they wrote (frontmatter, name shape, line limits) and list any problem under "Standards check". Spec plans name EARS notation and stamp the standards they follow in their header. agents --dry-run lists which tools read AGENTS.md natively (Codex, Claude Code, Kiro, Cursor, GitHub Copilot, Gemini CLI) and how.

v1.8.0

Minor Changes

• dec4a2a: Generated markdown now lives in .sinscribe/ with lowercase names, so nothing lands at the repo root anymore: .sinscribe/handoff.md, .sinscribe/pr-description.md, .sinscribe/agent-prompt.md, .sinscribe/project-documentation.md, and spec plans in .sinscribe/specs/<branch>/ with loop-prompt.md. An existing root HANDOFF.md is still read as context until the next save writes the new file. A plan still in specs/<branch>/ stops sinscribe plan with the exact git mv to move it, rather than starting a second plan.

v1.7.2

Patch Changes

• 190537e: Support kiro-cli 2.28 when Kiro explores the repository. 2.28 prints only the model's text on stdout and moves tool traffic to stderr, so the session context and spec plan failed with "explored but returned no answer". The answer now starts after the last completed tool call, and reads are taken from the stderr tool lines; kiro-cli 2.3.0 keeps working. When no answer is found, the error now quotes Kiro's last output. The read-only confinement was re-verified on 2.28: reads outside the repository and on denied paths are refused.

v1.7.1

Patch Changes

• 195b22a: Fix the issues found reviewing the AI session context and spec plan work. The read-only explorer for API-key providers can no longer write conversation history or evicted tool results into the repository. The session-context screen windows its exploration log, so long runs no longer freeze the terminal. Secret redaction keeps ordinary code references such as token: RefreshTokenResponse. Non-ASCII file names such as docs/diseño.md are listed correctly. In the spec plan: Modify and "Fix validator issues" revise the draft you reviewed instead of starting over; --feedback works in the interactive plan; a hand-edited approved stage is never regenerated over your edits; "Accept hand edits" works across several edited stages; a failed regeneration keeps your feedback and can go back to the draft; progress counts only [T-n] commits from this plan; and the loop prompt no longer has the agent edit tasks.md. Reads that were denied are no longer listed as files read. The Kiro answer is no longer cut short by lines that look like tool output. Claude's exhausted turn budget is reported clearly.

v1.7.0

Minor Changes

• e80cb95: Add "Generate session context with AI" to the menu. You give the direction — what the session is for — and the AI backs it with evidence from the repository: code and markdown documents (reports, notes, specs, handoffs), read-only with every provider: the claude CLI under --restricted, Kiro through an agent whose only tool is an untrusted fs_read confined to the repository, and a write-denied agent for API-key providers. Kiro's read-only explorer also grounds the spec plan's requirements and design. The draft shows its sources and open questions; refine the goal, add details or answers, send it back into the repository, edit it by hand, then approve — nothing is saved before that. A branch without a context now asks whether to generate it with AI or write it manually.

v1.6.0

Minor Changes

• 69f0a66: Add sinscribe plan and the "Spec plan (SDD)" menu item: spec-driven development for the current branch. Requirements → design → tasks → handoff are generated from the session context, reviewed and approved in order, and written to specs/<branch>/ with an index.md that tracks approval and staleness. Requirements and design can read the repository first, read-only (the claude CLI under --restricted with Read/Glob/Grep, or a write-denied agent for API-key providers), with a single-shot fallback and --no-explore. tasks.md is validated for AC coverage and dangling references, and approving it writes LOOP_PROMPT.md, a task-by-task contract for your coding agent. --approve, --sync and --loop-prompt run offline.

v1.5.0

Minor Changes

• e915255: 🐛 Fix OpenCode Go requests failing with 400 Request is missing x-opencode-session. OpenCode Go now routes only clients that identify
  themselves, so every request sends User-Agent: sinscribe/<version>,
  x-opencode-client: sinscribe and an x-opencode-session id: one per
  single-shot run, and the conversation's thread id for agentic commands, so
  chat keeps one session across turns. The settings Test connection probe
  sends the same headers.

  ✨ Add the claude-cli provider: drives your signed-in Claude Code CLI
  (claude -p --output-format stream-json) the way kiro-cli drives Kiro, so
  pr/commit/branch/prompt can use a Claude subscription with no API key.
  It runs with no tools and isolated from your Claude Code settings, hooks, MCP
  servers, skills and repository CLAUDE.md; models are the CLI aliases
  sonnet (default), haiku, opus and fable.

Patch Changes

• e414997: 📝 Compact the documentation and harden the private-files boundary.

  .gitignore is now deny-by-default for .claude/. Three gaps closed:
  .claude/settings.local.json was excluded only by a developer's global
  gitignore, so a fresh clone or CI would have tracked it; .claude/rules/,
  .claude/agents/ and .claude/commands/ were not excluded at all; and
  HANDOFF.md — which sinscribe prompt writes at the repository root and
  deliberately never self-ignores — was trackable in a public repo. The
  root-anchored /ARCHITECTURE.md entry is unchanged, and docs/ARCHITECTURE.md
  is verified still tracked. docs/CONTRIBUTING.md gains the matching house rule:
  run git check-ignore -v <path> before adding a doc or dotfile.

  The documentation drops from 1,633 to 1,292 lines with no information removed,
  by deleting restatement rather than content. docs/testing/COVERAGE-PLAN.md was
  a session report rather than documentation — it is gone, but its two durable
  testing conventions (mock only at true I/O boundaries; redirect homedir() in
  any test touching ~/.sinscribe) moved into docs/CONTRIBUTING.md. DESIGN.md
  is now decisions-only: its command surface, template schema and folder tree were
  duplicates of README.md and docs/ARCHITECTURE.md and now link to them.
  docs/ARCHITECTURE.md's "Extension points" pointed at the same three recipes
  docs/CONTRIBUTING.md already carried in a better form, so it now points there
  and keeps only the shape those procedures assume.

  Three stale facts are corrected against src/: base-ref resolution also probes
  origin/develop and develop (src/git/repo.ts); the diff marker is
  [diff truncated to N bytes], not [truncated: N more files] (src/git/diff.ts);
  and the README's status note read "Stable (v1.0.0)" while the package was at
  1.4.1. CHANGELOG.md loses a clause that advertised the existence and location
  of untracked local notes, and three bullets that began with # — the exact
  defect docs/CONTRIBUTING.md warns about — are fixed.