Activity

1. Gambit-Checkmate commented on Aug 22, 2026

   

   Gambit-Checkmate

   on Aug 22, 2026

   More actions

   Hi, I’d like to investigate the missing speckit-converge entry in the manifest generated by init. I can reproduce it with a clean installation, add a regression test for the generated manifest, and make the smallest fix that keeps existing manifests compatible. Is this issue open for an external PR?
2. mnriem

   added

   bug-assess

   on Sep 1, 2026
3. mnriem commented on Sep 1, 2026

   

   mnriem

   on Sep 1, 2026

   Last edited by mnriem

   Collaborator

   More actions

   @Gambit-Checkmate Absolutely, go for it. I have triggered our bug assess pipeline so you will get some additional background
4. github-actions commented on Sep 1, 2026

   

   github-actionsBot

   on Sep 1, 2026 – with GitHub Actions

   Contributor

   More actions

   Bug assessment — converge-missing-manifest: Valid · severity medium

   ---

   Bug Assessment: converge-missing-manifest

   • Slug: converge-missing-manifest
   • Created: 2026-09-01T13:43:44Z
   • Source: issue [Bug]: specify init writes speckit.manifest.json without the speckit-converge skill it just installed #4273
   • Verdict: valid
   • Severity: medium

   Report (summarized)

   After running specify init --here --integration claude --force with spec-kit v1.0.1 (upgrading from v0.5.0), the file .specify/integrations/claude.manifest.json (referred to as speckit.manifest.json in the report — see note below) contains 9 speckit-* SKILL.md entries but is missing speckit-converge/SKILL.md, even though that file was successfully installed at .claude/skills/speckit-converge/SKILL.md. Observed on two independent in-place upgrades.

   Note on filename: .specify/integrations/speckit.manifest.json is the shared infrastructure manifest (tracks scripts/templates). The per-integration manifest for claude lives at .specify/integrations/claude.manifest.json. Based on the description ("9 speckit-* SKILL.md entries"), the reporter is looking at claude.manifest.json and using "speckit.manifest.json" generically.

   Symptom

   specify init --here --integration claude --force installs all 10 speckit-* SKILL.md files to .claude/skills/ (including speckit-converge/SKILL.md) but the resulting claude.manifest.json records only 9 of them — speckit-converge/SKILL.md is absent. Any tooling that consults the manifest to classify files (pristine / modified / missing / untracked) will treat speckit-converge/SKILL.md as untracked, preventing clean uninstall, upgrade diff-detection, and hash-based freshness checks for that skill.

   Reproduction

   1. Create a project with spec-kit v0.5.0: specify init myproject --integration claude
   2. Upgrade the CLI to v1.0.1.
   3. From inside myproject/, run specify init --here --integration claude --force.
   4. Inspect .specify/integrations/claude.manifest.json — check files keys for speckit-*/SKILL.md entries.
   5. Observe that .claude/skills/speckit-converge/SKILL.md exists on disk but has no corresponding entry in the manifest.

   [NEEDS CLARIFICATION: Whether the same omission occurs on a fresh specify init (no prior install) with v1.0.1 or only on --force re-inits of existing projects.]

   Suspected Code Paths

   • src/specify_cli/integrations/base.py:1725–1818 (SkillsIntegration.setup() loop) — the main loop that calls write_file_and_record() for each skill. Converge is index 4 of _CORE_COMMAND_TEMPLATE_ORDER (line 49); the loop should process all 10 templates without exclusion.
   • src/specify_cli/integrations/base.py:558–576 (write_file_and_record()) — writes the file, then calls manifest.record_existing(rel). If record_existing raises (e.g. for a symlinked path component) and this exception is swallowed somewhere, the file is on disk but not in the manifest.
   • src/specify_cli/integrations/manifest.py:168–233 (record_existing()) — raises ValueError if any path component traversed from self.project_root is a symlink, or if the file is not a regular file. This is a plausible silent-failure site if an outer context swallows the exception for one file.
   • src/specify_cli/commands/init.py:665–698 — the init entry point for integration setup. Creates a fresh IntegrationManifest, calls setup(), then manifest.save(). After manifest.save() (line 698), _register_extensions_for_agent and _register_presets_for_agent are called when --force is set (lines 700–722). These post-save calls do not affect the already-saved integration manifest but they do run after the file is saved, creating a potential window.
   • src/specify_cli/integrations/base.py:466–479 (list_command_templates()) — lists all .md files from shared_commands_dir(). Converge is present in templates/commands/converge.md and would be returned; the sort key is distinct (_CORE_COMMAND_TEMPLATE_RANK["converge"] == 4). No exclusion here.

   Root Cause Hypothesis

   Converge was introduced in commit 0c29d890 (feat: add /speckit.converge command, June 2026), making it the most recently added core command when v1.0.1 shipped. The SkillsIntegration.setup() loop has no try/except around the per-skill write; any exception would abort setup entirely, not silently skip one file. This rules out most exception-swallowing theories.

   The most credible scenario is an environmental interaction specific to the upgrade path: on a v0.5.0 project, the .claude/skills/ directory existed with 9 subdirectories pre-installed. When specify init --force runs setup() over an already-populated tree, dest.parent.mkdir(parents=True, exist_ok=True) silently succeeds for all 10 skill directories. For converge specifically, if the newly created speckit-converge/ directory resolves differently (e.g., a stale inode, platform-specific casing, or a symlink left by a prior install) and manifest.record_existing() subsequently raises ValueError for that path after the file is written, then...

   However, setup() does not have a per-iteration try/except — a ValueError from record_existing would propagate to init.py's outer try/except Exception handler (line 936), which would print an error and exit. The user would not see a "success with 9 files" result. This means the exception-path theory requires a currently-invisible swallowing site.

   Most likely: The bug is a missing coverage gap — write_file_and_record calls manifest.record_existing(rel) where rel is computed as dest.resolve().relative_to(project_root.resolve()) (a Path object). record_existing re-walks from manifest.project_root to verify no symlinks. If project_root passed to write_file_and_record differs from manifest.project_root (even by trailing slash or symlink resolution), the reconstructed absolute path for converge's SKILL.md may not match what was written, so _sha256(abs_path) operates on a slightly different path. The save() then writes only the keys that were successfully recorded.

   Confidence: medium. The code path is clear, but reproducing it requires the specific v0.5.0→v1.0.1 upgrade environment. A targeted regression test would either confirm or rule out this theory.

   Proposed Remediation

   Preferred: Add an explicit assertion / test that verifies the count of manifest entries equals the count of installed SKILL.md files after setup() in both fresh-install and re-init (--force) scenarios. Then add the same check inside SkillsIntegration.setup() as a postcondition assertion (or a warning log) that counts len([f for f in created if f.name == "SKILL.md"]) against len(manifest.files filtered to SKILL.md paths). If they diverge, raise so the bug surfaces loudly rather than producing a silently-partial manifest.

   Additionally, in write_file_and_record, replace the manifest.record_existing(rel) call with the more direct manifest.record_file(rel, ...) pattern (which writes content and records in one step, computing the hash from the in-memory bytes rather than re-reading from disk). This eliminates the window between write and record and avoids any potential path-resolution discrepancy.

   Alternative: Add a post-setup reconciliation step in SkillsIntegration.setup() after the loop: iterate skills_dir.rglob("SKILL.md"), and for any SKILL.md found under a speckit-* directory that is NOT yet in the manifest, call manifest.record_existing() on it. This is a belt-and-suspenders backstop that self-heals silently omitted entries.

   Files likely to change:

   • src/specify_cli/integrations/base.py (SkillsIntegration.setup loop — add postcondition or switch to record_file)
   • tests/integrations/test_integration_base_skills.py (add a test that verifies exactly N SKILL.md entries appear in manifest.files, exercising re-init over a pre-populated skills directory)
   • tests/integrations/test_integration_claude.py (add the same assertion for the Claude-specific subclass)

   Tests to add or update:

   • A test that pre-populates a tmp_path with 9 SKILL.md stubs under .claude/skills/speckit-*/ (simulating an existing v0.5.0 install without converge) and then calls setup() with a fresh manifest — verifies all 10 skills including converge appear in manifest.files after setup().
   • A test that calls setup() twice with a fresh manifest each time (simulating --force reinit) and asserts both manifest objects record the same 10 entries.
   • Extend test_all_files_tracked_in_manifest to also check the inverse: that len([k for k in m.files if k.endswith("SKILL.md")]) equals len(templates) (currently the test only checks files in created are tracked, not that ALL templates produced entries).

   Risks & Considerations

   • Uninstall breakage: specify integration uninstall claude will not remove speckit-converge/SKILL.md because it is not in the manifest. The directory .claude/skills/speckit-converge/ is also not removed. This is already occurring for any project that hit this bug.
   • Upgrade stale-file pass: The integration upgrade stale-file pass (Phase 2 in _migrate_commands.py:844–867) computes stale_keys = old_files - new_files. If the old manifest is missing converge, the stale set will never include converge — so a subsequent upgrade won't try to remove a stale converge skill. This is benign but means the broken state persists silently across upgrades.
   • No data loss: The converge skill file is fully intact on disk; only the manifest entry is missing. Recovery is manual: re-run specify integration upgrade claude or specify init --here --force once the bug is fixed.
   • Blast radius: Affects every skills-based integration (not just claude) that ran specify init --force over a pre-existing project after converge was added in v1.0.1.

   Open Questions

   • [NEEDS CLARIFICATION: Does a fresh specify init (no prior .claude/skills/ directory) on v1.0.1 also omit converge, or is it strictly a --force/upgrade-path issue?]
   • [NEEDS CLARIFICATION: Is the missing entry consistently the converge skill, or could it be whichever skill happens to fall at position 4 in the template order? If the issue is positional (e.g., an off-by-one in a loop boundary), earlier or later commands would also be affected.]

   ---

   Posted on behalf of @mnriem by GitHub Copilot (model: claude-sonnet-4.6)

   Generated by 🐛 Assess Bug from Labeled Issue for issue #4273 · 2.8K AIC · ⌖ 23.4 AIC · ⊞ 34.2K · ◷
5. github-actions

   added

   severity-medium

   on Sep 1, 2026
6. mnriem

   added

   triage-must-haveVerdict: high-value, important work for Spec Kit — do firstVerdict: high-value, important work for Spec Kit — do first

   bug-fixTrigger the bug-fix agentic workflowTrigger the bug-fix agentic workflow

   on Oct 15, 2026
7. github-actions

   added a commit that references this issue on Oct 15, 2026

   Fix converge-missing-manifest: track generated skill files

   ...

   bb3f143
8. github-actions commented on Oct 15, 2026

   

   github-actionsBot

   on Oct 15, 2026 – with GitHub Actions

   Contributor

   More actions

   Pull request created: #4990

   Generated by Fix Bug from Labeled Issue · copilot · gpt52codex · 3.34 AIC · ⌖ 0.696 AIC · ⊞ 24.6K
9. github-actions commented on Oct 15, 2026

   

   github-actionsBot

   on Oct 15, 2026 – with GitHub Actions

   Contributor

   More actions

   Draft fix PR created on fix/4273-converge-missing-manifest: converge-missing-manifest now records generated skill files atomically and verifies every SKILL.md is present in the manifest. Please review the draft and validate it in the stage-3 bug-test workflow, if configured.

   Assisted-by: GitHub Copilot (model: gpt-5.2-codex, autonomous)

   Generated by 🛠️ Fix Bug from Labeled Issue for #4273 · copilot · gpt52codex · 3.34 AIC · ⌖ 0.696 AIC · ⊞ 24.6K · ◷
10. github-actions

    added

    fix-proposed

    on Oct 15, 2026
11. mnriem

    added

    bug-test

    on Oct 19, 2026
12. github-actions

    added

    tests-passing

    on Oct 19, 2026
13. github-actions commented on Oct 19, 2026

    

    github-actionsBot

    on Oct 19, 2026 – with GitHub Actions

    Contributor

    More actions

    Bug test — converge-missing-manifest: ✅ passing · 58 passed, 0 failed, 0 skipped · fix from linked PR #4990 (head bb3f14322fcd9fb7618d63e281557caa7d4a3e48)

    ---

    Bug Test Report: specify init omits speckit-converge from the manifest

    • Slug: converge-missing-manifest
    • Date: 2026-10-19T12:19:05Z
    • Source issue: [Bug]: specify init writes speckit.manifest.json without the speckit-converge skill it just installed #4273
    • Fix under test: linked PR [bug-fix] Fix converge-missing-manifest: track generated skill files #4990 (head bb3f14322fcd9fb7618d63e281557caa7d4a3e48)
    • Test command: timeout 600 python3 -m pytest tests/integrations/test_integration_base_skills.py tests/integrations/test_integration_claude.py
    • Scope: focused subset
    • Result: passing

    Summary

    The proposed fix's focused integration tests all passed. The regression coverage confirms that generated skill files, including speckit-converge/SKILL.md, are recorded in the manifest both for a fresh setup and when setup runs over an existing nine-skill installation.

    Test Results

    Metric	Count
    Passed	58
    Failed	0
    Skipped	0
    Errored	0
    Duration	approximately 1.83 seconds

    Failures (if any)

    None.

    Verification vs. Historical Fix

    Partially matches. Historical PR #4405 overlaps with the generated fix on the core remediation: routing generated integration files through IntegrationManifest.record_file, plus Claude manifest coverage. The generated PR #4990 additionally includes a SkillsIntegration.setup() postcondition and base integration tests for the generated-to-manifest inverse relationship and re-init over a pre-existing nine-skill installation. PR #4405 is closed and was not merged; no historical implementation was used as the test revision.

    Notes & Caveats

    • Tests were run against the linked PR head, not the base branch.
    • The repository is Python-based (pyproject.toml with pytest configuration). Pytest and runtime dependencies were already available; no dependency installation was needed.
    • Only the focused integration test files identified by the bug assessment and fix PR were run; the full suite was not run.

    Generated by 🧪 Test a Bug Fix from a Labeled Issue for #4273 · copilot · gpt52codex · 1.99 AIC · ⌖ 0.628 AIC · ⊞ 24.9K · ◷