github/github-spec-kit
1
0
Fork 0
mirror of https://github.com/github/spec-kit.git synced 2026-07-03 12:28:06 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(claude): stop forking /speckit-analyze to prevent long-session freezes (#3188)

Browse Source 
PR #2511 added `context: fork` + `agent: general-purpose` to the generated
speckit-analyze SKILL.md on the assumption that its heavy reads collapse to a
short summary. In practice /speckit-analyze returns a 300-500 line report that
is injected back into the main conversation. In long sessions each subsequent
fork inherits that growing context, compounding overhead until the chat
freezes (#3185).

Empty FORK_CONTEXT_COMMANDS so no command opts into context: fork, restoring
direct in-session execution for analyze. The injection mechanism is retained
so a command can be re-enabled once it genuinely returns a compact result.

Fixes #3185

Assisted-by: GitHub Copilot (model: Claude Opus 4.8, autonomous)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This commit is contained in:
Manfred Riem
2026-06-26 12:11:54 -05:00
committed by GitHub
parent 49cc05384a
commit c49966da4d
2 changed files with 49 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
src/specify_cli/integrations/claude/__init__.py
View File

@@ -22,13 +22,17 @@ ARGUMENT_HINTS: dict[str, str] = {
}
# Per-command frontmatter overrides for skills that should run in a forked
# subagent context. Read-only analysis commands are good candidates: the
# heavy reads (spec/plan/tasks artefacts) collapse to a short summary,
# so isolating them keeps the main conversation context clean.
# See https://code.claude.com/docs/en/skills#run-skills-in-a-subagent
FORK_CONTEXT_COMMANDS: dict[str, dict[str, str]] = {
"analyze": {"context": "fork", "agent": "general-purpose"},
}
# subagent context. See https://code.claude.com/docs/en/skills#run-skills-in-a-subagent
#
# This is intentionally empty. ``analyze`` was previously forked (added in
# #2511) on the assumption that its heavy reads collapse to a short summary,
# but in practice ``/speckit-analyze`` returns a 300-500 line report that is
# injected back into the main conversation. In long sessions each subsequent
# fork inherits that growing context, compounding overhead until the chat
# freezes (#3185). Until a command genuinely returns a compact result, no
# command opts into ``context: fork``. The injection mechanism below stays in
# place so a future command can be added here when that holds true.
FORK_CONTEXT_COMMANDS: dict[str, dict[str, str]] = {}
class ClaudeIntegration(SkillsIntegration):