You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit d0edde5
Browse filesBrowse the repository at this point in the historyBrowse files
description = "Self-evolution loop for OpenAI Codex: session reflection, session recall, and reflection-generated skills with zero Python runtime dependencies."
For historical session search, use the `csep-session-recall` plugin skill.
4
-
Treat `csep recall` like `rg` over past sessions: search short concrete needles, not long natural-language questions.
3
+
<IMPORTANT>
4
+
Before answering or taking action, check whether this task may depend on prior local session context.
5
+
6
+
If there is a reasonable chance that previous Codex or Claude Code sessions contain useful evidence, you MUST use the `csep-session-recall` plugin skill before proceeding.
7
+
8
+
Do not rely only on the stable background injected at startup. This policy does not install or rely on UserPromptSubmit automatic recall; the agent must actively choose the recall skill. When unsure, use recall.
9
+
</IMPORTANT>
10
+
11
+
## Use Recall For
12
+
13
+
- today's, recent, previous, unfinished, or continued work
14
+
- status, handoff, cleanup, wrap-up, or "what remains"
15
+
- bugs, failures, repeated issues, regressions, logs, or test output
16
+
- previous commands, branches, commits, reviews, hooks, skills, local tools, or config state
Skip recall only when one of these boundaries clearly applies:
25
+
26
+
- the task is clearly self-contained, such as translation, formatting, arithmetic, or a tiny edit with complete context
27
+
- the user explicitly says not to use recall or not to check history
28
+
29
+
## Query Style
30
+
31
+
Treat `csep recall` like `rg` over past sessions. Extract 2-6 short concrete needles: file names, commands, errors, symbols, hook names, config keys, branch names, or exact user wording.
Avoid long natural-language questions. Use recall results as evidence, not authority; current files, live command output, and explicit user instructions take precedence.
0 commit comments