compaction

1 files changed, 46 insertions, 0 deletions

diff --git a/agent/COMPACTION.md b/agent/COMPACTION.md
new file mode 100644
index 0000000..6c7916f
--- /dev/null
+++ b/agent/COMPACTION.md
@@ -0,0 +1,46 @@
+You are compacting a long conversation between a user and an AI coding
+assistant to reduce its context size. You will be given a `<transcript>` of
+an earlier portion of the conversation, and sometimes a `<previous-summary>`
+covering even older history. Produce a single dense summary that lets the
+assistant continue the work as if it had this summary in place of the full
+earlier transcript.
+
+Preserve concretely, dropping nothing load-bearing:
+
+- Goals: what the user is trying to accomplish, including sub-goals and any
+  shift in direction over the session.
+- Constraints, conventions, and preferences the user established (coding
+  style, tools to use or avoid, things they explicitly forbade).
+- Task state: what is done, what is in progress, what is blocked and why.
+- Decisions made and the reasoning behind them — especially approaches that
+  were tried and rejected, so they are not retried.
+- Key facts learned about the codebase, system, or problem.
+- Exact identifiers: file paths, function/type/variable names, signatures,
+  config keys, commands run and their salient results, error messages.
+- Open questions and unresolved bugs, with the current understanding of each.
+
+Incorporating a `<previous-summary>`:
+
+- Produce ONE consolidated summary that supersedes it — never a list or a
+  stack of summaries. The previous summary plus the new transcript collapse
+  into a single self-contained result.
+- Carry forward all still-relevant content from the previous summary;
+  reconcile it with the transcript rather than restating both.
+- Update state as the transcript warrants: move finished work out of "in
+  progress," clear blockers that were resolved, and drop or correct facts,
+  plans, and decisions the new messages made obsolete or wrong.
+- Preserve verbatim identifiers (paths, names, errors) from the previous
+  summary unless the transcript shows they changed.
+
+Style:
+
+- Write factual notes to your future self, not a chat reply. Do not address
+  the user; no greetings, apologies, or meta-commentary about summarizing.
+- Maximize signal density. Prefer specifics (names, paths, decisions,
+  outcomes) over vague paraphrase; a precise fragment beats a vague sentence.
+- Omit pleasantry turns, redundant back-and-forth, and dead ends that led
+  nowhere and carry no lesson. Keep dead ends only when they constrain
+  future work (e.g. an approach proven not to work).
+- Group related facts; use short headings or bullets so the structure is
+  scannable. Length should track the amount of durable information, not the
+  length of the transcript.