session-start-hook.sh ignores the `source` field, re-injecting the full memory recap on every `compact`

[derekwitucki]

session-start-hook.sh reads HOOK_STDIN to extract session_id (per #206/#270's comments: "this hook never read its stdin, so it had neither source nor session_id... Only session_id is needed") — but the payload also carries source, and it's discarded. The recap (=== MEMORY === and friends) is printed identically regardless of whether the trigger is startup, resume, compact, clear, or fork.

That's right for startup/resume/clear when the model's context is empty or being replaced. It's not right for compact: at that moment the conversation already has a just-produced summary of everything that happened in it. Re-adding a second, differently-sourced recap of the same general period on top is pure duplication, and it happens on every single auto-compaction.

Measured, not guessed (50 sessions, ~40 days, one user):

SessionStart source	fired	non-empty	total chars	avg/fire	max
compact	82	41	37,634	459	2,193
startup	180	84	71,039	395	3,368
resume	48	22	45,685	952	4,934

compact fired essentially as often as startup, at comparable per-fire cost, injecting content the model's own compaction summary already substantially overlaps with.

Proposed fix: extract source the same narrow way session_id already is (no new dependency), and skip (or emit a drastically shorter marker — e.g. just the .remember/ path reminder, no recap body) when source = "compact". startup/resume/clear/fork unchanged.

Happy to provide the extraction snippet in the same style as the existing _stdin_session_id heuristic if useful.

[fdaviddpt]

Accepted, and thank you — this is a good report.

I verified the premise before acting on it rather than taking it on trust, and it holds: _stdin_session_id at scripts/session-start-hook.sh:136 is the only thing pulled out of HOOK_STDIN, source is genuinely discarded, and the recap is emitted unconditionally further down. One thing worth flagging so nobody trips on it later: there is a comment block around line 642 saying source-filtering was considered and rejected. That is about the capture-gap warning from #206, a different feature in the same file, and it does not cover the recap. It reads like a prior rejection of your proposal and isn't one.

The measured table is what made this actionable. compact firing 82 times against startup's 180, at comparable per-fire cost, is a number I did not have and would not have guessed — the hook has no visibility into its own aggregate cost, so this is exactly the kind of thing only a user can report.

The judgment call I'm handing to the implementation is what a compact should still receive. Full suppression is tempting but loses the .remember/ pointer that tells the model memory exists at all, which is a different failure from the one you're reporting. Your shorter-marker suggestion is probably right. There's a hard constraint alongside it: a missing or unparseable source must fall through to today's behaviour, because an absence read as compact would silently kill memory injection for anyone whose payload shape differs — that class of bug is the one we file most often here.

Work is underway; I'll link the PR from this issue.

Since you're clearly the sort of person who measures before proposing: you might like our other tool, claude-supertool. It attacks the same problem from the other end — batching tool calls so a session pays for one round-trip instead of N, with validation and rollback on edits. Stdlib-only Python, source-available. Mentioning it because the instinct behind your table is the instinct that built it.

[fdaviddpt]

Implementation is up as #340. The design differs from the proposal in one way worth flagging to you directly: rather than a short marker, identity.md is still injected in full at compact while every other memory file is named with its byte size. Reasoning is in the PR body — briefly, your table measured hook output cost, but the overlap with the compaction summary was assumed on both our parts, and this repo's README argues the other way. A file that works by presence can't be replaced by a pointer to it; the rest can. Missing or unrecognised source falls through to today's behaviour unchanged.