summarize: the role=system summary is placed where the provider rejects it (measured 400)

[amiddavid]

summarize emits its summary as a role: "system" message between msgs[0] and the kept tail
(components/offload/summarize.go:247 and :300, no provider branch). That placement is
deliberate and the reasoning is sound — an operator directive added mid-conversation leaves the
cached prefix intact, where editing the top-level system block would invalidate everything
after it. The problem is only the placement, and it is a hard provider rejection.

Measured

Probed directly against the gateway (benchmark endpoint, not Context Guru), identical ~3k-token
cached system prefix, varying only messages:

shape	claude-sonnet-5	claude-opus-5
[user] (warm the prefix)	write=3064 read=0	write=3063 read=0
[user] again	write=0 read=3064	write=0 read=3063
[user, system, user] — what summarize emits	HTTP 400	HTTP 400
[user, system, assistant, user]	write=0 read=3064	write=0 read=3063
[user] after	write=0 read=3064	write=0 read=3063

messages.1: role 'system' must precede an 'assistant' message or end the array;
the directive-only form (content: [] with output_config) is accepted at any position

Two conclusions. It is not a model-support restriction — both models accept a
mid-conversation system message. And the cache rationale is confirmed: in the legal shape the
prefix bills read in full, so nothing is being folded into system by the gateway.

Scope

Any turn where summarize acts and the kept tail begins with a non-assistant message emits an
invalid request. #110 narrowed this incidentally — snapping the tail off a split tool exchange
lands on the declaring assistant message — but did not close it: keep_last: 1 on a transcript
ending in a user turn still produces [user, summary, user].

Why #110 does not fix it

The obvious fix — require the kept tail to begin with an assistant message — was implemented and
reverted, because it over-reaches. On a transcript with no assistant message near the
boundary the snap walks back to the span start and summarize declines entirely; that regressed
four existing tests (TestSummarizeRestructures, TestSummarizeCountChangeLossless,
TestSummarizeModelErrorFailsOpen, TestSummarizeReusesCheckpoint, all reporting
skipped=true / calls=0). Silently not compacting is a worse failure than the one being
fixed.

So this needs a decision rather than a patch:

1. Fall back to role: "user" for the summary when the tail cannot start with an assistant
   message. Always legal, no placement constraint. Changes how the model weights the summary
   and gives up the operator-channel/injection-resistance property in the fallback case.
2. Always use role: "user". Simplest and uniform; gives up the property everywhere. Note
   the prefix benefit is a property of where the message sits, not of its role, so this likely
   keeps the cache win — worth confirming with the same probe.
3. Emit the summary as the directive-only form the error message mentions (content: [] with
   output_config), accepted at any position — needs checking whether a summary can be carried
   that way at all.
4. Snap the tail to an assistant message when one exists, else fall back to (1). Keeps the
   property in the common agent-loop case, correct in all cases, most moving parts.

(2) or (4) look best; (4) is the one that preserves the intent. Whichever is chosen, the
placement assertion written for #110 should come back — it caught the residual keep_last: 1
case immediately and is already drafted.

Acceptance

• No emitted wire contains a role: "system" message followed by a non-assistant message.
• Verified against the probe shapes above, not only in unit tests.
• The four tests listed above still pass, i.e. summarize still acts on transcripts with no
  assistant message near the boundary.

[amiddavid]

Superseded — the fix already existed, on PR #80 only.

80e95d5 (2026-08-21, feat/coref-compaction, not on main) is exactly this fix: emit the summary as a user-role message rather than system, in both the fresh-summary and checkpoint-replay paths, with summarize_role_test.go asserting no system role appears anywhere except index 0. It was found the same way — LOCA-bench against a real API — and reports the identical error I measured independently here.

That also explains the discrepancy that surfaced this: a run built from feat/coref-compaction shows the summary as a user message, while main still emits system.

It is option (2) from the list above, and the commit message settles the open question in it: a user-role message is "valid and conventional -- it is what Claude Code's own compaction does." My worry that the prefix benefit might be lost does not apply — the benefit comes from the message's position, not its role, and the position is unchanged.

Cherry-picked into #110 as 3681e1c. The commit also records the measurement blind spot that let it ship, which is the part worth keeping: every measurement replayed through /compact, which runs the pipeline and returns the rewritten body without forwarding upstream — "a body no provider ever validates cannot fail schema validation."

Closing once #110 merges.

[amiddavid]

Fixed in #110, now merged. 2edb9d4 emits the summary as a user-role message in both the fresh-summary and checkpoint-replay paths (cherry-pick of 80e95d5), with summarize_role_test.go asserting no system role appears anywhere except index 0.

The design question this issue posed is settled by that commit rather than by the options listed above: a user-role summary is the answer, and the cached-prefix benefit is preserved because it comes from the message's POSITION, which does not change — only the role does. Measured against the gateway: [user, system, user] is HTTP 400 on both claude-sonnet-5 and claude-opus-5, while the fixed shape bills read=3064 / write=0.