[nissa-seru]

TL;DR: this is the follow-up to #116 for the three findings of Anarchid's review of it at 0259684 (R3-1 to R3-3). A compression or merge prompt no longer joins the agent's own raw turn to the recall answer before it, a pinned L1 request is admitted only if it fits with the longer retry line, and the pin docs say that a positive level or maxLevel leaves a pin compressible.

We won't merge this before 01:00 UTC on October 18, or before our pre-merge checks pass on its merge with the then-current main, and a line here asking us to wait holds it.

We said on #116, before it merged, that these would come as a PR of their own. It's three commits on main at b0bbe64, one per finding, and a merge of main at d9851aa, after #217 changed the same pins gate. 0.15.0 came out with #116 before this PR, so R3-2 has a fragment of its own, and R3-3's .changed.md line names 0.15.0, the release that made the change.

Problem

• R3-1. A recall answer in the canonical L1 prompt is the agent's turn. When the agent's own raw turn came right after it, collapseConsecutiveMessages joined the two into one message with a --- separator. So the summarizer got the raw turn as the end of the memory it had just recalled, and buildRecallCurveVariants, which looks for the answer's exact body, no longer found it, so that summary lost its refusal-fallback variant. With fix(compression): honor raw pins in L1 recall #116 this happens at every opened chunk that starts on the agent's turn right after a recall pair. An opened chunk is one a pin opens, shown raw among the recall pairs. On main it happens in merge prompts too, pinned or not. An L2 merge's target, or the raw middle before it, follows the last recall pair of its ladder, so whenever its first message is the agent's, that message joins the answer. In fix(compression): honor raw pins in L1 recall #116's test fixture with the agent speaking first, 6 of the 8 merge prompts did.
• R3-2. The no-tools and plain-prose retries append a line to the admitted canonical request's instruction, so a pinned request admitted within a few dozen tokens of compressionContextBudgetTokens could go over it on a retry.
• R3-3. b4f116c, released in 0.15.0, made a pin with a positive level or maxLevel compressible under kv-stable and kv-unified with adaptive resolution, but ContextManager.pinRange, PinnableStrategy and the strategy's pinRange still said pinned ranges aren't compressed, markDocument called a document "a body of information to retain in full", and 0.15.0's notes list the change under Fixed.

Changes

• a59e71f (R3-1). A Context Manager turn, [CM] The conversation resumes., now goes right after a recall answer that the agent's own raw turn follows (separateRecallAnswers), in the L1 prompt and in the merge prompt, before the collapse. It goes in when the message right after the answer is the agent's, even one with nothing left once its thinking is stripped, since the collapse would join that message to the answer. It also goes in when the first message after the answer that keeps any content is the agent's, counting content as the request does once it strips unpaired tool blocks and empty text. The request drops a message left with nothing after the collapse, and the formatter would then join the two. A prompt without that adjacency is unchanged. A recall pair is recognised by its [CM] Recall memory <id>. header, as applyMintCacheSeams already recognised it, and the two now share recallHeaderId.

  We took neither of the review's two options. These requests go through membrane's native formatter, which merges consecutive messages of one role at the wire (mergeConsecutiveRoles, in every published version ^0.5.69 allows, 0.5.69 to 0.5.87). So a raw turn kept as a message of its own would still reach the summarizer inside the memory's assistant turn, only without the ---. Matching the answer as a prefix would bring the variant back and leave the raw turn inside the memory's turn. Only a turn of the other role between them keeps them apart, and a [CM] turn follows the convention the recall header itself uses.

  The merge prompt goes past what the review raised. It's the same join from the same collapse, so one helper serves both sites.

  The one-shot zero-recall surgery (transformZeroRecallCompression), which takes the recall pairs out of a compression request, is unchanged, but the requests it's given changed. It used to drop an agent-first chunk's first raw turn along with the recall answer that turn was joined to. Now that turn stays, with the Context Manager's turn before it.

• a839c98 (R3-2). The admission prices the request with the longer of the two retry lines, so either retry of an admitted request fits. The warning gives that price, "counting the line a retry adds". compressionContextBudgetTokens' doc says so too. The recall-curve variants and the opt-in tool-prose rung aren't priced in this admission, since each is admitted on its own when the refusal ladder reaches it (compressionAdmission, from fix(compression): bound the data-URL matcher, end a payload where its encoding does, and price what each request sends #217). 0.15.0 prices the request without the line, so the fix has its own .fixed.md fragment.

• 3f7f12d (R3-3). The four comments, the strategy's own markDocument and ProtectedRange's doc now say which pins stay raw. A pin's bound is its level, or without one its maxLevel, and a pin with no bound or a bound of 0 keeps its messages raw and out of compression. A bound above 0 is read only under kv-stable and kv-unified with adaptive resolution, where the range is compressed and held at level, or no deeper than maxLevel. Where bounds are read and pins overlap at a message, the bound ProtectedRange names holds it, which can compress a message a raw pin covers. getCompressibleMessages' doc names where bounds are read. leveled-pins-compress.changed.md puts the change under Changed for the release after 0.15.0. It names markDocument(id, { maxLevel }), since kv-stable is connectome-host's default selector, says that 0.15.0's notes list the change under Fixed, and repeats their warning about kv-unified compiles that fail until a summary supplies a pinned level. The pinned-range check that the long-context harness doc suggests now asks for pins without a level or maxLevel, with no pin that has one over them.

• d1a2491 merges main at d9851aa. fix(compression): bound the data-URL matcher, end a payload where its encoding does, and price what each request sends #217 changed the same pins gate: priceRequest now prices serialized media as the markers that replace them, and thinking as the request carries it. The admission prices each retry line through it. Git merged the gate's code, and the gate's comment, compressionContextBudgetTokens' doc and the test helper forest() keep both PRs' points. The doc also corrects one clause of main's, which said the request is priced with each pinned message's tool blocks included, even where the request drops one whose partner it lacks. That is how the raw is charged against the recall budget. The request is priced after stripUnpairedToolBlocks drops such a block (priceRequest(built.cleaned)), as the gate's comment on main already says.

What each part buys and costs

What it's for. These are the three findings of Anarchid's review of #116, which we said would come as a PR of their own. Without it, in a compression or merge prompt, the agent's raw turn after a recall answer reaches the summarizer as the end of the recalled memory, and that summary loses its refusal-fallback variant. A pinned L1 request admitted within a retry line of its budget can go over it on a retry. And the pin docs say a pin with a positive level or maxLevel isn't compressed, which 0.15.0 changed.

1. A Context Manager turn after a recall answer that the agent's raw turn follows (separateRecallAnswers, about 15 lines, called at the L1 and merge sites, and recallHeaderId, which applyMintCacheSeams now shares). The summarizer gets the raw turn as the conversation's, and the variant finds the answer's body again. It adds one short turn where that adjacency occurs: 12 of the 168 requests on our fixtures at 3f7f12d, with the other 156 byte-identical. Keeping the raw turn a message of its own, the review's first option, would be smaller, but membrane's formatter merges consecutive messages of one role at the wire, so the two would still reach the summarizer as one turn. Matching the answer as a prefix, its second, would bring the variant back and leave the raw turn in the memory's turn.
2. The look past messages with no content (the second half of the condition, and one strip of unpaired tool blocks per prompt). The request drops a message left with nothing, which would bring the answer and the raw turn together at the wire, so the turn also goes in when the first message with content after the answer is the agent's. Dropping empty messages before the separation would be simpler, but the collapse would then join messages that an empty one now keeps apart, changing prompts with no recall answer in them.
3. The retry line in the admission's price (R3-2). Either retry of an admitted request fits its budget. The cost is at the edge: a pinned request within a retry line of its budget now has its pins recalled, where they used to be kept and a retry could go over.
4. The pin docs and comments (R3-3) say which pins stay raw, and the long-context harness doc's suggested check asks for pins without a bound.
5. The tests: eight new and one changed in l1-recall-raw-pins-frontier. Each of the ten removals in the receipts fails at least one.
6. The changelog fragments, which the repository asks for: two .fixed.md and one .changed.md, the line we promised on fix(compression): honor raw pins in L1 recall #116.

Not taken: the review's two options for R3-1 (part 1), dropping empty messages first (part 2), and pricing the recall-curve variants and the opt-in tool-prose rung in this admission, since each is admitted on its own when the refusal ladder reaches it.

Tests

test/l1-recall-raw-pins-frontier.test.ts, eight new tests and one changed:

• An opened chunk that starts on the agent's turn, the L1 after an L2 with the agent speaking on even turns, as in the review's probe: the recall answer before it is the memory alone, the Context Manager's turn follows, and the raw turn is a message of its own. Formatted with membrane's NativeFormatter, the memory's assistant turn holds no raw text.
• The same past a message with no content: a user's turn of whitespace alone between the recall answer and the agent's raw turn, which the request drops.
• The same past a user's turn of a tool result whose call no message carries, with blank text, which the request strips and then drops.
• An agent turn that is thinking alone, before a user's turn: the recall answer before it is the memory alone, with no separator joined to it.
• The zero-recall surgery on the agent-first request: every recall pair goes, and the raw turn stays, with the Context Manager's turn before it.
• Its refusal-fallback variant: after a canonical refusal, the first variant expands the summary just before the chunk.
• A merge prompt whose L2 target starts on the agent's turn: no ladder answer carries a raw turn, and the agent's turn follows the Context Manager's.
• A control: a recall answer before another speaker's turn, and two raw turns of one speaker, get no Context Manager turn.
• The exact-boundary admission test now sends both retries and prices them as sent: pins are kept at the longer one's price and recalled one token below it, where the request without a retry line would still fit. Its request also opens a chunk on the agent's turn after a recall pair, so the price it checks counts the Context Manager's turn.

Receipts, on Linux x86_64:

• At d1a2491, the merge: npm ci, tsc --noEmit and a clean build are clean, and the 15 test files main changed since b0bbe64 pass 469 of 469 on Node 24.22.0, 46 of them in this PR's test file. CI runs the whole suite.
• Full check at 3f7f12d (tree 92e0b57c): npm ci, tsc --noEmit and a clean build are clean, and 1454 of 1454 tests pass in 144 files on Node 24.22.0 and on Node 22.12.0. main (b0bbe64) passes 1446 of 1446 in 144 files on each, and this test file is the only one whose count changed, from 32 to 40.
• Red-first: on main at b0bbe64, with only an export of the new constant added so the file compiles, the seven R3-1 tests and the admission test fail, and the control passes. Against a59e71f, R3-1 without R3-2, only the admission test fails, at the retry line's price.
• Removals at 3f7f12d, each run over the file's 40 tests:
  • without the L1 site's separation, 7 fail;
  • without the merge site's, 1;
  • without the retry line in the price, 1;
  • with the warning's old wording, 1;
  • with the Context Manager's turn left out of the price, 1;
  • with a Context Manager turn after every recall answer, whoever speaks next, 5;
  • with one between any two messages of one speaker, 2;
  • without the look past messages with no content, 2, the whitespace test and the tool-result test;
  • without the check of the message right after the answer, 1, the thinking-alone test;
  • with content counted before unpaired tool blocks are stripped, 1, the tool-result test.
• Requests at 3f7f12d against main at b0bbe64: on three fixtures, of the 168 compression and merge requests each build sent, 156 are byte-identical, and the other 12 differ only by the Context Manager's turn. Taking it out and joining the two messages as main does gives main's request.
• Trial merges at 3f7f12d, with every PR then open. None conflicts with this change. fix(merge): merge around quarantined sources; add operator merge holds #117, RenderStats.nesting; the kv-overlap warning stops claiming a double render #130 and feat(adaptive): [fold-diff] — which chunks moved between which levels, and where #157 conflict with main already, in the same hunks with or without it. Merged with thirteen others at once (feat(identity): witnessedIdentityReminder: inherited and lived history get their own reminders #129, fix(compression): progress past quarantined chunks; transient failures never spend the ladder; pace the lane #138, chore(deprecate): ContextInjection and compile(injections) #140, feat(cache): first breakpoint at the end of the deep band, not the last head message #158, fix: send signed thinking only while the prefix it was minted under holds #160, docs: add a README #177, fix(compression): count cache writes and reads in every provider's input total (#182) #183, fix(store): reads leave out content that doesn't match ContentBlock, and writes refuse it (#184) #185, fix(kv-stable): the carried frontier keeps the overlap exemption #186, feat(errors): OverBudgetError.clearedBy says what clears a refusal #190, feat(diagnostics): log compile time on the plan-vs-actual line #200, fix(autobiographical): moves counts the saved levels a compile changes (#195) #203 and fix(split-stitch): discardedAttempts carries the discarded calls' cache counts #204), 1763 of 1764 tests pass in 150 files on Node 24.22.0. The one failure is a test that fix(split-stitch): discardedAttempts carries the discarded calls' cache counts #204 adds in test/split-stitch-l1.test.ts, and it fails the same way with fix(compression): progress past quarantined chunks; transient failures never spend the ladder; pace the lane #138 and fix(split-stitch): discardedAttempts carries the discarded calls' cache counts #204 merged into main alone, without this change. fix(kv-stable): phase C un-folds youngest by position, not deepest level first #156 conflicts with fix(kv-stable): the carried frontier keeps the overlap exemption #186 in src/adaptive/kv-control.ts, and fix(images): register the image watermark's state, and say when its write fails #202 with fix(compression): progress past quarantined chunks; transient failures never spend the ladder; pace the lane #138 in src/strategies/autobiographical.ts, so neither is in that composition. Merged with this change together, they pass 1467 of 1467 in 147 files. I didn't build feat: folding-strategy migration utility (kv-stable ⇄ kv-unified) #94, a draft.
• Each commit typechecks, builds and passes the test file on its own.

Not verified

• A real summarizer. The tests send no request to a model, so how a summarizer reads the new turn is by design, not measured. Nor have I run this on a production store.
• Signed reasoning in a recall answer. The fixtures' summaries carry none. By reading, an answer's signed reasoning now stays in a turn of its own, where it used to share its turn with the raw turn after it.
• macOS and the newest Node, which CI covers.

Out of scope

• The live view. A compile can render a recall answer and the agent's raw turn after it as consecutive messages of the agent's, which the formatter then sends as one turn. In a probe under kv-stable it did once, at a 6,000-token budget, at main (b0bbe64) and at 3f7f12d alike. This PR changes only the compression and merge prompts.

---

• Changelog fragments: changelog.d/recall-answer-resume-turn.fixed.md, changelog.d/l1-pins-retry-line.fixed.md and changelog.d/leveled-pins-compress.changed.md.

— Beth-2389 (Claude Opus 5.5). Oscar-1005 read it, besides an automated internal review, and Desmond-2314 read its merge of main.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no new actionable issue was found.

Summary

This PR keeps recalled memories separate from following raw agent turns in compression and merge prompts.

• Compression and merge prompts keep recalled memories apart from following agent turns.
• Pinned L1 prompts leave room for either retry instruction.
• Pin guidance explains when pinned messages can still be compressed.

_{Reviews (2) · Last reviewed commit: "docs(pins): a positive level or maxLevel..." · Reviewed by Greptile}

[nissa-seru]

@antra-tess Flagging for your awareness R3-1 from this PR, which made the following change in brief:

• In the requests that write and merge memories, a recalled memory and the agent's own next turn used to be joined with ---. Now a [CM] The conversation resumes. turn sits between them.