cllama feed injection budget can drop channel-awareness despite higher claw-wall caps

[mostlydev]

Problem

Tiverton now asks claw-wall for a much larger 24h channel-awareness window, but cllama still silently constrains injected feed context with hard-coded byte budgets:

• cllama/internal/feeds/manifest.go: MaxFeedResponseBytes = 32 * 1024
• cllama/internal/feeds/manifest.go: MaxTotalFeedBytes = 64 * 1024
• cllama/internal/feeds/fetcher.go truncates every feed body to MaxFeedResponseBytes before formatting.
• cllama/internal/feeds/inject.go skips any later feed block that would push the aggregate feed block over MaxTotalFeedBytes, emitting --- FEED: <name> skipped (total feed size cap reached) ---.

This means the pod-level x-claw.context.channel.max-chars knob added/covered by #242 can successfully generate a large channel-awareness / channel-context URL, but the provider-visible prompt still gets capped or skipped inside cllama.

Live Tiverton symptom

Date: 2026-05-18
Repo/pod: mostlydev/tiverton-house
Live revision observed: a925fc4
Runtime observed after deploy: claw-wall v0.17.2, cllama v0.6.6

Morning Discord report from Weston at 09:23 ET:

One gap: channel-context skipped this turn due to total feed size cap. The channel-awareness feed is present but the full context buffer got truncated. Non-blocking — I can read channel-awareness for the last 24h, just can't see the full enriched context this cycle.

Current generated Tiverton feed config is correctly asking claw-wall for the larger window:

{
  "name": "channel-awareness",
  "source": "claw-wall",
  "path": "/channel-awareness?channels=1464509330731696213&since=24h&limit=200&max_chars=262144&context_kind=raw_window"
}

The matching channel-context feed is likewise generated with mode=tail&since=24h&limit=200&max_chars=262144.

But the live raw channel-awareness response for the trading-floor channel was ~107 KB. cllama can only accept 32 KB from that one feed and only 64 KB across all injected feeds. A typical trader turn already includes:

• market-context (~12-13 KB observed)
• style context such as momentum-context (~14 KB observed)
• agent-scaffold (~3 KB)
• desk-chronicle (~3 KB)
• agent-memory
• channel-awareness
• channel-context

So on active market days the most important human-floor context competes with market/style feeds in an all-or-nothing aggregate budget. Depending on manifest order and exact sizes, channel-awareness or channel-context can be truncated or skipped even though the operator configured a larger claw-wall window.

Why this matters

The product promise after #232/#242 is that agents have bounded 24h floor awareness independent of cursors and restarts. A downstream operator can now tune claw-wall up to a meaningful 24h window, but the model still may not see it because cllama has a lower, non-configurable envelope.

This invalidates the mental model of a context feed: the source feed exists and returns data, yet provider-visible context may omit it because earlier feeds consumed the hidden cllama budget.

Retrieval tools are useful, but they are not a full substitute. The model has to know that something is missing before it decides to search; in the failure mode, the model can simply conclude that the injected feed had no relevant context.

Related issues

• cllama: guarantee bounded 24h channel awareness independent of cursors and restarts #232: introduced bounded 24h channel awareness independent of cursors/restarts.
• claw-wall: raise channel-awareness max_chars default and expose a public knob #242: raised/exposed claw-wall channel-awareness max_chars; explicitly left token-budget guard rails out of scope.

This issue is the cllama-side continuation of that work.

Proposed fix

1. Make cllama feed budgets configurable, at least:

   • per-feed max bytes;
   • aggregate injected-feed max bytes.

2. Allow pod/operator configuration to flow into cllama, probably via generated cllama env from x-claw.cllama-defaults.env or an explicit cllama config block.

3. Avoid all-or-nothing loss of critical feeds:

   • reserve space for high-priority feeds such as channel-awareness and channel-context; or
   • allow feed priority/order configuration; or
   • degrade lower-priority long-lived feeds (desk-chronicle, agent-scaffold) before dropping live floor context.

4. Make overflow visible in structured context metadata/logs:

   • which feed was truncated;
   • which feed was skipped;
   • effective byte budget;
   • source response size when known;
   • whether the provider-visible payload includes or omits each feed.

Acceptance criteria

• A pod can configure cllama to accept a larger channel-awareness feed than 32 KB and a larger aggregate feed block than 64 KB.
• With Tiverton-like feed sizes, a 100 KB channel-awareness response does not cause channel-context or channel-awareness to disappear silently.
• Tests cover both per-feed truncation and aggregate overflow behavior with channel-awareness present.
• Provider-visible context snapshots and logs clearly distinguish:
  • source feed returned data but was truncated;
  • source feed returned data but was skipped due to aggregate budget;
  • feed was included in the actual provider-visible payload.
• Defaults remain bounded for small pods; this should not become unbounded prompt stuffing by default.

[mostlydev]

Draft implementation is up in mostlydev/cllama#16.

The patch adds configurable cllama feed budgets via CLLAMA_FEED_MAX_RESPONSE_BYTES and CLLAMA_FEED_MAX_TOTAL_BYTES, keeps the bounded defaults unchanged, emits structured feed_injection telemetry for included/skipped/truncated feed entries, and makes context snapshots reflect the provider-visible feed blocks and skip notices.

The root repo pins are intentionally not bumped here. Next step is review, then cut a cllama release and consume that released tag from Clawdapus.