[Workspace] system-prompt anchor never reaches protocol-v1 harnesses: Claude Code and Codex lose absolute-cwd grounding on all platforms

[kcao-gss]

Describe the bug

The [Workspace] system-prompt section never reaches protocol-v1 (legacy-framing) harnesses on any platform. Because upstream ACP's current protocol version is 1, that includes the shipping Claude Code and Codex adapters — so most real Buzz agent runs lose the absolute-cwd grounding, not an edge case.

workspace_section() (pool.rs:1165-1177) has exactly one producer, framed_system_prompt (pool.rs:1137-1158), whose output is consumed by two routes only:

1. pool.rs:828-839 — session/new, wrapped in session_new_system_prompt (pool.rs:185-195):

   if is_goose || protocol_version < 2 { None } else { prompt }

   For a v1 agent the combined prompt is built and then discarded.
2. pool.rs:841-858 — Goose's session_set_goose_system_prompt extension, which is version-independent, so Goose is fine.

The legacy path assembles its own prompt and structurally cannot render the section: prepend_base_for_legacy (pool.rs:1090-1101, called at pool.rs:1593 initial message and pool.rs:1731 heartbeat) emits [Base] and [Channel Canvas] only, and takes no cwd argument. Per-turn assembly in format_prompt (queue.rs:1429-1443) is the same, and FormatPromptArgs (queue.rs:1352-1375) has no cwd field at all.

Why protocol-v1 is the live path, not a relic. Buzz requests protocolVersion: 2 (acp.rs:126) and stores the response, defaulting to 1 when absent (lib.rs:3775-3776, lib.rs:3864: unwrap_or(1)). The v2 request is self-described as non-standard — acp.rs:540-541:

// Requesting version 2 is an intentional temporary pin — we are squatting on ACP v2 ahead of the upstream ACP RFD. Revisit when that RFD merges.

and again in buzz-agent/src/lib.rs:280-283. Upstream ACP is at version 1, and the spec requires an agent that does not support the requested version to answer with the latest it does support. So a spec-compliant upstream adapter answers 1.

Against KNOWN_ACP_RUNTIMES (discovery.rs:66-198):

Runtime	Gets [Workspace]?	Why
Claude Code (claude-agent-acp, discovery.rs:101-132)	No	negotiates v1 → has_system_prompt_support false (pool.rs:173-183) → legacy framing
Codex (codex-acp, discovery.rs:133-165)	No	same
Goose (discovery.rs:67-100)	Yes	version-independent extension (pool.rs:845); lost only if the probe caches Some(false) (pool.rs:849-855)
Buzz Agent (discovery.rs:166-197)	Yes	pins PROTOCOL_VERSION = 2 (buzz-agent/src/config.rs:3), min(requested, 2) (buzz-agent/src/lib.rs:284)

Bounding the impact honestly: [Base] is still delivered on the legacy path — queue.rs:1431 per turn, pool.rs:1593 initial, pool.rs:1731 heartbeat — and it is default-on (lib.rs:1539-1544, opt out only via --no-base-prompt / BUZZ_ACP_NO_BASE_PROMPT, config.rs:409-410). So base_prompt.md:84-100's Workspace Layout table, REPOS/ and OUTBOX/ included, plus the "never run find over $HOME" line at :100, do arrive. What is lost is only the absolute-path anchor ("Your absolute working directory is <cwd>... this is where you already are"). Same shape as #3124, much wider blast radius. Expected symptom is a probabilistic uptick in path-guessing and wasted exploration, not a functional failure.

Steps to reproduce

Code-level, no harness needed:

$ rg -n 'cwd' crates/buzz-acp/src/queue.rs | rg -n 'FormatPromptArgs' -A2
# FormatPromptArgs has no cwd field -> the per-turn legacy path cannot render [Workspace]

workspace_section has one caller (pool.rs:1154), reachable only from create_session_and_apply_model (pool.rs:820); neither legacy assembly point calls it.

End-to-end: run a managed Claude Code or Codex agent on any platform and inspect the prompt the harness receives (desktop transcript observer, or the harness log). Expect [Base] with no [Workspace] block. Compare against a buzz-agent agent on the same host, which does get it.

Expected behavior

Protocol-v1 harnesses receive the same workspace anchor as v2 harnesses.

Version and platform

• Buzz version: main @ 87b3fcd3. All platforms — this is not OS-specific.
• Affected since [Workspace] was introduced in fix(desktop): ground agent workspace, migrate legacy nest, configurable repos_dir #1194 (2026-06-23).
• Found by code inspection while confirming Windows agents never receive the [Workspace] system-prompt section: workspace_section() gates on a POSIX-only starts_with('/') predicate #3124. Not a field report.

Logs / additional context

Origin — an omission, not a decision. workspace_section came in with 1011cea2 ("fix(desktop): ground agent workspace, migrate legacy nest, configurable repos_dir", #1194, 2026-06-23) — the only commit git log -S workspace_section -- crates/buzz-acp/ returns. Inside the crate it touched only pool.rs and base_prompt.md; it did not touch queue.rs, and within pool.rs it did not touch prepend_base_for_legacy — the diff sat in framed_system_prompt, then inside an explicit if agent.protocol_version >= 2 { … } else { … }, and the legacy arm was left alone. All seven tests added target framed_system_prompt/workspace_section in isolation. The doc comment added at pool.rs:1151-1153 reasons about persona-only agents and the / fallback but never mentions protocol version. prepend_base_for_legacy predates it (f0858824, #981, 2026-06-12; later 73cc31cc #1277, 3b9b371f #1755, 64b8fea6 #2034) and no commit in that set considers [Workspace]. (Inference from diffs and messages: never added, rather than consciously excluded.)

Fix shape — three insertion points, one shared renderer:

1. queue.rs:1352-1375 — add cwd: Option<&'a str> to FormatPromptArgs; pool.rs:1771-1784 already has ctx.cwd in scope.
2. queue.rs:1429-1432 — push the workspace section ahead of base_section(bp), matching framed_system_prompt's ordering.
3. pool.rs:1090-1101 — add a cwd parameter for the initial-message (pool.rs:1593) and heartbeat (pool.rs:1731) call sites. Missing the heartbeat site causes exactly the drift the pool.rs:1088-1089 doc comment says the helper exists to prevent.

workspace_section is private in pool.rs and would need to move to a shared module or be re-exported.

Risks worth weighing before doing it:

• Per-turn re-injection is real — format_prompt runs on every batch flush. But [Base] (~150 lines) plus [System], [Team Instructions], [Agent Memory — core], and [Channel Canvas] are already re-sent every legacy turn. [Workspace] is one ~70-word paragraph, well under 1% of existing per-turn overhead.
• Preserve the Some(base_prompt) coupling. framed_system_prompt emits [Workspace] only when a base prompt exists (pool.rs:1154-1157); a persona-only legacy agent must not get an anchor with no layout to go with it — the case pool.rs:1151-1153 warns about.
• MAX_SYSTEM_PROMPT_BYTES does not apply (the legacy path uses no system prompt). The observer size trimmer (queue.rs:1396-1402) handles a new section gracefully, but it adds a leaf to the desktop "Prompt context" panel — worth a visual check.

Relationship to #3124 — independent, neither subsumes the other. #3124 fixes whether the renderer returns Some (the POSIX-only predicate), which helps callers that already invoke it. This issue is that the legacy path never invokes it. A Windows Claude Code agent is broken for both reasons; a macOS Claude Code agent for this one alone. They could land in one PR — the natural implementation here is "call the predicate-fixed workspace_section from the legacy assembly points" — but closing one must not close the other.

Open question for maintainers: given the acp.rs:540-541 note about squatting on v2 ahead of the upstream RFD, is the intended long-term answer to fix the legacy path, or to converge on upstream versioning so the v1/v2 split disappears? That choice decides whether this is worth fixing directly or is better absorbed into the RFD follow-up.

Duplicate search: searched issues for "workspace prompt", "OUTBOX", and PRs for "base_prompt workspace" — none found. Nearest is #2334 (base prompt shown as session title in harness UIs), a display concern.