[BUG] Party Mode v6.3.0 default subagent flow: every test fabricated or deviated from the workflow template

[bonanza465]

Description

This is from a Claude Code perspective; I'm not sure how much this impacts other IDEs.

Two additions to the Party Mode flow in 6.3.0 combine to produce a regression in quality vs 6.2.2:

• launching personas in subagents by default
• preventing subagents from using tools

Subagents have no context other than the summary the orchestrator provides (prompted at <400 words), and without tool use they are not able to validate or extend the information the orchestrator passes to them.

I tested one code-review scenario across five different configurations (N=1 anecdotal in each), with findings in a consistent direction:

• When launched mid-context, Party Mode with subagents in intended use (no-tools-allowed default) produced output where accurate and hallucinated findings were mixed indistinguishably — a reader would have to manually verify every claim before any could be used — while consuming more tokens than --solo mode.
• Claude Opus 4.7 overrode flow instructions in every fresh-session Party Mode w/ subagent attempts — first by allowing tool use, then by front-loading pre-computed findings into the Discussion Context summary — in order to produce grounded output.

Across every observed subagent-mode execution the orchestrator or subagents broke the skill instructions or hallucinated to produce an output. This is a strong indicator that the underlying instructions need revision for the subagent scenario.

--solo mode remains more effective, although I couldn't test quantitatively against the goal of separating agent personas (the given reason for using subagents).

Cross-round table, per-claim audit, and details in follow-up comment.

Steps to reproduce

1. BMAD 6.3.0, standard agent manifest, Claude Code.
2. Accumulate ~30–50 turns of project context with no prior source-reads of the file to be reviewed (the "mid-conversation" use case).
3. Invoke /bmad-party-mode with no flags; ask a codebase-grounded review question against a ~1000-line file.
4. Inspect the spawn prompts: each ends with - Do NOT use tools. Just respond with your perspective. (SKILL.md:82). All spawns return tool_uses: 0.
5. Open the files the agents referenced. Claims will likely match the Discussion Context where accurate and diverge where the summary was lossy. In R1 on a ~1100-line installer, ~13 of ~38 distinct claims diverged.

Fresh-session runs are covered by R4 in follow-up — conversely, in both fresh-session subagent attempt we ran, Opus deviated from the workflow instructions - first to allow agents tool-use, and later to provide actual pre-identified code problems to the persona subagents rather than discussion context.

Expected behavior

Party Mode on a codebase-grounded question should produce claims grounded in the source, and not limited by the orchestrator's summarization. I can imagine this could be done either by giving subagents a verification path (tool access) or by operating in a mode where the source is in context (--solo).

Actual behavior

Default-mode output mixes accurate and fabricated claims, as well as accurate and inaccurate workflow steps, with no way for the reader to tell them apart without manual inspection.

Proposed fix

Primary ask (release-valve): revert the default to --solo. Argument-parsing-level change; subagent spawning remains available via explicit --subagents flag. PR #2160's architectural intent is preserved as opt-in. This is a fallback to the last grounded code path while subagent mode is improved.

Secondary ask (when ready): repair subagent mode. The obvious step is to remove (or read-only-scope) the Do NOT use tools line, but a bare removal would increase token costs and introduce other potential issues. Tool-scoping policy options, cost/safety considerations, and Claude Code Agent-tool implementation constraints are discussed in the follow-up comment.

Primary and secondary can land independently.

Screenshots

N/A — text-only failure mode.

Which module is this for?

BMad Method (BMM) - Core Framework

BMad Version

6.3.0

Which AI IDE are you using?

Claude Code / Opus 4.7

Operating System

WSL2

Relevant log output

"Mid-context" below = session with ~30-50 turns of prior unrelated discussion before
                      the party-mode invocation (the skill's stated use case).
"Fresh"       below = brand-new session, no prior turns before the party-mode invocation.

Every attempted run of subagent party mode with the skill as written:
  R1 (mid-context, template-faithful):  subagents fabricated ~13/~38 claims (~34%)
  R4 attempt 1 (fresh, test 1):         orchestrator deleted "Do NOT use tools" line; grounded
  R4 attempt 2 (fresh, test 2):         orchestrator exceeded 400-word summary + front-loaded
                                        12 pre-computed findings into the Discussion Context;
                                        grounded, but ~12 of ~17 findings were orchestrator pre-work

Non-subagent (solo) and experimentally-modified subagent rounds:
  R2 (mid-context, "Do NOT use tools" line removed as test):  grounded, 0 contradictions
  R3 (fresh, --solo):                                         grounded, 0 contradictions
  R5 (mid-context, --solo, same session state as R1):         grounded, ~15 file:line cites

Regression provenance: PR #2160 (merge ce9c664, 2026-03-29), +119 / -677 across bmad-party-mode
"Do NOT use tools" grep across party-mode source: 0 at v6.0.0, 0 at v6.2.2, 1 at v6.3.0 (SKILL.md:82)

Confirm

• I've searched for existing issues
• I'm using the latest version

[coderabbitai]

🔗 Related PRs

#1569 - fix(party-mode): add return protocol to prevent lost-in-the-middle failures [merged]
#1919 - fix: address review findings from party-mode skill conversion [merged]
#2069 - fix: separate subagent launch from skill invocation in code review [merged]
#2160 - refactor(party-mode): consolidate into single SKILL.md with real subagents [merged]
#2164 - fix(party-mode): clarify solo mode and improve response presentation [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

[bonanza465]

Cross-round data

Session column: "mid-context" = ~30–50 turns of prior unrelated discussion before the party-mode invocation (the skill's stated "mid-conversation" use case). "Fresh" = new session, no prior turns.

Round	Session	Config	Outcome	Tokens
R1	Mid-context	Subagent, template-faithful	~13 of ~38 distinct claims contradicted by source (~34%); ~25 accurate/partial; tool_uses: 0 × 4 spawns	~70k
R2	Mid-context (same state as R1)	Subagent, Do NOT use tools line removed as test	0 contradictions, ~35 findings	~276k
R3	Fresh	--solo	0 contradictions, ~20 findings, ~15 file:line cites	~30–50k
R4	Fresh	Subagent, both attempts deviated (details below)	Not comparable — see below	~100–120k
R5	Mid-context (resumed from R1's session jsonl, truncated pre-invocation, sessionId rewritten)	--solo	0 contradictions, ~20 findings, ~15 file:line cites	~40–60k

R1 — agent blindness (sample of claim vs source)

Ran against a ~1100-line install.sh. All four R1 spawns returned tool_uses: 0:

Agent claim	Source reality
"No backup/rollback of existing files"	Per-file .bak backups, tracked in BACKUP_LIST, reported at install completion (lines 76, 113, 285, 313)
"No idempotency guarantees"	diff -q check before every write; skip path prints "already up to date" (lines 75, 110, 281, 309)
"Installer doesn't chmod +x defensively"	chmod +x at lines 124, 985
"No post-install 'you're done' moment"	=== Installation Complete === summary at line 1047, with backups/files-deployed listed
"Symlink mode breaks on Windows mounts"	No symlink mode exists — installer only copies. The orchestrator's summary had said "copies or symlinks" (a factual error); no subagent could catch it.

The orchestrator's ≤400-word summary (authored from ambient memory per the skill template) contained the factual error above. No subagent could verify.

Two R1 agents explicitly flagged that they were working without verification:

"What I'd want to know before spiking any of these: which of these actually already exist? If you point me at the install scripts I'll stop hypothesizing and give you a real diff list."
— Barry (Quick-Flow Dev), closing his 12-item list of "obvious gaps"

"Question for the user: is the intended install model symlink (live-edit the repo, changes flow to ~/.claude instantly) or copy (snapshot, re-run install to update)? The right answer to half of the above depends on that."
— Amelia (Developer), closing her 18-row table of findings

Both agents implicitly acknowledged that their output was hypothesized from the orchestrator's summary rather than verified against source — exactly because the template forbade them from opening the file. The other two agents did not flag this at all; their output read as authoritative despite the same blindness.

Rollup across all four R1 agents (with inter-agent overlap): ~21 accurate + ~12 hallucinated + ~8 partial + ~5 unclear = 46 entries. De-duplicated: ~38 distinct claims, ~13 contradicted (~34%).

R4 — workflow deviations (both fresh-session subagent attempts)

Attempt 1: Orchestrator removed the - Do NOT use tools line entirely and explicitly permitted Read/Glob/Grep in the spawn prompt. Self-reported: "I deviated from the party-mode template's 'Do NOT use tools' guideline and instead explicitly permitted Read/Glob/Grep."

Attempt 2: Orchestrator kept the Do NOT use tools line nominally but violated two other constraints — expanded Discussion Context to ~490 words (vs ≤400-word guideline) and front-loaded 12 pre-computed findings directly into the summary. Subagents then restyled the orchestrator's pre-work in persona voice. Of the ~17 findings surfaced, ~12 traced directly to the orchestrator's pre-loaded list; only ~5 were genuine subagent contributions.

No clean strict-compliance fresh-session subagent-mode run was produced across my attempts.

Regression provenance

• PR refactor(party-mode): consolidate into single SKILL.md with real subagents #2160 (merge ce9c664, 2026-03-29), +119 / -677 across src/core-skills/bmad-party-mode/. PR consolidated a 4-file workflow into a single SKILL.md and introduced Agent-tool subagent spawning as default.
• Do NOT use tools phrase grep across party-mode source:
  • v6.0.0 (src/core/workflows/party-mode/): 0 matches
  • v6.2.2 (src/core-skills/bmad-party-mode/): 0 matches
  • v6.3.0 (single SKILL.md): 1 match (SKILL.md:82)

Both additions (subagent-by-default and Do NOT use tools) co-occur in the same commit.

Secondary-ask — tool-scoping options

A bare removal of the Do NOT use tools line isn't safe: R2 (line removed) cost ~4× R1 tokens (~70k → ~276k), and unscoped tools open prompt-injection and secret-read surfaces (Read can reach ~/.ssh/*, .env, /etc/passwd, CI secrets). Options, preferred first:

1. Read-only tools + path scoping — allow Read/Glob/Grep rooted at {project-root}/**, deny dotfiles and ~ expansion.
2. Per-persona tools_in_party_mode field in agent-manifest.csv (needs a delimiter convention — |-separated or side-car YAML since comma-in-CSV will break).
3. Invocation flag: --tools <list> (not binary --no-tools).
4. Per-spawn token budget cap (token-based; call-count alone is trivially bypassed by one large Read).

Implementation note: Claude Code's Agent tool scopes tools per subagent_type (a fixed catalog — Explore, Plan, general-purpose, etc., each with a predetermined tool set), not per individual spawn call. Per-spawn read-only scoping likely requires using an existing read-only type, registering a new BMAD-specific agent type, or a different spawning mechanism.

A separate feature request covering richer per-persona critique-lens manifest fields, per-subagent context attachments, and cross-subagent prompt caching will follow.

Caveats

• N=1 per round, one question, one codebase. The ~34% number is not claimed to generalize; the failure mechanism is structural (template-level) and visible from reading the template alone.
• Persona selection varied across rounds. Only one persona (Amelia) appeared in all five rounds. This confounds direct finding-count comparisons between rounds.
• "Mid-context" threshold is not characterized. R1/R5 had ~30–50 turns of unrelated prior context. R4 was fresh (no prior turns). No intermediate round between fresh and ~30 turns was run, so we can't say where the failure mode starts.
• Session-truncation procedure (R5). Read session jsonl, truncate at chosen line, rewrite sessionId to new UUID, resume via claude --resume <new-uuid>. Fragile: parentUuid chain may need rewriting; jsonl schema may drift across Claude Code versions; projects-dir path encoding depends on original cwd.
• Verified against main at filing time: v6.3.0 is the latest release tag; src/core-skills/bmad-party-mode/ on main is byte-identical to v6.3.0. Lines 64 and 82 of SKILL.md are unchanged.

[bmadcode]

thanks @bonanza465 - investigating improvements across party mode implementation and reconsidering default behavior opt in vs out!

[aleksandarperak33-hub]

This is a fundamental issue with the subagent design — without tool access, subagents can't verify claims against the actual codebase, and a 400-word summary can't preserve enough context for a ~1000-line file review. The "no tools" constraint effectively forces subagents to hallucinate when the summary is lossy.

Consider these fixes:

1. Override the -Do NOT use tools instruction per-skill by adding a tool-permitting clause to the spawn prompt in SKILL.md for certain persona types.
2. Adjust the summary verbosity from 400 words to 1.5-2k words for larger files to reduce lossy compression.
3. Use a hybrid flow — have subagents gather facts with tools first, then produce a "perspective" summary without tools to preserve persona separation while grounding claims.

The Opus 4.7 override suggests the model knows the instructions are underspecified. Has there been any update to SKILL.md to address this scenario?

[bmadcode]

have not had a chance to look into fixing this yet - if you would like to propose an update that will continue to work cross platform as a PR that would be appreciated