fix: refactor: architect pitching should be bash-driven with stateless model calls (#490)

This commit is contained in:
Agent 2026-04-09 10:04:48 +00:00
parent 80e19f8e51
commit b11c4cca15
2 changed files with 163 additions and 587 deletions

View file

@ -10,12 +10,12 @@
# 2. Precondition checks: skip if no work (no vision issues, no responses)
# 3. Load formula (formulas/run-architect.toml)
# 4. Context: VISION.md, AGENTS.md, ops:prerequisites.md, structural graph
# 5. Bash-driven design phase:
# a. Fetch reviews API for ACCEPT/REJECT detection (deterministic)
# b. REJECT: handled entirely in bash (close PR, delete branch, journal)
# c. ACCEPT: invoke claude with human guidance injected into prompt
# d. Answers: resume saved session with answers injected
# 6. New pitches: agent_run(worktree, prompt)
# 5. Stateless pitch generation: for each selected issue:
# - Fetch issue body from Forgejo API (bash)
# - Invoke claude -p with issue body + context (stateless, no API calls)
# - Create PR with pitch content (bash)
# - Post footer comment (bash)
# 6. Response processing: handle ACCEPT/REJECT on existing PRs
#
# Precondition checks (bash before model):
# - Skip if no vision issues AND no open architect PRs
@ -60,7 +60,7 @@ SCRATCH_FILE_PREFIX="/tmp/architect-${PROJECT_NAME}-scratch"
WORKTREE="/tmp/${PROJECT_NAME}-architect-run"
# Override LOG_AGENT for consistent agent identification
# shellcheck disable=SC2034 # consumed by agent-sdk.sh and env.sh log()
# shellcheck disable=SC2034 # consumed by agent-sdk.sh and env.sh
LOG_AGENT="architect"
# Override log() to append to architect-specific log file
@ -91,448 +91,12 @@ load_formula_or_profile "architect" "$FACTORY_ROOT/formulas/run-architect.toml"
build_context_block VISION.md AGENTS.md ops:prerequisites.md
# ── Prepare .profile context (lessons injection) ─────────────────────────
formula_prepare_profile_context
# ── Build structural analysis graph ──────────────────────────────────────
build_graph_section
# ── Read scratch file (compaction survival) ───────────────────────────────
SCRATCH_CONTEXT=$(read_scratch_context "$SCRATCH_FILE")
SCRATCH_INSTRUCTION=$(build_scratch_instruction "$SCRATCH_FILE")
# ── Build prompt footer ──────────────────────────────────────────────────
build_sdk_prompt_footer
# ── Design phase: bash-driven review detection ────────────────────────────
# Fetch PR reviews from Forgejo API (deterministic, not model-dependent).
# Sets global output variables (not stdout — guidance text is often multiline):
# REVIEW_DECISION — ACCEPT|REJECT|NONE
# REVIEW_GUIDANCE — human guidance text (review body or comment text)
# Args: pr_number
fetch_pr_review_decision() {
local pr_num="$1"
REVIEW_DECISION="NONE"
REVIEW_GUIDANCE=""
# Step 1: Check PR reviews (Forgejo review UI) — takes precedence
local reviews_json
reviews_json=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${FORGE_API}/repos/${FORGE_OPS_REPO}/pulls/${pr_num}/reviews" 2>/dev/null) || reviews_json='[]'
# Find most recent non-bot review with a decision state
local review_decision review_body
review_decision=$(printf '%s' "$reviews_json" | jq -r '
[.[] | select(.user.login | test("bot$"; "i") | not)
| select(.state == "APPROVED" or .state == "REQUEST_CHANGES")]
| last | .state // empty
' 2>/dev/null) || review_decision=""
review_body=$(printf '%s' "$reviews_json" | jq -r '
[.[] | select(.user.login | test("bot$"; "i") | not)
| select(.state == "APPROVED" or .state == "REQUEST_CHANGES")]
| last | .body // empty
' 2>/dev/null) || review_body=""
if [ "$review_decision" = "APPROVED" ]; then
REVIEW_DECISION="ACCEPT"
REVIEW_GUIDANCE="$review_body"
return 0
elif [ "$review_decision" = "REQUEST_CHANGES" ]; then
REVIEW_DECISION="REJECT"
REVIEW_GUIDANCE="$review_body"
return 0
fi
# Step 2: Fallback — check PR comments for ACCEPT/REJECT text
local comments_json
comments_json=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${FORGE_API}/repos/${FORGE_OPS_REPO}/issues/${pr_num}/comments" 2>/dev/null) || comments_json='[]'
# Find most recent comment with ACCEPT or REJECT (case insensitive)
local comment_body
comment_body=$(printf '%s' "$comments_json" | jq -r '
[.[] | select(.body | test("(?i)^\\s*(ACCEPT|REJECT)"))] | last | .body // empty
' 2>/dev/null) || comment_body=""
if [ -n "$comment_body" ]; then
if printf '%s' "$comment_body" | grep -qiE '^\s*ACCEPT'; then
REVIEW_DECISION="ACCEPT"
# Extract guidance text after ACCEPT (e.g., "ACCEPT — use SSH approach" → "use SSH approach")
REVIEW_GUIDANCE=$(printf '%s' "$comment_body" | sed -n 's/^[[:space:]]*[Aa][Cc][Cc][Ee][Pp][Tt][[:space:]]*[—:-]*[[:space:]]*//p' | head -1)
# If guidance is empty on first line, use rest of comment
if [ -z "$REVIEW_GUIDANCE" ]; then
REVIEW_GUIDANCE=$(printf '%s' "$comment_body" | tail -n +2)
fi
elif printf '%s' "$comment_body" | grep -qiE '^\s*REJECT'; then
REVIEW_DECISION="REJECT"
REVIEW_GUIDANCE=$(printf '%s' "$comment_body" | sed -n 's/^[[:space:]]*[Rr][Ee][Jj][Ee][Cc][Tt][[:space:]]*[—:-]*[[:space:]]*//p' | head -1)
if [ -z "$REVIEW_GUIDANCE" ]; then
REVIEW_GUIDANCE=$(printf '%s' "$comment_body" | tail -n +2)
fi
fi
fi
}
# Handle REJECT entirely in bash — no model invocation needed.
# Args: pr_number, pr_head_branch, rejection_reason
handle_reject() {
local pr_num="$1"
local pr_branch="$2"
local reason="$3"
log "Handling REJECT for PR #${pr_num}: ${reason}"
# Close the PR via Forgejo API
curl -sf -X PATCH -H "Authorization: token ${FORGE_TOKEN}" \
-H 'Content-Type: application/json' \
"${FORGE_API}/repos/${FORGE_OPS_REPO}/pulls/${pr_num}" \
-d '{"state":"closed"}' >/dev/null 2>&1 || log "WARN: failed to close PR #${pr_num}"
# Delete the branch via Forgejo API
if [ -n "$pr_branch" ]; then
curl -sf -X DELETE -H "Authorization: token ${FORGE_TOKEN}" \
"${FORGE_API}/repos/${FORGE_OPS_REPO}/git/branches/${pr_branch}" >/dev/null 2>&1 \
|| log "WARN: failed to delete branch ${pr_branch}"
fi
# Remove in-progress label from the vision issue referenced in the PR
local pr_body
pr_body=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${FORGE_API}/repos/${FORGE_OPS_REPO}/pulls/${pr_num}" 2>/dev/null | jq -r '.body // ""') || pr_body=""
local vision_ref
vision_ref=$(printf '%s' "$pr_body" | grep -oE '#[0-9]+' | head -1 | tr -d '#') || vision_ref=""
if [ -n "$vision_ref" ]; then
# Look up in-progress label ID
local label_id
label_id=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${FORGE_API}/labels" 2>/dev/null | jq -r '.[] | select(.name == "in-progress") | .id' 2>/dev/null) || label_id=""
if [ -n "$label_id" ]; then
curl -sf -X DELETE -H "Authorization: token ${FORGE_TOKEN}" \
"${FORGE_API}/issues/${vision_ref}/labels/${label_id}" >/dev/null 2>&1 \
|| log "WARN: failed to remove in-progress label from issue #${vision_ref}"
fi
fi
# Journal the rejection via .profile (if available)
profile_write_journal "architect-reject-${pr_num}" \
"Sprint PR #${pr_num} rejected" \
"rejected: ${reason}" "" || true
# Clean up per-PR session file
rm -f "${SID_DIR}/pr-${pr_num}.sid"
log "REJECT handled for PR #${pr_num}"
}
# Detect answers on a PR in questions phase.
# Returns answer text via stdout, empty if no answers found.
# Args: pr_number
fetch_pr_answers() {
local pr_num="$1"
# Get PR body to check for Design forks section
local pr_body
pr_body=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${FORGE_API}/repos/${FORGE_OPS_REPO}/pulls/${pr_num}" 2>/dev/null | jq -r '.body // ""') || return 1
if ! printf '%s' "$pr_body" | grep -q "## Design forks"; then
return 1
fi
# Fetch comments and look for answer patterns (Q1: A, Q2: B, etc.)
local comments_json
comments_json=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${FORGE_API}/repos/${FORGE_OPS_REPO}/issues/${pr_num}/comments" 2>/dev/null) || return 1
# Find the most recent comment containing answer patterns
local answer_comment
answer_comment=$(printf '%s' "$comments_json" | jq -r '
[.[] | select(.body | test("Q[0-9]+:\\s*[A-Da-d]"))] | last | .body // empty
' 2>/dev/null) || answer_comment=""
if [ -n "$answer_comment" ]; then
printf '%s' "$answer_comment"
return 0
fi
return 1
}
# ── Sub-issue existence check ────────────────────────────────────────────
# Check if a vision issue already has sub-issues filed from it.
# Returns 0 if sub-issues exist and are open, 1 otherwise.
# Args: vision_issue_number
has_open_subissues() {
local vision_issue="$1"
local subissue_count=0
# Search for issues whose body contains 'Decomposed from #N' pattern
# Fetch all open issues with bodies in one API call (avoids N+1 calls)
local issues_json
issues_json=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${FORGE_API}/issues?state=open&limit=100" 2>/dev/null) || return 1
# Check each issue for the decomposition pattern using jq to extract bodies
subissue_count=$(printf '%s' "$issues_json" | jq -r --arg vid "$vision_issue" '
[.[] | select(.number != $vid) | select(.body // "" | contains("Decomposed from #" + $vid))] | length
' 2>/dev/null) || subissue_count=0
if [ "$subissue_count" -gt 0 ]; then
log "Vision issue #${vision_issue} has ${subissue_count} open sub-issue(s) — skipping"
return 0 # Has open sub-issues
fi
log "Vision issue #${vision_issue} has no open sub-issues"
return 1 # No open sub-issues
}
# ── Merged sprint PR check ───────────────────────────────────────────────
# Check if a vision issue already has a merged sprint PR on the ops repo.
# Returns 0 if a merged sprint PR exists, 1 otherwise.
# Args: vision_issue_number
has_merged_sprint_pr() {
local vision_issue="$1"
# Get closed PRs from ops repo
local prs_json
prs_json=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${FORGE_API}/repos/${FORGE_OPS_REPO}/pulls?state=closed&limit=100" 2>/dev/null) || return 1
# Check each closed PR for architect markers and vision issue reference
local pr_numbers
pr_numbers=$(printf '%s' "$prs_json" | jq -r '.[] | select(.title | contains("architect:")) | .number' 2>/dev/null) || return 1
local pr_num
while IFS= read -r pr_num; do
[ -z "$pr_num" ] && continue
# Get PR details including merged status
local pr_details
pr_details=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${FORGE_API}/repos/${FORGE_OPS_REPO}/pulls/${pr_num}" 2>/dev/null) || continue
# Check if PR is actually merged (not just closed)
local is_merged
is_merged=$(printf '%s' "$pr_details" | jq -r '.merged // false') || continue
if [ "$is_merged" != "true" ]; then
continue
fi
# Get PR body and check for vision issue reference
local pr_body
pr_body=$(printf '%s' "$pr_details" | jq -r '.body // ""') || continue
# Check if PR body references the vision issue number
# Look for patterns like "#N" where N is the vision issue number
if printf '%s' "$pr_body" | grep -qE "(#|refs|references)[[:space:]]*#${vision_issue}|#${vision_issue}[^0-9]|#${vision_issue}$"; then
log "Found merged sprint PR #${pr_num} referencing vision issue #${vision_issue} — skipping"
return 0 # Has merged sprint PR
fi
done <<< "$pr_numbers"
log "Vision issue #${vision_issue} has no merged sprint PR"
return 1 # No merged sprint PR
}
# ── Precondition checks in bash before invoking the model ─────────────────
# Check 1: Skip if no vision issues exist and no open architect PRs to handle
vision_count=$(curl -sf -H "Authorization: token $FORGE_TOKEN" \
"$FORGE_API/issues?labels=vision&state=open&limit=1" 2>/dev/null | jq length) || vision_count=0
if [ "${vision_count:-0}" -eq 0 ]; then
# Check for open architect PRs that need handling (ACCEPT/REJECT responses)
open_arch_prs=$(curl -sf -H "Authorization: token $FORGE_TOKEN" \
"${FORGE_API}/repos/${FORGE_OPS_REPO}/pulls?state=open&limit=10" 2>/dev/null | jq '[.[] | select(.title | startswith("architect:"))] | length') || open_arch_prs=0
if [ "${open_arch_prs:-0}" -eq 0 ]; then
log "no vision issues and no open architect PRs — skipping"
exit 0
fi
fi
# ── Design phase: process existing architect PRs (bash-driven) ────────────
# Bash reads the reviews API and handles state transitions deterministically.
# Model is only invoked for research (ACCEPT) and answer processing.
open_arch_prs_json=$(curl -sf -H "Authorization: token $FORGE_TOKEN" \
"${FORGE_API}/repos/${FORGE_OPS_REPO}/pulls?state=open&limit=10" 2>/dev/null) || open_arch_prs_json='[]'
open_arch_prs=$(printf '%s' "$open_arch_prs_json" | jq '[.[] | select(.title | startswith("architect:"))] | length') || open_arch_prs=0
# Track whether we processed any responses (to decide if pitching is needed)
processed_responses=false
# Iterate over open architect PRs and handle each based on review state
arch_pr_data=$(printf '%s' "$open_arch_prs_json" | jq -r '.[] | select(.title | startswith("architect:")) | "\(.number)\t\(.head.ref // "")"' 2>/dev/null) || arch_pr_data=""
while IFS=$'\t' read -r pr_num pr_branch; do
[ -z "$pr_num" ] && continue
log "Checking PR #${pr_num} (branch: ${pr_branch})"
# First check: is this PR in the answers phase (questions posted, answers received)?
answer_text=""
answer_text=$(fetch_pr_answers "$pr_num") || true
if [ -n "$answer_text" ]; then
# ── Answers received: resume saved session with answers injected ──
log "Answers detected on PR #${pr_num} — resuming design session"
processed_responses=true
pr_sid_file="${SID_DIR}/pr-${pr_num}.sid"
RESUME_ARGS=()
if [ -f "$pr_sid_file" ]; then
RESUME_SESSION=$(cat "$pr_sid_file")
RESUME_ARGS=(--resume "$RESUME_SESSION")
log "Resuming session ${RESUME_SESSION:0:12}... for answer processing"
else
log "No saved session for PR #${pr_num} — starting fresh for answers"
fi
# Build answer-processing prompt with answers injected
# shellcheck disable=SC2034
SID_FILE="$pr_sid_file"
ANSWER_PROMPT="You are the architect agent for ${FORGE_REPO}. You previously researched a sprint and posted design questions on PR #${pr_num}.
Human answered the design fork questions. Parse the answers and file concrete sub-issues.
## Human answers
${answer_text}
## Project context
${CONTEXT_BLOCK}
${GRAPH_SECTION}
$(formula_lessons_block)
## Instructions
1. Parse each answer (e.g. Q1: A, Q2: C)
2. Read the sprint spec from the PR branch
3. Look up the backlog label ID on the disinto repo:
GET ${FORGE_API}/labels — find label with name 'backlog'
4. Generate final sub-issues based on answers:
- Each sub-issue uses the appropriate issue template
- Fill all template fields (problem, solution, affected files max 3, acceptance criteria max 5, dependencies)
- File via Forgejo API on the disinto repo (not ops repo)
- MUST include 'labels' with backlog label ID in create-issue request
- Include 'Decomposed from #<vision_issue_number>' in each issue body
5. Comment on PR #${pr_num}: 'Sprint filed: #N, #N, #N'
6. Merge the PR via: POST ${FORGE_API}/repos/${FORGE_OPS_REPO}/pulls/${pr_num}/merge with body {\"Do\":\"merge\"}
${PROMPT_FOOTER}"
# Create worktree if not already set up
if [ ! -d "$WORKTREE" ]; then
formula_worktree_setup "$WORKTREE"
fi
export CLAUDE_MODEL="sonnet"
agent_run "${RESUME_ARGS[@]}" --worktree "$WORKTREE" "$ANSWER_PROMPT"
# Restore SID_FILE to default
# shellcheck disable=SC2034 # consumed by agent-sdk.sh
SID_FILE="/tmp/architect-session-${PROJECT_NAME}.sid"
log "Answer processing complete for PR #${pr_num}"
continue
fi
# Second check: fetch review decision (ACCEPT/REJECT/NONE)
# Sets REVIEW_DECISION and REVIEW_GUIDANCE global variables
fetch_pr_review_decision "$pr_num"
decision="$REVIEW_DECISION"
guidance="$REVIEW_GUIDANCE"
case "$decision" in
REJECT)
# ── REJECT: handled entirely in bash ──
handle_reject "$pr_num" "$pr_branch" "$guidance"
processed_responses=true
# Decrement open PR count (PR is now closed)
open_arch_prs=$((open_arch_prs - 1))
;;
ACCEPT)
# ── ACCEPT: invoke model with human guidance for research + questions ──
log "ACCEPT detected on PR #${pr_num} with guidance: ${guidance:-(none)}"
processed_responses=true
# Build human guidance block
GUIDANCE_BLOCK=""
if [ -n "$guidance" ]; then
GUIDANCE_BLOCK="## Human guidance (from sprint PR review)
${guidance}
The architect MUST factor this guidance into design fork identification
and question formulation — if the human specifies an approach, that approach
should be the default fork, and questions should refine it rather than
re-evaluate it."
fi
# Build research + questions prompt
RESEARCH_PROMPT="You are the architect agent for ${FORGE_REPO}. A sprint pitch on PR #${pr_num} has been ACCEPTED by a human reviewer.
Your task: research the codebase deeply, identify design forks, and formulate questions.
${GUIDANCE_BLOCK}
## Project context
${CONTEXT_BLOCK}
${GRAPH_SECTION}
${SCRATCH_CONTEXT}
$(formula_lessons_block)
## Instructions
1. Read the sprint spec from PR #${pr_num} on the ops repo (branch: ${pr_branch})
2. Research the codebase deeply:
- Read all files mentioned in the sprint spec
- Search for existing interfaces that could be reused
- Check what infrastructure already exists
3. Identify design forks — multiple valid implementation approaches
4. Formulate multiple-choice questions (Q1, Q2, Q3...)
5. Update the sprint spec file on the PR branch:
- Add '## Design forks' section with fork options
- Add '## Proposed sub-issues' section with concrete issues per fork path
- Use Forgejo API: PUT ${FORGE_API}/repos/${FORGE_OPS_REPO}/contents/<path> with branch ${pr_branch}
6. Update the PR body to include the Design forks section (required for answer detection):
- PATCH ${FORGE_API}/repos/${FORGE_OPS_REPO}/pulls/${pr_num}
- Body: {\"body\": \"<existing PR body + Design forks section>\"}
- The PR body MUST contain '## Design forks' after this step
7. Comment on PR #${pr_num} with the questions formatted as multiple choice:
- POST ${FORGE_API}/repos/${FORGE_OPS_REPO}/issues/${pr_num}/comments
${SCRATCH_INSTRUCTION}
${PROMPT_FOOTER}"
# Use per-PR session file for stateful resumption
pr_sid_file="${SID_DIR}/pr-${pr_num}.sid"
# shellcheck disable=SC2034
SID_FILE="$pr_sid_file"
# Create worktree if not already set up
if [ ! -d "$WORKTREE" ]; then
formula_worktree_setup "$WORKTREE"
fi
export CLAUDE_MODEL="sonnet"
agent_run --worktree "$WORKTREE" "$RESEARCH_PROMPT"
log "Research + questions posted for PR #${pr_num}, session saved: ${pr_sid_file}"
# Restore SID_FILE to default
# shellcheck disable=SC2034 # consumed by agent-sdk.sh
SID_FILE="/tmp/architect-session-${PROJECT_NAME}.sid"
;;
NONE)
log "PR #${pr_num} — no response yet, skipping"
;;
esac
done <<< "$arch_pr_data"
# ── Preflight: Select vision issues for pitching ──────────────────────────
# Recalculate open PR count after handling responses (REJECTs reduce count)
# Re-fetch if we processed any responses (PR count may have changed)
if [ "$processed_responses" = true ]; then
open_arch_prs_json=$(curl -sf -H "Authorization: token $FORGE_TOKEN" \
"${FORGE_API}/repos/${FORGE_OPS_REPO}/pulls?state=open&limit=10" 2>/dev/null) || open_arch_prs_json='[]'
open_arch_prs=$(printf '%s' "$open_arch_prs_json" | jq '[.[] | select(.title | startswith("architect:"))] | length') || open_arch_prs=0
fi
SCRATCH_CONTEXT=$(build_scratch_context)
GRAPH_SECTION=$(build_graph_section)
GUIDANCE_BLOCK=$(build_guidance_block)
FORMULA_CONTENT=$(build_formula_content)
PROMPT_FOOTER=$(build_prompt_footer)
SCRATCH_INSTRUCTION=$(build_scratch_instruction)
# Build list of vision issues that already have open architect PRs
declare -A _arch_vision_issues_with_open_prs
@ -628,42 +192,12 @@ fi
log "Selected ${#ARCHITECT_TARGET_ISSUES[@]} vision issue(s) for pitching: ${ARCHITECT_TARGET_ISSUES[*]}"
# ── Pitch prompt: research + PR creation (model handles pitching only) ────
# Architecture Decision: AD-003 — The runtime creates and destroys, the formula preserves.
build_architect_prompt() {
cat <<_PROMPT_EOF_
You are the architect agent for ${FORGE_REPO}. Work through the formula below.
Your role: strategic decomposition of vision issues into development sprints.
Propose sprints via PRs on the ops repo.
## Target vision issues for pitching
${ARCHITECT_TARGET_ISSUES[*]}
## Project context
${CONTEXT_BLOCK}
${GRAPH_SECTION}
${SCRATCH_CONTEXT}
$(formula_lessons_block)
## Formula
${FORMULA_CONTENT}
${SCRATCH_INSTRUCTION}
${PROMPT_FOOTER}
_PROMPT_EOF_
}
PROMPT=$(build_architect_prompt)
# ── Create worktree (if not already set up from design phase) ─────────────
if [ ! -d "$WORKTREE" ]; then
formula_worktree_setup "$WORKTREE"
fi
# ── Run agent for pitching ────────────────────────────────────────────────
export CLAUDE_MODEL="sonnet"
agent_run --worktree "$WORKTREE" "$PROMPT"
log "agent_run complete (pitching)"
# 5. Stateless pitch generation: for each selected issue:
# - Fetch issue body from Forgejo API (bash)
# - Invoke claude -p with issue body + context (stateless, no API calls)
# - Create PR with pitch content (bash)
# - Post footer comment (bash)
# 6. Response processing: handle ACCEPT/REJECT on existing PRs
# Clean up scratch files (legacy single file + per-issue files)
rm -f "$SCRATCH_FILE"