Table of Contents

Session 2

What Was Done

Forgejo Project Infrastructure
Gitea MCP Setup

Discoveries and Observations
Decisions Made
Raw Thinking
What's Next

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

Session 2

Date: 2026-04-06 Focus: Project planning infrastructure — Forgejo issues, milestones, project board, Gitea MCP setup Duration estimate: ~2 hours

What Was Done

Forgejo Project Infrastructure

Created the full issue/milestone/project structure on Forgejo to track the PLAN.md roadmap:

9 milestones created (IDs 11–19), one per phase:

Phase 1: Confidence Tracking
Phase 2: Survey Pass
Phase 3: Investigation Planning
Phase 4: External Knowledge Tools
Phase 5: Scale-Tiered Synthesis
Phase 6: Multi-Level Synthesis
Phase 7: Hypothesis-Driven Synthesis
Phase 8: Refinement Pass
Phase 9: Dynamic Report Structure

36 issues created (#1–#36), each with full descriptions drawn from PLAN.md:

#1–#3: Phase 1 (confidence fields, prompt updates, cache query)
#4–#7: Phase 2 (survey prompt, _run_survey(), wiring, size threshold)
#8–#11: Phase 3 (planning prompt, _run_planning(), dynamic turns, cache persistence)
#12–#17: Phase 4 (search API decision, search.py, package_lookup, fetch_url, ask_user, --no-external)
#18–#20: Phase 5 (tier classification, small-tier synthesis, depth prompts)
#21–#22: Phase 6 (grouping pass, two-level xlarge grouping)
#23–#24: Phase 7 (hypothesis synthesis prompt, think tool)
#25–#27: Phase 8 (refinement pass, --refine-depth, confidence prioritization)
#28–#29: Phase 9 (superset output fields, dynamic report renderer)
#30–#36: Backlog suggestions (config file, structured logging, --security, output formats, --model, watch+AI, PDF readers)

Project board ("Agentic Investigation Engine") created via web UI simulation (the Forgejo 9.0.3 API doesn't expose a projects endpoint). Used login session + CSRF token to POST the form directly. Board has 4 columns: Backlog / To Do / In Progress / Done. All 36 issues added to Backlog.

Gitea MCP Setup

Discovered the official gitea-mcp server (v1.1.0). Installed to ~/.local/bin/gitea-mcp. Registered globally via claude mcp add -s user which writes to ~/.claude.json (not settings.json — important distinction). Patched env vars (GITEA_HOST, GITEA_ACCESS_TOKEN) directly into ~/.claude.json since claude mcp add doesn't accept -e KEY=value when -- separator is used.

Uses forgejo_mcp_token (the claude-code user's token), not the archeious personal token.

Confirmed working: claude mcp list shows gitea: ✓ Connected. MCP tools verified live against Forgejo.

Discoveries and Observations

Forgejo 9.0.3's REST API (/api/v1/repos/{owner}/{repo}/projects) returns 404 — project boards are UI-only in this version. Web simulation (login → CSRF → POST form) was the only path.
claude mcp add stores config in ~/.claude.json, not ~/.claude/settings.json. Manually editing settings.json with mcpServers is ignored by the CLI.
The -e KEY=value flag for claude mcp add breaks when a -- separator is also present. Workaround: pass env vars in shell environment for the add command, then patch ~/.claude.json directly.
Wiki operations stay file-based (docs/wiki/) per CLAUDE.md — mcp__gitea__wiki_read/write should not be used.

Decisions Made

forgejo_mcp_token for MCP, not archeious token — the tfvars had a dedicated claude-code user token; using that keeps MCP operations under a distinct identity.
Backlog issues have no milestone — the 7 "Additional Suggestions" issues (#30–#36) are intentionally milestone-free so they don't inflate phase progress counts.
4-column board kept as-is — Basic kanban template gave Backlog/To Do/In Progress/Done, which is better than the originally planned 3 columns.

Raw Thinking

The project board issue was interesting — the API returning 404 while has_projects: true is a common Forgejo/Gitea gap. The web simulation approach (grab CSRF, POST form) is fragile but it worked. Worth noting that future Forgejo upgrades may expose the API properly.

The MCP setup was messier than expected. The claude mcp add CLI has a flag parsing quirk where -e KEY=value conflicts with -- separator. This is worth documenting for future setups. The fix (patch JSON directly) works but isn't ideal.

Phase 1 (#1–#3) is genuinely the right starting point — it's pure instrumentation (add fields, update a prompt, add a cache method), no behavior change, and it unlocks the confidence-aware features in Phases 5 and 8. The 3 issues are each ~15–30 minute tasks, well-sized for a focused session.

The hypothesis-driven synthesis change (Phase 7, #23) is probably the highest leverage single change in the whole plan — small prompt diff, large output quality difference. Worth considering doing it earlier than Phase 7 suggests, since it doesn't depend on any prior phase.

What's Next

Phase 1 is ready to start — branch feat/confidence-tracking, work through #1 → #2 → #3 in order
After Phase 1: Phase 2 (survey pass) is the most architecturally interesting next step
Consider pulling Phase 7 (#23 hypothesis-driven synthesis) forward — it's self-contained and high-leverage