Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Git Context Controller (GCC) - Manages agent memory as a versioned file system under .GCC/. This skill should be used when working on multi-step projects that benefit from structured memory persistence, milestone tracking, branching for alternative approaches, and cross-session context recovery. Triggers on /gcc commands or natural language like 'commit this progress', 'branch to try an alternative', 'merge results', 'recover context'.
.claude/skills/davila7-gcc/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 74% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 4% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 23% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 137% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 525% | 0% |
GCC transforms agent memory from a passive token stream into a structured, versioned file system under .GCC/. Inspired by Git, it provides four operations — COMMIT, BRANCH, MERGE, CONTEXT — to persist milestones, explore alternatives in isolation, synthesize results, and recover historical context efficiently.
On first use, check if .GCC/ exists in the project root. If not, run scripts/gcc_init.sh to create the directory structure:
.GCC/
├── main.md # Global roadmap and objectives
├── metadata.yaml # Infrastructure state (branches, file tree, config)
├── commit.md # Commit history for main branch
├── log.md # OTA execution log for main branch
└── branches/ # Isolated workspaces for experiments
└── <branch-name>/
├── commit.md
├── log.md
└── summary.mdFor detailed file format specifications, read references/file_formats.md.
GCC behavior is controlled via metadata.yaml:
proactive_commits: true — Automatically suggest commits after completing coherent sub-tasksproactive_commits: false — Only commit when explicitly requestedToggle with: "enable/disable proactive commits" or by editing metadata.yaml.
Persist a milestone on the current branch.
Triggers: /gcc commit <summary>, "commit this progress", "save this milestone", "checkpoint"
Procedure:
commit.md to determine the next commit numbercommit.md with:[C004])summary.md if on a branch, or from main.md)log.md recording the commit actionmetadata.yaml file tree if files were created/modifiedmain.mdProactive behavior: When proactive_commits: true, suggest a commit after:
Create an isolated workspace for exploring an alternative approach.
Triggers: /gcc branch <name>, "branch to try...", "explore alternative...", "experiment with..."
Procedure:
.GCC/branches/<branch-name>/ directorysummary.md with: purpose, parent branch, creation date, key hypothesescommit.md and log.md for the branchmetadata.yaml to register the new branchmain.md Active Branches sectionlog.mdFrom this point, all COMMITs and OTA logs go to the branch-specific files until a MERGE or explicit branch switch.
Integrate a completed branch back into the main flow.
Triggers: /gcc merge <branch>, "merge results from...", "integrate the experiment", "branch X is done"
Procedure:
summary.md and commit.md to understand outcomescommit.md summarizing:main.md:metadata.yaml: set branch status to merged or abandonedlog.mdRetrieve historical memory at different resolution levels.
Triggers: /gcc context <flag>, "what did we do on...", "recover context", "show me the history", "where were we"
Flags:
--branch [name] — Read summary.md and latest commits for a specific branch (or current branch if no name). Provides high-level understanding of what happened and why.--log [n] — Read last N entries (default 20) from the current branch's log.md. Provides fine-grained OTA traces for debugging or resuming interrupted work.--metadata — Read metadata.yaml to recover project structure: file tree, dependencies, active branches, configuration.--full — Read main.md for the complete project roadmap, all milestones, and active branches. Use for cross-session recovery or handoff to another agent.When no flag is specified, default to --branch for the current active branch.
Throughout all work (not just during explicit commands), maintain the OTA execution log:
Append entries to the active branch's log.md. Keep a maximum of 50 entries; when exceeding, remove the oldest entries. Each entry includes a sequential ID, timestamp, and branch name.
Log OTA entries at meaningful decision points — not every single action, but significant observations, strategy changes, and outcomes.
When starting a new session on an existing project with .GCC/:
metadata.yaml to understand project state and active branchesmain.md for the global roadmap and objectives| User says | Command | |---|---| | "save/checkpoint/persist this" | COMMIT | | "try a different approach" | BRANCH | | "that experiment worked, integrate it" | MERGE | | "where were we?" / "what's the status?" | CONTEXT --full | | "what happened on branch X?" | CONTEXT --branch X | | "show recent activity" | CONTEXT --log | | "what files do we have?" | CONTEXT --metadata | | "enable/disable auto-commits" | Toggle proactive_commits in metadata.yaml |
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 7,824 | 6,134 | -22% | 1 | 1 | 0% | 1,146 | 2,116 | +85% | 0 | 0 | — |
case-02 | fail→pass | 9,940 | 7,494 | -25% | 1 | 1 | 0% | 1,698 | 2,957 | +74% | 0 | 0 | — |
case-03 | fail→fail | 4,733 | 5,015 | +6% | 1 | 1 | 0% | 663 | 1,900 | +187% | 0 | 0 | — |
case-04 | fail→pass | 11,736 | 4,312 | -63% | 1 | 1 | 0% | 2,015 | 2,098 | +4% | 0 | 0 | — |
case-18 | fail→pass | 13,391 | 4,753 | -65% | 1 | 1 | 0% | 1,843 | 2,276 | +23% | 0 | 0 | — |
case-05 | fail→fail | 3,942 | 6,272 | +59% | 1 | 1 | 0% | 517 | 2,497 | +383% | 0 | 0 | — |
case-06 | fail→pass | 7,378 | 7,444 | +1% | 1 | 1 | 0% | 1,103 | 2,615 | +137% | 0 | 0 | — |
case-07 | fail→pass | 3,507 | 9,247 | +164% | 1 | 1 | 0% | 468 | 2,925 | +525% | 0 | 0 | — |
case-08 | fail→pass | 5,114 | 8,444 | +65% | 1 | 1 | 0% | 753 | 2,254 | +199% | 0 | 0 | — |
case-09 | fail→pass | 6,412 | 5,560 | -13% | 1 | 1 | 0% | 916 | 2,430 | +165% | 0 | 0 | — |
case-10 | fail→pass | 2,849 | 3,881 | +36% | 1 | 1 | 0% | 384 | 1,974 | +414% | 0 | 0 | — |
case-11 | fail→pass | 8,026 | 6,675 | -17% | 1 | 1 | 0% | 1,239 | 2,341 | +89% | 0 | 0 | — |
case-12 | pass→pass | 10,666 | 3,791 | -64% | 1 | 1 | 0% | 1,753 | 2,168 | +24% | 0 | 0 | — |
case-13 | fail→pass | 8,652 | 6,345 | -27% | 1 | 1 | 0% | 1,457 | 2,461 | +69% | 0 | 0 | — |
case-14 | fail→fail | 5,308 | 5,232 | -1% | 1 | 1 | 0% | 776 | 2,261 | +191% | 0 | 0 | — |
case-15 | fail→pass | 14,458 | 4,442 | -69% | 1 | 1 | 0% | 2,460 | 2,159 | -12% | 0 | 0 | — |
case-16 | fail→fail | 3,174 | 8,963 | +182% | 1 | 1 | 0% | 368 | 2,875 | +681% | 0 | 0 | — |
case-17 | pass→pass | 3,066 | 4,824 | +57% | 1 | 1 | 0% | 419 | 2,268 | +441% | 0 | 0 | — |
case-19 | pass→pass | 8,559 | 4,067 | -52% | 1 | 1 | 0% | 1,309 | 2,145 | +64% | 0 | 0 | — |
case-20 | pass→pass | 4,708 | 3,740 | -21% | 1 | 1 | 0% | 617 | 2,012 | +226% | 0 | 0 | — |
case-21 | pass→pass | 7,278 | 6,281 | -14% | 1 | 1 | 0% | 1,431 | 2,623 | +83% | 0 | 0 | — |
case-22 | pass→pass | 4,151 | 4,880 | +18% | 1 | 1 | 0% | 725 | 2,043 | +182% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +50 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.